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 = READA_FORWARD;
 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 = READA_FORWARD;
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 = READA_FORWARD;
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 = READA_FORWARD;
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->fs_info->creating_free_space_tree)
2930                 return 0;
2931
2932         if (root == root->fs_info->extent_root)
2933                 root = root->fs_info->tree_root;
2934
2935         delayed_refs = &trans->transaction->delayed_refs;
2936         if (count == 0)
2937                 count = atomic_read(&delayed_refs->num_entries) * 2;
2938
2939 again:
2940 #ifdef SCRAMBLE_DELAYED_REFS
2941         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2942 #endif
2943         trans->can_flush_pending_bgs = false;
2944         ret = __btrfs_run_delayed_refs(trans, root, count);
2945         if (ret < 0) {
2946                 btrfs_abort_transaction(trans, root, ret);
2947                 return ret;
2948         }
2949
2950         if (run_all) {
2951                 if (!list_empty(&trans->new_bgs))
2952                         btrfs_create_pending_block_groups(trans, root);
2953
2954                 spin_lock(&delayed_refs->lock);
2955                 node = rb_first(&delayed_refs->href_root);
2956                 if (!node) {
2957                         spin_unlock(&delayed_refs->lock);
2958                         goto out;
2959                 }
2960                 count = (unsigned long)-1;
2961
2962                 while (node) {
2963                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2964                                         href_node);
2965                         if (btrfs_delayed_ref_is_head(&head->node)) {
2966                                 struct btrfs_delayed_ref_node *ref;
2967
2968                                 ref = &head->node;
2969                                 atomic_inc(&ref->refs);
2970
2971                                 spin_unlock(&delayed_refs->lock);
2972                                 /*
2973                                  * Mutex was contended, block until it's
2974                                  * released and try again
2975                                  */
2976                                 mutex_lock(&head->mutex);
2977                                 mutex_unlock(&head->mutex);
2978
2979                                 btrfs_put_delayed_ref(ref);
2980                                 cond_resched();
2981                                 goto again;
2982                         } else {
2983                                 WARN_ON(1);
2984                         }
2985                         node = rb_next(node);
2986                 }
2987                 spin_unlock(&delayed_refs->lock);
2988                 cond_resched();
2989                 goto again;
2990         }
2991 out:
2992         assert_qgroups_uptodate(trans);
2993         trans->can_flush_pending_bgs = can_flush_pending_bgs;
2994         return 0;
2995 }
2996
2997 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2998                                 struct btrfs_root *root,
2999                                 u64 bytenr, u64 num_bytes, u64 flags,
3000                                 int level, int is_data)
3001 {
3002         struct btrfs_delayed_extent_op *extent_op;
3003         int ret;
3004
3005         extent_op = btrfs_alloc_delayed_extent_op();
3006         if (!extent_op)
3007                 return -ENOMEM;
3008
3009         extent_op->flags_to_set = flags;
3010         extent_op->update_flags = true;
3011         extent_op->update_key = false;
3012         extent_op->is_data = is_data ? true : false;
3013         extent_op->level = level;
3014
3015         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
3016                                           num_bytes, extent_op);
3017         if (ret)
3018                 btrfs_free_delayed_extent_op(extent_op);
3019         return ret;
3020 }
3021
3022 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
3023                                       struct btrfs_root *root,
3024                                       struct btrfs_path *path,
3025                                       u64 objectid, u64 offset, u64 bytenr)
3026 {
3027         struct btrfs_delayed_ref_head *head;
3028         struct btrfs_delayed_ref_node *ref;
3029         struct btrfs_delayed_data_ref *data_ref;
3030         struct btrfs_delayed_ref_root *delayed_refs;
3031         int ret = 0;
3032
3033         delayed_refs = &trans->transaction->delayed_refs;
3034         spin_lock(&delayed_refs->lock);
3035         head = btrfs_find_delayed_ref_head(trans, bytenr);
3036         if (!head) {
3037                 spin_unlock(&delayed_refs->lock);
3038                 return 0;
3039         }
3040
3041         if (!mutex_trylock(&head->mutex)) {
3042                 atomic_inc(&head->node.refs);
3043                 spin_unlock(&delayed_refs->lock);
3044
3045                 btrfs_release_path(path);
3046
3047                 /*
3048                  * Mutex was contended, block until it's released and let
3049                  * caller try again
3050                  */
3051                 mutex_lock(&head->mutex);
3052                 mutex_unlock(&head->mutex);
3053                 btrfs_put_delayed_ref(&head->node);
3054                 return -EAGAIN;
3055         }
3056         spin_unlock(&delayed_refs->lock);
3057
3058         spin_lock(&head->lock);
3059         list_for_each_entry(ref, &head->ref_list, list) {
3060                 /* If it's a shared ref we know a cross reference exists */
3061                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3062                         ret = 1;
3063                         break;
3064                 }
3065
3066                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3067
3068                 /*
3069                  * If our ref doesn't match the one we're currently looking at
3070                  * then we have a cross reference.
3071                  */
3072                 if (data_ref->root != root->root_key.objectid ||
3073                     data_ref->objectid != objectid ||
3074                     data_ref->offset != offset) {
3075                         ret = 1;
3076                         break;
3077                 }
3078         }
3079         spin_unlock(&head->lock);
3080         mutex_unlock(&head->mutex);
3081         return ret;
3082 }
3083
3084 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
3085                                         struct btrfs_root *root,
3086                                         struct btrfs_path *path,
3087                                         u64 objectid, u64 offset, u64 bytenr)
3088 {
3089         struct btrfs_root *extent_root = root->fs_info->extent_root;
3090         struct extent_buffer *leaf;
3091         struct btrfs_extent_data_ref *ref;
3092         struct btrfs_extent_inline_ref *iref;
3093         struct btrfs_extent_item *ei;
3094         struct btrfs_key key;
3095         u32 item_size;
3096         int ret;
3097
3098         key.objectid = bytenr;
3099         key.offset = (u64)-1;
3100         key.type = BTRFS_EXTENT_ITEM_KEY;
3101
3102         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3103         if (ret < 0)
3104                 goto out;
3105         BUG_ON(ret == 0); /* Corruption */
3106
3107         ret = -ENOENT;
3108         if (path->slots[0] == 0)
3109                 goto out;
3110
3111         path->slots[0]--;
3112         leaf = path->nodes[0];
3113         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3114
3115         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3116                 goto out;
3117
3118         ret = 1;
3119         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3120 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3121         if (item_size < sizeof(*ei)) {
3122                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3123                 goto out;
3124         }
3125 #endif
3126         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3127
3128         if (item_size != sizeof(*ei) +
3129             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3130                 goto out;
3131
3132         if (btrfs_extent_generation(leaf, ei) <=
3133             btrfs_root_last_snapshot(&root->root_item))
3134                 goto out;
3135
3136         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3137         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3138             BTRFS_EXTENT_DATA_REF_KEY)
3139                 goto out;
3140
3141         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3142         if (btrfs_extent_refs(leaf, ei) !=
3143             btrfs_extent_data_ref_count(leaf, ref) ||
3144             btrfs_extent_data_ref_root(leaf, ref) !=
3145             root->root_key.objectid ||
3146             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3147             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3148                 goto out;
3149
3150         ret = 0;
3151 out:
3152         return ret;
3153 }
3154
3155 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3156                           struct btrfs_root *root,
3157                           u64 objectid, u64 offset, u64 bytenr)
3158 {
3159         struct btrfs_path *path;
3160         int ret;
3161         int ret2;
3162
3163         path = btrfs_alloc_path();
3164         if (!path)
3165                 return -ENOENT;
3166
3167         do {
3168                 ret = check_committed_ref(trans, root, path, objectid,
3169                                           offset, bytenr);
3170                 if (ret && ret != -ENOENT)
3171                         goto out;
3172
3173                 ret2 = check_delayed_ref(trans, root, path, objectid,
3174                                          offset, bytenr);
3175         } while (ret2 == -EAGAIN);
3176
3177         if (ret2 && ret2 != -ENOENT) {
3178                 ret = ret2;
3179                 goto out;
3180         }
3181
3182         if (ret != -ENOENT || ret2 != -ENOENT)
3183                 ret = 0;
3184 out:
3185         btrfs_free_path(path);
3186         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3187                 WARN_ON(ret > 0);
3188         return ret;
3189 }
3190
3191 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3192                            struct btrfs_root *root,
3193                            struct extent_buffer *buf,
3194                            int full_backref, int inc)
3195 {
3196         u64 bytenr;
3197         u64 num_bytes;
3198         u64 parent;
3199         u64 ref_root;
3200         u32 nritems;
3201         struct btrfs_key key;
3202         struct btrfs_file_extent_item *fi;
3203         int i;
3204         int level;
3205         int ret = 0;
3206         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3207                             u64, u64, u64, u64, u64, u64);
3208
3209
3210         if (btrfs_test_is_dummy_root(root))
3211                 return 0;
3212
3213         ref_root = btrfs_header_owner(buf);
3214         nritems = btrfs_header_nritems(buf);
3215         level = btrfs_header_level(buf);
3216
3217         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3218                 return 0;
3219
3220         if (inc)
3221                 process_func = btrfs_inc_extent_ref;
3222         else
3223                 process_func = btrfs_free_extent;
3224
3225         if (full_backref)
3226                 parent = buf->start;
3227         else
3228                 parent = 0;
3229
3230         for (i = 0; i < nritems; i++) {
3231                 if (level == 0) {
3232                         btrfs_item_key_to_cpu(buf, &key, i);
3233                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3234                                 continue;
3235                         fi = btrfs_item_ptr(buf, i,
3236                                             struct btrfs_file_extent_item);
3237                         if (btrfs_file_extent_type(buf, fi) ==
3238                             BTRFS_FILE_EXTENT_INLINE)
3239                                 continue;
3240                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3241                         if (bytenr == 0)
3242                                 continue;
3243
3244                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3245                         key.offset -= btrfs_file_extent_offset(buf, fi);
3246                         ret = process_func(trans, root, bytenr, num_bytes,
3247                                            parent, ref_root, key.objectid,
3248                                            key.offset);
3249                         if (ret)
3250                                 goto fail;
3251                 } else {
3252                         bytenr = btrfs_node_blockptr(buf, i);
3253                         num_bytes = root->nodesize;
3254                         ret = process_func(trans, root, bytenr, num_bytes,
3255                                            parent, ref_root, level - 1, 0);
3256                         if (ret)
3257                                 goto fail;
3258                 }
3259         }
3260         return 0;
3261 fail:
3262         return ret;
3263 }
3264
3265 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3266                   struct extent_buffer *buf, int full_backref)
3267 {
3268         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3269 }
3270
3271 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3272                   struct extent_buffer *buf, int full_backref)
3273 {
3274         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3275 }
3276
3277 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3278                                  struct btrfs_root *root,
3279                                  struct btrfs_path *path,
3280                                  struct btrfs_block_group_cache *cache)
3281 {
3282         int ret;
3283         struct btrfs_root *extent_root = root->fs_info->extent_root;
3284         unsigned long bi;
3285         struct extent_buffer *leaf;
3286
3287         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3288         if (ret) {
3289                 if (ret > 0)
3290                         ret = -ENOENT;
3291                 goto fail;
3292         }
3293
3294         leaf = path->nodes[0];
3295         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3296         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3297         btrfs_mark_buffer_dirty(leaf);
3298 fail:
3299         btrfs_release_path(path);
3300         return ret;
3301
3302 }
3303
3304 static struct btrfs_block_group_cache *
3305 next_block_group(struct btrfs_root *root,
3306                  struct btrfs_block_group_cache *cache)
3307 {
3308         struct rb_node *node;
3309
3310         spin_lock(&root->fs_info->block_group_cache_lock);
3311
3312         /* If our block group was removed, we need a full search. */
3313         if (RB_EMPTY_NODE(&cache->cache_node)) {
3314                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3315
3316                 spin_unlock(&root->fs_info->block_group_cache_lock);
3317                 btrfs_put_block_group(cache);
3318                 cache = btrfs_lookup_first_block_group(root->fs_info,
3319                                                        next_bytenr);
3320                 return cache;
3321         }
3322         node = rb_next(&cache->cache_node);
3323         btrfs_put_block_group(cache);
3324         if (node) {
3325                 cache = rb_entry(node, struct btrfs_block_group_cache,
3326                                  cache_node);
3327                 btrfs_get_block_group(cache);
3328         } else
3329                 cache = NULL;
3330         spin_unlock(&root->fs_info->block_group_cache_lock);
3331         return cache;
3332 }
3333
3334 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3335                             struct btrfs_trans_handle *trans,
3336                             struct btrfs_path *path)
3337 {
3338         struct btrfs_root *root = block_group->fs_info->tree_root;
3339         struct inode *inode = NULL;
3340         u64 alloc_hint = 0;
3341         int dcs = BTRFS_DC_ERROR;
3342         u64 num_pages = 0;
3343         int retries = 0;
3344         int ret = 0;
3345
3346         /*
3347          * If this block group is smaller than 100 megs don't bother caching the
3348          * block group.
3349          */
3350         if (block_group->key.offset < (100 * SZ_1M)) {
3351                 spin_lock(&block_group->lock);
3352                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3353                 spin_unlock(&block_group->lock);
3354                 return 0;
3355         }
3356
3357         if (trans->aborted)
3358                 return 0;
3359 again:
3360         inode = lookup_free_space_inode(root, block_group, path);
3361         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3362                 ret = PTR_ERR(inode);
3363                 btrfs_release_path(path);
3364                 goto out;
3365         }
3366
3367         if (IS_ERR(inode)) {
3368                 BUG_ON(retries);
3369                 retries++;
3370
3371                 if (block_group->ro)
3372                         goto out_free;
3373
3374                 ret = create_free_space_inode(root, trans, block_group, path);
3375                 if (ret)
3376                         goto out_free;
3377                 goto again;
3378         }
3379
3380         /* We've already setup this transaction, go ahead and exit */
3381         if (block_group->cache_generation == trans->transid &&
3382             i_size_read(inode)) {
3383                 dcs = BTRFS_DC_SETUP;
3384                 goto out_put;
3385         }
3386
3387         /*
3388          * We want to set the generation to 0, that way if anything goes wrong
3389          * from here on out we know not to trust this cache when we load up next
3390          * time.
3391          */
3392         BTRFS_I(inode)->generation = 0;
3393         ret = btrfs_update_inode(trans, root, inode);
3394         if (ret) {
3395                 /*
3396                  * So theoretically we could recover from this, simply set the
3397                  * super cache generation to 0 so we know to invalidate the
3398                  * cache, but then we'd have to keep track of the block groups
3399                  * that fail this way so we know we _have_ to reset this cache
3400                  * before the next commit or risk reading stale cache.  So to
3401                  * limit our exposure to horrible edge cases lets just abort the
3402                  * transaction, this only happens in really bad situations
3403                  * anyway.
3404                  */
3405                 btrfs_abort_transaction(trans, root, ret);
3406                 goto out_put;
3407         }
3408         WARN_ON(ret);
3409
3410         if (i_size_read(inode) > 0) {
3411                 ret = btrfs_check_trunc_cache_free_space(root,
3412                                         &root->fs_info->global_block_rsv);
3413                 if (ret)
3414                         goto out_put;
3415
3416                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3417                 if (ret)
3418                         goto out_put;
3419         }
3420
3421         spin_lock(&block_group->lock);
3422         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3423             !btrfs_test_opt(root, SPACE_CACHE)) {
3424                 /*
3425                  * don't bother trying to write stuff out _if_
3426                  * a) we're not cached,
3427                  * b) we're with nospace_cache mount option.
3428                  */
3429                 dcs = BTRFS_DC_WRITTEN;
3430                 spin_unlock(&block_group->lock);
3431                 goto out_put;
3432         }
3433         spin_unlock(&block_group->lock);
3434
3435         /*
3436          * We hit an ENOSPC when setting up the cache in this transaction, just
3437          * skip doing the setup, we've already cleared the cache so we're safe.
3438          */
3439         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3440                 ret = -ENOSPC;
3441                 goto out_put;
3442         }
3443
3444         /*
3445          * Try to preallocate enough space based on how big the block group is.
3446          * Keep in mind this has to include any pinned space which could end up
3447          * taking up quite a bit since it's not folded into the other space
3448          * cache.
3449          */
3450         num_pages = div_u64(block_group->key.offset, SZ_256M);
3451         if (!num_pages)
3452                 num_pages = 1;
3453
3454         num_pages *= 16;
3455         num_pages *= PAGE_CACHE_SIZE;
3456
3457         ret = btrfs_check_data_free_space(inode, 0, num_pages);
3458         if (ret)
3459                 goto out_put;
3460
3461         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3462                                               num_pages, num_pages,
3463                                               &alloc_hint);
3464         /*
3465          * Our cache requires contiguous chunks so that we don't modify a bunch
3466          * of metadata or split extents when writing the cache out, which means
3467          * we can enospc if we are heavily fragmented in addition to just normal
3468          * out of space conditions.  So if we hit this just skip setting up any
3469          * other block groups for this transaction, maybe we'll unpin enough
3470          * space the next time around.
3471          */
3472         if (!ret)
3473                 dcs = BTRFS_DC_SETUP;
3474         else if (ret == -ENOSPC)
3475                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3476         btrfs_free_reserved_data_space(inode, 0, num_pages);
3477
3478 out_put:
3479         iput(inode);
3480 out_free:
3481         btrfs_release_path(path);
3482 out:
3483         spin_lock(&block_group->lock);
3484         if (!ret && dcs == BTRFS_DC_SETUP)
3485                 block_group->cache_generation = trans->transid;
3486         block_group->disk_cache_state = dcs;
3487         spin_unlock(&block_group->lock);
3488
3489         return ret;
3490 }
3491
3492 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3493                             struct btrfs_root *root)
3494 {
3495         struct btrfs_block_group_cache *cache, *tmp;
3496         struct btrfs_transaction *cur_trans = trans->transaction;
3497         struct btrfs_path *path;
3498
3499         if (list_empty(&cur_trans->dirty_bgs) ||
3500             !btrfs_test_opt(root, SPACE_CACHE))
3501                 return 0;
3502
3503         path = btrfs_alloc_path();
3504         if (!path)
3505                 return -ENOMEM;
3506
3507         /* Could add new block groups, use _safe just in case */
3508         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3509                                  dirty_list) {
3510                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3511                         cache_save_setup(cache, trans, path);
3512         }
3513
3514         btrfs_free_path(path);
3515         return 0;
3516 }
3517
3518 /*
3519  * transaction commit does final block group cache writeback during a
3520  * critical section where nothing is allowed to change the FS.  This is
3521  * required in order for the cache to actually match the block group,
3522  * but can introduce a lot of latency into the commit.
3523  *
3524  * So, btrfs_start_dirty_block_groups is here to kick off block group
3525  * cache IO.  There's a chance we'll have to redo some of it if the
3526  * block group changes again during the commit, but it greatly reduces
3527  * the commit latency by getting rid of the easy block groups while
3528  * we're still allowing others to join the commit.
3529  */
3530 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3531                                    struct btrfs_root *root)
3532 {
3533         struct btrfs_block_group_cache *cache;
3534         struct btrfs_transaction *cur_trans = trans->transaction;
3535         int ret = 0;
3536         int should_put;
3537         struct btrfs_path *path = NULL;
3538         LIST_HEAD(dirty);
3539         struct list_head *io = &cur_trans->io_bgs;
3540         int num_started = 0;
3541         int loops = 0;
3542
3543         spin_lock(&cur_trans->dirty_bgs_lock);
3544         if (list_empty(&cur_trans->dirty_bgs)) {
3545                 spin_unlock(&cur_trans->dirty_bgs_lock);
3546                 return 0;
3547         }
3548         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3549         spin_unlock(&cur_trans->dirty_bgs_lock);
3550
3551 again:
3552         /*
3553          * make sure all the block groups on our dirty list actually
3554          * exist
3555          */
3556         btrfs_create_pending_block_groups(trans, root);
3557
3558         if (!path) {
3559                 path = btrfs_alloc_path();
3560                 if (!path)
3561                         return -ENOMEM;
3562         }
3563
3564         /*
3565          * cache_write_mutex is here only to save us from balance or automatic
3566          * removal of empty block groups deleting this block group while we are
3567          * writing out the cache
3568          */
3569         mutex_lock(&trans->transaction->cache_write_mutex);
3570         while (!list_empty(&dirty)) {
3571                 cache = list_first_entry(&dirty,
3572                                          struct btrfs_block_group_cache,
3573                                          dirty_list);
3574                 /*
3575                  * this can happen if something re-dirties a block
3576                  * group that is already under IO.  Just wait for it to
3577                  * finish and then do it all again
3578                  */
3579                 if (!list_empty(&cache->io_list)) {
3580                         list_del_init(&cache->io_list);
3581                         btrfs_wait_cache_io(root, trans, cache,
3582                                             &cache->io_ctl, path,
3583                                             cache->key.objectid);
3584                         btrfs_put_block_group(cache);
3585                 }
3586
3587
3588                 /*
3589                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3590                  * if it should update the cache_state.  Don't delete
3591                  * until after we wait.
3592                  *
3593                  * Since we're not running in the commit critical section
3594                  * we need the dirty_bgs_lock to protect from update_block_group
3595                  */
3596                 spin_lock(&cur_trans->dirty_bgs_lock);
3597                 list_del_init(&cache->dirty_list);
3598                 spin_unlock(&cur_trans->dirty_bgs_lock);
3599
3600                 should_put = 1;
3601
3602                 cache_save_setup(cache, trans, path);
3603
3604                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3605                         cache->io_ctl.inode = NULL;
3606                         ret = btrfs_write_out_cache(root, trans, cache, path);
3607                         if (ret == 0 && cache->io_ctl.inode) {
3608                                 num_started++;
3609                                 should_put = 0;
3610
3611                                 /*
3612                                  * the cache_write_mutex is protecting
3613                                  * the io_list
3614                                  */
3615                                 list_add_tail(&cache->io_list, io);
3616                         } else {
3617                                 /*
3618                                  * if we failed to write the cache, the
3619                                  * generation will be bad and life goes on
3620                                  */
3621                                 ret = 0;
3622                         }
3623                 }
3624                 if (!ret) {
3625                         ret = write_one_cache_group(trans, root, path, cache);
3626                         /*
3627                          * Our block group might still be attached to the list
3628                          * of new block groups in the transaction handle of some
3629                          * other task (struct btrfs_trans_handle->new_bgs). This
3630                          * means its block group item isn't yet in the extent
3631                          * tree. If this happens ignore the error, as we will
3632                          * try again later in the critical section of the
3633                          * transaction commit.
3634                          */
3635                         if (ret == -ENOENT) {
3636                                 ret = 0;
3637                                 spin_lock(&cur_trans->dirty_bgs_lock);
3638                                 if (list_empty(&cache->dirty_list)) {
3639                                         list_add_tail(&cache->dirty_list,
3640                                                       &cur_trans->dirty_bgs);
3641                                         btrfs_get_block_group(cache);
3642                                 }
3643                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3644                         } else if (ret) {
3645                                 btrfs_abort_transaction(trans, root, ret);
3646                         }
3647                 }
3648
3649                 /* if its not on the io list, we need to put the block group */
3650                 if (should_put)
3651                         btrfs_put_block_group(cache);
3652
3653                 if (ret)
3654                         break;
3655
3656                 /*
3657                  * Avoid blocking other tasks for too long. It might even save
3658                  * us from writing caches for block groups that are going to be
3659                  * removed.
3660                  */
3661                 mutex_unlock(&trans->transaction->cache_write_mutex);
3662                 mutex_lock(&trans->transaction->cache_write_mutex);
3663         }
3664         mutex_unlock(&trans->transaction->cache_write_mutex);
3665
3666         /*
3667          * go through delayed refs for all the stuff we've just kicked off
3668          * and then loop back (just once)
3669          */
3670         ret = btrfs_run_delayed_refs(trans, root, 0);
3671         if (!ret && loops == 0) {
3672                 loops++;
3673                 spin_lock(&cur_trans->dirty_bgs_lock);
3674                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3675                 /*
3676                  * dirty_bgs_lock protects us from concurrent block group
3677                  * deletes too (not just cache_write_mutex).
3678                  */
3679                 if (!list_empty(&dirty)) {
3680                         spin_unlock(&cur_trans->dirty_bgs_lock);
3681                         goto again;
3682                 }
3683                 spin_unlock(&cur_trans->dirty_bgs_lock);
3684         }
3685
3686         btrfs_free_path(path);
3687         return ret;
3688 }
3689
3690 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3691                                    struct btrfs_root *root)
3692 {
3693         struct btrfs_block_group_cache *cache;
3694         struct btrfs_transaction *cur_trans = trans->transaction;
3695         int ret = 0;
3696         int should_put;
3697         struct btrfs_path *path;
3698         struct list_head *io = &cur_trans->io_bgs;
3699         int num_started = 0;
3700
3701         path = btrfs_alloc_path();
3702         if (!path)
3703                 return -ENOMEM;
3704
3705         /*
3706          * Even though we are in the critical section of the transaction commit,
3707          * we can still have concurrent tasks adding elements to this
3708          * transaction's list of dirty block groups. These tasks correspond to
3709          * endio free space workers started when writeback finishes for a
3710          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3711          * allocate new block groups as a result of COWing nodes of the root
3712          * tree when updating the free space inode. The writeback for the space
3713          * caches is triggered by an earlier call to
3714          * btrfs_start_dirty_block_groups() and iterations of the following
3715          * loop.
3716          * Also we want to do the cache_save_setup first and then run the
3717          * delayed refs to make sure we have the best chance at doing this all
3718          * in one shot.
3719          */
3720         spin_lock(&cur_trans->dirty_bgs_lock);
3721         while (!list_empty(&cur_trans->dirty_bgs)) {
3722                 cache = list_first_entry(&cur_trans->dirty_bgs,
3723                                          struct btrfs_block_group_cache,
3724                                          dirty_list);
3725
3726                 /*
3727                  * this can happen if cache_save_setup re-dirties a block
3728                  * group that is already under IO.  Just wait for it to
3729                  * finish and then do it all again
3730                  */
3731                 if (!list_empty(&cache->io_list)) {
3732                         spin_unlock(&cur_trans->dirty_bgs_lock);
3733                         list_del_init(&cache->io_list);
3734                         btrfs_wait_cache_io(root, trans, cache,
3735                                             &cache->io_ctl, path,
3736                                             cache->key.objectid);
3737                         btrfs_put_block_group(cache);
3738                         spin_lock(&cur_trans->dirty_bgs_lock);
3739                 }
3740
3741                 /*
3742                  * don't remove from the dirty list until after we've waited
3743                  * on any pending IO
3744                  */
3745                 list_del_init(&cache->dirty_list);
3746                 spin_unlock(&cur_trans->dirty_bgs_lock);
3747                 should_put = 1;
3748
3749                 cache_save_setup(cache, trans, path);
3750
3751                 if (!ret)
3752                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3753
3754                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3755                         cache->io_ctl.inode = NULL;
3756                         ret = btrfs_write_out_cache(root, trans, cache, path);
3757                         if (ret == 0 && cache->io_ctl.inode) {
3758                                 num_started++;
3759                                 should_put = 0;
3760                                 list_add_tail(&cache->io_list, io);
3761                         } else {
3762                                 /*
3763                                  * if we failed to write the cache, the
3764                                  * generation will be bad and life goes on
3765                                  */
3766                                 ret = 0;
3767                         }
3768                 }
3769                 if (!ret) {
3770                         ret = write_one_cache_group(trans, root, path, cache);
3771                         /*
3772                          * One of the free space endio workers might have
3773                          * created a new block group while updating a free space
3774                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3775                          * and hasn't released its transaction handle yet, in
3776                          * which case the new block group is still attached to
3777                          * its transaction handle and its creation has not
3778                          * finished yet (no block group item in the extent tree
3779                          * yet, etc). If this is the case, wait for all free
3780                          * space endio workers to finish and retry. This is a
3781                          * a very rare case so no need for a more efficient and
3782                          * complex approach.
3783                          */
3784                         if (ret == -ENOENT) {
3785                                 wait_event(cur_trans->writer_wait,
3786                                    atomic_read(&cur_trans->num_writers) == 1);
3787                                 ret = write_one_cache_group(trans, root, path,
3788                                                             cache);
3789                         }
3790                         if (ret)
3791                                 btrfs_abort_transaction(trans, root, ret);
3792                 }
3793
3794                 /* if its not on the io list, we need to put the block group */
3795                 if (should_put)
3796                         btrfs_put_block_group(cache);
3797                 spin_lock(&cur_trans->dirty_bgs_lock);
3798         }
3799         spin_unlock(&cur_trans->dirty_bgs_lock);
3800
3801         while (!list_empty(io)) {
3802                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3803                                          io_list);
3804                 list_del_init(&cache->io_list);
3805                 btrfs_wait_cache_io(root, trans, cache,
3806                                     &cache->io_ctl, path, cache->key.objectid);
3807                 btrfs_put_block_group(cache);
3808         }
3809
3810         btrfs_free_path(path);
3811         return ret;
3812 }
3813
3814 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3815 {
3816         struct btrfs_block_group_cache *block_group;
3817         int readonly = 0;
3818
3819         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3820         if (!block_group || block_group->ro)
3821                 readonly = 1;
3822         if (block_group)
3823                 btrfs_put_block_group(block_group);
3824         return readonly;
3825 }
3826
3827 static const char *alloc_name(u64 flags)
3828 {
3829         switch (flags) {
3830         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3831                 return "mixed";
3832         case BTRFS_BLOCK_GROUP_METADATA:
3833                 return "metadata";
3834         case BTRFS_BLOCK_GROUP_DATA:
3835                 return "data";
3836         case BTRFS_BLOCK_GROUP_SYSTEM:
3837                 return "system";
3838         default:
3839                 WARN_ON(1);
3840                 return "invalid-combination";
3841         };
3842 }
3843
3844 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3845                              u64 total_bytes, u64 bytes_used,
3846                              struct btrfs_space_info **space_info)
3847 {
3848         struct btrfs_space_info *found;
3849         int i;
3850         int factor;
3851         int ret;
3852
3853         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3854                      BTRFS_BLOCK_GROUP_RAID10))
3855                 factor = 2;
3856         else
3857                 factor = 1;
3858
3859         found = __find_space_info(info, flags);
3860         if (found) {
3861                 spin_lock(&found->lock);
3862                 found->total_bytes += total_bytes;
3863                 found->disk_total += total_bytes * factor;
3864                 found->bytes_used += bytes_used;
3865                 found->disk_used += bytes_used * factor;
3866                 if (total_bytes > 0)
3867                         found->full = 0;
3868                 spin_unlock(&found->lock);
3869                 *space_info = found;
3870                 return 0;
3871         }
3872         found = kzalloc(sizeof(*found), GFP_NOFS);
3873         if (!found)
3874                 return -ENOMEM;
3875
3876         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3877         if (ret) {
3878                 kfree(found);
3879                 return ret;
3880         }
3881
3882         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3883                 INIT_LIST_HEAD(&found->block_groups[i]);
3884         init_rwsem(&found->groups_sem);
3885         spin_lock_init(&found->lock);
3886         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3887         found->total_bytes = total_bytes;
3888         found->disk_total = total_bytes * factor;
3889         found->bytes_used = bytes_used;
3890         found->disk_used = bytes_used * factor;
3891         found->bytes_pinned = 0;
3892         found->bytes_reserved = 0;
3893         found->bytes_readonly = 0;
3894         found->bytes_may_use = 0;
3895         found->full = 0;
3896         found->max_extent_size = 0;
3897         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3898         found->chunk_alloc = 0;
3899         found->flush = 0;
3900         init_waitqueue_head(&found->wait);
3901         INIT_LIST_HEAD(&found->ro_bgs);
3902
3903         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3904                                     info->space_info_kobj, "%s",
3905                                     alloc_name(found->flags));
3906         if (ret) {
3907                 kfree(found);
3908                 return ret;
3909         }
3910
3911         *space_info = found;
3912         list_add_rcu(&found->list, &info->space_info);
3913         if (flags & BTRFS_BLOCK_GROUP_DATA)
3914                 info->data_sinfo = found;
3915
3916         return ret;
3917 }
3918
3919 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3920 {
3921         u64 extra_flags = chunk_to_extended(flags) &
3922                                 BTRFS_EXTENDED_PROFILE_MASK;
3923
3924         write_seqlock(&fs_info->profiles_lock);
3925         if (flags & BTRFS_BLOCK_GROUP_DATA)
3926                 fs_info->avail_data_alloc_bits |= extra_flags;
3927         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3928                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3929         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3930                 fs_info->avail_system_alloc_bits |= extra_flags;
3931         write_sequnlock(&fs_info->profiles_lock);
3932 }
3933
3934 /*
3935  * returns target flags in extended format or 0 if restripe for this
3936  * chunk_type is not in progress
3937  *
3938  * should be called with either volume_mutex or balance_lock held
3939  */
3940 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3941 {
3942         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3943         u64 target = 0;
3944
3945         if (!bctl)
3946                 return 0;
3947
3948         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3949             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3950                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3951         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3952                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3953                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3954         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3955                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3956                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3957         }
3958
3959         return target;
3960 }
3961
3962 /*
3963  * @flags: available profiles in extended format (see ctree.h)
3964  *
3965  * Returns reduced profile in chunk format.  If profile changing is in
3966  * progress (either running or paused) picks the target profile (if it's
3967  * already available), otherwise falls back to plain reducing.
3968  */
3969 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3970 {
3971         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3972         u64 target;
3973         u64 raid_type;
3974         u64 allowed = 0;
3975
3976         /*
3977          * see if restripe for this chunk_type is in progress, if so
3978          * try to reduce to the target profile
3979          */
3980         spin_lock(&root->fs_info->balance_lock);
3981         target = get_restripe_target(root->fs_info, flags);
3982         if (target) {
3983                 /* pick target profile only if it's already available */
3984                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3985                         spin_unlock(&root->fs_info->balance_lock);
3986                         return extended_to_chunk(target);
3987                 }
3988         }
3989         spin_unlock(&root->fs_info->balance_lock);
3990
3991         /* First, mask out the RAID levels which aren't possible */
3992         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
3993                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
3994                         allowed |= btrfs_raid_group[raid_type];
3995         }
3996         allowed &= flags;
3997
3998         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
3999                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4000         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4001                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4002         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4003                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4004         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4005                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4006         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4007                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4008
4009         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4010
4011         return extended_to_chunk(flags | allowed);
4012 }
4013
4014 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
4015 {
4016         unsigned seq;
4017         u64 flags;
4018
4019         do {
4020                 flags = orig_flags;
4021                 seq = read_seqbegin(&root->fs_info->profiles_lock);
4022
4023                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4024                         flags |= root->fs_info->avail_data_alloc_bits;
4025                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4026                         flags |= root->fs_info->avail_system_alloc_bits;
4027                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4028                         flags |= root->fs_info->avail_metadata_alloc_bits;
4029         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
4030
4031         return btrfs_reduce_alloc_profile(root, flags);
4032 }
4033
4034 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
4035 {
4036         u64 flags;
4037         u64 ret;
4038
4039         if (data)
4040                 flags = BTRFS_BLOCK_GROUP_DATA;
4041         else if (root == root->fs_info->chunk_root)
4042                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4043         else
4044                 flags = BTRFS_BLOCK_GROUP_METADATA;
4045
4046         ret = get_alloc_profile(root, flags);
4047         return ret;
4048 }
4049
4050 int btrfs_alloc_data_chunk_ondemand(struct inode *inode, u64 bytes)
4051 {
4052         struct btrfs_space_info *data_sinfo;
4053         struct btrfs_root *root = BTRFS_I(inode)->root;
4054         struct btrfs_fs_info *fs_info = root->fs_info;
4055         u64 used;
4056         int ret = 0;
4057         int need_commit = 2;
4058         int have_pinned_space;
4059
4060         /* make sure bytes are sectorsize aligned */
4061         bytes = ALIGN(bytes, root->sectorsize);
4062
4063         if (btrfs_is_free_space_inode(inode)) {
4064                 need_commit = 0;
4065                 ASSERT(current->journal_info);
4066         }
4067
4068         data_sinfo = fs_info->data_sinfo;
4069         if (!data_sinfo)
4070                 goto alloc;
4071
4072 again:
4073         /* make sure we have enough space to handle the data first */
4074         spin_lock(&data_sinfo->lock);
4075         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
4076                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
4077                 data_sinfo->bytes_may_use;
4078
4079         if (used + bytes > data_sinfo->total_bytes) {
4080                 struct btrfs_trans_handle *trans;
4081
4082                 /*
4083                  * if we don't have enough free bytes in this space then we need
4084                  * to alloc a new chunk.
4085                  */
4086                 if (!data_sinfo->full) {
4087                         u64 alloc_target;
4088
4089                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4090                         spin_unlock(&data_sinfo->lock);
4091 alloc:
4092                         alloc_target = btrfs_get_alloc_profile(root, 1);
4093                         /*
4094                          * It is ugly that we don't call nolock join
4095                          * transaction for the free space inode case here.
4096                          * But it is safe because we only do the data space
4097                          * reservation for the free space cache in the
4098                          * transaction context, the common join transaction
4099                          * just increase the counter of the current transaction
4100                          * handler, doesn't try to acquire the trans_lock of
4101                          * the fs.
4102                          */
4103                         trans = btrfs_join_transaction(root);
4104                         if (IS_ERR(trans))
4105                                 return PTR_ERR(trans);
4106
4107                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4108                                              alloc_target,
4109                                              CHUNK_ALLOC_NO_FORCE);
4110                         btrfs_end_transaction(trans, root);
4111                         if (ret < 0) {
4112                                 if (ret != -ENOSPC)
4113                                         return ret;
4114                                 else {
4115                                         have_pinned_space = 1;
4116                                         goto commit_trans;
4117                                 }
4118                         }
4119
4120                         if (!data_sinfo)
4121                                 data_sinfo = fs_info->data_sinfo;
4122
4123                         goto again;
4124                 }
4125
4126                 /*
4127                  * If we don't have enough pinned space to deal with this
4128                  * allocation, and no removed chunk in current transaction,
4129                  * don't bother committing the transaction.
4130                  */
4131                 have_pinned_space = percpu_counter_compare(
4132                         &data_sinfo->total_bytes_pinned,
4133                         used + bytes - data_sinfo->total_bytes);
4134                 spin_unlock(&data_sinfo->lock);
4135
4136                 /* commit the current transaction and try again */
4137 commit_trans:
4138                 if (need_commit &&
4139                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
4140                         need_commit--;
4141
4142                         if (need_commit > 0)
4143                                 btrfs_wait_ordered_roots(fs_info, -1);
4144
4145                         trans = btrfs_join_transaction(root);
4146                         if (IS_ERR(trans))
4147                                 return PTR_ERR(trans);
4148                         if (have_pinned_space >= 0 ||
4149                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4150                                      &trans->transaction->flags) ||
4151                             need_commit > 0) {
4152                                 ret = btrfs_commit_transaction(trans, root);
4153                                 if (ret)
4154                                         return ret;
4155                                 /*
4156                                  * make sure that all running delayed iput are
4157                                  * done
4158                                  */
4159                                 down_write(&root->fs_info->delayed_iput_sem);
4160                                 up_write(&root->fs_info->delayed_iput_sem);
4161                                 goto again;
4162                         } else {
4163                                 btrfs_end_transaction(trans, root);
4164                         }
4165                 }
4166
4167                 trace_btrfs_space_reservation(root->fs_info,
4168                                               "space_info:enospc",
4169                                               data_sinfo->flags, bytes, 1);
4170                 return -ENOSPC;
4171         }
4172         data_sinfo->bytes_may_use += bytes;
4173         trace_btrfs_space_reservation(root->fs_info, "space_info",
4174                                       data_sinfo->flags, bytes, 1);
4175         spin_unlock(&data_sinfo->lock);
4176
4177         return ret;
4178 }
4179
4180 /*
4181  * New check_data_free_space() with ability for precious data reservation
4182  * Will replace old btrfs_check_data_free_space(), but for patch split,
4183  * add a new function first and then replace it.
4184  */
4185 int btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4186 {
4187         struct btrfs_root *root = BTRFS_I(inode)->root;
4188         int ret;
4189
4190         /* align the range */
4191         len = round_up(start + len, root->sectorsize) -
4192               round_down(start, root->sectorsize);
4193         start = round_down(start, root->sectorsize);
4194
4195         ret = btrfs_alloc_data_chunk_ondemand(inode, len);
4196         if (ret < 0)
4197                 return ret;
4198
4199         /*
4200          * Use new btrfs_qgroup_reserve_data to reserve precious data space
4201          *
4202          * TODO: Find a good method to avoid reserve data space for NOCOW
4203          * range, but don't impact performance on quota disable case.
4204          */
4205         ret = btrfs_qgroup_reserve_data(inode, start, len);
4206         return ret;
4207 }
4208
4209 /*
4210  * Called if we need to clear a data reservation for this inode
4211  * Normally in a error case.
4212  *
4213  * This one will *NOT* use accurate qgroup reserved space API, just for case
4214  * which we can't sleep and is sure it won't affect qgroup reserved space.
4215  * Like clear_bit_hook().
4216  */
4217 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4218                                             u64 len)
4219 {
4220         struct btrfs_root *root = BTRFS_I(inode)->root;
4221         struct btrfs_space_info *data_sinfo;
4222
4223         /* Make sure the range is aligned to sectorsize */
4224         len = round_up(start + len, root->sectorsize) -
4225               round_down(start, root->sectorsize);
4226         start = round_down(start, root->sectorsize);
4227
4228         data_sinfo = root->fs_info->data_sinfo;
4229         spin_lock(&data_sinfo->lock);
4230         if (WARN_ON(data_sinfo->bytes_may_use < len))
4231                 data_sinfo->bytes_may_use = 0;
4232         else
4233                 data_sinfo->bytes_may_use -= len;
4234         trace_btrfs_space_reservation(root->fs_info, "space_info",
4235                                       data_sinfo->flags, len, 0);
4236         spin_unlock(&data_sinfo->lock);
4237 }
4238
4239 /*
4240  * Called if we need to clear a data reservation for this inode
4241  * Normally in a error case.
4242  *
4243  * This one will handle the per-indoe data rsv map for accurate reserved
4244  * space framework.
4245  */
4246 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4247 {
4248         btrfs_free_reserved_data_space_noquota(inode, start, len);
4249         btrfs_qgroup_free_data(inode, start, len);
4250 }
4251
4252 static void force_metadata_allocation(struct btrfs_fs_info *info)
4253 {
4254         struct list_head *head = &info->space_info;
4255         struct btrfs_space_info *found;
4256
4257         rcu_read_lock();
4258         list_for_each_entry_rcu(found, head, list) {
4259                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4260                         found->force_alloc = CHUNK_ALLOC_FORCE;
4261         }
4262         rcu_read_unlock();
4263 }
4264
4265 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4266 {
4267         return (global->size << 1);
4268 }
4269
4270 static int should_alloc_chunk(struct btrfs_root *root,
4271                               struct btrfs_space_info *sinfo, int force)
4272 {
4273         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4274         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4275         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4276         u64 thresh;
4277
4278         if (force == CHUNK_ALLOC_FORCE)
4279                 return 1;
4280
4281         /*
4282          * We need to take into account the global rsv because for all intents
4283          * and purposes it's used space.  Don't worry about locking the
4284          * global_rsv, it doesn't change except when the transaction commits.
4285          */
4286         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4287                 num_allocated += calc_global_rsv_need_space(global_rsv);
4288
4289         /*
4290          * in limited mode, we want to have some free space up to
4291          * about 1% of the FS size.
4292          */
4293         if (force == CHUNK_ALLOC_LIMITED) {
4294                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4295                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4296
4297                 if (num_bytes - num_allocated < thresh)
4298                         return 1;
4299         }
4300
4301         if (num_allocated + SZ_2M < div_factor(num_bytes, 8))
4302                 return 0;
4303         return 1;
4304 }
4305
4306 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4307 {
4308         u64 num_dev;
4309
4310         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4311                     BTRFS_BLOCK_GROUP_RAID0 |
4312                     BTRFS_BLOCK_GROUP_RAID5 |
4313                     BTRFS_BLOCK_GROUP_RAID6))
4314                 num_dev = root->fs_info->fs_devices->rw_devices;
4315         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4316                 num_dev = 2;
4317         else
4318                 num_dev = 1;    /* DUP or single */
4319
4320         return num_dev;
4321 }
4322
4323 /*
4324  * If @is_allocation is true, reserve space in the system space info necessary
4325  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4326  * removing a chunk.
4327  */
4328 void check_system_chunk(struct btrfs_trans_handle *trans,
4329                         struct btrfs_root *root,
4330                         u64 type)
4331 {
4332         struct btrfs_space_info *info;
4333         u64 left;
4334         u64 thresh;
4335         int ret = 0;
4336         u64 num_devs;
4337
4338         /*
4339          * Needed because we can end up allocating a system chunk and for an
4340          * atomic and race free space reservation in the chunk block reserve.
4341          */
4342         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4343
4344         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4345         spin_lock(&info->lock);
4346         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4347                 info->bytes_reserved - info->bytes_readonly -
4348                 info->bytes_may_use;
4349         spin_unlock(&info->lock);
4350
4351         num_devs = get_profile_num_devs(root, type);
4352
4353         /* num_devs device items to update and 1 chunk item to add or remove */
4354         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4355                 btrfs_calc_trans_metadata_size(root, 1);
4356
4357         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4358                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4359                         left, thresh, type);
4360                 dump_space_info(info, 0, 0);
4361         }
4362
4363         if (left < thresh) {
4364                 u64 flags;
4365
4366                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4367                 /*
4368                  * Ignore failure to create system chunk. We might end up not
4369                  * needing it, as we might not need to COW all nodes/leafs from
4370                  * the paths we visit in the chunk tree (they were already COWed
4371                  * or created in the current transaction for example).
4372                  */
4373                 ret = btrfs_alloc_chunk(trans, root, flags);
4374         }
4375
4376         if (!ret) {
4377                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4378                                           &root->fs_info->chunk_block_rsv,
4379                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4380                 if (!ret)
4381                         trans->chunk_bytes_reserved += thresh;
4382         }
4383 }
4384
4385 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4386                           struct btrfs_root *extent_root, u64 flags, int force)
4387 {
4388         struct btrfs_space_info *space_info;
4389         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4390         int wait_for_alloc = 0;
4391         int ret = 0;
4392
4393         /* Don't re-enter if we're already allocating a chunk */
4394         if (trans->allocating_chunk)
4395                 return -ENOSPC;
4396
4397         space_info = __find_space_info(extent_root->fs_info, flags);
4398         if (!space_info) {
4399                 ret = update_space_info(extent_root->fs_info, flags,
4400                                         0, 0, &space_info);
4401                 BUG_ON(ret); /* -ENOMEM */
4402         }
4403         BUG_ON(!space_info); /* Logic error */
4404
4405 again:
4406         spin_lock(&space_info->lock);
4407         if (force < space_info->force_alloc)
4408                 force = space_info->force_alloc;
4409         if (space_info->full) {
4410                 if (should_alloc_chunk(extent_root, space_info, force))
4411                         ret = -ENOSPC;
4412                 else
4413                         ret = 0;
4414                 spin_unlock(&space_info->lock);
4415                 return ret;
4416         }
4417
4418         if (!should_alloc_chunk(extent_root, space_info, force)) {
4419                 spin_unlock(&space_info->lock);
4420                 return 0;
4421         } else if (space_info->chunk_alloc) {
4422                 wait_for_alloc = 1;
4423         } else {
4424                 space_info->chunk_alloc = 1;
4425         }
4426
4427         spin_unlock(&space_info->lock);
4428
4429         mutex_lock(&fs_info->chunk_mutex);
4430
4431         /*
4432          * The chunk_mutex is held throughout the entirety of a chunk
4433          * allocation, so once we've acquired the chunk_mutex we know that the
4434          * other guy is done and we need to recheck and see if we should
4435          * allocate.
4436          */
4437         if (wait_for_alloc) {
4438                 mutex_unlock(&fs_info->chunk_mutex);
4439                 wait_for_alloc = 0;
4440                 goto again;
4441         }
4442
4443         trans->allocating_chunk = true;
4444
4445         /*
4446          * If we have mixed data/metadata chunks we want to make sure we keep
4447          * allocating mixed chunks instead of individual chunks.
4448          */
4449         if (btrfs_mixed_space_info(space_info))
4450                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4451
4452         /*
4453          * if we're doing a data chunk, go ahead and make sure that
4454          * we keep a reasonable number of metadata chunks allocated in the
4455          * FS as well.
4456          */
4457         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4458                 fs_info->data_chunk_allocations++;
4459                 if (!(fs_info->data_chunk_allocations %
4460                       fs_info->metadata_ratio))
4461                         force_metadata_allocation(fs_info);
4462         }
4463
4464         /*
4465          * Check if we have enough space in SYSTEM chunk because we may need
4466          * to update devices.
4467          */
4468         check_system_chunk(trans, extent_root, flags);
4469
4470         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4471         trans->allocating_chunk = false;
4472
4473         spin_lock(&space_info->lock);
4474         if (ret < 0 && ret != -ENOSPC)
4475                 goto out;
4476         if (ret)
4477                 space_info->full = 1;
4478         else
4479                 ret = 1;
4480
4481         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4482 out:
4483         space_info->chunk_alloc = 0;
4484         spin_unlock(&space_info->lock);
4485         mutex_unlock(&fs_info->chunk_mutex);
4486         /*
4487          * When we allocate a new chunk we reserve space in the chunk block
4488          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4489          * add new nodes/leafs to it if we end up needing to do it when
4490          * inserting the chunk item and updating device items as part of the
4491          * second phase of chunk allocation, performed by
4492          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4493          * large number of new block groups to create in our transaction
4494          * handle's new_bgs list to avoid exhausting the chunk block reserve
4495          * in extreme cases - like having a single transaction create many new
4496          * block groups when starting to write out the free space caches of all
4497          * the block groups that were made dirty during the lifetime of the
4498          * transaction.
4499          */
4500         if (trans->can_flush_pending_bgs &&
4501             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4502                 btrfs_create_pending_block_groups(trans, trans->root);
4503                 btrfs_trans_release_chunk_metadata(trans);
4504         }
4505         return ret;
4506 }
4507
4508 static int can_overcommit(struct btrfs_root *root,
4509                           struct btrfs_space_info *space_info, u64 bytes,
4510                           enum btrfs_reserve_flush_enum flush)
4511 {
4512         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4513         u64 profile = btrfs_get_alloc_profile(root, 0);
4514         u64 space_size;
4515         u64 avail;
4516         u64 used;
4517
4518         used = space_info->bytes_used + space_info->bytes_reserved +
4519                 space_info->bytes_pinned + space_info->bytes_readonly;
4520
4521         /*
4522          * We only want to allow over committing if we have lots of actual space
4523          * free, but if we don't have enough space to handle the global reserve
4524          * space then we could end up having a real enospc problem when trying
4525          * to allocate a chunk or some other such important allocation.
4526          */
4527         spin_lock(&global_rsv->lock);
4528         space_size = calc_global_rsv_need_space(global_rsv);
4529         spin_unlock(&global_rsv->lock);
4530         if (used + space_size >= space_info->total_bytes)
4531                 return 0;
4532
4533         used += space_info->bytes_may_use;
4534
4535         spin_lock(&root->fs_info->free_chunk_lock);
4536         avail = root->fs_info->free_chunk_space;
4537         spin_unlock(&root->fs_info->free_chunk_lock);
4538
4539         /*
4540          * If we have dup, raid1 or raid10 then only half of the free
4541          * space is actually useable.  For raid56, the space info used
4542          * doesn't include the parity drive, so we don't have to
4543          * change the math
4544          */
4545         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4546                        BTRFS_BLOCK_GROUP_RAID1 |
4547                        BTRFS_BLOCK_GROUP_RAID10))
4548                 avail >>= 1;
4549
4550         /*
4551          * If we aren't flushing all things, let us overcommit up to
4552          * 1/2th of the space. If we can flush, don't let us overcommit
4553          * too much, let it overcommit up to 1/8 of the space.
4554          */
4555         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4556                 avail >>= 3;
4557         else
4558                 avail >>= 1;
4559
4560         if (used + bytes < space_info->total_bytes + avail)
4561                 return 1;
4562         return 0;
4563 }
4564
4565 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4566                                          unsigned long nr_pages, int nr_items)
4567 {
4568         struct super_block *sb = root->fs_info->sb;
4569
4570         if (down_read_trylock(&sb->s_umount)) {
4571                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4572                 up_read(&sb->s_umount);
4573         } else {
4574                 /*
4575                  * We needn't worry the filesystem going from r/w to r/o though
4576                  * we don't acquire ->s_umount mutex, because the filesystem
4577                  * should guarantee the delalloc inodes list be empty after
4578                  * the filesystem is readonly(all dirty pages are written to
4579                  * the disk).
4580                  */
4581                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4582                 if (!current->journal_info)
4583                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4584         }
4585 }
4586
4587 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4588 {
4589         u64 bytes;
4590         int nr;
4591
4592         bytes = btrfs_calc_trans_metadata_size(root, 1);
4593         nr = (int)div64_u64(to_reclaim, bytes);
4594         if (!nr)
4595                 nr = 1;
4596         return nr;
4597 }
4598
4599 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4600
4601 /*
4602  * shrink metadata reservation for delalloc
4603  */
4604 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4605                             bool wait_ordered)
4606 {
4607         struct btrfs_block_rsv *block_rsv;
4608         struct btrfs_space_info *space_info;
4609         struct btrfs_trans_handle *trans;
4610         u64 delalloc_bytes;
4611         u64 max_reclaim;
4612         long time_left;
4613         unsigned long nr_pages;
4614         int loops;
4615         int items;
4616         enum btrfs_reserve_flush_enum flush;
4617
4618         /* Calc the number of the pages we need flush for space reservation */
4619         items = calc_reclaim_items_nr(root, to_reclaim);
4620         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4621
4622         trans = (struct btrfs_trans_handle *)current->journal_info;
4623         block_rsv = &root->fs_info->delalloc_block_rsv;
4624         space_info = block_rsv->space_info;
4625
4626         delalloc_bytes = percpu_counter_sum_positive(
4627                                                 &root->fs_info->delalloc_bytes);
4628         if (delalloc_bytes == 0) {
4629                 if (trans)
4630                         return;
4631                 if (wait_ordered)
4632                         btrfs_wait_ordered_roots(root->fs_info, items);
4633                 return;
4634         }
4635
4636         loops = 0;
4637         while (delalloc_bytes && loops < 3) {
4638                 max_reclaim = min(delalloc_bytes, to_reclaim);
4639                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4640                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4641                 /*
4642                  * We need to wait for the async pages to actually start before
4643                  * we do anything.
4644                  */
4645                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4646                 if (!max_reclaim)
4647                         goto skip_async;
4648
4649                 if (max_reclaim <= nr_pages)
4650                         max_reclaim = 0;
4651                 else
4652                         max_reclaim -= nr_pages;
4653
4654                 wait_event(root->fs_info->async_submit_wait,
4655                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4656                            (int)max_reclaim);
4657 skip_async:
4658                 if (!trans)
4659                         flush = BTRFS_RESERVE_FLUSH_ALL;
4660                 else
4661                         flush = BTRFS_RESERVE_NO_FLUSH;
4662                 spin_lock(&space_info->lock);
4663                 if (can_overcommit(root, space_info, orig, flush)) {
4664                         spin_unlock(&space_info->lock);
4665                         break;
4666                 }
4667                 spin_unlock(&space_info->lock);
4668
4669                 loops++;
4670                 if (wait_ordered && !trans) {
4671                         btrfs_wait_ordered_roots(root->fs_info, items);
4672                 } else {
4673                         time_left = schedule_timeout_killable(1);
4674                         if (time_left)
4675                                 break;
4676                 }
4677                 delalloc_bytes = percpu_counter_sum_positive(
4678                                                 &root->fs_info->delalloc_bytes);
4679         }
4680 }
4681
4682 /**
4683  * maybe_commit_transaction - possibly commit the transaction if its ok to
4684  * @root - the root we're allocating for
4685  * @bytes - the number of bytes we want to reserve
4686  * @force - force the commit
4687  *
4688  * This will check to make sure that committing the transaction will actually
4689  * get us somewhere and then commit the transaction if it does.  Otherwise it
4690  * will return -ENOSPC.
4691  */
4692 static int may_commit_transaction(struct btrfs_root *root,
4693                                   struct btrfs_space_info *space_info,
4694                                   u64 bytes, int force)
4695 {
4696         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4697         struct btrfs_trans_handle *trans;
4698
4699         trans = (struct btrfs_trans_handle *)current->journal_info;
4700         if (trans)
4701                 return -EAGAIN;
4702
4703         if (force)
4704                 goto commit;
4705
4706         /* See if there is enough pinned space to make this reservation */
4707         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4708                                    bytes) >= 0)
4709                 goto commit;
4710
4711         /*
4712          * See if there is some space in the delayed insertion reservation for
4713          * this reservation.
4714          */
4715         if (space_info != delayed_rsv->space_info)
4716                 return -ENOSPC;
4717
4718         spin_lock(&delayed_rsv->lock);
4719         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4720                                    bytes - delayed_rsv->size) >= 0) {
4721                 spin_unlock(&delayed_rsv->lock);
4722                 return -ENOSPC;
4723         }
4724         spin_unlock(&delayed_rsv->lock);
4725
4726 commit:
4727         trans = btrfs_join_transaction(root);
4728         if (IS_ERR(trans))
4729                 return -ENOSPC;
4730
4731         return btrfs_commit_transaction(trans, root);
4732 }
4733
4734 enum flush_state {
4735         FLUSH_DELAYED_ITEMS_NR  =       1,
4736         FLUSH_DELAYED_ITEMS     =       2,
4737         FLUSH_DELALLOC          =       3,
4738         FLUSH_DELALLOC_WAIT     =       4,
4739         ALLOC_CHUNK             =       5,
4740         COMMIT_TRANS            =       6,
4741 };
4742
4743 static int flush_space(struct btrfs_root *root,
4744                        struct btrfs_space_info *space_info, u64 num_bytes,
4745                        u64 orig_bytes, int state)
4746 {
4747         struct btrfs_trans_handle *trans;
4748         int nr;
4749         int ret = 0;
4750
4751         switch (state) {
4752         case FLUSH_DELAYED_ITEMS_NR:
4753         case FLUSH_DELAYED_ITEMS:
4754                 if (state == FLUSH_DELAYED_ITEMS_NR)
4755                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4756                 else
4757                         nr = -1;
4758
4759                 trans = btrfs_join_transaction(root);
4760                 if (IS_ERR(trans)) {
4761                         ret = PTR_ERR(trans);
4762                         break;
4763                 }
4764                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4765                 btrfs_end_transaction(trans, root);
4766                 break;
4767         case FLUSH_DELALLOC:
4768         case FLUSH_DELALLOC_WAIT:
4769                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4770                                 state == FLUSH_DELALLOC_WAIT);
4771                 break;
4772         case ALLOC_CHUNK:
4773                 trans = btrfs_join_transaction(root);
4774                 if (IS_ERR(trans)) {
4775                         ret = PTR_ERR(trans);
4776                         break;
4777                 }
4778                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4779                                      btrfs_get_alloc_profile(root, 0),
4780                                      CHUNK_ALLOC_NO_FORCE);
4781                 btrfs_end_transaction(trans, root);
4782                 if (ret == -ENOSPC)
4783                         ret = 0;
4784                 break;
4785         case COMMIT_TRANS:
4786                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4787                 break;
4788         default:
4789                 ret = -ENOSPC;
4790                 break;
4791         }
4792
4793         return ret;
4794 }
4795
4796 static inline u64
4797 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4798                                  struct btrfs_space_info *space_info)
4799 {
4800         u64 used;
4801         u64 expected;
4802         u64 to_reclaim;
4803
4804         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4805         spin_lock(&space_info->lock);
4806         if (can_overcommit(root, space_info, to_reclaim,
4807                            BTRFS_RESERVE_FLUSH_ALL)) {
4808                 to_reclaim = 0;
4809                 goto out;
4810         }
4811
4812         used = space_info->bytes_used + space_info->bytes_reserved +
4813                space_info->bytes_pinned + space_info->bytes_readonly +
4814                space_info->bytes_may_use;
4815         if (can_overcommit(root, space_info, SZ_1M, BTRFS_RESERVE_FLUSH_ALL))
4816                 expected = div_factor_fine(space_info->total_bytes, 95);
4817         else
4818                 expected = div_factor_fine(space_info->total_bytes, 90);
4819
4820         if (used > expected)
4821                 to_reclaim = used - expected;
4822         else
4823                 to_reclaim = 0;
4824         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4825                                      space_info->bytes_reserved);
4826 out:
4827         spin_unlock(&space_info->lock);
4828
4829         return to_reclaim;
4830 }
4831
4832 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4833                                         struct btrfs_fs_info *fs_info, u64 used)
4834 {
4835         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4836
4837         /* If we're just plain full then async reclaim just slows us down. */
4838         if (space_info->bytes_used >= thresh)
4839                 return 0;
4840
4841         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4842                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4843 }
4844
4845 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4846                                        struct btrfs_fs_info *fs_info,
4847                                        int flush_state)
4848 {
4849         u64 used;
4850
4851         spin_lock(&space_info->lock);
4852         /*
4853          * We run out of space and have not got any free space via flush_space,
4854          * so don't bother doing async reclaim.
4855          */
4856         if (flush_state > COMMIT_TRANS && space_info->full) {
4857                 spin_unlock(&space_info->lock);
4858                 return 0;
4859         }
4860
4861         used = space_info->bytes_used + space_info->bytes_reserved +
4862                space_info->bytes_pinned + space_info->bytes_readonly +
4863                space_info->bytes_may_use;
4864         if (need_do_async_reclaim(space_info, fs_info, used)) {
4865                 spin_unlock(&space_info->lock);
4866                 return 1;
4867         }
4868         spin_unlock(&space_info->lock);
4869
4870         return 0;
4871 }
4872
4873 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4874 {
4875         struct btrfs_fs_info *fs_info;
4876         struct btrfs_space_info *space_info;
4877         u64 to_reclaim;
4878         int flush_state;
4879
4880         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4881         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4882
4883         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4884                                                       space_info);
4885         if (!to_reclaim)
4886                 return;
4887
4888         flush_state = FLUSH_DELAYED_ITEMS_NR;
4889         do {
4890                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4891                             to_reclaim, flush_state);
4892                 flush_state++;
4893                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4894                                                  flush_state))
4895                         return;
4896         } while (flush_state < COMMIT_TRANS);
4897 }
4898
4899 void btrfs_init_async_reclaim_work(struct work_struct *work)
4900 {
4901         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4902 }
4903
4904 /**
4905  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4906  * @root - the root we're allocating for
4907  * @block_rsv - the block_rsv we're allocating for
4908  * @orig_bytes - the number of bytes we want
4909  * @flush - whether or not we can flush to make our reservation
4910  *
4911  * This will reserve orgi_bytes number of bytes from the space info associated
4912  * with the block_rsv.  If there is not enough space it will make an attempt to
4913  * flush out space to make room.  It will do this by flushing delalloc if
4914  * possible or committing the transaction.  If flush is 0 then no attempts to
4915  * regain reservations will be made and this will fail if there is not enough
4916  * space already.
4917  */
4918 static int reserve_metadata_bytes(struct btrfs_root *root,
4919                                   struct btrfs_block_rsv *block_rsv,
4920                                   u64 orig_bytes,
4921                                   enum btrfs_reserve_flush_enum flush)
4922 {
4923         struct btrfs_space_info *space_info = block_rsv->space_info;
4924         u64 used;
4925         u64 num_bytes = orig_bytes;
4926         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4927         int ret = 0;
4928         bool flushing = false;
4929
4930 again:
4931         ret = 0;
4932         spin_lock(&space_info->lock);
4933         /*
4934          * We only want to wait if somebody other than us is flushing and we
4935          * are actually allowed to flush all things.
4936          */
4937         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4938                space_info->flush) {
4939                 spin_unlock(&space_info->lock);
4940                 /*
4941                  * If we have a trans handle we can't wait because the flusher
4942                  * may have to commit the transaction, which would mean we would
4943                  * deadlock since we are waiting for the flusher to finish, but
4944                  * hold the current transaction open.
4945                  */
4946                 if (current->journal_info)
4947                         return -EAGAIN;
4948                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4949                 /* Must have been killed, return */
4950                 if (ret)
4951                         return -EINTR;
4952
4953                 spin_lock(&space_info->lock);
4954         }
4955
4956         ret = -ENOSPC;
4957         used = space_info->bytes_used + space_info->bytes_reserved +
4958                 space_info->bytes_pinned + space_info->bytes_readonly +
4959                 space_info->bytes_may_use;
4960
4961         /*
4962          * The idea here is that we've not already over-reserved the block group
4963          * then we can go ahead and save our reservation first and then start
4964          * flushing if we need to.  Otherwise if we've already overcommitted
4965          * lets start flushing stuff first and then come back and try to make
4966          * our reservation.
4967          */
4968         if (used <= space_info->total_bytes) {
4969                 if (used + orig_bytes <= space_info->total_bytes) {
4970                         space_info->bytes_may_use += orig_bytes;
4971                         trace_btrfs_space_reservation(root->fs_info,
4972                                 "space_info", space_info->flags, orig_bytes, 1);
4973                         ret = 0;
4974                 } else {
4975                         /*
4976                          * Ok set num_bytes to orig_bytes since we aren't
4977                          * overocmmitted, this way we only try and reclaim what
4978                          * we need.
4979                          */
4980                         num_bytes = orig_bytes;
4981                 }
4982         } else {
4983                 /*
4984                  * Ok we're over committed, set num_bytes to the overcommitted
4985                  * amount plus the amount of bytes that we need for this
4986                  * reservation.
4987                  */
4988                 num_bytes = used - space_info->total_bytes +
4989                         (orig_bytes * 2);
4990         }
4991
4992         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4993                 space_info->bytes_may_use += orig_bytes;
4994                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4995                                               space_info->flags, orig_bytes,
4996                                               1);
4997                 ret = 0;
4998         }
4999
5000         /*
5001          * Couldn't make our reservation, save our place so while we're trying
5002          * to reclaim space we can actually use it instead of somebody else
5003          * stealing it from us.
5004          *
5005          * We make the other tasks wait for the flush only when we can flush
5006          * all things.
5007          */
5008         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5009                 flushing = true;
5010                 space_info->flush = 1;
5011         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5012                 used += orig_bytes;
5013                 /*
5014                  * We will do the space reservation dance during log replay,
5015                  * which means we won't have fs_info->fs_root set, so don't do
5016                  * the async reclaim as we will panic.
5017                  */
5018                 if (!root->fs_info->log_root_recovering &&
5019                     need_do_async_reclaim(space_info, root->fs_info, used) &&
5020                     !work_busy(&root->fs_info->async_reclaim_work))
5021                         queue_work(system_unbound_wq,
5022                                    &root->fs_info->async_reclaim_work);
5023         }
5024         spin_unlock(&space_info->lock);
5025
5026         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5027                 goto out;
5028
5029         ret = flush_space(root, space_info, num_bytes, orig_bytes,
5030                           flush_state);
5031         flush_state++;
5032
5033         /*
5034          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
5035          * would happen. So skip delalloc flush.
5036          */
5037         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
5038             (flush_state == FLUSH_DELALLOC ||
5039              flush_state == FLUSH_DELALLOC_WAIT))
5040                 flush_state = ALLOC_CHUNK;
5041
5042         if (!ret)
5043                 goto again;
5044         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
5045                  flush_state < COMMIT_TRANS)
5046                 goto again;
5047         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
5048                  flush_state <= COMMIT_TRANS)
5049                 goto again;
5050
5051 out:
5052         if (ret == -ENOSPC &&
5053             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5054                 struct btrfs_block_rsv *global_rsv =
5055                         &root->fs_info->global_block_rsv;
5056
5057                 if (block_rsv != global_rsv &&
5058                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5059                         ret = 0;
5060         }
5061         if (ret == -ENOSPC)
5062                 trace_btrfs_space_reservation(root->fs_info,
5063                                               "space_info:enospc",
5064                                               space_info->flags, orig_bytes, 1);
5065         if (flushing) {
5066                 spin_lock(&space_info->lock);
5067                 space_info->flush = 0;
5068                 wake_up_all(&space_info->wait);
5069                 spin_unlock(&space_info->lock);
5070         }
5071         return ret;
5072 }
5073
5074 static struct btrfs_block_rsv *get_block_rsv(
5075                                         const struct btrfs_trans_handle *trans,
5076                                         const struct btrfs_root *root)
5077 {
5078         struct btrfs_block_rsv *block_rsv = NULL;
5079
5080         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5081             (root == root->fs_info->csum_root && trans->adding_csums) ||
5082              (root == root->fs_info->uuid_root))
5083                 block_rsv = trans->block_rsv;
5084
5085         if (!block_rsv)
5086                 block_rsv = root->block_rsv;
5087
5088         if (!block_rsv)
5089                 block_rsv = &root->fs_info->empty_block_rsv;
5090
5091         return block_rsv;
5092 }
5093
5094 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5095                                u64 num_bytes)
5096 {
5097         int ret = -ENOSPC;
5098         spin_lock(&block_rsv->lock);
5099         if (block_rsv->reserved >= num_bytes) {
5100                 block_rsv->reserved -= num_bytes;
5101                 if (block_rsv->reserved < block_rsv->size)
5102                         block_rsv->full = 0;
5103                 ret = 0;
5104         }
5105         spin_unlock(&block_rsv->lock);
5106         return ret;
5107 }
5108
5109 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5110                                 u64 num_bytes, int update_size)
5111 {
5112         spin_lock(&block_rsv->lock);
5113         block_rsv->reserved += num_bytes;
5114         if (update_size)
5115                 block_rsv->size += num_bytes;
5116         else if (block_rsv->reserved >= block_rsv->size)
5117                 block_rsv->full = 1;
5118         spin_unlock(&block_rsv->lock);
5119 }
5120
5121 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5122                              struct btrfs_block_rsv *dest, u64 num_bytes,
5123                              int min_factor)
5124 {
5125         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5126         u64 min_bytes;
5127
5128         if (global_rsv->space_info != dest->space_info)
5129                 return -ENOSPC;
5130
5131         spin_lock(&global_rsv->lock);
5132         min_bytes = div_factor(global_rsv->size, min_factor);
5133         if (global_rsv->reserved < min_bytes + num_bytes) {
5134                 spin_unlock(&global_rsv->lock);
5135                 return -ENOSPC;
5136         }
5137         global_rsv->reserved -= num_bytes;
5138         if (global_rsv->reserved < global_rsv->size)
5139                 global_rsv->full = 0;
5140         spin_unlock(&global_rsv->lock);
5141
5142         block_rsv_add_bytes(dest, num_bytes, 1);
5143         return 0;
5144 }
5145
5146 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5147                                     struct btrfs_block_rsv *block_rsv,
5148                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5149 {
5150         struct btrfs_space_info *space_info = block_rsv->space_info;
5151
5152         spin_lock(&block_rsv->lock);
5153         if (num_bytes == (u64)-1)
5154                 num_bytes = block_rsv->size;
5155         block_rsv->size -= num_bytes;
5156         if (block_rsv->reserved >= block_rsv->size) {
5157                 num_bytes = block_rsv->reserved - block_rsv->size;
5158                 block_rsv->reserved = block_rsv->size;
5159                 block_rsv->full = 1;
5160         } else {
5161                 num_bytes = 0;
5162         }
5163         spin_unlock(&block_rsv->lock);
5164
5165         if (num_bytes > 0) {
5166                 if (dest) {
5167                         spin_lock(&dest->lock);
5168                         if (!dest->full) {
5169                                 u64 bytes_to_add;
5170
5171                                 bytes_to_add = dest->size - dest->reserved;
5172                                 bytes_to_add = min(num_bytes, bytes_to_add);
5173                                 dest->reserved += bytes_to_add;
5174                                 if (dest->reserved >= dest->size)
5175                                         dest->full = 1;
5176                                 num_bytes -= bytes_to_add;
5177                         }
5178                         spin_unlock(&dest->lock);
5179                 }
5180                 if (num_bytes) {
5181                         spin_lock(&space_info->lock);
5182                         space_info->bytes_may_use -= num_bytes;
5183                         trace_btrfs_space_reservation(fs_info, "space_info",
5184                                         space_info->flags, num_bytes, 0);
5185                         spin_unlock(&space_info->lock);
5186                 }
5187         }
5188 }
5189
5190 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
5191                                    struct btrfs_block_rsv *dst, u64 num_bytes)
5192 {
5193         int ret;
5194
5195         ret = block_rsv_use_bytes(src, num_bytes);
5196         if (ret)
5197                 return ret;
5198
5199         block_rsv_add_bytes(dst, num_bytes, 1);
5200         return 0;
5201 }
5202
5203 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5204 {
5205         memset(rsv, 0, sizeof(*rsv));
5206         spin_lock_init(&rsv->lock);
5207         rsv->type = type;
5208 }
5209
5210 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5211                                               unsigned short type)
5212 {
5213         struct btrfs_block_rsv *block_rsv;
5214         struct btrfs_fs_info *fs_info = root->fs_info;
5215
5216         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5217         if (!block_rsv)
5218                 return NULL;
5219
5220         btrfs_init_block_rsv(block_rsv, type);
5221         block_rsv->space_info = __find_space_info(fs_info,
5222                                                   BTRFS_BLOCK_GROUP_METADATA);
5223         return block_rsv;
5224 }
5225
5226 void btrfs_free_block_rsv(struct btrfs_root *root,
5227                           struct btrfs_block_rsv *rsv)
5228 {
5229         if (!rsv)
5230                 return;
5231         btrfs_block_rsv_release(root, rsv, (u64)-1);
5232         kfree(rsv);
5233 }
5234
5235 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5236 {
5237         kfree(rsv);
5238 }
5239
5240 int btrfs_block_rsv_add(struct btrfs_root *root,
5241                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5242                         enum btrfs_reserve_flush_enum flush)
5243 {
5244         int ret;
5245
5246         if (num_bytes == 0)
5247                 return 0;
5248
5249         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5250         if (!ret) {
5251                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5252                 return 0;
5253         }
5254
5255         return ret;
5256 }
5257
5258 int btrfs_block_rsv_check(struct btrfs_root *root,
5259                           struct btrfs_block_rsv *block_rsv, int min_factor)
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 = div_factor(block_rsv->size, min_factor);
5269         if (block_rsv->reserved >= num_bytes)
5270                 ret = 0;
5271         spin_unlock(&block_rsv->lock);
5272
5273         return ret;
5274 }
5275
5276 int btrfs_block_rsv_refill(struct btrfs_root *root,
5277                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5278                            enum btrfs_reserve_flush_enum flush)
5279 {
5280         u64 num_bytes = 0;
5281         int ret = -ENOSPC;
5282
5283         if (!block_rsv)
5284                 return 0;
5285
5286         spin_lock(&block_rsv->lock);
5287         num_bytes = min_reserved;
5288         if (block_rsv->reserved >= num_bytes)
5289                 ret = 0;
5290         else
5291                 num_bytes -= block_rsv->reserved;
5292         spin_unlock(&block_rsv->lock);
5293
5294         if (!ret)
5295                 return 0;
5296
5297         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5298         if (!ret) {
5299                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5300                 return 0;
5301         }
5302
5303         return ret;
5304 }
5305
5306 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5307                             struct btrfs_block_rsv *dst_rsv,
5308                             u64 num_bytes)
5309 {
5310         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5311 }
5312
5313 void btrfs_block_rsv_release(struct btrfs_root *root,
5314                              struct btrfs_block_rsv *block_rsv,
5315                              u64 num_bytes)
5316 {
5317         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5318         if (global_rsv == block_rsv ||
5319             block_rsv->space_info != global_rsv->space_info)
5320                 global_rsv = NULL;
5321         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5322                                 num_bytes);
5323 }
5324
5325 /*
5326  * helper to calculate size of global block reservation.
5327  * the desired value is sum of space used by extent tree,
5328  * checksum tree and root tree
5329  */
5330 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5331 {
5332         struct btrfs_space_info *sinfo;
5333         u64 num_bytes;
5334         u64 meta_used;
5335         u64 data_used;
5336         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5337
5338         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5339         spin_lock(&sinfo->lock);
5340         data_used = sinfo->bytes_used;
5341         spin_unlock(&sinfo->lock);
5342
5343         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5344         spin_lock(&sinfo->lock);
5345         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5346                 data_used = 0;
5347         meta_used = sinfo->bytes_used;
5348         spin_unlock(&sinfo->lock);
5349
5350         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5351                     csum_size * 2;
5352         num_bytes += div_u64(data_used + meta_used, 50);
5353
5354         if (num_bytes * 3 > meta_used)
5355                 num_bytes = div_u64(meta_used, 3);
5356
5357         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5358 }
5359
5360 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5361 {
5362         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5363         struct btrfs_space_info *sinfo = block_rsv->space_info;
5364         u64 num_bytes;
5365
5366         num_bytes = calc_global_metadata_size(fs_info);
5367
5368         spin_lock(&sinfo->lock);
5369         spin_lock(&block_rsv->lock);
5370
5371         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5372
5373         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5374                     sinfo->bytes_reserved + sinfo->bytes_readonly +
5375                     sinfo->bytes_may_use;
5376
5377         if (sinfo->total_bytes > num_bytes) {
5378                 num_bytes = sinfo->total_bytes - num_bytes;
5379                 block_rsv->reserved += num_bytes;
5380                 sinfo->bytes_may_use += num_bytes;
5381                 trace_btrfs_space_reservation(fs_info, "space_info",
5382                                       sinfo->flags, num_bytes, 1);
5383         }
5384
5385         if (block_rsv->reserved >= block_rsv->size) {
5386                 num_bytes = block_rsv->reserved - block_rsv->size;
5387                 sinfo->bytes_may_use -= num_bytes;
5388                 trace_btrfs_space_reservation(fs_info, "space_info",
5389                                       sinfo->flags, num_bytes, 0);
5390                 block_rsv->reserved = block_rsv->size;
5391                 block_rsv->full = 1;
5392         }
5393
5394         spin_unlock(&block_rsv->lock);
5395         spin_unlock(&sinfo->lock);
5396 }
5397
5398 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5399 {
5400         struct btrfs_space_info *space_info;
5401
5402         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5403         fs_info->chunk_block_rsv.space_info = space_info;
5404
5405         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5406         fs_info->global_block_rsv.space_info = space_info;
5407         fs_info->delalloc_block_rsv.space_info = space_info;
5408         fs_info->trans_block_rsv.space_info = space_info;
5409         fs_info->empty_block_rsv.space_info = space_info;
5410         fs_info->delayed_block_rsv.space_info = space_info;
5411
5412         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5413         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5414         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5415         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5416         if (fs_info->quota_root)
5417                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5418         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5419
5420         update_global_block_rsv(fs_info);
5421 }
5422
5423 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5424 {
5425         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5426                                 (u64)-1);
5427         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5428         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5429         WARN_ON(fs_info->trans_block_rsv.size > 0);
5430         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5431         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5432         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5433         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5434         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5435 }
5436
5437 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5438                                   struct btrfs_root *root)
5439 {
5440         if (!trans->block_rsv)
5441                 return;
5442
5443         if (!trans->bytes_reserved)
5444                 return;
5445
5446         trace_btrfs_space_reservation(root->fs_info, "transaction",
5447                                       trans->transid, trans->bytes_reserved, 0);
5448         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5449         trans->bytes_reserved = 0;
5450 }
5451
5452 /*
5453  * To be called after all the new block groups attached to the transaction
5454  * handle have been created (btrfs_create_pending_block_groups()).
5455  */
5456 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5457 {
5458         struct btrfs_fs_info *fs_info = trans->root->fs_info;
5459
5460         if (!trans->chunk_bytes_reserved)
5461                 return;
5462
5463         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5464
5465         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5466                                 trans->chunk_bytes_reserved);
5467         trans->chunk_bytes_reserved = 0;
5468 }
5469
5470 /* Can only return 0 or -ENOSPC */
5471 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5472                                   struct inode *inode)
5473 {
5474         struct btrfs_root *root = BTRFS_I(inode)->root;
5475         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5476         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5477
5478         /*
5479          * We need to hold space in order to delete our orphan item once we've
5480          * added it, so this takes the reservation so we can release it later
5481          * when we are truly done with the orphan item.
5482          */
5483         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5484         trace_btrfs_space_reservation(root->fs_info, "orphan",
5485                                       btrfs_ino(inode), num_bytes, 1);
5486         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5487 }
5488
5489 void btrfs_orphan_release_metadata(struct inode *inode)
5490 {
5491         struct btrfs_root *root = BTRFS_I(inode)->root;
5492         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5493         trace_btrfs_space_reservation(root->fs_info, "orphan",
5494                                       btrfs_ino(inode), num_bytes, 0);
5495         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5496 }
5497
5498 /*
5499  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5500  * root: the root of the parent directory
5501  * rsv: block reservation
5502  * items: the number of items that we need do reservation
5503  * qgroup_reserved: used to return the reserved size in qgroup
5504  *
5505  * This function is used to reserve the space for snapshot/subvolume
5506  * creation and deletion. Those operations are different with the
5507  * common file/directory operations, they change two fs/file trees
5508  * and root tree, the number of items that the qgroup reserves is
5509  * different with the free space reservation. So we can not use
5510  * the space reseravtion mechanism in start_transaction().
5511  */
5512 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5513                                      struct btrfs_block_rsv *rsv,
5514                                      int items,
5515                                      u64 *qgroup_reserved,
5516                                      bool use_global_rsv)
5517 {
5518         u64 num_bytes;
5519         int ret;
5520         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5521
5522         if (root->fs_info->quota_enabled) {
5523                 /* One for parent inode, two for dir entries */
5524                 num_bytes = 3 * root->nodesize;
5525                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5526                 if (ret)
5527                         return ret;
5528         } else {
5529                 num_bytes = 0;
5530         }
5531
5532         *qgroup_reserved = num_bytes;
5533
5534         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5535         rsv->space_info = __find_space_info(root->fs_info,
5536                                             BTRFS_BLOCK_GROUP_METADATA);
5537         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5538                                   BTRFS_RESERVE_FLUSH_ALL);
5539
5540         if (ret == -ENOSPC && use_global_rsv)
5541                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5542
5543         if (ret && *qgroup_reserved)
5544                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5545
5546         return ret;
5547 }
5548
5549 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5550                                       struct btrfs_block_rsv *rsv,
5551                                       u64 qgroup_reserved)
5552 {
5553         btrfs_block_rsv_release(root, rsv, (u64)-1);
5554 }
5555
5556 /**
5557  * drop_outstanding_extent - drop an outstanding extent
5558  * @inode: the inode we're dropping the extent for
5559  * @num_bytes: the number of bytes we're relaseing.
5560  *
5561  * This is called when we are freeing up an outstanding extent, either called
5562  * after an error or after an extent is written.  This will return the number of
5563  * reserved extents that need to be freed.  This must be called with
5564  * BTRFS_I(inode)->lock held.
5565  */
5566 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5567 {
5568         unsigned drop_inode_space = 0;
5569         unsigned dropped_extents = 0;
5570         unsigned num_extents = 0;
5571
5572         num_extents = (unsigned)div64_u64(num_bytes +
5573                                           BTRFS_MAX_EXTENT_SIZE - 1,
5574                                           BTRFS_MAX_EXTENT_SIZE);
5575         ASSERT(num_extents);
5576         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5577         BTRFS_I(inode)->outstanding_extents -= num_extents;
5578
5579         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5580             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5581                                &BTRFS_I(inode)->runtime_flags))
5582                 drop_inode_space = 1;
5583
5584         /*
5585          * If we have more or the same amount of outsanding extents than we have
5586          * reserved then we need to leave the reserved extents count alone.
5587          */
5588         if (BTRFS_I(inode)->outstanding_extents >=
5589             BTRFS_I(inode)->reserved_extents)
5590                 return drop_inode_space;
5591
5592         dropped_extents = BTRFS_I(inode)->reserved_extents -
5593                 BTRFS_I(inode)->outstanding_extents;
5594         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5595         return dropped_extents + drop_inode_space;
5596 }
5597
5598 /**
5599  * calc_csum_metadata_size - return the amount of metada space that must be
5600  *      reserved/free'd for the given bytes.
5601  * @inode: the inode we're manipulating
5602  * @num_bytes: the number of bytes in question
5603  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5604  *
5605  * This adjusts the number of csum_bytes in the inode and then returns the
5606  * correct amount of metadata that must either be reserved or freed.  We
5607  * calculate how many checksums we can fit into one leaf and then divide the
5608  * number of bytes that will need to be checksumed by this value to figure out
5609  * how many checksums will be required.  If we are adding bytes then the number
5610  * may go up and we will return the number of additional bytes that must be
5611  * reserved.  If it is going down we will return the number of bytes that must
5612  * be freed.
5613  *
5614  * This must be called with BTRFS_I(inode)->lock held.
5615  */
5616 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5617                                    int reserve)
5618 {
5619         struct btrfs_root *root = BTRFS_I(inode)->root;
5620         u64 old_csums, num_csums;
5621
5622         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5623             BTRFS_I(inode)->csum_bytes == 0)
5624                 return 0;
5625
5626         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5627         if (reserve)
5628                 BTRFS_I(inode)->csum_bytes += num_bytes;
5629         else
5630                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5631         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5632
5633         /* No change, no need to reserve more */
5634         if (old_csums == num_csums)
5635                 return 0;
5636
5637         if (reserve)
5638                 return btrfs_calc_trans_metadata_size(root,
5639                                                       num_csums - old_csums);
5640
5641         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5642 }
5643
5644 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5645 {
5646         struct btrfs_root *root = BTRFS_I(inode)->root;
5647         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5648         u64 to_reserve = 0;
5649         u64 csum_bytes;
5650         unsigned nr_extents = 0;
5651         int extra_reserve = 0;
5652         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5653         int ret = 0;
5654         bool delalloc_lock = true;
5655         u64 to_free = 0;
5656         unsigned dropped;
5657
5658         /* If we are a free space inode we need to not flush since we will be in
5659          * the middle of a transaction commit.  We also don't need the delalloc
5660          * mutex since we won't race with anybody.  We need this mostly to make
5661          * lockdep shut its filthy mouth.
5662          */
5663         if (btrfs_is_free_space_inode(inode)) {
5664                 flush = BTRFS_RESERVE_NO_FLUSH;
5665                 delalloc_lock = false;
5666         }
5667
5668         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5669             btrfs_transaction_in_commit(root->fs_info))
5670                 schedule_timeout(1);
5671
5672         if (delalloc_lock)
5673                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5674
5675         num_bytes = ALIGN(num_bytes, root->sectorsize);
5676
5677         spin_lock(&BTRFS_I(inode)->lock);
5678         nr_extents = (unsigned)div64_u64(num_bytes +
5679                                          BTRFS_MAX_EXTENT_SIZE - 1,
5680                                          BTRFS_MAX_EXTENT_SIZE);
5681         BTRFS_I(inode)->outstanding_extents += nr_extents;
5682         nr_extents = 0;
5683
5684         if (BTRFS_I(inode)->outstanding_extents >
5685             BTRFS_I(inode)->reserved_extents)
5686                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5687                         BTRFS_I(inode)->reserved_extents;
5688
5689         /*
5690          * Add an item to reserve for updating the inode when we complete the
5691          * delalloc io.
5692          */
5693         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5694                       &BTRFS_I(inode)->runtime_flags)) {
5695                 nr_extents++;
5696                 extra_reserve = 1;
5697         }
5698
5699         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5700         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5701         csum_bytes = BTRFS_I(inode)->csum_bytes;
5702         spin_unlock(&BTRFS_I(inode)->lock);
5703
5704         if (root->fs_info->quota_enabled) {
5705                 ret = btrfs_qgroup_reserve_meta(root,
5706                                 nr_extents * root->nodesize);
5707                 if (ret)
5708                         goto out_fail;
5709         }
5710
5711         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5712         if (unlikely(ret)) {
5713                 btrfs_qgroup_free_meta(root, nr_extents * root->nodesize);
5714                 goto out_fail;
5715         }
5716
5717         spin_lock(&BTRFS_I(inode)->lock);
5718         if (extra_reserve) {
5719                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5720                         &BTRFS_I(inode)->runtime_flags);
5721                 nr_extents--;
5722         }
5723         BTRFS_I(inode)->reserved_extents += nr_extents;
5724         spin_unlock(&BTRFS_I(inode)->lock);
5725
5726         if (delalloc_lock)
5727                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5728
5729         if (to_reserve)
5730                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5731                                               btrfs_ino(inode), to_reserve, 1);
5732         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5733
5734         return 0;
5735
5736 out_fail:
5737         spin_lock(&BTRFS_I(inode)->lock);
5738         dropped = drop_outstanding_extent(inode, num_bytes);
5739         /*
5740          * If the inodes csum_bytes is the same as the original
5741          * csum_bytes then we know we haven't raced with any free()ers
5742          * so we can just reduce our inodes csum bytes and carry on.
5743          */
5744         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5745                 calc_csum_metadata_size(inode, num_bytes, 0);
5746         } else {
5747                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5748                 u64 bytes;
5749
5750                 /*
5751                  * This is tricky, but first we need to figure out how much we
5752                  * free'd from any free-ers that occured during this
5753                  * reservation, so we reset ->csum_bytes to the csum_bytes
5754                  * before we dropped our lock, and then call the free for the
5755                  * number of bytes that were freed while we were trying our
5756                  * reservation.
5757                  */
5758                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5759                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5760                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5761
5762
5763                 /*
5764                  * Now we need to see how much we would have freed had we not
5765                  * been making this reservation and our ->csum_bytes were not
5766                  * artificially inflated.
5767                  */
5768                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5769                 bytes = csum_bytes - orig_csum_bytes;
5770                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5771
5772                 /*
5773                  * Now reset ->csum_bytes to what it should be.  If bytes is
5774                  * more than to_free then we would have free'd more space had we
5775                  * not had an artificially high ->csum_bytes, so we need to free
5776                  * the remainder.  If bytes is the same or less then we don't
5777                  * need to do anything, the other free-ers did the correct
5778                  * thing.
5779                  */
5780                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5781                 if (bytes > to_free)
5782                         to_free = bytes - to_free;
5783                 else
5784                         to_free = 0;
5785         }
5786         spin_unlock(&BTRFS_I(inode)->lock);
5787         if (dropped)
5788                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5789
5790         if (to_free) {
5791                 btrfs_block_rsv_release(root, block_rsv, to_free);
5792                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5793                                               btrfs_ino(inode), to_free, 0);
5794         }
5795         if (delalloc_lock)
5796                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5797         return ret;
5798 }
5799
5800 /**
5801  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5802  * @inode: the inode to release the reservation for
5803  * @num_bytes: the number of bytes we're releasing
5804  *
5805  * This will release the metadata reservation for an inode.  This can be called
5806  * once we complete IO for a given set of bytes to release their metadata
5807  * reservations.
5808  */
5809 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5810 {
5811         struct btrfs_root *root = BTRFS_I(inode)->root;
5812         u64 to_free = 0;
5813         unsigned dropped;
5814
5815         num_bytes = ALIGN(num_bytes, root->sectorsize);
5816         spin_lock(&BTRFS_I(inode)->lock);
5817         dropped = drop_outstanding_extent(inode, num_bytes);
5818
5819         if (num_bytes)
5820                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5821         spin_unlock(&BTRFS_I(inode)->lock);
5822         if (dropped > 0)
5823                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5824
5825         if (btrfs_test_is_dummy_root(root))
5826                 return;
5827
5828         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5829                                       btrfs_ino(inode), to_free, 0);
5830
5831         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5832                                 to_free);
5833 }
5834
5835 /**
5836  * btrfs_delalloc_reserve_space - reserve data and metadata space for
5837  * delalloc
5838  * @inode: inode we're writing to
5839  * @start: start range we are writing to
5840  * @len: how long the range we are writing to
5841  *
5842  * TODO: This function will finally replace old btrfs_delalloc_reserve_space()
5843  *
5844  * This will do the following things
5845  *
5846  * o reserve space in data space info for num bytes
5847  *   and reserve precious corresponding qgroup space
5848  *   (Done in check_data_free_space)
5849  *
5850  * o reserve space for metadata space, based on the number of outstanding
5851  *   extents and how much csums will be needed
5852  *   also reserve metadata space in a per root over-reserve method.
5853  * o add to the inodes->delalloc_bytes
5854  * o add it to the fs_info's delalloc inodes list.
5855  *   (Above 3 all done in delalloc_reserve_metadata)
5856  *
5857  * Return 0 for success
5858  * Return <0 for error(-ENOSPC or -EQUOT)
5859  */
5860 int btrfs_delalloc_reserve_space(struct inode *inode, u64 start, u64 len)
5861 {
5862         int ret;
5863
5864         ret = btrfs_check_data_free_space(inode, start, len);
5865         if (ret < 0)
5866                 return ret;
5867         ret = btrfs_delalloc_reserve_metadata(inode, len);
5868         if (ret < 0)
5869                 btrfs_free_reserved_data_space(inode, start, len);
5870         return ret;
5871 }
5872
5873 /**
5874  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5875  * @inode: inode we're releasing space for
5876  * @start: start position of the space already reserved
5877  * @len: the len of the space already reserved
5878  *
5879  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5880  * called in the case that we don't need the metadata AND data reservations
5881  * anymore.  So if there is an error or we insert an inline extent.
5882  *
5883  * This function will release the metadata space that was not used and will
5884  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5885  * list if there are no delalloc bytes left.
5886  * Also it will handle the qgroup reserved space.
5887  */
5888 void btrfs_delalloc_release_space(struct inode *inode, u64 start, u64 len)
5889 {
5890         btrfs_delalloc_release_metadata(inode, len);
5891         btrfs_free_reserved_data_space(inode, start, len);
5892 }
5893
5894 static int update_block_group(struct btrfs_trans_handle *trans,
5895                               struct btrfs_root *root, u64 bytenr,
5896                               u64 num_bytes, int alloc)
5897 {
5898         struct btrfs_block_group_cache *cache = NULL;
5899         struct btrfs_fs_info *info = root->fs_info;
5900         u64 total = num_bytes;
5901         u64 old_val;
5902         u64 byte_in_group;
5903         int factor;
5904
5905         /* block accounting for super block */
5906         spin_lock(&info->delalloc_root_lock);
5907         old_val = btrfs_super_bytes_used(info->super_copy);
5908         if (alloc)
5909                 old_val += num_bytes;
5910         else
5911                 old_val -= num_bytes;
5912         btrfs_set_super_bytes_used(info->super_copy, old_val);
5913         spin_unlock(&info->delalloc_root_lock);
5914
5915         while (total) {
5916                 cache = btrfs_lookup_block_group(info, bytenr);
5917                 if (!cache)
5918                         return -ENOENT;
5919                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5920                                     BTRFS_BLOCK_GROUP_RAID1 |
5921                                     BTRFS_BLOCK_GROUP_RAID10))
5922                         factor = 2;
5923                 else
5924                         factor = 1;
5925                 /*
5926                  * If this block group has free space cache written out, we
5927                  * need to make sure to load it if we are removing space.  This
5928                  * is because we need the unpinning stage to actually add the
5929                  * space back to the block group, otherwise we will leak space.
5930                  */
5931                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5932                         cache_block_group(cache, 1);
5933
5934                 byte_in_group = bytenr - cache->key.objectid;
5935                 WARN_ON(byte_in_group > cache->key.offset);
5936
5937                 spin_lock(&cache->space_info->lock);
5938                 spin_lock(&cache->lock);
5939
5940                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5941                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5942                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5943
5944                 old_val = btrfs_block_group_used(&cache->item);
5945                 num_bytes = min(total, cache->key.offset - byte_in_group);
5946                 if (alloc) {
5947                         old_val += num_bytes;
5948                         btrfs_set_block_group_used(&cache->item, old_val);
5949                         cache->reserved -= num_bytes;
5950                         cache->space_info->bytes_reserved -= num_bytes;
5951                         cache->space_info->bytes_used += num_bytes;
5952                         cache->space_info->disk_used += num_bytes * factor;
5953                         spin_unlock(&cache->lock);
5954                         spin_unlock(&cache->space_info->lock);
5955                 } else {
5956                         old_val -= num_bytes;
5957                         btrfs_set_block_group_used(&cache->item, old_val);
5958                         cache->pinned += num_bytes;
5959                         cache->space_info->bytes_pinned += num_bytes;
5960                         cache->space_info->bytes_used -= num_bytes;
5961                         cache->space_info->disk_used -= num_bytes * factor;
5962                         spin_unlock(&cache->lock);
5963                         spin_unlock(&cache->space_info->lock);
5964
5965                         set_extent_dirty(info->pinned_extents,
5966                                          bytenr, bytenr + num_bytes - 1,
5967                                          GFP_NOFS | __GFP_NOFAIL);
5968                 }
5969
5970                 spin_lock(&trans->transaction->dirty_bgs_lock);
5971                 if (list_empty(&cache->dirty_list)) {
5972                         list_add_tail(&cache->dirty_list,
5973                                       &trans->transaction->dirty_bgs);
5974                                 trans->transaction->num_dirty_bgs++;
5975                         btrfs_get_block_group(cache);
5976                 }
5977                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5978
5979                 /*
5980                  * No longer have used bytes in this block group, queue it for
5981                  * deletion. We do this after adding the block group to the
5982                  * dirty list to avoid races between cleaner kthread and space
5983                  * cache writeout.
5984                  */
5985                 if (!alloc && old_val == 0) {
5986                         spin_lock(&info->unused_bgs_lock);
5987                         if (list_empty(&cache->bg_list)) {
5988                                 btrfs_get_block_group(cache);
5989                                 list_add_tail(&cache->bg_list,
5990                                               &info->unused_bgs);
5991                         }
5992                         spin_unlock(&info->unused_bgs_lock);
5993                 }
5994
5995                 btrfs_put_block_group(cache);
5996                 total -= num_bytes;
5997                 bytenr += num_bytes;
5998         }
5999         return 0;
6000 }
6001
6002 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
6003 {
6004         struct btrfs_block_group_cache *cache;
6005         u64 bytenr;
6006
6007         spin_lock(&root->fs_info->block_group_cache_lock);
6008         bytenr = root->fs_info->first_logical_byte;
6009         spin_unlock(&root->fs_info->block_group_cache_lock);
6010
6011         if (bytenr < (u64)-1)
6012                 return bytenr;
6013
6014         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
6015         if (!cache)
6016                 return 0;
6017
6018         bytenr = cache->key.objectid;
6019         btrfs_put_block_group(cache);
6020
6021         return bytenr;
6022 }
6023
6024 static int pin_down_extent(struct btrfs_root *root,
6025                            struct btrfs_block_group_cache *cache,
6026                            u64 bytenr, u64 num_bytes, int reserved)
6027 {
6028         spin_lock(&cache->space_info->lock);
6029         spin_lock(&cache->lock);
6030         cache->pinned += num_bytes;
6031         cache->space_info->bytes_pinned += num_bytes;
6032         if (reserved) {
6033                 cache->reserved -= num_bytes;
6034                 cache->space_info->bytes_reserved -= num_bytes;
6035         }
6036         spin_unlock(&cache->lock);
6037         spin_unlock(&cache->space_info->lock);
6038
6039         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
6040                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6041         if (reserved)
6042                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
6043         return 0;
6044 }
6045
6046 /*
6047  * this function must be called within transaction
6048  */
6049 int btrfs_pin_extent(struct btrfs_root *root,
6050                      u64 bytenr, u64 num_bytes, int reserved)
6051 {
6052         struct btrfs_block_group_cache *cache;
6053
6054         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6055         BUG_ON(!cache); /* Logic error */
6056
6057         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
6058
6059         btrfs_put_block_group(cache);
6060         return 0;
6061 }
6062
6063 /*
6064  * this function must be called within transaction
6065  */
6066 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
6067                                     u64 bytenr, u64 num_bytes)
6068 {
6069         struct btrfs_block_group_cache *cache;
6070         int ret;
6071
6072         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6073         if (!cache)
6074                 return -EINVAL;
6075
6076         /*
6077          * pull in the free space cache (if any) so that our pin
6078          * removes the free space from the cache.  We have load_only set
6079          * to one because the slow code to read in the free extents does check
6080          * the pinned extents.
6081          */
6082         cache_block_group(cache, 1);
6083
6084         pin_down_extent(root, cache, bytenr, num_bytes, 0);
6085
6086         /* remove us from the free space cache (if we're there at all) */
6087         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6088         btrfs_put_block_group(cache);
6089         return ret;
6090 }
6091
6092 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
6093 {
6094         int ret;
6095         struct btrfs_block_group_cache *block_group;
6096         struct btrfs_caching_control *caching_ctl;
6097
6098         block_group = btrfs_lookup_block_group(root->fs_info, start);
6099         if (!block_group)
6100                 return -EINVAL;
6101
6102         cache_block_group(block_group, 0);
6103         caching_ctl = get_caching_control(block_group);
6104
6105         if (!caching_ctl) {
6106                 /* Logic error */
6107                 BUG_ON(!block_group_cache_done(block_group));
6108                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6109         } else {
6110                 mutex_lock(&caching_ctl->mutex);
6111
6112                 if (start >= caching_ctl->progress) {
6113                         ret = add_excluded_extent(root, start, num_bytes);
6114                 } else if (start + num_bytes <= caching_ctl->progress) {
6115                         ret = btrfs_remove_free_space(block_group,
6116                                                       start, num_bytes);
6117                 } else {
6118                         num_bytes = caching_ctl->progress - start;
6119                         ret = btrfs_remove_free_space(block_group,
6120                                                       start, num_bytes);
6121                         if (ret)
6122                                 goto out_lock;
6123
6124                         num_bytes = (start + num_bytes) -
6125                                 caching_ctl->progress;
6126                         start = caching_ctl->progress;
6127                         ret = add_excluded_extent(root, start, num_bytes);
6128                 }
6129 out_lock:
6130                 mutex_unlock(&caching_ctl->mutex);
6131                 put_caching_control(caching_ctl);
6132         }
6133         btrfs_put_block_group(block_group);
6134         return ret;
6135 }
6136
6137 int btrfs_exclude_logged_extents(struct btrfs_root *log,
6138                                  struct extent_buffer *eb)
6139 {
6140         struct btrfs_file_extent_item *item;
6141         struct btrfs_key key;
6142         int found_type;
6143         int i;
6144
6145         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
6146                 return 0;
6147
6148         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6149                 btrfs_item_key_to_cpu(eb, &key, i);
6150                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6151                         continue;
6152                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6153                 found_type = btrfs_file_extent_type(eb, item);
6154                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6155                         continue;
6156                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6157                         continue;
6158                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6159                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6160                 __exclude_logged_extent(log, key.objectid, key.offset);
6161         }
6162
6163         return 0;
6164 }
6165
6166 /**
6167  * btrfs_update_reserved_bytes - update the block_group and space info counters
6168  * @cache:      The cache we are manipulating
6169  * @num_bytes:  The number of bytes in question
6170  * @reserve:    One of the reservation enums
6171  * @delalloc:   The blocks are allocated for the delalloc write
6172  *
6173  * This is called by the allocator when it reserves space, or by somebody who is
6174  * freeing space that was never actually used on disk.  For example if you
6175  * reserve some space for a new leaf in transaction A and before transaction A
6176  * commits you free that leaf, you call this with reserve set to 0 in order to
6177  * clear the reservation.
6178  *
6179  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
6180  * ENOSPC accounting.  For data we handle the reservation through clearing the
6181  * delalloc bits in the io_tree.  We have to do this since we could end up
6182  * allocating less disk space for the amount of data we have reserved in the
6183  * case of compression.
6184  *
6185  * If this is a reservation and the block group has become read only we cannot
6186  * make the reservation and return -EAGAIN, otherwise this function always
6187  * succeeds.
6188  */
6189 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
6190                                        u64 num_bytes, int reserve, int delalloc)
6191 {
6192         struct btrfs_space_info *space_info = cache->space_info;
6193         int ret = 0;
6194
6195         spin_lock(&space_info->lock);
6196         spin_lock(&cache->lock);
6197         if (reserve != RESERVE_FREE) {
6198                 if (cache->ro) {
6199                         ret = -EAGAIN;
6200                 } else {
6201                         cache->reserved += num_bytes;
6202                         space_info->bytes_reserved += num_bytes;
6203                         if (reserve == RESERVE_ALLOC) {
6204                                 trace_btrfs_space_reservation(cache->fs_info,
6205                                                 "space_info", space_info->flags,
6206                                                 num_bytes, 0);
6207                                 space_info->bytes_may_use -= num_bytes;
6208                         }
6209
6210                         if (delalloc)
6211                                 cache->delalloc_bytes += num_bytes;
6212                 }
6213         } else {
6214                 if (cache->ro)
6215                         space_info->bytes_readonly += num_bytes;
6216                 cache->reserved -= num_bytes;
6217                 space_info->bytes_reserved -= num_bytes;
6218
6219                 if (delalloc)
6220                         cache->delalloc_bytes -= num_bytes;
6221         }
6222         spin_unlock(&cache->lock);
6223         spin_unlock(&space_info->lock);
6224         return ret;
6225 }
6226
6227 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6228                                 struct btrfs_root *root)
6229 {
6230         struct btrfs_fs_info *fs_info = root->fs_info;
6231         struct btrfs_caching_control *next;
6232         struct btrfs_caching_control *caching_ctl;
6233         struct btrfs_block_group_cache *cache;
6234
6235         down_write(&fs_info->commit_root_sem);
6236
6237         list_for_each_entry_safe(caching_ctl, next,
6238                                  &fs_info->caching_block_groups, list) {
6239                 cache = caching_ctl->block_group;
6240                 if (block_group_cache_done(cache)) {
6241                         cache->last_byte_to_unpin = (u64)-1;
6242                         list_del_init(&caching_ctl->list);
6243                         put_caching_control(caching_ctl);
6244                 } else {
6245                         cache->last_byte_to_unpin = caching_ctl->progress;
6246                 }
6247         }
6248
6249         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6250                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6251         else
6252                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6253
6254         up_write(&fs_info->commit_root_sem);
6255
6256         update_global_block_rsv(fs_info);
6257 }
6258
6259 /*
6260  * Returns the free cluster for the given space info and sets empty_cluster to
6261  * what it should be based on the mount options.
6262  */
6263 static struct btrfs_free_cluster *
6264 fetch_cluster_info(struct btrfs_root *root, struct btrfs_space_info *space_info,
6265                    u64 *empty_cluster)
6266 {
6267         struct btrfs_free_cluster *ret = NULL;
6268         bool ssd = btrfs_test_opt(root, SSD);
6269
6270         *empty_cluster = 0;
6271         if (btrfs_mixed_space_info(space_info))
6272                 return ret;
6273
6274         if (ssd)
6275                 *empty_cluster = SZ_2M;
6276         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6277                 ret = &root->fs_info->meta_alloc_cluster;
6278                 if (!ssd)
6279                         *empty_cluster = SZ_64K;
6280         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) && ssd) {
6281                 ret = &root->fs_info->data_alloc_cluster;
6282         }
6283
6284         return ret;
6285 }
6286
6287 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
6288                               const bool return_free_space)
6289 {
6290         struct btrfs_fs_info *fs_info = root->fs_info;
6291         struct btrfs_block_group_cache *cache = NULL;
6292         struct btrfs_space_info *space_info;
6293         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6294         struct btrfs_free_cluster *cluster = NULL;
6295         u64 len;
6296         u64 total_unpinned = 0;
6297         u64 empty_cluster = 0;
6298         bool readonly;
6299
6300         while (start <= end) {
6301                 readonly = false;
6302                 if (!cache ||
6303                     start >= cache->key.objectid + cache->key.offset) {
6304                         if (cache)
6305                                 btrfs_put_block_group(cache);
6306                         total_unpinned = 0;
6307                         cache = btrfs_lookup_block_group(fs_info, start);
6308                         BUG_ON(!cache); /* Logic error */
6309
6310                         cluster = fetch_cluster_info(root,
6311                                                      cache->space_info,
6312                                                      &empty_cluster);
6313                         empty_cluster <<= 1;
6314                 }
6315
6316                 len = cache->key.objectid + cache->key.offset - start;
6317                 len = min(len, end + 1 - start);
6318
6319                 if (start < cache->last_byte_to_unpin) {
6320                         len = min(len, cache->last_byte_to_unpin - start);
6321                         if (return_free_space)
6322                                 btrfs_add_free_space(cache, start, len);
6323                 }
6324
6325                 start += len;
6326                 total_unpinned += len;
6327                 space_info = cache->space_info;
6328
6329                 /*
6330                  * If this space cluster has been marked as fragmented and we've
6331                  * unpinned enough in this block group to potentially allow a
6332                  * cluster to be created inside of it go ahead and clear the
6333                  * fragmented check.
6334                  */
6335                 if (cluster && cluster->fragmented &&
6336                     total_unpinned > empty_cluster) {
6337                         spin_lock(&cluster->lock);
6338                         cluster->fragmented = 0;
6339                         spin_unlock(&cluster->lock);
6340                 }
6341
6342                 spin_lock(&space_info->lock);
6343                 spin_lock(&cache->lock);
6344                 cache->pinned -= len;
6345                 space_info->bytes_pinned -= len;
6346                 space_info->max_extent_size = 0;
6347                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6348                 if (cache->ro) {
6349                         space_info->bytes_readonly += len;
6350                         readonly = true;
6351                 }
6352                 spin_unlock(&cache->lock);
6353                 if (!readonly && global_rsv->space_info == space_info) {
6354                         spin_lock(&global_rsv->lock);
6355                         if (!global_rsv->full) {
6356                                 len = min(len, global_rsv->size -
6357                                           global_rsv->reserved);
6358                                 global_rsv->reserved += len;
6359                                 space_info->bytes_may_use += len;
6360                                 if (global_rsv->reserved >= global_rsv->size)
6361                                         global_rsv->full = 1;
6362                         }
6363                         spin_unlock(&global_rsv->lock);
6364                 }
6365                 spin_unlock(&space_info->lock);
6366         }
6367
6368         if (cache)
6369                 btrfs_put_block_group(cache);
6370         return 0;
6371 }
6372
6373 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6374                                struct btrfs_root *root)
6375 {
6376         struct btrfs_fs_info *fs_info = root->fs_info;
6377         struct btrfs_block_group_cache *block_group, *tmp;
6378         struct list_head *deleted_bgs;
6379         struct extent_io_tree *unpin;
6380         u64 start;
6381         u64 end;
6382         int ret;
6383
6384         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6385                 unpin = &fs_info->freed_extents[1];
6386         else
6387                 unpin = &fs_info->freed_extents[0];
6388
6389         while (!trans->aborted) {
6390                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6391                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6392                                             EXTENT_DIRTY, NULL);
6393                 if (ret) {
6394                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6395                         break;
6396                 }
6397
6398                 if (btrfs_test_opt(root, DISCARD))
6399                         ret = btrfs_discard_extent(root, start,
6400                                                    end + 1 - start, NULL);
6401
6402                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
6403                 unpin_extent_range(root, start, end, true);
6404                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6405                 cond_resched();
6406         }
6407
6408         /*
6409          * Transaction is finished.  We don't need the lock anymore.  We
6410          * do need to clean up the block groups in case of a transaction
6411          * abort.
6412          */
6413         deleted_bgs = &trans->transaction->deleted_bgs;
6414         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6415                 u64 trimmed = 0;
6416
6417                 ret = -EROFS;
6418                 if (!trans->aborted)
6419                         ret = btrfs_discard_extent(root,
6420                                                    block_group->key.objectid,
6421                                                    block_group->key.offset,
6422                                                    &trimmed);
6423
6424                 list_del_init(&block_group->bg_list);
6425                 btrfs_put_block_group_trimming(block_group);
6426                 btrfs_put_block_group(block_group);
6427
6428                 if (ret) {
6429                         const char *errstr = btrfs_decode_error(ret);
6430                         btrfs_warn(fs_info,
6431                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6432                                    ret, errstr);
6433                 }
6434         }
6435
6436         return 0;
6437 }
6438
6439 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6440                              u64 owner, u64 root_objectid)
6441 {
6442         struct btrfs_space_info *space_info;
6443         u64 flags;
6444
6445         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6446                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6447                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6448                 else
6449                         flags = BTRFS_BLOCK_GROUP_METADATA;
6450         } else {
6451                 flags = BTRFS_BLOCK_GROUP_DATA;
6452         }
6453
6454         space_info = __find_space_info(fs_info, flags);
6455         BUG_ON(!space_info); /* Logic bug */
6456         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6457 }
6458
6459
6460 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6461                                 struct btrfs_root *root,
6462                                 struct btrfs_delayed_ref_node *node, u64 parent,
6463                                 u64 root_objectid, u64 owner_objectid,
6464                                 u64 owner_offset, int refs_to_drop,
6465                                 struct btrfs_delayed_extent_op *extent_op)
6466 {
6467         struct btrfs_key key;
6468         struct btrfs_path *path;
6469         struct btrfs_fs_info *info = root->fs_info;
6470         struct btrfs_root *extent_root = info->extent_root;
6471         struct extent_buffer *leaf;
6472         struct btrfs_extent_item *ei;
6473         struct btrfs_extent_inline_ref *iref;
6474         int ret;
6475         int is_data;
6476         int extent_slot = 0;
6477         int found_extent = 0;
6478         int num_to_del = 1;
6479         u32 item_size;
6480         u64 refs;
6481         u64 bytenr = node->bytenr;
6482         u64 num_bytes = node->num_bytes;
6483         int last_ref = 0;
6484         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6485                                                  SKINNY_METADATA);
6486
6487         path = btrfs_alloc_path();
6488         if (!path)
6489                 return -ENOMEM;
6490
6491         path->reada = READA_FORWARD;
6492         path->leave_spinning = 1;
6493
6494         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6495         BUG_ON(!is_data && refs_to_drop != 1);
6496
6497         if (is_data)
6498                 skinny_metadata = 0;
6499
6500         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6501                                     bytenr, num_bytes, parent,
6502                                     root_objectid, owner_objectid,
6503                                     owner_offset);
6504         if (ret == 0) {
6505                 extent_slot = path->slots[0];
6506                 while (extent_slot >= 0) {
6507                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6508                                               extent_slot);
6509                         if (key.objectid != bytenr)
6510                                 break;
6511                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6512                             key.offset == num_bytes) {
6513                                 found_extent = 1;
6514                                 break;
6515                         }
6516                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6517                             key.offset == owner_objectid) {
6518                                 found_extent = 1;
6519                                 break;
6520                         }
6521                         if (path->slots[0] - extent_slot > 5)
6522                                 break;
6523                         extent_slot--;
6524                 }
6525 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6526                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6527                 if (found_extent && item_size < sizeof(*ei))
6528                         found_extent = 0;
6529 #endif
6530                 if (!found_extent) {
6531                         BUG_ON(iref);
6532                         ret = remove_extent_backref(trans, extent_root, path,
6533                                                     NULL, refs_to_drop,
6534                                                     is_data, &last_ref);
6535                         if (ret) {
6536                                 btrfs_abort_transaction(trans, extent_root, ret);
6537                                 goto out;
6538                         }
6539                         btrfs_release_path(path);
6540                         path->leave_spinning = 1;
6541
6542                         key.objectid = bytenr;
6543                         key.type = BTRFS_EXTENT_ITEM_KEY;
6544                         key.offset = num_bytes;
6545
6546                         if (!is_data && skinny_metadata) {
6547                                 key.type = BTRFS_METADATA_ITEM_KEY;
6548                                 key.offset = owner_objectid;
6549                         }
6550
6551                         ret = btrfs_search_slot(trans, extent_root,
6552                                                 &key, path, -1, 1);
6553                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6554                                 /*
6555                                  * Couldn't find our skinny metadata item,
6556                                  * see if we have ye olde extent item.
6557                                  */
6558                                 path->slots[0]--;
6559                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6560                                                       path->slots[0]);
6561                                 if (key.objectid == bytenr &&
6562                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6563                                     key.offset == num_bytes)
6564                                         ret = 0;
6565                         }
6566
6567                         if (ret > 0 && skinny_metadata) {
6568                                 skinny_metadata = false;
6569                                 key.objectid = bytenr;
6570                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6571                                 key.offset = num_bytes;
6572                                 btrfs_release_path(path);
6573                                 ret = btrfs_search_slot(trans, extent_root,
6574                                                         &key, path, -1, 1);
6575                         }
6576
6577                         if (ret) {
6578                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6579                                         ret, bytenr);
6580                                 if (ret > 0)
6581                                         btrfs_print_leaf(extent_root,
6582                                                          path->nodes[0]);
6583                         }
6584                         if (ret < 0) {
6585                                 btrfs_abort_transaction(trans, extent_root, ret);
6586                                 goto out;
6587                         }
6588                         extent_slot = path->slots[0];
6589                 }
6590         } else if (WARN_ON(ret == -ENOENT)) {
6591                 btrfs_print_leaf(extent_root, path->nodes[0]);
6592                 btrfs_err(info,
6593                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6594                         bytenr, parent, root_objectid, owner_objectid,
6595                         owner_offset);
6596                 btrfs_abort_transaction(trans, extent_root, ret);
6597                 goto out;
6598         } else {
6599                 btrfs_abort_transaction(trans, extent_root, ret);
6600                 goto out;
6601         }
6602
6603         leaf = path->nodes[0];
6604         item_size = btrfs_item_size_nr(leaf, extent_slot);
6605 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6606         if (item_size < sizeof(*ei)) {
6607                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6608                 ret = convert_extent_item_v0(trans, extent_root, path,
6609                                              owner_objectid, 0);
6610                 if (ret < 0) {
6611                         btrfs_abort_transaction(trans, extent_root, ret);
6612                         goto out;
6613                 }
6614
6615                 btrfs_release_path(path);
6616                 path->leave_spinning = 1;
6617
6618                 key.objectid = bytenr;
6619                 key.type = BTRFS_EXTENT_ITEM_KEY;
6620                 key.offset = num_bytes;
6621
6622                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6623                                         -1, 1);
6624                 if (ret) {
6625                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6626                                 ret, bytenr);
6627                         btrfs_print_leaf(extent_root, path->nodes[0]);
6628                 }
6629                 if (ret < 0) {
6630                         btrfs_abort_transaction(trans, extent_root, ret);
6631                         goto out;
6632                 }
6633
6634                 extent_slot = path->slots[0];
6635                 leaf = path->nodes[0];
6636                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6637         }
6638 #endif
6639         BUG_ON(item_size < sizeof(*ei));
6640         ei = btrfs_item_ptr(leaf, extent_slot,
6641                             struct btrfs_extent_item);
6642         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6643             key.type == BTRFS_EXTENT_ITEM_KEY) {
6644                 struct btrfs_tree_block_info *bi;
6645                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6646                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6647                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6648         }
6649
6650         refs = btrfs_extent_refs(leaf, ei);
6651         if (refs < refs_to_drop) {
6652                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6653                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6654                 ret = -EINVAL;
6655                 btrfs_abort_transaction(trans, extent_root, ret);
6656                 goto out;
6657         }
6658         refs -= refs_to_drop;
6659
6660         if (refs > 0) {
6661                 if (extent_op)
6662                         __run_delayed_extent_op(extent_op, leaf, ei);
6663                 /*
6664                  * In the case of inline back ref, reference count will
6665                  * be updated by remove_extent_backref
6666                  */
6667                 if (iref) {
6668                         BUG_ON(!found_extent);
6669                 } else {
6670                         btrfs_set_extent_refs(leaf, ei, refs);
6671                         btrfs_mark_buffer_dirty(leaf);
6672                 }
6673                 if (found_extent) {
6674                         ret = remove_extent_backref(trans, extent_root, path,
6675                                                     iref, refs_to_drop,
6676                                                     is_data, &last_ref);
6677                         if (ret) {
6678                                 btrfs_abort_transaction(trans, extent_root, ret);
6679                                 goto out;
6680                         }
6681                 }
6682                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6683                                  root_objectid);
6684         } else {
6685                 if (found_extent) {
6686                         BUG_ON(is_data && refs_to_drop !=
6687                                extent_data_ref_count(path, iref));
6688                         if (iref) {
6689                                 BUG_ON(path->slots[0] != extent_slot);
6690                         } else {
6691                                 BUG_ON(path->slots[0] != extent_slot + 1);
6692                                 path->slots[0] = extent_slot;
6693                                 num_to_del = 2;
6694                         }
6695                 }
6696
6697                 last_ref = 1;
6698                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6699                                       num_to_del);
6700                 if (ret) {
6701                         btrfs_abort_transaction(trans, extent_root, ret);
6702                         goto out;
6703                 }
6704                 btrfs_release_path(path);
6705
6706                 if (is_data) {
6707                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6708                         if (ret) {
6709                                 btrfs_abort_transaction(trans, extent_root, ret);
6710                                 goto out;
6711                         }
6712                 }
6713
6714                 ret = add_to_free_space_tree(trans, root->fs_info, bytenr,
6715                                              num_bytes);
6716                 if (ret) {
6717                         btrfs_abort_transaction(trans, extent_root, ret);
6718                         goto out;
6719                 }
6720
6721                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6722                 if (ret) {
6723                         btrfs_abort_transaction(trans, extent_root, ret);
6724                         goto out;
6725                 }
6726         }
6727         btrfs_release_path(path);
6728
6729 out:
6730         btrfs_free_path(path);
6731         return ret;
6732 }
6733
6734 /*
6735  * when we free an block, it is possible (and likely) that we free the last
6736  * delayed ref for that extent as well.  This searches the delayed ref tree for
6737  * a given extent, and if there are no other delayed refs to be processed, it
6738  * removes it from the tree.
6739  */
6740 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6741                                       struct btrfs_root *root, u64 bytenr)
6742 {
6743         struct btrfs_delayed_ref_head *head;
6744         struct btrfs_delayed_ref_root *delayed_refs;
6745         int ret = 0;
6746
6747         delayed_refs = &trans->transaction->delayed_refs;
6748         spin_lock(&delayed_refs->lock);
6749         head = btrfs_find_delayed_ref_head(trans, bytenr);
6750         if (!head)
6751                 goto out_delayed_unlock;
6752
6753         spin_lock(&head->lock);
6754         if (!list_empty(&head->ref_list))
6755                 goto out;
6756
6757         if (head->extent_op) {
6758                 if (!head->must_insert_reserved)
6759                         goto out;
6760                 btrfs_free_delayed_extent_op(head->extent_op);
6761                 head->extent_op = NULL;
6762         }
6763
6764         /*
6765          * waiting for the lock here would deadlock.  If someone else has it
6766          * locked they are already in the process of dropping it anyway
6767          */
6768         if (!mutex_trylock(&head->mutex))
6769                 goto out;
6770
6771         /*
6772          * at this point we have a head with no other entries.  Go
6773          * ahead and process it.
6774          */
6775         head->node.in_tree = 0;
6776         rb_erase(&head->href_node, &delayed_refs->href_root);
6777
6778         atomic_dec(&delayed_refs->num_entries);
6779
6780         /*
6781          * we don't take a ref on the node because we're removing it from the
6782          * tree, so we just steal the ref the tree was holding.
6783          */
6784         delayed_refs->num_heads--;
6785         if (head->processing == 0)
6786                 delayed_refs->num_heads_ready--;
6787         head->processing = 0;
6788         spin_unlock(&head->lock);
6789         spin_unlock(&delayed_refs->lock);
6790
6791         BUG_ON(head->extent_op);
6792         if (head->must_insert_reserved)
6793                 ret = 1;
6794
6795         mutex_unlock(&head->mutex);
6796         btrfs_put_delayed_ref(&head->node);
6797         return ret;
6798 out:
6799         spin_unlock(&head->lock);
6800
6801 out_delayed_unlock:
6802         spin_unlock(&delayed_refs->lock);
6803         return 0;
6804 }
6805
6806 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6807                            struct btrfs_root *root,
6808                            struct extent_buffer *buf,
6809                            u64 parent, int last_ref)
6810 {
6811         int pin = 1;
6812         int ret;
6813
6814         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6815                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6816                                         buf->start, buf->len,
6817                                         parent, root->root_key.objectid,
6818                                         btrfs_header_level(buf),
6819                                         BTRFS_DROP_DELAYED_REF, NULL);
6820                 BUG_ON(ret); /* -ENOMEM */
6821         }
6822
6823         if (!last_ref)
6824                 return;
6825
6826         if (btrfs_header_generation(buf) == trans->transid) {
6827                 struct btrfs_block_group_cache *cache;
6828
6829                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6830                         ret = check_ref_cleanup(trans, root, buf->start);
6831                         if (!ret)
6832                                 goto out;
6833                 }
6834
6835                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6836
6837                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6838                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6839                         btrfs_put_block_group(cache);
6840                         goto out;
6841                 }
6842
6843                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6844
6845                 btrfs_add_free_space(cache, buf->start, buf->len);
6846                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6847                 btrfs_put_block_group(cache);
6848                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6849                 pin = 0;
6850         }
6851 out:
6852         if (pin)
6853                 add_pinned_bytes(root->fs_info, buf->len,
6854                                  btrfs_header_level(buf),
6855                                  root->root_key.objectid);
6856
6857         /*
6858          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6859          * anymore.
6860          */
6861         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6862 }
6863
6864 /* Can return -ENOMEM */
6865 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6866                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6867                       u64 owner, u64 offset)
6868 {
6869         int ret;
6870         struct btrfs_fs_info *fs_info = root->fs_info;
6871
6872         if (btrfs_test_is_dummy_root(root))
6873                 return 0;
6874
6875         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6876
6877         /*
6878          * tree log blocks never actually go into the extent allocation
6879          * tree, just update pinning info and exit early.
6880          */
6881         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6882                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6883                 /* unlocks the pinned mutex */
6884                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6885                 ret = 0;
6886         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6887                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6888                                         num_bytes,
6889                                         parent, root_objectid, (int)owner,
6890                                         BTRFS_DROP_DELAYED_REF, NULL);
6891         } else {
6892                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6893                                                 num_bytes,
6894                                                 parent, root_objectid, owner,
6895                                                 offset, 0,
6896                                                 BTRFS_DROP_DELAYED_REF, NULL);
6897         }
6898         return ret;
6899 }
6900
6901 /*
6902  * when we wait for progress in the block group caching, its because
6903  * our allocation attempt failed at least once.  So, we must sleep
6904  * and let some progress happen before we try again.
6905  *
6906  * This function will sleep at least once waiting for new free space to
6907  * show up, and then it will check the block group free space numbers
6908  * for our min num_bytes.  Another option is to have it go ahead
6909  * and look in the rbtree for a free extent of a given size, but this
6910  * is a good start.
6911  *
6912  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6913  * any of the information in this block group.
6914  */
6915 static noinline void
6916 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6917                                 u64 num_bytes)
6918 {
6919         struct btrfs_caching_control *caching_ctl;
6920
6921         caching_ctl = get_caching_control(cache);
6922         if (!caching_ctl)
6923                 return;
6924
6925         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6926                    (cache->free_space_ctl->free_space >= num_bytes));
6927
6928         put_caching_control(caching_ctl);
6929 }
6930
6931 static noinline int
6932 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6933 {
6934         struct btrfs_caching_control *caching_ctl;
6935         int ret = 0;
6936
6937         caching_ctl = get_caching_control(cache);
6938         if (!caching_ctl)
6939                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6940
6941         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6942         if (cache->cached == BTRFS_CACHE_ERROR)
6943                 ret = -EIO;
6944         put_caching_control(caching_ctl);
6945         return ret;
6946 }
6947
6948 int __get_raid_index(u64 flags)
6949 {
6950         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6951                 return BTRFS_RAID_RAID10;
6952         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6953                 return BTRFS_RAID_RAID1;
6954         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6955                 return BTRFS_RAID_DUP;
6956         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6957                 return BTRFS_RAID_RAID0;
6958         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6959                 return BTRFS_RAID_RAID5;
6960         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6961                 return BTRFS_RAID_RAID6;
6962
6963         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6964 }
6965
6966 int get_block_group_index(struct btrfs_block_group_cache *cache)
6967 {
6968         return __get_raid_index(cache->flags);
6969 }
6970
6971 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6972         [BTRFS_RAID_RAID10]     = "raid10",
6973         [BTRFS_RAID_RAID1]      = "raid1",
6974         [BTRFS_RAID_DUP]        = "dup",
6975         [BTRFS_RAID_RAID0]      = "raid0",
6976         [BTRFS_RAID_SINGLE]     = "single",
6977         [BTRFS_RAID_RAID5]      = "raid5",
6978         [BTRFS_RAID_RAID6]      = "raid6",
6979 };
6980
6981 static const char *get_raid_name(enum btrfs_raid_types type)
6982 {
6983         if (type >= BTRFS_NR_RAID_TYPES)
6984                 return NULL;
6985
6986         return btrfs_raid_type_names[type];
6987 }
6988
6989 enum btrfs_loop_type {
6990         LOOP_CACHING_NOWAIT = 0,
6991         LOOP_CACHING_WAIT = 1,
6992         LOOP_ALLOC_CHUNK = 2,
6993         LOOP_NO_EMPTY_SIZE = 3,
6994 };
6995
6996 static inline void
6997 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6998                        int delalloc)
6999 {
7000         if (delalloc)
7001                 down_read(&cache->data_rwsem);
7002 }
7003
7004 static inline void
7005 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7006                        int delalloc)
7007 {
7008         btrfs_get_block_group(cache);
7009         if (delalloc)
7010                 down_read(&cache->data_rwsem);
7011 }
7012
7013 static struct btrfs_block_group_cache *
7014 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7015                    struct btrfs_free_cluster *cluster,
7016                    int delalloc)
7017 {
7018         struct btrfs_block_group_cache *used_bg;
7019         bool locked = false;
7020 again:
7021         spin_lock(&cluster->refill_lock);
7022         if (locked) {
7023                 if (used_bg == cluster->block_group)
7024                         return used_bg;
7025
7026                 up_read(&used_bg->data_rwsem);
7027                 btrfs_put_block_group(used_bg);
7028         }
7029
7030         used_bg = cluster->block_group;
7031         if (!used_bg)
7032                 return NULL;
7033
7034         if (used_bg == block_group)
7035                 return used_bg;
7036
7037         btrfs_get_block_group(used_bg);
7038
7039         if (!delalloc)
7040                 return used_bg;
7041
7042         if (down_read_trylock(&used_bg->data_rwsem))
7043                 return used_bg;
7044
7045         spin_unlock(&cluster->refill_lock);
7046         down_read(&used_bg->data_rwsem);
7047         locked = true;
7048         goto again;
7049 }
7050
7051 static inline void
7052 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7053                          int delalloc)
7054 {
7055         if (delalloc)
7056                 up_read(&cache->data_rwsem);
7057         btrfs_put_block_group(cache);
7058 }
7059
7060 /*
7061  * walks the btree of allocated extents and find a hole of a given size.
7062  * The key ins is changed to record the hole:
7063  * ins->objectid == start position
7064  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7065  * ins->offset == the size of the hole.
7066  * Any available blocks before search_start are skipped.
7067  *
7068  * If there is no suitable free space, we will record the max size of
7069  * the free space extent currently.
7070  */
7071 static noinline int find_free_extent(struct btrfs_root *orig_root,
7072                                      u64 num_bytes, u64 empty_size,
7073                                      u64 hint_byte, struct btrfs_key *ins,
7074                                      u64 flags, int delalloc)
7075 {
7076         int ret = 0;
7077         struct btrfs_root *root = orig_root->fs_info->extent_root;
7078         struct btrfs_free_cluster *last_ptr = NULL;
7079         struct btrfs_block_group_cache *block_group = NULL;
7080         u64 search_start = 0;
7081         u64 max_extent_size = 0;
7082         u64 empty_cluster = 0;
7083         struct btrfs_space_info *space_info;
7084         int loop = 0;
7085         int index = __get_raid_index(flags);
7086         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
7087                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
7088         bool failed_cluster_refill = false;
7089         bool failed_alloc = false;
7090         bool use_cluster = true;
7091         bool have_caching_bg = false;
7092         bool orig_have_caching_bg = false;
7093         bool full_search = false;
7094
7095         WARN_ON(num_bytes < root->sectorsize);
7096         ins->type = BTRFS_EXTENT_ITEM_KEY;
7097         ins->objectid = 0;
7098         ins->offset = 0;
7099
7100         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
7101
7102         space_info = __find_space_info(root->fs_info, flags);
7103         if (!space_info) {
7104                 btrfs_err(root->fs_info, "No space info for %llu", flags);
7105                 return -ENOSPC;
7106         }
7107
7108         /*
7109          * If our free space is heavily fragmented we may not be able to make
7110          * big contiguous allocations, so instead of doing the expensive search
7111          * for free space, simply return ENOSPC with our max_extent_size so we
7112          * can go ahead and search for a more manageable chunk.
7113          *
7114          * If our max_extent_size is large enough for our allocation simply
7115          * disable clustering since we will likely not be able to find enough
7116          * space to create a cluster and induce latency trying.
7117          */
7118         if (unlikely(space_info->max_extent_size)) {
7119                 spin_lock(&space_info->lock);
7120                 if (space_info->max_extent_size &&
7121                     num_bytes > space_info->max_extent_size) {
7122                         ins->offset = space_info->max_extent_size;
7123                         spin_unlock(&space_info->lock);
7124                         return -ENOSPC;
7125                 } else if (space_info->max_extent_size) {
7126                         use_cluster = false;
7127                 }
7128                 spin_unlock(&space_info->lock);
7129         }
7130
7131         last_ptr = fetch_cluster_info(orig_root, space_info, &empty_cluster);
7132         if (last_ptr) {
7133                 spin_lock(&last_ptr->lock);
7134                 if (last_ptr->block_group)
7135                         hint_byte = last_ptr->window_start;
7136                 if (last_ptr->fragmented) {
7137                         /*
7138                          * We still set window_start so we can keep track of the
7139                          * last place we found an allocation to try and save
7140                          * some time.
7141                          */
7142                         hint_byte = last_ptr->window_start;
7143                         use_cluster = false;
7144                 }
7145                 spin_unlock(&last_ptr->lock);
7146         }
7147
7148         search_start = max(search_start, first_logical_byte(root, 0));
7149         search_start = max(search_start, hint_byte);
7150         if (search_start == hint_byte) {
7151                 block_group = btrfs_lookup_block_group(root->fs_info,
7152                                                        search_start);
7153                 /*
7154                  * we don't want to use the block group if it doesn't match our
7155                  * allocation bits, or if its not cached.
7156                  *
7157                  * However if we are re-searching with an ideal block group
7158                  * picked out then we don't care that the block group is cached.
7159                  */
7160                 if (block_group && block_group_bits(block_group, flags) &&
7161                     block_group->cached != BTRFS_CACHE_NO) {
7162                         down_read(&space_info->groups_sem);
7163                         if (list_empty(&block_group->list) ||
7164                             block_group->ro) {
7165                                 /*
7166                                  * someone is removing this block group,
7167                                  * we can't jump into the have_block_group
7168                                  * target because our list pointers are not
7169                                  * valid
7170                                  */
7171                                 btrfs_put_block_group(block_group);
7172                                 up_read(&space_info->groups_sem);
7173                         } else {
7174                                 index = get_block_group_index(block_group);
7175                                 btrfs_lock_block_group(block_group, delalloc);
7176                                 goto have_block_group;
7177                         }
7178                 } else if (block_group) {
7179                         btrfs_put_block_group(block_group);
7180                 }
7181         }
7182 search:
7183         have_caching_bg = false;
7184         if (index == 0 || index == __get_raid_index(flags))
7185                 full_search = true;
7186         down_read(&space_info->groups_sem);
7187         list_for_each_entry(block_group, &space_info->block_groups[index],
7188                             list) {
7189                 u64 offset;
7190                 int cached;
7191
7192                 btrfs_grab_block_group(block_group, delalloc);
7193                 search_start = block_group->key.objectid;
7194
7195                 /*
7196                  * this can happen if we end up cycling through all the
7197                  * raid types, but we want to make sure we only allocate
7198                  * for the proper type.
7199                  */
7200                 if (!block_group_bits(block_group, flags)) {
7201                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7202                                 BTRFS_BLOCK_GROUP_RAID1 |
7203                                 BTRFS_BLOCK_GROUP_RAID5 |
7204                                 BTRFS_BLOCK_GROUP_RAID6 |
7205                                 BTRFS_BLOCK_GROUP_RAID10;
7206
7207                         /*
7208                          * if they asked for extra copies and this block group
7209                          * doesn't provide them, bail.  This does allow us to
7210                          * fill raid0 from raid1.
7211                          */
7212                         if ((flags & extra) && !(block_group->flags & extra))
7213                                 goto loop;
7214                 }
7215
7216 have_block_group:
7217                 cached = block_group_cache_done(block_group);
7218                 if (unlikely(!cached)) {
7219                         have_caching_bg = true;
7220                         ret = cache_block_group(block_group, 0);
7221                         BUG_ON(ret < 0);
7222                         ret = 0;
7223                 }
7224
7225                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7226                         goto loop;
7227                 if (unlikely(block_group->ro))
7228                         goto loop;
7229
7230                 /*
7231                  * Ok we want to try and use the cluster allocator, so
7232                  * lets look there
7233                  */
7234                 if (last_ptr && use_cluster) {
7235                         struct btrfs_block_group_cache *used_block_group;
7236                         unsigned long aligned_cluster;
7237                         /*
7238                          * the refill lock keeps out other
7239                          * people trying to start a new cluster
7240                          */
7241                         used_block_group = btrfs_lock_cluster(block_group,
7242                                                               last_ptr,
7243                                                               delalloc);
7244                         if (!used_block_group)
7245                                 goto refill_cluster;
7246
7247                         if (used_block_group != block_group &&
7248                             (used_block_group->ro ||
7249                              !block_group_bits(used_block_group, flags)))
7250                                 goto release_cluster;
7251
7252                         offset = btrfs_alloc_from_cluster(used_block_group,
7253                                                 last_ptr,
7254                                                 num_bytes,
7255                                                 used_block_group->key.objectid,
7256                                                 &max_extent_size);
7257                         if (offset) {
7258                                 /* we have a block, we're done */
7259                                 spin_unlock(&last_ptr->refill_lock);
7260                                 trace_btrfs_reserve_extent_cluster(root,
7261                                                 used_block_group,
7262                                                 search_start, num_bytes);
7263                                 if (used_block_group != block_group) {
7264                                         btrfs_release_block_group(block_group,
7265                                                                   delalloc);
7266                                         block_group = used_block_group;
7267                                 }
7268                                 goto checks;
7269                         }
7270
7271                         WARN_ON(last_ptr->block_group != used_block_group);
7272 release_cluster:
7273                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7274                          * set up a new clusters, so lets just skip it
7275                          * and let the allocator find whatever block
7276                          * it can find.  If we reach this point, we
7277                          * will have tried the cluster allocator
7278                          * plenty of times and not have found
7279                          * anything, so we are likely way too
7280                          * fragmented for the clustering stuff to find
7281                          * anything.
7282                          *
7283                          * However, if the cluster is taken from the
7284                          * current block group, release the cluster
7285                          * first, so that we stand a better chance of
7286                          * succeeding in the unclustered
7287                          * allocation.  */
7288                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7289                             used_block_group != block_group) {
7290                                 spin_unlock(&last_ptr->refill_lock);
7291                                 btrfs_release_block_group(used_block_group,
7292                                                           delalloc);
7293                                 goto unclustered_alloc;
7294                         }
7295
7296                         /*
7297                          * this cluster didn't work out, free it and
7298                          * start over
7299                          */
7300                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7301
7302                         if (used_block_group != block_group)
7303                                 btrfs_release_block_group(used_block_group,
7304                                                           delalloc);
7305 refill_cluster:
7306                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7307                                 spin_unlock(&last_ptr->refill_lock);
7308                                 goto unclustered_alloc;
7309                         }
7310
7311                         aligned_cluster = max_t(unsigned long,
7312                                                 empty_cluster + empty_size,
7313                                               block_group->full_stripe_len);
7314
7315                         /* allocate a cluster in this block group */
7316                         ret = btrfs_find_space_cluster(root, block_group,
7317                                                        last_ptr, search_start,
7318                                                        num_bytes,
7319                                                        aligned_cluster);
7320                         if (ret == 0) {
7321                                 /*
7322                                  * now pull our allocation out of this
7323                                  * cluster
7324                                  */
7325                                 offset = btrfs_alloc_from_cluster(block_group,
7326                                                         last_ptr,
7327                                                         num_bytes,
7328                                                         search_start,
7329                                                         &max_extent_size);
7330                                 if (offset) {
7331                                         /* we found one, proceed */
7332                                         spin_unlock(&last_ptr->refill_lock);
7333                                         trace_btrfs_reserve_extent_cluster(root,
7334                                                 block_group, search_start,
7335                                                 num_bytes);
7336                                         goto checks;
7337                                 }
7338                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7339                                    && !failed_cluster_refill) {
7340                                 spin_unlock(&last_ptr->refill_lock);
7341
7342                                 failed_cluster_refill = true;
7343                                 wait_block_group_cache_progress(block_group,
7344                                        num_bytes + empty_cluster + empty_size);
7345                                 goto have_block_group;
7346                         }
7347
7348                         /*
7349                          * at this point we either didn't find a cluster
7350                          * or we weren't able to allocate a block from our
7351                          * cluster.  Free the cluster we've been trying
7352                          * to use, and go to the next block group
7353                          */
7354                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7355                         spin_unlock(&last_ptr->refill_lock);
7356                         goto loop;
7357                 }
7358
7359 unclustered_alloc:
7360                 /*
7361                  * We are doing an unclustered alloc, set the fragmented flag so
7362                  * we don't bother trying to setup a cluster again until we get
7363                  * more space.
7364                  */
7365                 if (unlikely(last_ptr)) {
7366                         spin_lock(&last_ptr->lock);
7367                         last_ptr->fragmented = 1;
7368                         spin_unlock(&last_ptr->lock);
7369                 }
7370                 spin_lock(&block_group->free_space_ctl->tree_lock);
7371                 if (cached &&
7372                     block_group->free_space_ctl->free_space <
7373                     num_bytes + empty_cluster + empty_size) {
7374                         if (block_group->free_space_ctl->free_space >
7375                             max_extent_size)
7376                                 max_extent_size =
7377                                         block_group->free_space_ctl->free_space;
7378                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7379                         goto loop;
7380                 }
7381                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7382
7383                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7384                                                     num_bytes, empty_size,
7385                                                     &max_extent_size);
7386                 /*
7387                  * If we didn't find a chunk, and we haven't failed on this
7388                  * block group before, and this block group is in the middle of
7389                  * caching and we are ok with waiting, then go ahead and wait
7390                  * for progress to be made, and set failed_alloc to true.
7391                  *
7392                  * If failed_alloc is true then we've already waited on this
7393                  * block group once and should move on to the next block group.
7394                  */
7395                 if (!offset && !failed_alloc && !cached &&
7396                     loop > LOOP_CACHING_NOWAIT) {
7397                         wait_block_group_cache_progress(block_group,
7398                                                 num_bytes + empty_size);
7399                         failed_alloc = true;
7400                         goto have_block_group;
7401                 } else if (!offset) {
7402                         goto loop;
7403                 }
7404 checks:
7405                 search_start = ALIGN(offset, root->stripesize);
7406
7407                 /* move on to the next group */
7408                 if (search_start + num_bytes >
7409                     block_group->key.objectid + block_group->key.offset) {
7410                         btrfs_add_free_space(block_group, offset, num_bytes);
7411                         goto loop;
7412                 }
7413
7414                 if (offset < search_start)
7415                         btrfs_add_free_space(block_group, offset,
7416                                              search_start - offset);
7417                 BUG_ON(offset > search_start);
7418
7419                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7420                                                   alloc_type, delalloc);
7421                 if (ret == -EAGAIN) {
7422                         btrfs_add_free_space(block_group, offset, num_bytes);
7423                         goto loop;
7424                 }
7425
7426                 /* we are all good, lets return */
7427                 ins->objectid = search_start;
7428                 ins->offset = num_bytes;
7429
7430                 trace_btrfs_reserve_extent(orig_root, block_group,
7431                                            search_start, num_bytes);
7432                 btrfs_release_block_group(block_group, delalloc);
7433                 break;
7434 loop:
7435                 failed_cluster_refill = false;
7436                 failed_alloc = false;
7437                 BUG_ON(index != get_block_group_index(block_group));
7438                 btrfs_release_block_group(block_group, delalloc);
7439         }
7440         up_read(&space_info->groups_sem);
7441
7442         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7443                 && !orig_have_caching_bg)
7444                 orig_have_caching_bg = true;
7445
7446         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7447                 goto search;
7448
7449         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7450                 goto search;
7451
7452         /*
7453          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7454          *                      caching kthreads as we move along
7455          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7456          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7457          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7458          *                      again
7459          */
7460         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7461                 index = 0;
7462                 if (loop == LOOP_CACHING_NOWAIT) {
7463                         /*
7464                          * We want to skip the LOOP_CACHING_WAIT step if we
7465                          * don't have any unached bgs and we've alrelady done a
7466                          * full search through.
7467                          */
7468                         if (orig_have_caching_bg || !full_search)
7469                                 loop = LOOP_CACHING_WAIT;
7470                         else
7471                                 loop = LOOP_ALLOC_CHUNK;
7472                 } else {
7473                         loop++;
7474                 }
7475
7476                 if (loop == LOOP_ALLOC_CHUNK) {
7477                         struct btrfs_trans_handle *trans;
7478                         int exist = 0;
7479
7480                         trans = current->journal_info;
7481                         if (trans)
7482                                 exist = 1;
7483                         else
7484                                 trans = btrfs_join_transaction(root);
7485
7486                         if (IS_ERR(trans)) {
7487                                 ret = PTR_ERR(trans);
7488                                 goto out;
7489                         }
7490
7491                         ret = do_chunk_alloc(trans, root, flags,
7492                                              CHUNK_ALLOC_FORCE);
7493
7494                         /*
7495                          * If we can't allocate a new chunk we've already looped
7496                          * through at least once, move on to the NO_EMPTY_SIZE
7497                          * case.
7498                          */
7499                         if (ret == -ENOSPC)
7500                                 loop = LOOP_NO_EMPTY_SIZE;
7501
7502                         /*
7503                          * Do not bail out on ENOSPC since we
7504                          * can do more things.
7505                          */
7506                         if (ret < 0 && ret != -ENOSPC)
7507                                 btrfs_abort_transaction(trans,
7508                                                         root, ret);
7509                         else
7510                                 ret = 0;
7511                         if (!exist)
7512                                 btrfs_end_transaction(trans, root);
7513                         if (ret)
7514                                 goto out;
7515                 }
7516
7517                 if (loop == LOOP_NO_EMPTY_SIZE) {
7518                         /*
7519                          * Don't loop again if we already have no empty_size and
7520                          * no empty_cluster.
7521                          */
7522                         if (empty_size == 0 &&
7523                             empty_cluster == 0) {
7524                                 ret = -ENOSPC;
7525                                 goto out;
7526                         }
7527                         empty_size = 0;
7528                         empty_cluster = 0;
7529                 }
7530
7531                 goto search;
7532         } else if (!ins->objectid) {
7533                 ret = -ENOSPC;
7534         } else if (ins->objectid) {
7535                 if (!use_cluster && last_ptr) {
7536                         spin_lock(&last_ptr->lock);
7537                         last_ptr->window_start = ins->objectid;
7538                         spin_unlock(&last_ptr->lock);
7539                 }
7540                 ret = 0;
7541         }
7542 out:
7543         if (ret == -ENOSPC) {
7544                 spin_lock(&space_info->lock);
7545                 space_info->max_extent_size = max_extent_size;
7546                 spin_unlock(&space_info->lock);
7547                 ins->offset = max_extent_size;
7548         }
7549         return ret;
7550 }
7551
7552 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7553                             int dump_block_groups)
7554 {
7555         struct btrfs_block_group_cache *cache;
7556         int index = 0;
7557
7558         spin_lock(&info->lock);
7559         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7560                info->flags,
7561                info->total_bytes - info->bytes_used - info->bytes_pinned -
7562                info->bytes_reserved - info->bytes_readonly,
7563                (info->full) ? "" : "not ");
7564         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7565                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7566                info->total_bytes, info->bytes_used, info->bytes_pinned,
7567                info->bytes_reserved, info->bytes_may_use,
7568                info->bytes_readonly);
7569         spin_unlock(&info->lock);
7570
7571         if (!dump_block_groups)
7572                 return;
7573
7574         down_read(&info->groups_sem);
7575 again:
7576         list_for_each_entry(cache, &info->block_groups[index], list) {
7577                 spin_lock(&cache->lock);
7578                 printk(KERN_INFO "BTRFS: "
7579                            "block group %llu has %llu bytes, "
7580                            "%llu used %llu pinned %llu reserved %s\n",
7581                        cache->key.objectid, cache->key.offset,
7582                        btrfs_block_group_used(&cache->item), cache->pinned,
7583                        cache->reserved, cache->ro ? "[readonly]" : "");
7584                 btrfs_dump_free_space(cache, bytes);
7585                 spin_unlock(&cache->lock);
7586         }
7587         if (++index < BTRFS_NR_RAID_TYPES)
7588                 goto again;
7589         up_read(&info->groups_sem);
7590 }
7591
7592 int btrfs_reserve_extent(struct btrfs_root *root,
7593                          u64 num_bytes, u64 min_alloc_size,
7594                          u64 empty_size, u64 hint_byte,
7595                          struct btrfs_key *ins, int is_data, int delalloc)
7596 {
7597         bool final_tried = num_bytes == min_alloc_size;
7598         u64 flags;
7599         int ret;
7600
7601         flags = btrfs_get_alloc_profile(root, is_data);
7602 again:
7603         WARN_ON(num_bytes < root->sectorsize);
7604         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7605                                flags, delalloc);
7606
7607         if (ret == -ENOSPC) {
7608                 if (!final_tried && ins->offset) {
7609                         num_bytes = min(num_bytes >> 1, ins->offset);
7610                         num_bytes = round_down(num_bytes, root->sectorsize);
7611                         num_bytes = max(num_bytes, min_alloc_size);
7612                         if (num_bytes == min_alloc_size)
7613                                 final_tried = true;
7614                         goto again;
7615                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7616                         struct btrfs_space_info *sinfo;
7617
7618                         sinfo = __find_space_info(root->fs_info, flags);
7619                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7620                                 flags, num_bytes);
7621                         if (sinfo)
7622                                 dump_space_info(sinfo, num_bytes, 1);
7623                 }
7624         }
7625
7626         return ret;
7627 }
7628
7629 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7630                                         u64 start, u64 len,
7631                                         int pin, int delalloc)
7632 {
7633         struct btrfs_block_group_cache *cache;
7634         int ret = 0;
7635
7636         cache = btrfs_lookup_block_group(root->fs_info, start);
7637         if (!cache) {
7638                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7639                         start);
7640                 return -ENOSPC;
7641         }
7642
7643         if (pin)
7644                 pin_down_extent(root, cache, start, len, 1);
7645         else {
7646                 if (btrfs_test_opt(root, DISCARD))
7647                         ret = btrfs_discard_extent(root, start, len, NULL);
7648                 btrfs_add_free_space(cache, start, len);
7649                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7650         }
7651
7652         btrfs_put_block_group(cache);
7653
7654         trace_btrfs_reserved_extent_free(root, start, len);
7655
7656         return ret;
7657 }
7658
7659 int btrfs_free_reserved_extent(struct btrfs_root *root,
7660                                u64 start, u64 len, int delalloc)
7661 {
7662         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7663 }
7664
7665 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7666                                        u64 start, u64 len)
7667 {
7668         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7669 }
7670
7671 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7672                                       struct btrfs_root *root,
7673                                       u64 parent, u64 root_objectid,
7674                                       u64 flags, u64 owner, u64 offset,
7675                                       struct btrfs_key *ins, int ref_mod)
7676 {
7677         int ret;
7678         struct btrfs_fs_info *fs_info = root->fs_info;
7679         struct btrfs_extent_item *extent_item;
7680         struct btrfs_extent_inline_ref *iref;
7681         struct btrfs_path *path;
7682         struct extent_buffer *leaf;
7683         int type;
7684         u32 size;
7685
7686         if (parent > 0)
7687                 type = BTRFS_SHARED_DATA_REF_KEY;
7688         else
7689                 type = BTRFS_EXTENT_DATA_REF_KEY;
7690
7691         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7692
7693         path = btrfs_alloc_path();
7694         if (!path)
7695                 return -ENOMEM;
7696
7697         path->leave_spinning = 1;
7698         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7699                                       ins, size);
7700         if (ret) {
7701                 btrfs_free_path(path);
7702                 return ret;
7703         }
7704
7705         leaf = path->nodes[0];
7706         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7707                                      struct btrfs_extent_item);
7708         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7709         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7710         btrfs_set_extent_flags(leaf, extent_item,
7711                                flags | BTRFS_EXTENT_FLAG_DATA);
7712
7713         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7714         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7715         if (parent > 0) {
7716                 struct btrfs_shared_data_ref *ref;
7717                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7718                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7719                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7720         } else {
7721                 struct btrfs_extent_data_ref *ref;
7722                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7723                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7724                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7725                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7726                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7727         }
7728
7729         btrfs_mark_buffer_dirty(path->nodes[0]);
7730         btrfs_free_path(path);
7731
7732         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
7733                                           ins->offset);
7734         if (ret)
7735                 return ret;
7736
7737         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7738         if (ret) { /* -ENOENT, logic error */
7739                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7740                         ins->objectid, ins->offset);
7741                 BUG();
7742         }
7743         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7744         return ret;
7745 }
7746
7747 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7748                                      struct btrfs_root *root,
7749                                      u64 parent, u64 root_objectid,
7750                                      u64 flags, struct btrfs_disk_key *key,
7751                                      int level, struct btrfs_key *ins)
7752 {
7753         int ret;
7754         struct btrfs_fs_info *fs_info = root->fs_info;
7755         struct btrfs_extent_item *extent_item;
7756         struct btrfs_tree_block_info *block_info;
7757         struct btrfs_extent_inline_ref *iref;
7758         struct btrfs_path *path;
7759         struct extent_buffer *leaf;
7760         u32 size = sizeof(*extent_item) + sizeof(*iref);
7761         u64 num_bytes = ins->offset;
7762         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7763                                                  SKINNY_METADATA);
7764
7765         if (!skinny_metadata)
7766                 size += sizeof(*block_info);
7767
7768         path = btrfs_alloc_path();
7769         if (!path) {
7770                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7771                                                    root->nodesize);
7772                 return -ENOMEM;
7773         }
7774
7775         path->leave_spinning = 1;
7776         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7777                                       ins, size);
7778         if (ret) {
7779                 btrfs_free_path(path);
7780                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7781                                                    root->nodesize);
7782                 return ret;
7783         }
7784
7785         leaf = path->nodes[0];
7786         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7787                                      struct btrfs_extent_item);
7788         btrfs_set_extent_refs(leaf, extent_item, 1);
7789         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7790         btrfs_set_extent_flags(leaf, extent_item,
7791                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7792
7793         if (skinny_metadata) {
7794                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7795                 num_bytes = root->nodesize;
7796         } else {
7797                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7798                 btrfs_set_tree_block_key(leaf, block_info, key);
7799                 btrfs_set_tree_block_level(leaf, block_info, level);
7800                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7801         }
7802
7803         if (parent > 0) {
7804                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7805                 btrfs_set_extent_inline_ref_type(leaf, iref,
7806                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7807                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7808         } else {
7809                 btrfs_set_extent_inline_ref_type(leaf, iref,
7810                                                  BTRFS_TREE_BLOCK_REF_KEY);
7811                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7812         }
7813
7814         btrfs_mark_buffer_dirty(leaf);
7815         btrfs_free_path(path);
7816
7817         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
7818                                           num_bytes);
7819         if (ret)
7820                 return ret;
7821
7822         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7823                                  1);
7824         if (ret) { /* -ENOENT, logic error */
7825                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7826                         ins->objectid, ins->offset);
7827                 BUG();
7828         }
7829
7830         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7831         return ret;
7832 }
7833
7834 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7835                                      struct btrfs_root *root,
7836                                      u64 root_objectid, u64 owner,
7837                                      u64 offset, u64 ram_bytes,
7838                                      struct btrfs_key *ins)
7839 {
7840         int ret;
7841
7842         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7843
7844         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7845                                          ins->offset, 0,
7846                                          root_objectid, owner, offset,
7847                                          ram_bytes, BTRFS_ADD_DELAYED_EXTENT,
7848                                          NULL);
7849         return ret;
7850 }
7851
7852 /*
7853  * this is used by the tree logging recovery code.  It records that
7854  * an extent has been allocated and makes sure to clear the free
7855  * space cache bits as well
7856  */
7857 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7858                                    struct btrfs_root *root,
7859                                    u64 root_objectid, u64 owner, u64 offset,
7860                                    struct btrfs_key *ins)
7861 {
7862         int ret;
7863         struct btrfs_block_group_cache *block_group;
7864
7865         /*
7866          * Mixed block groups will exclude before processing the log so we only
7867          * need to do the exlude dance if this fs isn't mixed.
7868          */
7869         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7870                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7871                 if (ret)
7872                         return ret;
7873         }
7874
7875         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7876         if (!block_group)
7877                 return -EINVAL;
7878
7879         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7880                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7881         BUG_ON(ret); /* logic error */
7882         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7883                                          0, owner, offset, ins, 1);
7884         btrfs_put_block_group(block_group);
7885         return ret;
7886 }
7887
7888 static struct extent_buffer *
7889 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7890                       u64 bytenr, int level)
7891 {
7892         struct extent_buffer *buf;
7893
7894         buf = btrfs_find_create_tree_block(root, bytenr);
7895         if (!buf)
7896                 return ERR_PTR(-ENOMEM);
7897         btrfs_set_header_generation(buf, trans->transid);
7898         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7899         btrfs_tree_lock(buf);
7900         clean_tree_block(trans, root->fs_info, buf);
7901         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7902
7903         btrfs_set_lock_blocking(buf);
7904         set_extent_buffer_uptodate(buf);
7905
7906         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7907                 buf->log_index = root->log_transid % 2;
7908                 /*
7909                  * we allow two log transactions at a time, use different
7910                  * EXENT bit to differentiate dirty pages.
7911                  */
7912                 if (buf->log_index == 0)
7913                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7914                                         buf->start + buf->len - 1, GFP_NOFS);
7915                 else
7916                         set_extent_new(&root->dirty_log_pages, buf->start,
7917                                         buf->start + buf->len - 1, GFP_NOFS);
7918         } else {
7919                 buf->log_index = -1;
7920                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7921                          buf->start + buf->len - 1, GFP_NOFS);
7922         }
7923         trans->blocks_used++;
7924         /* this returns a buffer locked for blocking */
7925         return buf;
7926 }
7927
7928 static struct btrfs_block_rsv *
7929 use_block_rsv(struct btrfs_trans_handle *trans,
7930               struct btrfs_root *root, u32 blocksize)
7931 {
7932         struct btrfs_block_rsv *block_rsv;
7933         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7934         int ret;
7935         bool global_updated = false;
7936
7937         block_rsv = get_block_rsv(trans, root);
7938
7939         if (unlikely(block_rsv->size == 0))
7940                 goto try_reserve;
7941 again:
7942         ret = block_rsv_use_bytes(block_rsv, blocksize);
7943         if (!ret)
7944                 return block_rsv;
7945
7946         if (block_rsv->failfast)
7947                 return ERR_PTR(ret);
7948
7949         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7950                 global_updated = true;
7951                 update_global_block_rsv(root->fs_info);
7952                 goto again;
7953         }
7954
7955         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7956                 static DEFINE_RATELIMIT_STATE(_rs,
7957                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7958                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7959                 if (__ratelimit(&_rs))
7960                         WARN(1, KERN_DEBUG
7961                                 "BTRFS: block rsv returned %d\n", ret);
7962         }
7963 try_reserve:
7964         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7965                                      BTRFS_RESERVE_NO_FLUSH);
7966         if (!ret)
7967                 return block_rsv;
7968         /*
7969          * If we couldn't reserve metadata bytes try and use some from
7970          * the global reserve if its space type is the same as the global
7971          * reservation.
7972          */
7973         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7974             block_rsv->space_info == global_rsv->space_info) {
7975                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7976                 if (!ret)
7977                         return global_rsv;
7978         }
7979         return ERR_PTR(ret);
7980 }
7981
7982 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7983                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7984 {
7985         block_rsv_add_bytes(block_rsv, blocksize, 0);
7986         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7987 }
7988
7989 /*
7990  * finds a free extent and does all the dirty work required for allocation
7991  * returns the tree buffer or an ERR_PTR on error.
7992  */
7993 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7994                                         struct btrfs_root *root,
7995                                         u64 parent, u64 root_objectid,
7996                                         struct btrfs_disk_key *key, int level,
7997                                         u64 hint, u64 empty_size)
7998 {
7999         struct btrfs_key ins;
8000         struct btrfs_block_rsv *block_rsv;
8001         struct extent_buffer *buf;
8002         struct btrfs_delayed_extent_op *extent_op;
8003         u64 flags = 0;
8004         int ret;
8005         u32 blocksize = root->nodesize;
8006         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8007                                                  SKINNY_METADATA);
8008
8009         if (btrfs_test_is_dummy_root(root)) {
8010                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8011                                             level);
8012                 if (!IS_ERR(buf))
8013                         root->alloc_bytenr += blocksize;
8014                 return buf;
8015         }
8016
8017         block_rsv = use_block_rsv(trans, root, blocksize);
8018         if (IS_ERR(block_rsv))
8019                 return ERR_CAST(block_rsv);
8020
8021         ret = btrfs_reserve_extent(root, blocksize, blocksize,
8022                                    empty_size, hint, &ins, 0, 0);
8023         if (ret)
8024                 goto out_unuse;
8025
8026         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8027         if (IS_ERR(buf)) {
8028                 ret = PTR_ERR(buf);
8029                 goto out_free_reserved;
8030         }
8031
8032         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8033                 if (parent == 0)
8034                         parent = ins.objectid;
8035                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8036         } else
8037                 BUG_ON(parent > 0);
8038
8039         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8040                 extent_op = btrfs_alloc_delayed_extent_op();
8041                 if (!extent_op) {
8042                         ret = -ENOMEM;
8043                         goto out_free_buf;
8044                 }
8045                 if (key)
8046                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8047                 else
8048                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8049                 extent_op->flags_to_set = flags;
8050                 extent_op->update_key = skinny_metadata ? false : true;
8051                 extent_op->update_flags = true;
8052                 extent_op->is_data = false;
8053                 extent_op->level = level;
8054
8055                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
8056                                                  ins.objectid, ins.offset,
8057                                                  parent, root_objectid, level,
8058                                                  BTRFS_ADD_DELAYED_EXTENT,
8059                                                  extent_op);
8060                 if (ret)
8061                         goto out_free_delayed;
8062         }
8063         return buf;
8064
8065 out_free_delayed:
8066         btrfs_free_delayed_extent_op(extent_op);
8067 out_free_buf:
8068         free_extent_buffer(buf);
8069 out_free_reserved:
8070         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
8071 out_unuse:
8072         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
8073         return ERR_PTR(ret);
8074 }
8075
8076 struct walk_control {
8077         u64 refs[BTRFS_MAX_LEVEL];
8078         u64 flags[BTRFS_MAX_LEVEL];
8079         struct btrfs_key update_progress;
8080         int stage;
8081         int level;
8082         int shared_level;
8083         int update_ref;
8084         int keep_locks;
8085         int reada_slot;
8086         int reada_count;
8087         int for_reloc;
8088 };
8089
8090 #define DROP_REFERENCE  1
8091 #define UPDATE_BACKREF  2
8092
8093 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8094                                      struct btrfs_root *root,
8095                                      struct walk_control *wc,
8096                                      struct btrfs_path *path)
8097 {
8098         u64 bytenr;
8099         u64 generation;
8100         u64 refs;
8101         u64 flags;
8102         u32 nritems;
8103         u32 blocksize;
8104         struct btrfs_key key;
8105         struct extent_buffer *eb;
8106         int ret;
8107         int slot;
8108         int nread = 0;
8109
8110         if (path->slots[wc->level] < wc->reada_slot) {
8111                 wc->reada_count = wc->reada_count * 2 / 3;
8112                 wc->reada_count = max(wc->reada_count, 2);
8113         } else {
8114                 wc->reada_count = wc->reada_count * 3 / 2;
8115                 wc->reada_count = min_t(int, wc->reada_count,
8116                                         BTRFS_NODEPTRS_PER_BLOCK(root));
8117         }
8118
8119         eb = path->nodes[wc->level];
8120         nritems = btrfs_header_nritems(eb);
8121         blocksize = root->nodesize;
8122
8123         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8124                 if (nread >= wc->reada_count)
8125                         break;
8126
8127                 cond_resched();
8128                 bytenr = btrfs_node_blockptr(eb, slot);
8129                 generation = btrfs_node_ptr_generation(eb, slot);
8130
8131                 if (slot == path->slots[wc->level])
8132                         goto reada;
8133
8134                 if (wc->stage == UPDATE_BACKREF &&
8135                     generation <= root->root_key.offset)
8136                         continue;
8137
8138                 /* We don't lock the tree block, it's OK to be racy here */
8139                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
8140                                                wc->level - 1, 1, &refs,
8141                                                &flags);
8142                 /* We don't care about errors in readahead. */
8143                 if (ret < 0)
8144                         continue;
8145                 BUG_ON(refs == 0);
8146
8147                 if (wc->stage == DROP_REFERENCE) {
8148                         if (refs == 1)
8149                                 goto reada;
8150
8151                         if (wc->level == 1 &&
8152                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8153                                 continue;
8154                         if (!wc->update_ref ||
8155                             generation <= root->root_key.offset)
8156                                 continue;
8157                         btrfs_node_key_to_cpu(eb, &key, slot);
8158                         ret = btrfs_comp_cpu_keys(&key,
8159                                                   &wc->update_progress);
8160                         if (ret < 0)
8161                                 continue;
8162                 } else {
8163                         if (wc->level == 1 &&
8164                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8165                                 continue;
8166                 }
8167 reada:
8168                 readahead_tree_block(root, bytenr);
8169                 nread++;
8170         }
8171         wc->reada_slot = slot;
8172 }
8173
8174 /*
8175  * These may not be seen by the usual inc/dec ref code so we have to
8176  * add them here.
8177  */
8178 static int record_one_subtree_extent(struct btrfs_trans_handle *trans,
8179                                      struct btrfs_root *root, u64 bytenr,
8180                                      u64 num_bytes)
8181 {
8182         struct btrfs_qgroup_extent_record *qrecord;
8183         struct btrfs_delayed_ref_root *delayed_refs;
8184
8185         qrecord = kmalloc(sizeof(*qrecord), GFP_NOFS);
8186         if (!qrecord)
8187                 return -ENOMEM;
8188
8189         qrecord->bytenr = bytenr;
8190         qrecord->num_bytes = num_bytes;
8191         qrecord->old_roots = NULL;
8192
8193         delayed_refs = &trans->transaction->delayed_refs;
8194         spin_lock(&delayed_refs->lock);
8195         if (btrfs_qgroup_insert_dirty_extent(delayed_refs, qrecord))
8196                 kfree(qrecord);
8197         spin_unlock(&delayed_refs->lock);
8198
8199         return 0;
8200 }
8201
8202 static int account_leaf_items(struct btrfs_trans_handle *trans,
8203                               struct btrfs_root *root,
8204                               struct extent_buffer *eb)
8205 {
8206         int nr = btrfs_header_nritems(eb);
8207         int i, extent_type, ret;
8208         struct btrfs_key key;
8209         struct btrfs_file_extent_item *fi;
8210         u64 bytenr, num_bytes;
8211
8212         /* We can be called directly from walk_up_proc() */
8213         if (!root->fs_info->quota_enabled)
8214                 return 0;
8215
8216         for (i = 0; i < nr; i++) {
8217                 btrfs_item_key_to_cpu(eb, &key, i);
8218
8219                 if (key.type != BTRFS_EXTENT_DATA_KEY)
8220                         continue;
8221
8222                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
8223                 /* filter out non qgroup-accountable extents  */
8224                 extent_type = btrfs_file_extent_type(eb, fi);
8225
8226                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
8227                         continue;
8228
8229                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
8230                 if (!bytenr)
8231                         continue;
8232
8233                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
8234
8235                 ret = record_one_subtree_extent(trans, root, bytenr, num_bytes);
8236                 if (ret)
8237                         return ret;
8238         }
8239         return 0;
8240 }
8241
8242 /*
8243  * Walk up the tree from the bottom, freeing leaves and any interior
8244  * nodes which have had all slots visited. If a node (leaf or
8245  * interior) is freed, the node above it will have it's slot
8246  * incremented. The root node will never be freed.
8247  *
8248  * At the end of this function, we should have a path which has all
8249  * slots incremented to the next position for a search. If we need to
8250  * read a new node it will be NULL and the node above it will have the
8251  * correct slot selected for a later read.
8252  *
8253  * If we increment the root nodes slot counter past the number of
8254  * elements, 1 is returned to signal completion of the search.
8255  */
8256 static int adjust_slots_upwards(struct btrfs_root *root,
8257                                 struct btrfs_path *path, int root_level)
8258 {
8259         int level = 0;
8260         int nr, slot;
8261         struct extent_buffer *eb;
8262
8263         if (root_level == 0)
8264                 return 1;
8265
8266         while (level <= root_level) {
8267                 eb = path->nodes[level];
8268                 nr = btrfs_header_nritems(eb);
8269                 path->slots[level]++;
8270                 slot = path->slots[level];
8271                 if (slot >= nr || level == 0) {
8272                         /*
8273                          * Don't free the root -  we will detect this
8274                          * condition after our loop and return a
8275                          * positive value for caller to stop walking the tree.
8276                          */
8277                         if (level != root_level) {
8278                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8279                                 path->locks[level] = 0;
8280
8281                                 free_extent_buffer(eb);
8282                                 path->nodes[level] = NULL;
8283                                 path->slots[level] = 0;
8284                         }
8285                 } else {
8286                         /*
8287                          * We have a valid slot to walk back down
8288                          * from. Stop here so caller can process these
8289                          * new nodes.
8290                          */
8291                         break;
8292                 }
8293
8294                 level++;
8295         }
8296
8297         eb = path->nodes[root_level];
8298         if (path->slots[root_level] >= btrfs_header_nritems(eb))
8299                 return 1;
8300
8301         return 0;
8302 }
8303
8304 /*
8305  * root_eb is the subtree root and is locked before this function is called.
8306  */
8307 static int account_shared_subtree(struct btrfs_trans_handle *trans,
8308                                   struct btrfs_root *root,
8309                                   struct extent_buffer *root_eb,
8310                                   u64 root_gen,
8311                                   int root_level)
8312 {
8313         int ret = 0;
8314         int level;
8315         struct extent_buffer *eb = root_eb;
8316         struct btrfs_path *path = NULL;
8317
8318         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
8319         BUG_ON(root_eb == NULL);
8320
8321         if (!root->fs_info->quota_enabled)
8322                 return 0;
8323
8324         if (!extent_buffer_uptodate(root_eb)) {
8325                 ret = btrfs_read_buffer(root_eb, root_gen);
8326                 if (ret)
8327                         goto out;
8328         }
8329
8330         if (root_level == 0) {
8331                 ret = account_leaf_items(trans, root, root_eb);
8332                 goto out;
8333         }
8334
8335         path = btrfs_alloc_path();
8336         if (!path)
8337                 return -ENOMEM;
8338
8339         /*
8340          * Walk down the tree.  Missing extent blocks are filled in as
8341          * we go. Metadata is accounted every time we read a new
8342          * extent block.
8343          *
8344          * When we reach a leaf, we account for file extent items in it,
8345          * walk back up the tree (adjusting slot pointers as we go)
8346          * and restart the search process.
8347          */
8348         extent_buffer_get(root_eb); /* For path */
8349         path->nodes[root_level] = root_eb;
8350         path->slots[root_level] = 0;
8351         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
8352 walk_down:
8353         level = root_level;
8354         while (level >= 0) {
8355                 if (path->nodes[level] == NULL) {
8356                         int parent_slot;
8357                         u64 child_gen;
8358                         u64 child_bytenr;
8359
8360                         /* We need to get child blockptr/gen from
8361                          * parent before we can read it. */
8362                         eb = path->nodes[level + 1];
8363                         parent_slot = path->slots[level + 1];
8364                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8365                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8366
8367                         eb = read_tree_block(root, child_bytenr, child_gen);
8368                         if (IS_ERR(eb)) {
8369                                 ret = PTR_ERR(eb);
8370                                 goto out;
8371                         } else if (!extent_buffer_uptodate(eb)) {
8372                                 free_extent_buffer(eb);
8373                                 ret = -EIO;
8374                                 goto out;
8375                         }
8376
8377                         path->nodes[level] = eb;
8378                         path->slots[level] = 0;
8379
8380                         btrfs_tree_read_lock(eb);
8381                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8382                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8383
8384                         ret = record_one_subtree_extent(trans, root, child_bytenr,
8385                                                         root->nodesize);
8386                         if (ret)
8387                                 goto out;
8388                 }
8389
8390                 if (level == 0) {
8391                         ret = account_leaf_items(trans, root, path->nodes[level]);
8392                         if (ret)
8393                                 goto out;
8394
8395                         /* Nonzero return here means we completed our search */
8396                         ret = adjust_slots_upwards(root, path, root_level);
8397                         if (ret)
8398                                 break;
8399
8400                         /* Restart search with new slots */
8401                         goto walk_down;
8402                 }
8403
8404                 level--;
8405         }
8406
8407         ret = 0;
8408 out:
8409         btrfs_free_path(path);
8410
8411         return ret;
8412 }
8413
8414 /*
8415  * helper to process tree block while walking down the tree.
8416  *
8417  * when wc->stage == UPDATE_BACKREF, this function updates
8418  * back refs for pointers in the block.
8419  *
8420  * NOTE: return value 1 means we should stop walking down.
8421  */
8422 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8423                                    struct btrfs_root *root,
8424                                    struct btrfs_path *path,
8425                                    struct walk_control *wc, int lookup_info)
8426 {
8427         int level = wc->level;
8428         struct extent_buffer *eb = path->nodes[level];
8429         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8430         int ret;
8431
8432         if (wc->stage == UPDATE_BACKREF &&
8433             btrfs_header_owner(eb) != root->root_key.objectid)
8434                 return 1;
8435
8436         /*
8437          * when reference count of tree block is 1, it won't increase
8438          * again. once full backref flag is set, we never clear it.
8439          */
8440         if (lookup_info &&
8441             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8442              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8443                 BUG_ON(!path->locks[level]);
8444                 ret = btrfs_lookup_extent_info(trans, root,
8445                                                eb->start, level, 1,
8446                                                &wc->refs[level],
8447                                                &wc->flags[level]);
8448                 BUG_ON(ret == -ENOMEM);
8449                 if (ret)
8450                         return ret;
8451                 BUG_ON(wc->refs[level] == 0);
8452         }
8453
8454         if (wc->stage == DROP_REFERENCE) {
8455                 if (wc->refs[level] > 1)
8456                         return 1;
8457
8458                 if (path->locks[level] && !wc->keep_locks) {
8459                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8460                         path->locks[level] = 0;
8461                 }
8462                 return 0;
8463         }
8464
8465         /* wc->stage == UPDATE_BACKREF */
8466         if (!(wc->flags[level] & flag)) {
8467                 BUG_ON(!path->locks[level]);
8468                 ret = btrfs_inc_ref(trans, root, eb, 1);
8469                 BUG_ON(ret); /* -ENOMEM */
8470                 ret = btrfs_dec_ref(trans, root, eb, 0);
8471                 BUG_ON(ret); /* -ENOMEM */
8472                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8473                                                   eb->len, flag,
8474                                                   btrfs_header_level(eb), 0);
8475                 BUG_ON(ret); /* -ENOMEM */
8476                 wc->flags[level] |= flag;
8477         }
8478
8479         /*
8480          * the block is shared by multiple trees, so it's not good to
8481          * keep the tree lock
8482          */
8483         if (path->locks[level] && level > 0) {
8484                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8485                 path->locks[level] = 0;
8486         }
8487         return 0;
8488 }
8489
8490 /*
8491  * helper to process tree block pointer.
8492  *
8493  * when wc->stage == DROP_REFERENCE, this function checks
8494  * reference count of the block pointed to. if the block
8495  * is shared and we need update back refs for the subtree
8496  * rooted at the block, this function changes wc->stage to
8497  * UPDATE_BACKREF. if the block is shared and there is no
8498  * need to update back, this function drops the reference
8499  * to the block.
8500  *
8501  * NOTE: return value 1 means we should stop walking down.
8502  */
8503 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8504                                  struct btrfs_root *root,
8505                                  struct btrfs_path *path,
8506                                  struct walk_control *wc, int *lookup_info)
8507 {
8508         u64 bytenr;
8509         u64 generation;
8510         u64 parent;
8511         u32 blocksize;
8512         struct btrfs_key key;
8513         struct extent_buffer *next;
8514         int level = wc->level;
8515         int reada = 0;
8516         int ret = 0;
8517         bool need_account = false;
8518
8519         generation = btrfs_node_ptr_generation(path->nodes[level],
8520                                                path->slots[level]);
8521         /*
8522          * if the lower level block was created before the snapshot
8523          * was created, we know there is no need to update back refs
8524          * for the subtree
8525          */
8526         if (wc->stage == UPDATE_BACKREF &&
8527             generation <= root->root_key.offset) {
8528                 *lookup_info = 1;
8529                 return 1;
8530         }
8531
8532         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8533         blocksize = root->nodesize;
8534
8535         next = btrfs_find_tree_block(root->fs_info, bytenr);
8536         if (!next) {
8537                 next = btrfs_find_create_tree_block(root, bytenr);
8538                 if (!next)
8539                         return -ENOMEM;
8540                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8541                                                level - 1);
8542                 reada = 1;
8543         }
8544         btrfs_tree_lock(next);
8545         btrfs_set_lock_blocking(next);
8546
8547         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8548                                        &wc->refs[level - 1],
8549                                        &wc->flags[level - 1]);
8550         if (ret < 0) {
8551                 btrfs_tree_unlock(next);
8552                 return ret;
8553         }
8554
8555         if (unlikely(wc->refs[level - 1] == 0)) {
8556                 btrfs_err(root->fs_info, "Missing references.");
8557                 BUG();
8558         }
8559         *lookup_info = 0;
8560
8561         if (wc->stage == DROP_REFERENCE) {
8562                 if (wc->refs[level - 1] > 1) {
8563                         need_account = true;
8564                         if (level == 1 &&
8565                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8566                                 goto skip;
8567
8568                         if (!wc->update_ref ||
8569                             generation <= root->root_key.offset)
8570                                 goto skip;
8571
8572                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8573                                               path->slots[level]);
8574                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8575                         if (ret < 0)
8576                                 goto skip;
8577
8578                         wc->stage = UPDATE_BACKREF;
8579                         wc->shared_level = level - 1;
8580                 }
8581         } else {
8582                 if (level == 1 &&
8583                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8584                         goto skip;
8585         }
8586
8587         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8588                 btrfs_tree_unlock(next);
8589                 free_extent_buffer(next);
8590                 next = NULL;
8591                 *lookup_info = 1;
8592         }
8593
8594         if (!next) {
8595                 if (reada && level == 1)
8596                         reada_walk_down(trans, root, wc, path);
8597                 next = read_tree_block(root, bytenr, generation);
8598                 if (IS_ERR(next)) {
8599                         return PTR_ERR(next);
8600                 } else if (!extent_buffer_uptodate(next)) {
8601                         free_extent_buffer(next);
8602                         return -EIO;
8603                 }
8604                 btrfs_tree_lock(next);
8605                 btrfs_set_lock_blocking(next);
8606         }
8607
8608         level--;
8609         BUG_ON(level != btrfs_header_level(next));
8610         path->nodes[level] = next;
8611         path->slots[level] = 0;
8612         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8613         wc->level = level;
8614         if (wc->level == 1)
8615                 wc->reada_slot = 0;
8616         return 0;
8617 skip:
8618         wc->refs[level - 1] = 0;
8619         wc->flags[level - 1] = 0;
8620         if (wc->stage == DROP_REFERENCE) {
8621                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8622                         parent = path->nodes[level]->start;
8623                 } else {
8624                         BUG_ON(root->root_key.objectid !=
8625                                btrfs_header_owner(path->nodes[level]));
8626                         parent = 0;
8627                 }
8628
8629                 if (need_account) {
8630                         ret = account_shared_subtree(trans, root, next,
8631                                                      generation, level - 1);
8632                         if (ret) {
8633                                 btrfs_err_rl(root->fs_info,
8634                                         "Error "
8635                                         "%d accounting shared subtree. Quota "
8636                                         "is out of sync, rescan required.",
8637                                         ret);
8638                         }
8639                 }
8640                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8641                                 root->root_key.objectid, level - 1, 0);
8642                 BUG_ON(ret); /* -ENOMEM */
8643         }
8644         btrfs_tree_unlock(next);
8645         free_extent_buffer(next);
8646         *lookup_info = 1;
8647         return 1;
8648 }
8649
8650 /*
8651  * helper to process tree block while walking up the tree.
8652  *
8653  * when wc->stage == DROP_REFERENCE, this function drops
8654  * reference count on the block.
8655  *
8656  * when wc->stage == UPDATE_BACKREF, this function changes
8657  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8658  * to UPDATE_BACKREF previously while processing the block.
8659  *
8660  * NOTE: return value 1 means we should stop walking up.
8661  */
8662 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8663                                  struct btrfs_root *root,
8664                                  struct btrfs_path *path,
8665                                  struct walk_control *wc)
8666 {
8667         int ret;
8668         int level = wc->level;
8669         struct extent_buffer *eb = path->nodes[level];
8670         u64 parent = 0;
8671
8672         if (wc->stage == UPDATE_BACKREF) {
8673                 BUG_ON(wc->shared_level < level);
8674                 if (level < wc->shared_level)
8675                         goto out;
8676
8677                 ret = find_next_key(path, level + 1, &wc->update_progress);
8678                 if (ret > 0)
8679                         wc->update_ref = 0;
8680
8681                 wc->stage = DROP_REFERENCE;
8682                 wc->shared_level = -1;
8683                 path->slots[level] = 0;
8684
8685                 /*
8686                  * check reference count again if the block isn't locked.
8687                  * we should start walking down the tree again if reference
8688                  * count is one.
8689                  */
8690                 if (!path->locks[level]) {
8691                         BUG_ON(level == 0);
8692                         btrfs_tree_lock(eb);
8693                         btrfs_set_lock_blocking(eb);
8694                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8695
8696                         ret = btrfs_lookup_extent_info(trans, root,
8697                                                        eb->start, level, 1,
8698                                                        &wc->refs[level],
8699                                                        &wc->flags[level]);
8700                         if (ret < 0) {
8701                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8702                                 path->locks[level] = 0;
8703                                 return ret;
8704                         }
8705                         BUG_ON(wc->refs[level] == 0);
8706                         if (wc->refs[level] == 1) {
8707                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8708                                 path->locks[level] = 0;
8709                                 return 1;
8710                         }
8711                 }
8712         }
8713
8714         /* wc->stage == DROP_REFERENCE */
8715         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8716
8717         if (wc->refs[level] == 1) {
8718                 if (level == 0) {
8719                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8720                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8721                         else
8722                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8723                         BUG_ON(ret); /* -ENOMEM */
8724                         ret = account_leaf_items(trans, root, eb);
8725                         if (ret) {
8726                                 btrfs_err_rl(root->fs_info,
8727                                         "error "
8728                                         "%d accounting leaf items. Quota "
8729                                         "is out of sync, rescan required.",
8730                                         ret);
8731                         }
8732                 }
8733                 /* make block locked assertion in clean_tree_block happy */
8734                 if (!path->locks[level] &&
8735                     btrfs_header_generation(eb) == trans->transid) {
8736                         btrfs_tree_lock(eb);
8737                         btrfs_set_lock_blocking(eb);
8738                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8739                 }
8740                 clean_tree_block(trans, root->fs_info, eb);
8741         }
8742
8743         if (eb == root->node) {
8744                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8745                         parent = eb->start;
8746                 else
8747                         BUG_ON(root->root_key.objectid !=
8748                                btrfs_header_owner(eb));
8749         } else {
8750                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8751                         parent = path->nodes[level + 1]->start;
8752                 else
8753                         BUG_ON(root->root_key.objectid !=
8754                                btrfs_header_owner(path->nodes[level + 1]));
8755         }
8756
8757         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8758 out:
8759         wc->refs[level] = 0;
8760         wc->flags[level] = 0;
8761         return 0;
8762 }
8763
8764 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8765                                    struct btrfs_root *root,
8766                                    struct btrfs_path *path,
8767                                    struct walk_control *wc)
8768 {
8769         int level = wc->level;
8770         int lookup_info = 1;
8771         int ret;
8772
8773         while (level >= 0) {
8774                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8775                 if (ret > 0)
8776                         break;
8777
8778                 if (level == 0)
8779                         break;
8780
8781                 if (path->slots[level] >=
8782                     btrfs_header_nritems(path->nodes[level]))
8783                         break;
8784
8785                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8786                 if (ret > 0) {
8787                         path->slots[level]++;
8788                         continue;
8789                 } else if (ret < 0)
8790                         return ret;
8791                 level = wc->level;
8792         }
8793         return 0;
8794 }
8795
8796 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8797                                  struct btrfs_root *root,
8798                                  struct btrfs_path *path,
8799                                  struct walk_control *wc, int max_level)
8800 {
8801         int level = wc->level;
8802         int ret;
8803
8804         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8805         while (level < max_level && path->nodes[level]) {
8806                 wc->level = level;
8807                 if (path->slots[level] + 1 <
8808                     btrfs_header_nritems(path->nodes[level])) {
8809                         path->slots[level]++;
8810                         return 0;
8811                 } else {
8812                         ret = walk_up_proc(trans, root, path, wc);
8813                         if (ret > 0)
8814                                 return 0;
8815
8816                         if (path->locks[level]) {
8817                                 btrfs_tree_unlock_rw(path->nodes[level],
8818                                                      path->locks[level]);
8819                                 path->locks[level] = 0;
8820                         }
8821                         free_extent_buffer(path->nodes[level]);
8822                         path->nodes[level] = NULL;
8823                         level++;
8824                 }
8825         }
8826         return 1;
8827 }
8828
8829 /*
8830  * drop a subvolume tree.
8831  *
8832  * this function traverses the tree freeing any blocks that only
8833  * referenced by the tree.
8834  *
8835  * when a shared tree block is found. this function decreases its
8836  * reference count by one. if update_ref is true, this function
8837  * also make sure backrefs for the shared block and all lower level
8838  * blocks are properly updated.
8839  *
8840  * If called with for_reloc == 0, may exit early with -EAGAIN
8841  */
8842 int btrfs_drop_snapshot(struct btrfs_root *root,
8843                          struct btrfs_block_rsv *block_rsv, int update_ref,
8844                          int for_reloc)
8845 {
8846         struct btrfs_path *path;
8847         struct btrfs_trans_handle *trans;
8848         struct btrfs_root *tree_root = root->fs_info->tree_root;
8849         struct btrfs_root_item *root_item = &root->root_item;
8850         struct walk_control *wc;
8851         struct btrfs_key key;
8852         int err = 0;
8853         int ret;
8854         int level;
8855         bool root_dropped = false;
8856
8857         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8858
8859         path = btrfs_alloc_path();
8860         if (!path) {
8861                 err = -ENOMEM;
8862                 goto out;
8863         }
8864
8865         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8866         if (!wc) {
8867                 btrfs_free_path(path);
8868                 err = -ENOMEM;
8869                 goto out;
8870         }
8871
8872         trans = btrfs_start_transaction(tree_root, 0);
8873         if (IS_ERR(trans)) {
8874                 err = PTR_ERR(trans);
8875                 goto out_free;
8876         }
8877
8878         if (block_rsv)
8879                 trans->block_rsv = block_rsv;
8880
8881         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8882                 level = btrfs_header_level(root->node);
8883                 path->nodes[level] = btrfs_lock_root_node(root);
8884                 btrfs_set_lock_blocking(path->nodes[level]);
8885                 path->slots[level] = 0;
8886                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8887                 memset(&wc->update_progress, 0,
8888                        sizeof(wc->update_progress));
8889         } else {
8890                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8891                 memcpy(&wc->update_progress, &key,
8892                        sizeof(wc->update_progress));
8893
8894                 level = root_item->drop_level;
8895                 BUG_ON(level == 0);
8896                 path->lowest_level = level;
8897                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8898                 path->lowest_level = 0;
8899                 if (ret < 0) {
8900                         err = ret;
8901                         goto out_end_trans;
8902                 }
8903                 WARN_ON(ret > 0);
8904
8905                 /*
8906                  * unlock our path, this is safe because only this
8907                  * function is allowed to delete this snapshot
8908                  */
8909                 btrfs_unlock_up_safe(path, 0);
8910
8911                 level = btrfs_header_level(root->node);
8912                 while (1) {
8913                         btrfs_tree_lock(path->nodes[level]);
8914                         btrfs_set_lock_blocking(path->nodes[level]);
8915                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8916
8917                         ret = btrfs_lookup_extent_info(trans, root,
8918                                                 path->nodes[level]->start,
8919                                                 level, 1, &wc->refs[level],
8920                                                 &wc->flags[level]);
8921                         if (ret < 0) {
8922                                 err = ret;
8923                                 goto out_end_trans;
8924                         }
8925                         BUG_ON(wc->refs[level] == 0);
8926
8927                         if (level == root_item->drop_level)
8928                                 break;
8929
8930                         btrfs_tree_unlock(path->nodes[level]);
8931                         path->locks[level] = 0;
8932                         WARN_ON(wc->refs[level] != 1);
8933                         level--;
8934                 }
8935         }
8936
8937         wc->level = level;
8938         wc->shared_level = -1;
8939         wc->stage = DROP_REFERENCE;
8940         wc->update_ref = update_ref;
8941         wc->keep_locks = 0;
8942         wc->for_reloc = for_reloc;
8943         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8944
8945         while (1) {
8946
8947                 ret = walk_down_tree(trans, root, path, wc);
8948                 if (ret < 0) {
8949                         err = ret;
8950                         break;
8951                 }
8952
8953                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8954                 if (ret < 0) {
8955                         err = ret;
8956                         break;
8957                 }
8958
8959                 if (ret > 0) {
8960                         BUG_ON(wc->stage != DROP_REFERENCE);
8961                         break;
8962                 }
8963
8964                 if (wc->stage == DROP_REFERENCE) {
8965                         level = wc->level;
8966                         btrfs_node_key(path->nodes[level],
8967                                        &root_item->drop_progress,
8968                                        path->slots[level]);
8969                         root_item->drop_level = level;
8970                 }
8971
8972                 BUG_ON(wc->level == 0);
8973                 if (btrfs_should_end_transaction(trans, tree_root) ||
8974                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8975                         ret = btrfs_update_root(trans, tree_root,
8976                                                 &root->root_key,
8977                                                 root_item);
8978                         if (ret) {
8979                                 btrfs_abort_transaction(trans, tree_root, ret);
8980                                 err = ret;
8981                                 goto out_end_trans;
8982                         }
8983
8984                         btrfs_end_transaction_throttle(trans, tree_root);
8985                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8986                                 pr_debug("BTRFS: drop snapshot early exit\n");
8987                                 err = -EAGAIN;
8988                                 goto out_free;
8989                         }
8990
8991                         trans = btrfs_start_transaction(tree_root, 0);
8992                         if (IS_ERR(trans)) {
8993                                 err = PTR_ERR(trans);
8994                                 goto out_free;
8995                         }
8996                         if (block_rsv)
8997                                 trans->block_rsv = block_rsv;
8998                 }
8999         }
9000         btrfs_release_path(path);
9001         if (err)
9002                 goto out_end_trans;
9003
9004         ret = btrfs_del_root(trans, tree_root, &root->root_key);
9005         if (ret) {
9006                 btrfs_abort_transaction(trans, tree_root, ret);
9007                 goto out_end_trans;
9008         }
9009
9010         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9011                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9012                                       NULL, NULL);
9013                 if (ret < 0) {
9014                         btrfs_abort_transaction(trans, tree_root, ret);
9015                         err = ret;
9016                         goto out_end_trans;
9017                 } else if (ret > 0) {
9018                         /* if we fail to delete the orphan item this time
9019                          * around, it'll get picked up the next time.
9020                          *
9021                          * The most common failure here is just -ENOENT.
9022                          */
9023                         btrfs_del_orphan_item(trans, tree_root,
9024                                               root->root_key.objectid);
9025                 }
9026         }
9027
9028         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9029                 btrfs_add_dropped_root(trans, root);
9030         } else {
9031                 free_extent_buffer(root->node);
9032                 free_extent_buffer(root->commit_root);
9033                 btrfs_put_fs_root(root);
9034         }
9035         root_dropped = true;
9036 out_end_trans:
9037         btrfs_end_transaction_throttle(trans, tree_root);
9038 out_free:
9039         kfree(wc);
9040         btrfs_free_path(path);
9041 out:
9042         /*
9043          * So if we need to stop dropping the snapshot for whatever reason we
9044          * need to make sure to add it back to the dead root list so that we
9045          * keep trying to do the work later.  This also cleans up roots if we
9046          * don't have it in the radix (like when we recover after a power fail
9047          * or unmount) so we don't leak memory.
9048          */
9049         if (!for_reloc && root_dropped == false)
9050                 btrfs_add_dead_root(root);
9051         if (err && err != -EAGAIN)
9052                 btrfs_std_error(root->fs_info, err, NULL);
9053         return err;
9054 }
9055
9056 /*
9057  * drop subtree rooted at tree block 'node'.
9058  *
9059  * NOTE: this function will unlock and release tree block 'node'
9060  * only used by relocation code
9061  */
9062 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9063                         struct btrfs_root *root,
9064                         struct extent_buffer *node,
9065                         struct extent_buffer *parent)
9066 {
9067         struct btrfs_path *path;
9068         struct walk_control *wc;
9069         int level;
9070         int parent_level;
9071         int ret = 0;
9072         int wret;
9073
9074         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9075
9076         path = btrfs_alloc_path();
9077         if (!path)
9078                 return -ENOMEM;
9079
9080         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9081         if (!wc) {
9082                 btrfs_free_path(path);
9083                 return -ENOMEM;
9084         }
9085
9086         btrfs_assert_tree_locked(parent);
9087         parent_level = btrfs_header_level(parent);
9088         extent_buffer_get(parent);
9089         path->nodes[parent_level] = parent;
9090         path->slots[parent_level] = btrfs_header_nritems(parent);
9091
9092         btrfs_assert_tree_locked(node);
9093         level = btrfs_header_level(node);
9094         path->nodes[level] = node;
9095         path->slots[level] = 0;
9096         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9097
9098         wc->refs[parent_level] = 1;
9099         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9100         wc->level = level;
9101         wc->shared_level = -1;
9102         wc->stage = DROP_REFERENCE;
9103         wc->update_ref = 0;
9104         wc->keep_locks = 1;
9105         wc->for_reloc = 1;
9106         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9107
9108         while (1) {
9109                 wret = walk_down_tree(trans, root, path, wc);
9110                 if (wret < 0) {
9111                         ret = wret;
9112                         break;
9113                 }
9114
9115                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9116                 if (wret < 0)
9117                         ret = wret;
9118                 if (wret != 0)
9119                         break;
9120         }
9121
9122         kfree(wc);
9123         btrfs_free_path(path);
9124         return ret;
9125 }
9126
9127 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
9128 {
9129         u64 num_devices;
9130         u64 stripped;
9131
9132         /*
9133          * if restripe for this chunk_type is on pick target profile and
9134          * return, otherwise do the usual balance
9135          */
9136         stripped = get_restripe_target(root->fs_info, flags);
9137         if (stripped)
9138                 return extended_to_chunk(stripped);
9139
9140         num_devices = root->fs_info->fs_devices->rw_devices;
9141
9142         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9143                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9144                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9145
9146         if (num_devices == 1) {
9147                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9148                 stripped = flags & ~stripped;
9149
9150                 /* turn raid0 into single device chunks */
9151                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9152                         return stripped;
9153
9154                 /* turn mirroring into duplication */
9155                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9156                              BTRFS_BLOCK_GROUP_RAID10))
9157                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9158         } else {
9159                 /* they already had raid on here, just return */
9160                 if (flags & stripped)
9161                         return flags;
9162
9163                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9164                 stripped = flags & ~stripped;
9165
9166                 /* switch duplicated blocks with raid1 */
9167                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9168                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9169
9170                 /* this is drive concat, leave it alone */
9171         }
9172
9173         return flags;
9174 }
9175
9176 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9177 {
9178         struct btrfs_space_info *sinfo = cache->space_info;
9179         u64 num_bytes;
9180         u64 min_allocable_bytes;
9181         int ret = -ENOSPC;
9182
9183         /*
9184          * We need some metadata space and system metadata space for
9185          * allocating chunks in some corner cases until we force to set
9186          * it to be readonly.
9187          */
9188         if ((sinfo->flags &
9189              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9190             !force)
9191                 min_allocable_bytes = SZ_1M;
9192         else
9193                 min_allocable_bytes = 0;
9194
9195         spin_lock(&sinfo->lock);
9196         spin_lock(&cache->lock);
9197
9198         if (cache->ro) {
9199                 cache->ro++;
9200                 ret = 0;
9201                 goto out;
9202         }
9203
9204         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9205                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9206
9207         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
9208             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
9209             min_allocable_bytes <= sinfo->total_bytes) {
9210                 sinfo->bytes_readonly += num_bytes;
9211                 cache->ro++;
9212                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9213                 ret = 0;
9214         }
9215 out:
9216         spin_unlock(&cache->lock);
9217         spin_unlock(&sinfo->lock);
9218         return ret;
9219 }
9220
9221 int btrfs_inc_block_group_ro(struct btrfs_root *root,
9222                              struct btrfs_block_group_cache *cache)
9223
9224 {
9225         struct btrfs_trans_handle *trans;
9226         u64 alloc_flags;
9227         int ret;
9228
9229 again:
9230         trans = btrfs_join_transaction(root);
9231         if (IS_ERR(trans))
9232                 return PTR_ERR(trans);
9233
9234         /*
9235          * we're not allowed to set block groups readonly after the dirty
9236          * block groups cache has started writing.  If it already started,
9237          * back off and let this transaction commit
9238          */
9239         mutex_lock(&root->fs_info->ro_block_group_mutex);
9240         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9241                 u64 transid = trans->transid;
9242
9243                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
9244                 btrfs_end_transaction(trans, root);
9245
9246                 ret = btrfs_wait_for_commit(root, transid);
9247                 if (ret)
9248                         return ret;
9249                 goto again;
9250         }
9251
9252         /*
9253          * if we are changing raid levels, try to allocate a corresponding
9254          * block group with the new raid level.
9255          */
9256         alloc_flags = update_block_group_flags(root, cache->flags);
9257         if (alloc_flags != cache->flags) {
9258                 ret = do_chunk_alloc(trans, root, alloc_flags,
9259                                      CHUNK_ALLOC_FORCE);
9260                 /*
9261                  * ENOSPC is allowed here, we may have enough space
9262                  * already allocated at the new raid level to
9263                  * carry on
9264                  */
9265                 if (ret == -ENOSPC)
9266                         ret = 0;
9267                 if (ret < 0)
9268                         goto out;
9269         }
9270
9271         ret = inc_block_group_ro(cache, 0);
9272         if (!ret)
9273                 goto out;
9274         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
9275         ret = do_chunk_alloc(trans, root, alloc_flags,
9276                              CHUNK_ALLOC_FORCE);
9277         if (ret < 0)
9278                 goto out;
9279         ret = inc_block_group_ro(cache, 0);
9280 out:
9281         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9282                 alloc_flags = update_block_group_flags(root, cache->flags);
9283                 lock_chunks(root->fs_info->chunk_root);
9284                 check_system_chunk(trans, root, alloc_flags);
9285                 unlock_chunks(root->fs_info->chunk_root);
9286         }
9287         mutex_unlock(&root->fs_info->ro_block_group_mutex);
9288
9289         btrfs_end_transaction(trans, root);
9290         return ret;
9291 }
9292
9293 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9294                             struct btrfs_root *root, u64 type)
9295 {
9296         u64 alloc_flags = get_alloc_profile(root, type);
9297         return do_chunk_alloc(trans, root, alloc_flags,
9298                               CHUNK_ALLOC_FORCE);
9299 }
9300
9301 /*
9302  * helper to account the unused space of all the readonly block group in the
9303  * space_info. takes mirrors into account.
9304  */
9305 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9306 {
9307         struct btrfs_block_group_cache *block_group;
9308         u64 free_bytes = 0;
9309         int factor;
9310
9311         /* It's df, we don't care if it's racey */
9312         if (list_empty(&sinfo->ro_bgs))
9313                 return 0;
9314
9315         spin_lock(&sinfo->lock);
9316         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9317                 spin_lock(&block_group->lock);
9318
9319                 if (!block_group->ro) {
9320                         spin_unlock(&block_group->lock);
9321                         continue;
9322                 }
9323
9324                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9325                                           BTRFS_BLOCK_GROUP_RAID10 |
9326                                           BTRFS_BLOCK_GROUP_DUP))
9327                         factor = 2;
9328                 else
9329                         factor = 1;
9330
9331                 free_bytes += (block_group->key.offset -
9332                                btrfs_block_group_used(&block_group->item)) *
9333                                factor;
9334
9335                 spin_unlock(&block_group->lock);
9336         }
9337         spin_unlock(&sinfo->lock);
9338
9339         return free_bytes;
9340 }
9341
9342 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9343                               struct btrfs_block_group_cache *cache)
9344 {
9345         struct btrfs_space_info *sinfo = cache->space_info;
9346         u64 num_bytes;
9347
9348         BUG_ON(!cache->ro);
9349
9350         spin_lock(&sinfo->lock);
9351         spin_lock(&cache->lock);
9352         if (!--cache->ro) {
9353                 num_bytes = cache->key.offset - cache->reserved -
9354                             cache->pinned - cache->bytes_super -
9355                             btrfs_block_group_used(&cache->item);
9356                 sinfo->bytes_readonly -= num_bytes;
9357                 list_del_init(&cache->ro_list);
9358         }
9359         spin_unlock(&cache->lock);
9360         spin_unlock(&sinfo->lock);
9361 }
9362
9363 /*
9364  * checks to see if its even possible to relocate this block group.
9365  *
9366  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9367  * ok to go ahead and try.
9368  */
9369 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9370 {
9371         struct btrfs_block_group_cache *block_group;
9372         struct btrfs_space_info *space_info;
9373         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9374         struct btrfs_device *device;
9375         struct btrfs_trans_handle *trans;
9376         u64 min_free;
9377         u64 dev_min = 1;
9378         u64 dev_nr = 0;
9379         u64 target;
9380         int index;
9381         int full = 0;
9382         int ret = 0;
9383
9384         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9385
9386         /* odd, couldn't find the block group, leave it alone */
9387         if (!block_group)
9388                 return -1;
9389
9390         min_free = btrfs_block_group_used(&block_group->item);
9391
9392         /* no bytes used, we're good */
9393         if (!min_free)
9394                 goto out;
9395
9396         space_info = block_group->space_info;
9397         spin_lock(&space_info->lock);
9398
9399         full = space_info->full;
9400
9401         /*
9402          * if this is the last block group we have in this space, we can't
9403          * relocate it unless we're able to allocate a new chunk below.
9404          *
9405          * Otherwise, we need to make sure we have room in the space to handle
9406          * all of the extents from this block group.  If we can, we're good
9407          */
9408         if ((space_info->total_bytes != block_group->key.offset) &&
9409             (space_info->bytes_used + space_info->bytes_reserved +
9410              space_info->bytes_pinned + space_info->bytes_readonly +
9411              min_free < space_info->total_bytes)) {
9412                 spin_unlock(&space_info->lock);
9413                 goto out;
9414         }
9415         spin_unlock(&space_info->lock);
9416
9417         /*
9418          * ok we don't have enough space, but maybe we have free space on our
9419          * devices to allocate new chunks for relocation, so loop through our
9420          * alloc devices and guess if we have enough space.  if this block
9421          * group is going to be restriped, run checks against the target
9422          * profile instead of the current one.
9423          */
9424         ret = -1;
9425
9426         /*
9427          * index:
9428          *      0: raid10
9429          *      1: raid1
9430          *      2: dup
9431          *      3: raid0
9432          *      4: single
9433          */
9434         target = get_restripe_target(root->fs_info, block_group->flags);
9435         if (target) {
9436                 index = __get_raid_index(extended_to_chunk(target));
9437         } else {
9438                 /*
9439                  * this is just a balance, so if we were marked as full
9440                  * we know there is no space for a new chunk
9441                  */
9442                 if (full)
9443                         goto out;
9444
9445                 index = get_block_group_index(block_group);
9446         }
9447
9448         if (index == BTRFS_RAID_RAID10) {
9449                 dev_min = 4;
9450                 /* Divide by 2 */
9451                 min_free >>= 1;
9452         } else if (index == BTRFS_RAID_RAID1) {
9453                 dev_min = 2;
9454         } else if (index == BTRFS_RAID_DUP) {
9455                 /* Multiply by 2 */
9456                 min_free <<= 1;
9457         } else if (index == BTRFS_RAID_RAID0) {
9458                 dev_min = fs_devices->rw_devices;
9459                 min_free = div64_u64(min_free, dev_min);
9460         }
9461
9462         /* We need to do this so that we can look at pending chunks */
9463         trans = btrfs_join_transaction(root);
9464         if (IS_ERR(trans)) {
9465                 ret = PTR_ERR(trans);
9466                 goto out;
9467         }
9468
9469         mutex_lock(&root->fs_info->chunk_mutex);
9470         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9471                 u64 dev_offset;
9472
9473                 /*
9474                  * check to make sure we can actually find a chunk with enough
9475                  * space to fit our block group in.
9476                  */
9477                 if (device->total_bytes > device->bytes_used + min_free &&
9478                     !device->is_tgtdev_for_dev_replace) {
9479                         ret = find_free_dev_extent(trans, device, min_free,
9480                                                    &dev_offset, NULL);
9481                         if (!ret)
9482                                 dev_nr++;
9483
9484                         if (dev_nr >= dev_min)
9485                                 break;
9486
9487                         ret = -1;
9488                 }
9489         }
9490         mutex_unlock(&root->fs_info->chunk_mutex);
9491         btrfs_end_transaction(trans, root);
9492 out:
9493         btrfs_put_block_group(block_group);
9494         return ret;
9495 }
9496
9497 static int find_first_block_group(struct btrfs_root *root,
9498                 struct btrfs_path *path, struct btrfs_key *key)
9499 {
9500         int ret = 0;
9501         struct btrfs_key found_key;
9502         struct extent_buffer *leaf;
9503         int slot;
9504
9505         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9506         if (ret < 0)
9507                 goto out;
9508
9509         while (1) {
9510                 slot = path->slots[0];
9511                 leaf = path->nodes[0];
9512                 if (slot >= btrfs_header_nritems(leaf)) {
9513                         ret = btrfs_next_leaf(root, path);
9514                         if (ret == 0)
9515                                 continue;
9516                         if (ret < 0)
9517                                 goto out;
9518                         break;
9519                 }
9520                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9521
9522                 if (found_key.objectid >= key->objectid &&
9523                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9524                         ret = 0;
9525                         goto out;
9526                 }
9527                 path->slots[0]++;
9528         }
9529 out:
9530         return ret;
9531 }
9532
9533 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9534 {
9535         struct btrfs_block_group_cache *block_group;
9536         u64 last = 0;
9537
9538         while (1) {
9539                 struct inode *inode;
9540
9541                 block_group = btrfs_lookup_first_block_group(info, last);
9542                 while (block_group) {
9543                         spin_lock(&block_group->lock);
9544                         if (block_group->iref)
9545                                 break;
9546                         spin_unlock(&block_group->lock);
9547                         block_group = next_block_group(info->tree_root,
9548                                                        block_group);
9549                 }
9550                 if (!block_group) {
9551                         if (last == 0)
9552                                 break;
9553                         last = 0;
9554                         continue;
9555                 }
9556
9557                 inode = block_group->inode;
9558                 block_group->iref = 0;
9559                 block_group->inode = NULL;
9560                 spin_unlock(&block_group->lock);
9561                 iput(inode);
9562                 last = block_group->key.objectid + block_group->key.offset;
9563                 btrfs_put_block_group(block_group);
9564         }
9565 }
9566
9567 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9568 {
9569         struct btrfs_block_group_cache *block_group;
9570         struct btrfs_space_info *space_info;
9571         struct btrfs_caching_control *caching_ctl;
9572         struct rb_node *n;
9573
9574         down_write(&info->commit_root_sem);
9575         while (!list_empty(&info->caching_block_groups)) {
9576                 caching_ctl = list_entry(info->caching_block_groups.next,
9577                                          struct btrfs_caching_control, list);
9578                 list_del(&caching_ctl->list);
9579                 put_caching_control(caching_ctl);
9580         }
9581         up_write(&info->commit_root_sem);
9582
9583         spin_lock(&info->unused_bgs_lock);
9584         while (!list_empty(&info->unused_bgs)) {
9585                 block_group = list_first_entry(&info->unused_bgs,
9586                                                struct btrfs_block_group_cache,
9587                                                bg_list);
9588                 list_del_init(&block_group->bg_list);
9589                 btrfs_put_block_group(block_group);
9590         }
9591         spin_unlock(&info->unused_bgs_lock);
9592
9593         spin_lock(&info->block_group_cache_lock);
9594         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9595                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9596                                        cache_node);
9597                 rb_erase(&block_group->cache_node,
9598                          &info->block_group_cache_tree);
9599                 RB_CLEAR_NODE(&block_group->cache_node);
9600                 spin_unlock(&info->block_group_cache_lock);
9601
9602                 down_write(&block_group->space_info->groups_sem);
9603                 list_del(&block_group->list);
9604                 up_write(&block_group->space_info->groups_sem);
9605
9606                 if (block_group->cached == BTRFS_CACHE_STARTED)
9607                         wait_block_group_cache_done(block_group);
9608
9609                 /*
9610                  * We haven't cached this block group, which means we could
9611                  * possibly have excluded extents on this block group.
9612                  */
9613                 if (block_group->cached == BTRFS_CACHE_NO ||
9614                     block_group->cached == BTRFS_CACHE_ERROR)
9615                         free_excluded_extents(info->extent_root, block_group);
9616
9617                 btrfs_remove_free_space_cache(block_group);
9618                 btrfs_put_block_group(block_group);
9619
9620                 spin_lock(&info->block_group_cache_lock);
9621         }
9622         spin_unlock(&info->block_group_cache_lock);
9623
9624         /* now that all the block groups are freed, go through and
9625          * free all the space_info structs.  This is only called during
9626          * the final stages of unmount, and so we know nobody is
9627          * using them.  We call synchronize_rcu() once before we start,
9628          * just to be on the safe side.
9629          */
9630         synchronize_rcu();
9631
9632         release_global_block_rsv(info);
9633
9634         while (!list_empty(&info->space_info)) {
9635                 int i;
9636
9637                 space_info = list_entry(info->space_info.next,
9638                                         struct btrfs_space_info,
9639                                         list);
9640                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9641                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9642                             space_info->bytes_reserved > 0 ||
9643                             space_info->bytes_may_use > 0)) {
9644                                 dump_space_info(space_info, 0, 0);
9645                         }
9646                 }
9647                 list_del(&space_info->list);
9648                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9649                         struct kobject *kobj;
9650                         kobj = space_info->block_group_kobjs[i];
9651                         space_info->block_group_kobjs[i] = NULL;
9652                         if (kobj) {
9653                                 kobject_del(kobj);
9654                                 kobject_put(kobj);
9655                         }
9656                 }
9657                 kobject_del(&space_info->kobj);
9658                 kobject_put(&space_info->kobj);
9659         }
9660         return 0;
9661 }
9662
9663 static void __link_block_group(struct btrfs_space_info *space_info,
9664                                struct btrfs_block_group_cache *cache)
9665 {
9666         int index = get_block_group_index(cache);
9667         bool first = false;
9668
9669         down_write(&space_info->groups_sem);
9670         if (list_empty(&space_info->block_groups[index]))
9671                 first = true;
9672         list_add_tail(&cache->list, &space_info->block_groups[index]);
9673         up_write(&space_info->groups_sem);
9674
9675         if (first) {
9676                 struct raid_kobject *rkobj;
9677                 int ret;
9678
9679                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9680                 if (!rkobj)
9681                         goto out_err;
9682                 rkobj->raid_type = index;
9683                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9684                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9685                                   "%s", get_raid_name(index));
9686                 if (ret) {
9687                         kobject_put(&rkobj->kobj);
9688                         goto out_err;
9689                 }
9690                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9691         }
9692
9693         return;
9694 out_err:
9695         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9696 }
9697
9698 static struct btrfs_block_group_cache *
9699 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9700 {
9701         struct btrfs_block_group_cache *cache;
9702
9703         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9704         if (!cache)
9705                 return NULL;
9706
9707         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9708                                         GFP_NOFS);
9709         if (!cache->free_space_ctl) {
9710                 kfree(cache);
9711                 return NULL;
9712         }
9713
9714         cache->key.objectid = start;
9715         cache->key.offset = size;
9716         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9717
9718         cache->sectorsize = root->sectorsize;
9719         cache->fs_info = root->fs_info;
9720         cache->full_stripe_len = btrfs_full_stripe_len(root,
9721                                                &root->fs_info->mapping_tree,
9722                                                start);
9723         set_free_space_tree_thresholds(cache);
9724
9725         atomic_set(&cache->count, 1);
9726         spin_lock_init(&cache->lock);
9727         init_rwsem(&cache->data_rwsem);
9728         INIT_LIST_HEAD(&cache->list);
9729         INIT_LIST_HEAD(&cache->cluster_list);
9730         INIT_LIST_HEAD(&cache->bg_list);
9731         INIT_LIST_HEAD(&cache->ro_list);
9732         INIT_LIST_HEAD(&cache->dirty_list);
9733         INIT_LIST_HEAD(&cache->io_list);
9734         btrfs_init_free_space_ctl(cache);
9735         atomic_set(&cache->trimming, 0);
9736         mutex_init(&cache->free_space_lock);
9737
9738         return cache;
9739 }
9740
9741 int btrfs_read_block_groups(struct btrfs_root *root)
9742 {
9743         struct btrfs_path *path;
9744         int ret;
9745         struct btrfs_block_group_cache *cache;
9746         struct btrfs_fs_info *info = root->fs_info;
9747         struct btrfs_space_info *space_info;
9748         struct btrfs_key key;
9749         struct btrfs_key found_key;
9750         struct extent_buffer *leaf;
9751         int need_clear = 0;
9752         u64 cache_gen;
9753
9754         root = info->extent_root;
9755         key.objectid = 0;
9756         key.offset = 0;
9757         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9758         path = btrfs_alloc_path();
9759         if (!path)
9760                 return -ENOMEM;
9761         path->reada = READA_FORWARD;
9762
9763         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9764         if (btrfs_test_opt(root, SPACE_CACHE) &&
9765             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9766                 need_clear = 1;
9767         if (btrfs_test_opt(root, CLEAR_CACHE))
9768                 need_clear = 1;
9769
9770         while (1) {
9771                 ret = find_first_block_group(root, path, &key);
9772                 if (ret > 0)
9773                         break;
9774                 if (ret != 0)
9775                         goto error;
9776
9777                 leaf = path->nodes[0];
9778                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9779
9780                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9781                                                        found_key.offset);
9782                 if (!cache) {
9783                         ret = -ENOMEM;
9784                         goto error;
9785                 }
9786
9787                 if (need_clear) {
9788                         /*
9789                          * When we mount with old space cache, we need to
9790                          * set BTRFS_DC_CLEAR and set dirty flag.
9791                          *
9792                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9793                          *    truncate the old free space cache inode and
9794                          *    setup a new one.
9795                          * b) Setting 'dirty flag' makes sure that we flush
9796                          *    the new space cache info onto disk.
9797                          */
9798                         if (btrfs_test_opt(root, SPACE_CACHE))
9799                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9800                 }
9801
9802                 read_extent_buffer(leaf, &cache->item,
9803                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9804                                    sizeof(cache->item));
9805                 cache->flags = btrfs_block_group_flags(&cache->item);
9806
9807                 key.objectid = found_key.objectid + found_key.offset;
9808                 btrfs_release_path(path);
9809
9810                 /*
9811                  * We need to exclude the super stripes now so that the space
9812                  * info has super bytes accounted for, otherwise we'll think
9813                  * we have more space than we actually do.
9814                  */
9815                 ret = exclude_super_stripes(root, cache);
9816                 if (ret) {
9817                         /*
9818                          * We may have excluded something, so call this just in
9819                          * case.
9820                          */
9821                         free_excluded_extents(root, cache);
9822                         btrfs_put_block_group(cache);
9823                         goto error;
9824                 }
9825
9826                 /*
9827                  * check for two cases, either we are full, and therefore
9828                  * don't need to bother with the caching work since we won't
9829                  * find any space, or we are empty, and we can just add all
9830                  * the space in and be done with it.  This saves us _alot_ of
9831                  * time, particularly in the full case.
9832                  */
9833                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9834                         cache->last_byte_to_unpin = (u64)-1;
9835                         cache->cached = BTRFS_CACHE_FINISHED;
9836                         free_excluded_extents(root, cache);
9837                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9838                         cache->last_byte_to_unpin = (u64)-1;
9839                         cache->cached = BTRFS_CACHE_FINISHED;
9840                         add_new_free_space(cache, root->fs_info,
9841                                            found_key.objectid,
9842                                            found_key.objectid +
9843                                            found_key.offset);
9844                         free_excluded_extents(root, cache);
9845                 }
9846
9847                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9848                 if (ret) {
9849                         btrfs_remove_free_space_cache(cache);
9850                         btrfs_put_block_group(cache);
9851                         goto error;
9852                 }
9853
9854                 ret = update_space_info(info, cache->flags, found_key.offset,
9855                                         btrfs_block_group_used(&cache->item),
9856                                         &space_info);
9857                 if (ret) {
9858                         btrfs_remove_free_space_cache(cache);
9859                         spin_lock(&info->block_group_cache_lock);
9860                         rb_erase(&cache->cache_node,
9861                                  &info->block_group_cache_tree);
9862                         RB_CLEAR_NODE(&cache->cache_node);
9863                         spin_unlock(&info->block_group_cache_lock);
9864                         btrfs_put_block_group(cache);
9865                         goto error;
9866                 }
9867
9868                 cache->space_info = space_info;
9869                 spin_lock(&cache->space_info->lock);
9870                 cache->space_info->bytes_readonly += cache->bytes_super;
9871                 spin_unlock(&cache->space_info->lock);
9872
9873                 __link_block_group(space_info, cache);
9874
9875                 set_avail_alloc_bits(root->fs_info, cache->flags);
9876                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9877                         inc_block_group_ro(cache, 1);
9878                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9879                         spin_lock(&info->unused_bgs_lock);
9880                         /* Should always be true but just in case. */
9881                         if (list_empty(&cache->bg_list)) {
9882                                 btrfs_get_block_group(cache);
9883                                 list_add_tail(&cache->bg_list,
9884                                               &info->unused_bgs);
9885                         }
9886                         spin_unlock(&info->unused_bgs_lock);
9887                 }
9888         }
9889
9890         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9891                 if (!(get_alloc_profile(root, space_info->flags) &
9892                       (BTRFS_BLOCK_GROUP_RAID10 |
9893                        BTRFS_BLOCK_GROUP_RAID1 |
9894                        BTRFS_BLOCK_GROUP_RAID5 |
9895                        BTRFS_BLOCK_GROUP_RAID6 |
9896                        BTRFS_BLOCK_GROUP_DUP)))
9897                         continue;
9898                 /*
9899                  * avoid allocating from un-mirrored block group if there are
9900                  * mirrored block groups.
9901                  */
9902                 list_for_each_entry(cache,
9903                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9904                                 list)
9905                         inc_block_group_ro(cache, 1);
9906                 list_for_each_entry(cache,
9907                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9908                                 list)
9909                         inc_block_group_ro(cache, 1);
9910         }
9911
9912         init_global_block_rsv(info);
9913         ret = 0;
9914 error:
9915         btrfs_free_path(path);
9916         return ret;
9917 }
9918
9919 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9920                                        struct btrfs_root *root)
9921 {
9922         struct btrfs_block_group_cache *block_group, *tmp;
9923         struct btrfs_root *extent_root = root->fs_info->extent_root;
9924         struct btrfs_block_group_item item;
9925         struct btrfs_key key;
9926         int ret = 0;
9927         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
9928
9929         trans->can_flush_pending_bgs = false;
9930         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9931                 if (ret)
9932                         goto next;
9933
9934                 spin_lock(&block_group->lock);
9935                 memcpy(&item, &block_group->item, sizeof(item));
9936                 memcpy(&key, &block_group->key, sizeof(key));
9937                 spin_unlock(&block_group->lock);
9938
9939                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9940                                         sizeof(item));
9941                 if (ret)
9942                         btrfs_abort_transaction(trans, extent_root, ret);
9943                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9944                                                key.objectid, key.offset);
9945                 if (ret)
9946                         btrfs_abort_transaction(trans, extent_root, ret);
9947                 add_block_group_free_space(trans, root->fs_info, block_group);
9948                 /* already aborted the transaction if it failed. */
9949 next:
9950                 list_del_init(&block_group->bg_list);
9951         }
9952         trans->can_flush_pending_bgs = can_flush_pending_bgs;
9953 }
9954
9955 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9956                            struct btrfs_root *root, u64 bytes_used,
9957                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9958                            u64 size)
9959 {
9960         int ret;
9961         struct btrfs_root *extent_root;
9962         struct btrfs_block_group_cache *cache;
9963
9964         extent_root = root->fs_info->extent_root;
9965
9966         btrfs_set_log_full_commit(root->fs_info, trans);
9967
9968         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9969         if (!cache)
9970                 return -ENOMEM;
9971
9972         btrfs_set_block_group_used(&cache->item, bytes_used);
9973         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9974         btrfs_set_block_group_flags(&cache->item, type);
9975
9976         cache->flags = type;
9977         cache->last_byte_to_unpin = (u64)-1;
9978         cache->cached = BTRFS_CACHE_FINISHED;
9979         cache->needs_free_space = 1;
9980         ret = exclude_super_stripes(root, cache);
9981         if (ret) {
9982                 /*
9983                  * We may have excluded something, so call this just in
9984                  * case.
9985                  */
9986                 free_excluded_extents(root, cache);
9987                 btrfs_put_block_group(cache);
9988                 return ret;
9989         }
9990
9991         add_new_free_space(cache, root->fs_info, chunk_offset,
9992                            chunk_offset + size);
9993
9994         free_excluded_extents(root, cache);
9995
9996 #ifdef CONFIG_BTRFS_DEBUG
9997         if (btrfs_should_fragment_free_space(root, cache)) {
9998                 u64 new_bytes_used = size - bytes_used;
9999
10000                 bytes_used += new_bytes_used >> 1;
10001                 fragment_free_space(root, cache);
10002         }
10003 #endif
10004         /*
10005          * Call to ensure the corresponding space_info object is created and
10006          * assigned to our block group, but don't update its counters just yet.
10007          * We want our bg to be added to the rbtree with its ->space_info set.
10008          */
10009         ret = update_space_info(root->fs_info, cache->flags, 0, 0,
10010                                 &cache->space_info);
10011         if (ret) {
10012                 btrfs_remove_free_space_cache(cache);
10013                 btrfs_put_block_group(cache);
10014                 return ret;
10015         }
10016
10017         ret = btrfs_add_block_group_cache(root->fs_info, cache);
10018         if (ret) {
10019                 btrfs_remove_free_space_cache(cache);
10020                 btrfs_put_block_group(cache);
10021                 return ret;
10022         }
10023
10024         /*
10025          * Now that our block group has its ->space_info set and is inserted in
10026          * the rbtree, update the space info's counters.
10027          */
10028         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
10029                                 &cache->space_info);
10030         if (ret) {
10031                 btrfs_remove_free_space_cache(cache);
10032                 spin_lock(&root->fs_info->block_group_cache_lock);
10033                 rb_erase(&cache->cache_node,
10034                          &root->fs_info->block_group_cache_tree);
10035                 RB_CLEAR_NODE(&cache->cache_node);
10036                 spin_unlock(&root->fs_info->block_group_cache_lock);
10037                 btrfs_put_block_group(cache);
10038                 return ret;
10039         }
10040         update_global_block_rsv(root->fs_info);
10041
10042         spin_lock(&cache->space_info->lock);
10043         cache->space_info->bytes_readonly += cache->bytes_super;
10044         spin_unlock(&cache->space_info->lock);
10045
10046         __link_block_group(cache->space_info, cache);
10047
10048         list_add_tail(&cache->bg_list, &trans->new_bgs);
10049
10050         set_avail_alloc_bits(extent_root->fs_info, type);
10051
10052         return 0;
10053 }
10054
10055 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10056 {
10057         u64 extra_flags = chunk_to_extended(flags) &
10058                                 BTRFS_EXTENDED_PROFILE_MASK;
10059
10060         write_seqlock(&fs_info->profiles_lock);
10061         if (flags & BTRFS_BLOCK_GROUP_DATA)
10062                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10063         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10064                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10065         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10066                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10067         write_sequnlock(&fs_info->profiles_lock);
10068 }
10069
10070 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10071                              struct btrfs_root *root, u64 group_start,
10072                              struct extent_map *em)
10073 {
10074         struct btrfs_path *path;
10075         struct btrfs_block_group_cache *block_group;
10076         struct btrfs_free_cluster *cluster;
10077         struct btrfs_root *tree_root = root->fs_info->tree_root;
10078         struct btrfs_key key;
10079         struct inode *inode;
10080         struct kobject *kobj = NULL;
10081         int ret;
10082         int index;
10083         int factor;
10084         struct btrfs_caching_control *caching_ctl = NULL;
10085         bool remove_em;
10086
10087         root = root->fs_info->extent_root;
10088
10089         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
10090         BUG_ON(!block_group);
10091         BUG_ON(!block_group->ro);
10092
10093         /*
10094          * Free the reserved super bytes from this block group before
10095          * remove it.
10096          */
10097         free_excluded_extents(root, block_group);
10098
10099         memcpy(&key, &block_group->key, sizeof(key));
10100         index = get_block_group_index(block_group);
10101         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10102                                   BTRFS_BLOCK_GROUP_RAID1 |
10103                                   BTRFS_BLOCK_GROUP_RAID10))
10104                 factor = 2;
10105         else
10106                 factor = 1;
10107
10108         /* make sure this block group isn't part of an allocation cluster */
10109         cluster = &root->fs_info->data_alloc_cluster;
10110         spin_lock(&cluster->refill_lock);
10111         btrfs_return_cluster_to_free_space(block_group, cluster);
10112         spin_unlock(&cluster->refill_lock);
10113
10114         /*
10115          * make sure this block group isn't part of a metadata
10116          * allocation cluster
10117          */
10118         cluster = &root->fs_info->meta_alloc_cluster;
10119         spin_lock(&cluster->refill_lock);
10120         btrfs_return_cluster_to_free_space(block_group, cluster);
10121         spin_unlock(&cluster->refill_lock);
10122
10123         path = btrfs_alloc_path();
10124         if (!path) {
10125                 ret = -ENOMEM;
10126                 goto out;
10127         }
10128
10129         /*
10130          * get the inode first so any iput calls done for the io_list
10131          * aren't the final iput (no unlinks allowed now)
10132          */
10133         inode = lookup_free_space_inode(tree_root, block_group, path);
10134
10135         mutex_lock(&trans->transaction->cache_write_mutex);
10136         /*
10137          * make sure our free spache cache IO is done before remove the
10138          * free space inode
10139          */
10140         spin_lock(&trans->transaction->dirty_bgs_lock);
10141         if (!list_empty(&block_group->io_list)) {
10142                 list_del_init(&block_group->io_list);
10143
10144                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10145
10146                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10147                 btrfs_wait_cache_io(root, trans, block_group,
10148                                     &block_group->io_ctl, path,
10149                                     block_group->key.objectid);
10150                 btrfs_put_block_group(block_group);
10151                 spin_lock(&trans->transaction->dirty_bgs_lock);
10152         }
10153
10154         if (!list_empty(&block_group->dirty_list)) {
10155                 list_del_init(&block_group->dirty_list);
10156                 btrfs_put_block_group(block_group);
10157         }
10158         spin_unlock(&trans->transaction->dirty_bgs_lock);
10159         mutex_unlock(&trans->transaction->cache_write_mutex);
10160
10161         if (!IS_ERR(inode)) {
10162                 ret = btrfs_orphan_add(trans, inode);
10163                 if (ret) {
10164                         btrfs_add_delayed_iput(inode);
10165                         goto out;
10166                 }
10167                 clear_nlink(inode);
10168                 /* One for the block groups ref */
10169                 spin_lock(&block_group->lock);
10170                 if (block_group->iref) {
10171                         block_group->iref = 0;
10172                         block_group->inode = NULL;
10173                         spin_unlock(&block_group->lock);
10174                         iput(inode);
10175                 } else {
10176                         spin_unlock(&block_group->lock);
10177                 }
10178                 /* One for our lookup ref */
10179                 btrfs_add_delayed_iput(inode);
10180         }
10181
10182         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10183         key.offset = block_group->key.objectid;
10184         key.type = 0;
10185
10186         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10187         if (ret < 0)
10188                 goto out;
10189         if (ret > 0)
10190                 btrfs_release_path(path);
10191         if (ret == 0) {
10192                 ret = btrfs_del_item(trans, tree_root, path);
10193                 if (ret)
10194                         goto out;
10195                 btrfs_release_path(path);
10196         }
10197
10198         spin_lock(&root->fs_info->block_group_cache_lock);
10199         rb_erase(&block_group->cache_node,
10200                  &root->fs_info->block_group_cache_tree);
10201         RB_CLEAR_NODE(&block_group->cache_node);
10202
10203         if (root->fs_info->first_logical_byte == block_group->key.objectid)
10204                 root->fs_info->first_logical_byte = (u64)-1;
10205         spin_unlock(&root->fs_info->block_group_cache_lock);
10206
10207         down_write(&block_group->space_info->groups_sem);
10208         /*
10209          * we must use list_del_init so people can check to see if they
10210          * are still on the list after taking the semaphore
10211          */
10212         list_del_init(&block_group->list);
10213         if (list_empty(&block_group->space_info->block_groups[index])) {
10214                 kobj = block_group->space_info->block_group_kobjs[index];
10215                 block_group->space_info->block_group_kobjs[index] = NULL;
10216                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
10217         }
10218         up_write(&block_group->space_info->groups_sem);
10219         if (kobj) {
10220                 kobject_del(kobj);
10221                 kobject_put(kobj);
10222         }
10223
10224         if (block_group->has_caching_ctl)
10225                 caching_ctl = get_caching_control(block_group);
10226         if (block_group->cached == BTRFS_CACHE_STARTED)
10227                 wait_block_group_cache_done(block_group);
10228         if (block_group->has_caching_ctl) {
10229                 down_write(&root->fs_info->commit_root_sem);
10230                 if (!caching_ctl) {
10231                         struct btrfs_caching_control *ctl;
10232
10233                         list_for_each_entry(ctl,
10234                                     &root->fs_info->caching_block_groups, list)
10235                                 if (ctl->block_group == block_group) {
10236                                         caching_ctl = ctl;
10237                                         atomic_inc(&caching_ctl->count);
10238                                         break;
10239                                 }
10240                 }
10241                 if (caching_ctl)
10242                         list_del_init(&caching_ctl->list);
10243                 up_write(&root->fs_info->commit_root_sem);
10244                 if (caching_ctl) {
10245                         /* Once for the caching bgs list and once for us. */
10246                         put_caching_control(caching_ctl);
10247                         put_caching_control(caching_ctl);
10248                 }
10249         }
10250
10251         spin_lock(&trans->transaction->dirty_bgs_lock);
10252         if (!list_empty(&block_group->dirty_list)) {
10253                 WARN_ON(1);
10254         }
10255         if (!list_empty(&block_group->io_list)) {
10256                 WARN_ON(1);
10257         }
10258         spin_unlock(&trans->transaction->dirty_bgs_lock);
10259         btrfs_remove_free_space_cache(block_group);
10260
10261         spin_lock(&block_group->space_info->lock);
10262         list_del_init(&block_group->ro_list);
10263
10264         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
10265                 WARN_ON(block_group->space_info->total_bytes
10266                         < block_group->key.offset);
10267                 WARN_ON(block_group->space_info->bytes_readonly
10268                         < block_group->key.offset);
10269                 WARN_ON(block_group->space_info->disk_total
10270                         < block_group->key.offset * factor);
10271         }
10272         block_group->space_info->total_bytes -= block_group->key.offset;
10273         block_group->space_info->bytes_readonly -= block_group->key.offset;
10274         block_group->space_info->disk_total -= block_group->key.offset * factor;
10275
10276         spin_unlock(&block_group->space_info->lock);
10277
10278         memcpy(&key, &block_group->key, sizeof(key));
10279
10280         lock_chunks(root);
10281         if (!list_empty(&em->list)) {
10282                 /* We're in the transaction->pending_chunks list. */
10283                 free_extent_map(em);
10284         }
10285         spin_lock(&block_group->lock);
10286         block_group->removed = 1;
10287         /*
10288          * At this point trimming can't start on this block group, because we
10289          * removed the block group from the tree fs_info->block_group_cache_tree
10290          * so no one can't find it anymore and even if someone already got this
10291          * block group before we removed it from the rbtree, they have already
10292          * incremented block_group->trimming - if they didn't, they won't find
10293          * any free space entries because we already removed them all when we
10294          * called btrfs_remove_free_space_cache().
10295          *
10296          * And we must not remove the extent map from the fs_info->mapping_tree
10297          * to prevent the same logical address range and physical device space
10298          * ranges from being reused for a new block group. This is because our
10299          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10300          * completely transactionless, so while it is trimming a range the
10301          * currently running transaction might finish and a new one start,
10302          * allowing for new block groups to be created that can reuse the same
10303          * physical device locations unless we take this special care.
10304          *
10305          * There may also be an implicit trim operation if the file system
10306          * is mounted with -odiscard. The same protections must remain
10307          * in place until the extents have been discarded completely when
10308          * the transaction commit has completed.
10309          */
10310         remove_em = (atomic_read(&block_group->trimming) == 0);
10311         /*
10312          * Make sure a trimmer task always sees the em in the pinned_chunks list
10313          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10314          * before checking block_group->removed).
10315          */
10316         if (!remove_em) {
10317                 /*
10318                  * Our em might be in trans->transaction->pending_chunks which
10319                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10320                  * and so is the fs_info->pinned_chunks list.
10321                  *
10322                  * So at this point we must be holding the chunk_mutex to avoid
10323                  * any races with chunk allocation (more specifically at
10324                  * volumes.c:contains_pending_extent()), to ensure it always
10325                  * sees the em, either in the pending_chunks list or in the
10326                  * pinned_chunks list.
10327                  */
10328                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10329         }
10330         spin_unlock(&block_group->lock);
10331
10332         if (remove_em) {
10333                 struct extent_map_tree *em_tree;
10334
10335                 em_tree = &root->fs_info->mapping_tree.map_tree;
10336                 write_lock(&em_tree->lock);
10337                 /*
10338                  * The em might be in the pending_chunks list, so make sure the
10339                  * chunk mutex is locked, since remove_extent_mapping() will
10340                  * delete us from that list.
10341                  */
10342                 remove_extent_mapping(em_tree, em);
10343                 write_unlock(&em_tree->lock);
10344                 /* once for the tree */
10345                 free_extent_map(em);
10346         }
10347
10348         unlock_chunks(root);
10349
10350         ret = remove_block_group_free_space(trans, root->fs_info, block_group);
10351         if (ret)
10352                 goto out;
10353
10354         btrfs_put_block_group(block_group);
10355         btrfs_put_block_group(block_group);
10356
10357         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10358         if (ret > 0)
10359                 ret = -EIO;
10360         if (ret < 0)
10361                 goto out;
10362
10363         ret = btrfs_del_item(trans, root, path);
10364 out:
10365         btrfs_free_path(path);
10366         return ret;
10367 }
10368
10369 struct btrfs_trans_handle *
10370 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10371                                      const u64 chunk_offset)
10372 {
10373         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10374         struct extent_map *em;
10375         struct map_lookup *map;
10376         unsigned int num_items;
10377
10378         read_lock(&em_tree->lock);
10379         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10380         read_unlock(&em_tree->lock);
10381         ASSERT(em && em->start == chunk_offset);
10382
10383         /*
10384          * We need to reserve 3 + N units from the metadata space info in order
10385          * to remove a block group (done at btrfs_remove_chunk() and at
10386          * btrfs_remove_block_group()), which are used for:
10387          *
10388          * 1 unit for adding the free space inode's orphan (located in the tree
10389          * of tree roots).
10390          * 1 unit for deleting the block group item (located in the extent
10391          * tree).
10392          * 1 unit for deleting the free space item (located in tree of tree
10393          * roots).
10394          * N units for deleting N device extent items corresponding to each
10395          * stripe (located in the device tree).
10396          *
10397          * In order to remove a block group we also need to reserve units in the
10398          * system space info in order to update the chunk tree (update one or
10399          * more device items and remove one chunk item), but this is done at
10400          * btrfs_remove_chunk() through a call to check_system_chunk().
10401          */
10402         map = (struct map_lookup *)em->bdev;
10403         num_items = 3 + map->num_stripes;
10404         free_extent_map(em);
10405
10406         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10407                                                            num_items, 1);
10408 }
10409
10410 /*
10411  * Process the unused_bgs list and remove any that don't have any allocated
10412  * space inside of them.
10413  */
10414 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10415 {
10416         struct btrfs_block_group_cache *block_group;
10417         struct btrfs_space_info *space_info;
10418         struct btrfs_root *root = fs_info->extent_root;
10419         struct btrfs_trans_handle *trans;
10420         int ret = 0;
10421
10422         if (!fs_info->open)
10423                 return;
10424
10425         spin_lock(&fs_info->unused_bgs_lock);
10426         while (!list_empty(&fs_info->unused_bgs)) {
10427                 u64 start, end;
10428                 int trimming;
10429
10430                 block_group = list_first_entry(&fs_info->unused_bgs,
10431                                                struct btrfs_block_group_cache,
10432                                                bg_list);
10433                 list_del_init(&block_group->bg_list);
10434
10435                 space_info = block_group->space_info;
10436
10437                 if (ret || btrfs_mixed_space_info(space_info)) {
10438                         btrfs_put_block_group(block_group);
10439                         continue;
10440                 }
10441                 spin_unlock(&fs_info->unused_bgs_lock);
10442
10443                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10444
10445                 /* Don't want to race with allocators so take the groups_sem */
10446                 down_write(&space_info->groups_sem);
10447                 spin_lock(&block_group->lock);
10448                 if (block_group->reserved ||
10449                     btrfs_block_group_used(&block_group->item) ||
10450                     block_group->ro ||
10451                     list_is_singular(&block_group->list)) {
10452                         /*
10453                          * We want to bail if we made new allocations or have
10454                          * outstanding allocations in this block group.  We do
10455                          * the ro check in case balance is currently acting on
10456                          * this block group.
10457                          */
10458                         spin_unlock(&block_group->lock);
10459                         up_write(&space_info->groups_sem);
10460                         goto next;
10461                 }
10462                 spin_unlock(&block_group->lock);
10463
10464                 /* We don't want to force the issue, only flip if it's ok. */
10465                 ret = inc_block_group_ro(block_group, 0);
10466                 up_write(&space_info->groups_sem);
10467                 if (ret < 0) {
10468                         ret = 0;
10469                         goto next;
10470                 }
10471
10472                 /*
10473                  * Want to do this before we do anything else so we can recover
10474                  * properly if we fail to join the transaction.
10475                  */
10476                 trans = btrfs_start_trans_remove_block_group(fs_info,
10477                                                      block_group->key.objectid);
10478                 if (IS_ERR(trans)) {
10479                         btrfs_dec_block_group_ro(root, block_group);
10480                         ret = PTR_ERR(trans);
10481                         goto next;
10482                 }
10483
10484                 /*
10485                  * We could have pending pinned extents for this block group,
10486                  * just delete them, we don't care about them anymore.
10487                  */
10488                 start = block_group->key.objectid;
10489                 end = start + block_group->key.offset - 1;
10490                 /*
10491                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10492                  * btrfs_finish_extent_commit(). If we are at transaction N,
10493                  * another task might be running finish_extent_commit() for the
10494                  * previous transaction N - 1, and have seen a range belonging
10495                  * to the block group in freed_extents[] before we were able to
10496                  * clear the whole block group range from freed_extents[]. This
10497                  * means that task can lookup for the block group after we
10498                  * unpinned it from freed_extents[] and removed it, leading to
10499                  * a BUG_ON() at btrfs_unpin_extent_range().
10500                  */
10501                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10502                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10503                                   EXTENT_DIRTY, GFP_NOFS);
10504                 if (ret) {
10505                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10506                         btrfs_dec_block_group_ro(root, block_group);
10507                         goto end_trans;
10508                 }
10509                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10510                                   EXTENT_DIRTY, GFP_NOFS);
10511                 if (ret) {
10512                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10513                         btrfs_dec_block_group_ro(root, block_group);
10514                         goto end_trans;
10515                 }
10516                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10517
10518                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10519                 spin_lock(&space_info->lock);
10520                 spin_lock(&block_group->lock);
10521
10522                 space_info->bytes_pinned -= block_group->pinned;
10523                 space_info->bytes_readonly += block_group->pinned;
10524                 percpu_counter_add(&space_info->total_bytes_pinned,
10525                                    -block_group->pinned);
10526                 block_group->pinned = 0;
10527
10528                 spin_unlock(&block_group->lock);
10529                 spin_unlock(&space_info->lock);
10530
10531                 /* DISCARD can flip during remount */
10532                 trimming = btrfs_test_opt(root, DISCARD);
10533
10534                 /* Implicit trim during transaction commit. */
10535                 if (trimming)
10536                         btrfs_get_block_group_trimming(block_group);
10537
10538                 /*
10539                  * Btrfs_remove_chunk will abort the transaction if things go
10540                  * horribly wrong.
10541                  */
10542                 ret = btrfs_remove_chunk(trans, root,
10543                                          block_group->key.objectid);
10544
10545                 if (ret) {
10546                         if (trimming)
10547                                 btrfs_put_block_group_trimming(block_group);
10548                         goto end_trans;
10549                 }
10550
10551                 /*
10552                  * If we're not mounted with -odiscard, we can just forget
10553                  * about this block group. Otherwise we'll need to wait
10554                  * until transaction commit to do the actual discard.
10555                  */
10556                 if (trimming) {
10557                         spin_lock(&fs_info->unused_bgs_lock);
10558                         /*
10559                          * A concurrent scrub might have added us to the list
10560                          * fs_info->unused_bgs, so use a list_move operation
10561                          * to add the block group to the deleted_bgs list.
10562                          */
10563                         list_move(&block_group->bg_list,
10564                                   &trans->transaction->deleted_bgs);
10565                         spin_unlock(&fs_info->unused_bgs_lock);
10566                         btrfs_get_block_group(block_group);
10567                 }
10568 end_trans:
10569                 btrfs_end_transaction(trans, root);
10570 next:
10571                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10572                 btrfs_put_block_group(block_group);
10573                 spin_lock(&fs_info->unused_bgs_lock);
10574         }
10575         spin_unlock(&fs_info->unused_bgs_lock);
10576 }
10577
10578 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10579 {
10580         struct btrfs_space_info *space_info;
10581         struct btrfs_super_block *disk_super;
10582         u64 features;
10583         u64 flags;
10584         int mixed = 0;
10585         int ret;
10586
10587         disk_super = fs_info->super_copy;
10588         if (!btrfs_super_root(disk_super))
10589                 return 1;
10590
10591         features = btrfs_super_incompat_flags(disk_super);
10592         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10593                 mixed = 1;
10594
10595         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10596         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10597         if (ret)
10598                 goto out;
10599
10600         if (mixed) {
10601                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10602                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10603         } else {
10604                 flags = BTRFS_BLOCK_GROUP_METADATA;
10605                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10606                 if (ret)
10607                         goto out;
10608
10609                 flags = BTRFS_BLOCK_GROUP_DATA;
10610                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10611         }
10612 out:
10613         return ret;
10614 }
10615
10616 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10617 {
10618         return unpin_extent_range(root, start, end, false);
10619 }
10620
10621 /*
10622  * It used to be that old block groups would be left around forever.
10623  * Iterating over them would be enough to trim unused space.  Since we
10624  * now automatically remove them, we also need to iterate over unallocated
10625  * space.
10626  *
10627  * We don't want a transaction for this since the discard may take a
10628  * substantial amount of time.  We don't require that a transaction be
10629  * running, but we do need to take a running transaction into account
10630  * to ensure that we're not discarding chunks that were released in
10631  * the current transaction.
10632  *
10633  * Holding the chunks lock will prevent other threads from allocating
10634  * or releasing chunks, but it won't prevent a running transaction
10635  * from committing and releasing the memory that the pending chunks
10636  * list head uses.  For that, we need to take a reference to the
10637  * transaction.
10638  */
10639 static int btrfs_trim_free_extents(struct btrfs_device *device,
10640                                    u64 minlen, u64 *trimmed)
10641 {
10642         u64 start = 0, len = 0;
10643         int ret;
10644
10645         *trimmed = 0;
10646
10647         /* Not writeable = nothing to do. */
10648         if (!device->writeable)
10649                 return 0;
10650
10651         /* No free space = nothing to do. */
10652         if (device->total_bytes <= device->bytes_used)
10653                 return 0;
10654
10655         ret = 0;
10656
10657         while (1) {
10658                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
10659                 struct btrfs_transaction *trans;
10660                 u64 bytes;
10661
10662                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10663                 if (ret)
10664                         return ret;
10665
10666                 down_read(&fs_info->commit_root_sem);
10667
10668                 spin_lock(&fs_info->trans_lock);
10669                 trans = fs_info->running_transaction;
10670                 if (trans)
10671                         atomic_inc(&trans->use_count);
10672                 spin_unlock(&fs_info->trans_lock);
10673
10674                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10675                                                  &start, &len);
10676                 if (trans)
10677                         btrfs_put_transaction(trans);
10678
10679                 if (ret) {
10680                         up_read(&fs_info->commit_root_sem);
10681                         mutex_unlock(&fs_info->chunk_mutex);
10682                         if (ret == -ENOSPC)
10683                                 ret = 0;
10684                         break;
10685                 }
10686
10687                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10688                 up_read(&fs_info->commit_root_sem);
10689                 mutex_unlock(&fs_info->chunk_mutex);
10690
10691                 if (ret)
10692                         break;
10693
10694                 start += len;
10695                 *trimmed += bytes;
10696
10697                 if (fatal_signal_pending(current)) {
10698                         ret = -ERESTARTSYS;
10699                         break;
10700                 }
10701
10702                 cond_resched();
10703         }
10704
10705         return ret;
10706 }
10707
10708 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10709 {
10710         struct btrfs_fs_info *fs_info = root->fs_info;
10711         struct btrfs_block_group_cache *cache = NULL;
10712         struct btrfs_device *device;
10713         struct list_head *devices;
10714         u64 group_trimmed;
10715         u64 start;
10716         u64 end;
10717         u64 trimmed = 0;
10718         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10719         int ret = 0;
10720
10721         /*
10722          * try to trim all FS space, our block group may start from non-zero.
10723          */
10724         if (range->len == total_bytes)
10725                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10726         else
10727                 cache = btrfs_lookup_block_group(fs_info, range->start);
10728
10729         while (cache) {
10730                 if (cache->key.objectid >= (range->start + range->len)) {
10731                         btrfs_put_block_group(cache);
10732                         break;
10733                 }
10734
10735                 start = max(range->start, cache->key.objectid);
10736                 end = min(range->start + range->len,
10737                                 cache->key.objectid + cache->key.offset);
10738
10739                 if (end - start >= range->minlen) {
10740                         if (!block_group_cache_done(cache)) {
10741                                 ret = cache_block_group(cache, 0);
10742                                 if (ret) {
10743                                         btrfs_put_block_group(cache);
10744                                         break;
10745                                 }
10746                                 ret = wait_block_group_cache_done(cache);
10747                                 if (ret) {
10748                                         btrfs_put_block_group(cache);
10749                                         break;
10750                                 }
10751                         }
10752                         ret = btrfs_trim_block_group(cache,
10753                                                      &group_trimmed,
10754                                                      start,
10755                                                      end,
10756                                                      range->minlen);
10757
10758                         trimmed += group_trimmed;
10759                         if (ret) {
10760                                 btrfs_put_block_group(cache);
10761                                 break;
10762                         }
10763                 }
10764
10765                 cache = next_block_group(fs_info->tree_root, cache);
10766         }
10767
10768         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
10769         devices = &root->fs_info->fs_devices->alloc_list;
10770         list_for_each_entry(device, devices, dev_alloc_list) {
10771                 ret = btrfs_trim_free_extents(device, range->minlen,
10772                                               &group_trimmed);
10773                 if (ret)
10774                         break;
10775
10776                 trimmed += group_trimmed;
10777         }
10778         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
10779
10780         range->len = trimmed;
10781         return ret;
10782 }
10783
10784 /*
10785  * btrfs_{start,end}_write_no_snapshoting() are similar to
10786  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10787  * data into the page cache through nocow before the subvolume is snapshoted,
10788  * but flush the data into disk after the snapshot creation, or to prevent
10789  * operations while snapshoting is ongoing and that cause the snapshot to be
10790  * inconsistent (writes followed by expanding truncates for example).
10791  */
10792 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10793 {
10794         percpu_counter_dec(&root->subv_writers->counter);
10795         /*
10796          * Make sure counter is updated before we wake up waiters.
10797          */
10798         smp_mb();
10799         if (waitqueue_active(&root->subv_writers->wait))
10800                 wake_up(&root->subv_writers->wait);
10801 }
10802
10803 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10804 {
10805         if (atomic_read(&root->will_be_snapshoted))
10806                 return 0;
10807
10808         percpu_counter_inc(&root->subv_writers->counter);
10809         /*
10810          * Make sure counter is updated before we check for snapshot creation.
10811          */
10812         smp_mb();
10813         if (atomic_read(&root->will_be_snapshoted)) {
10814                 btrfs_end_write_no_snapshoting(root);
10815                 return 0;
10816         }
10817         return 1;
10818 }