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);
 235         set_extent_bits(&root->fs_info->freed_extents[1],
 236                         start, end, EXTENT_UPTODATE);
 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);
 250         clear_extent_bits(&root->fs_info->freed_extents[1],
 251                           start, end, EXTENT_UPTODATE);
 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 COWed 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         /*
2046          * Avoid races with device replace and make sure our bbio has devices
2047          * associated to its stripes that don't go away while we are discarding.
2048          */
2049         btrfs_bio_counter_inc_blocked(root->fs_info);
2050         /* Tell the block device(s) that the sectors can be discarded */
2051         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
2052                               bytenr, &num_bytes, &bbio, 0);
2053         /* Error condition is -ENOMEM */
2054         if (!ret) {
2055                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2056                 int i;
2057
2058
2059                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2060                         u64 bytes;
2061                         if (!stripe->dev->can_discard)
2062                                 continue;
2063
2064                         ret = btrfs_issue_discard(stripe->dev->bdev,
2065                                                   stripe->physical,
2066                                                   stripe->length,
2067                                                   &bytes);
2068                         if (!ret)
2069                                 discarded_bytes += bytes;
2070                         else if (ret != -EOPNOTSUPP)
2071                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2072
2073                         /*
2074                          * Just in case we get back EOPNOTSUPP for some reason,
2075                          * just ignore the return value so we don't screw up
2076                          * people calling discard_extent.
2077                          */
2078                         ret = 0;
2079                 }
2080                 btrfs_put_bbio(bbio);
2081         }
2082         btrfs_bio_counter_dec(root->fs_info);
2083
2084         if (actual_bytes)
2085                 *actual_bytes = discarded_bytes;
2086
2087
2088         if (ret == -EOPNOTSUPP)
2089                 ret = 0;
2090         return ret;
2091 }
2092
2093 /* Can return -ENOMEM */
2094 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2095                          struct btrfs_root *root,
2096                          u64 bytenr, u64 num_bytes, u64 parent,
2097                          u64 root_objectid, u64 owner, u64 offset)
2098 {
2099         int ret;
2100         struct btrfs_fs_info *fs_info = root->fs_info;
2101
2102         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2103                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2104
2105         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2106                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2107                                         num_bytes,
2108                                         parent, root_objectid, (int)owner,
2109                                         BTRFS_ADD_DELAYED_REF, NULL);
2110         } else {
2111                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2112                                         num_bytes, parent, root_objectid,
2113                                         owner, offset, 0,
2114                                         BTRFS_ADD_DELAYED_REF, NULL);
2115         }
2116         return ret;
2117 }
2118
2119 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2120                                   struct btrfs_root *root,
2121                                   struct btrfs_delayed_ref_node *node,
2122                                   u64 parent, u64 root_objectid,
2123                                   u64 owner, u64 offset, int refs_to_add,
2124                                   struct btrfs_delayed_extent_op *extent_op)
2125 {
2126         struct btrfs_fs_info *fs_info = root->fs_info;
2127         struct btrfs_path *path;
2128         struct extent_buffer *leaf;
2129         struct btrfs_extent_item *item;
2130         struct btrfs_key key;
2131         u64 bytenr = node->bytenr;
2132         u64 num_bytes = node->num_bytes;
2133         u64 refs;
2134         int ret;
2135
2136         path = btrfs_alloc_path();
2137         if (!path)
2138                 return -ENOMEM;
2139
2140         path->reada = READA_FORWARD;
2141         path->leave_spinning = 1;
2142         /* this will setup the path even if it fails to insert the back ref */
2143         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2144                                            bytenr, num_bytes, parent,
2145                                            root_objectid, owner, offset,
2146                                            refs_to_add, extent_op);
2147         if ((ret < 0 && ret != -EAGAIN) || !ret)
2148                 goto out;
2149
2150         /*
2151          * Ok we had -EAGAIN which means we didn't have space to insert and
2152          * inline extent ref, so just update the reference count and add a
2153          * normal backref.
2154          */
2155         leaf = path->nodes[0];
2156         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2157         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2158         refs = btrfs_extent_refs(leaf, item);
2159         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2160         if (extent_op)
2161                 __run_delayed_extent_op(extent_op, leaf, item);
2162
2163         btrfs_mark_buffer_dirty(leaf);
2164         btrfs_release_path(path);
2165
2166         path->reada = READA_FORWARD;
2167         path->leave_spinning = 1;
2168         /* now insert the actual backref */
2169         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2170                                     path, bytenr, parent, root_objectid,
2171                                     owner, offset, refs_to_add);
2172         if (ret)
2173                 btrfs_abort_transaction(trans, root, ret);
2174 out:
2175         btrfs_free_path(path);
2176         return ret;
2177 }
2178
2179 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2180                                 struct btrfs_root *root,
2181                                 struct btrfs_delayed_ref_node *node,
2182                                 struct btrfs_delayed_extent_op *extent_op,
2183                                 int insert_reserved)
2184 {
2185         int ret = 0;
2186         struct btrfs_delayed_data_ref *ref;
2187         struct btrfs_key ins;
2188         u64 parent = 0;
2189         u64 ref_root = 0;
2190         u64 flags = 0;
2191
2192         ins.objectid = node->bytenr;
2193         ins.offset = node->num_bytes;
2194         ins.type = BTRFS_EXTENT_ITEM_KEY;
2195
2196         ref = btrfs_delayed_node_to_data_ref(node);
2197         trace_run_delayed_data_ref(node, ref, node->action);
2198
2199         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2200                 parent = ref->parent;
2201         ref_root = ref->root;
2202
2203         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2204                 if (extent_op)
2205                         flags |= extent_op->flags_to_set;
2206                 ret = alloc_reserved_file_extent(trans, root,
2207                                                  parent, ref_root, flags,
2208                                                  ref->objectid, ref->offset,
2209                                                  &ins, node->ref_mod);
2210         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2211                 ret = __btrfs_inc_extent_ref(trans, root, node, parent,
2212                                              ref_root, ref->objectid,
2213                                              ref->offset, node->ref_mod,
2214                                              extent_op);
2215         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2216                 ret = __btrfs_free_extent(trans, root, node, parent,
2217                                           ref_root, ref->objectid,
2218                                           ref->offset, node->ref_mod,
2219                                           extent_op);
2220         } else {
2221                 BUG();
2222         }
2223         return ret;
2224 }
2225
2226 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2227                                     struct extent_buffer *leaf,
2228                                     struct btrfs_extent_item *ei)
2229 {
2230         u64 flags = btrfs_extent_flags(leaf, ei);
2231         if (extent_op->update_flags) {
2232                 flags |= extent_op->flags_to_set;
2233                 btrfs_set_extent_flags(leaf, ei, flags);
2234         }
2235
2236         if (extent_op->update_key) {
2237                 struct btrfs_tree_block_info *bi;
2238                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2239                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2240                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2241         }
2242 }
2243
2244 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2245                                  struct btrfs_root *root,
2246                                  struct btrfs_delayed_ref_node *node,
2247                                  struct btrfs_delayed_extent_op *extent_op)
2248 {
2249         struct btrfs_key key;
2250         struct btrfs_path *path;
2251         struct btrfs_extent_item *ei;
2252         struct extent_buffer *leaf;
2253         u32 item_size;
2254         int ret;
2255         int err = 0;
2256         int metadata = !extent_op->is_data;
2257
2258         if (trans->aborted)
2259                 return 0;
2260
2261         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2262                 metadata = 0;
2263
2264         path = btrfs_alloc_path();
2265         if (!path)
2266                 return -ENOMEM;
2267
2268         key.objectid = node->bytenr;
2269
2270         if (metadata) {
2271                 key.type = BTRFS_METADATA_ITEM_KEY;
2272                 key.offset = extent_op->level;
2273         } else {
2274                 key.type = BTRFS_EXTENT_ITEM_KEY;
2275                 key.offset = node->num_bytes;
2276         }
2277
2278 again:
2279         path->reada = READA_FORWARD;
2280         path->leave_spinning = 1;
2281         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2282                                 path, 0, 1);
2283         if (ret < 0) {
2284                 err = ret;
2285                 goto out;
2286         }
2287         if (ret > 0) {
2288                 if (metadata) {
2289                         if (path->slots[0] > 0) {
2290                                 path->slots[0]--;
2291                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2292                                                       path->slots[0]);
2293                                 if (key.objectid == node->bytenr &&
2294                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2295                                     key.offset == node->num_bytes)
2296                                         ret = 0;
2297                         }
2298                         if (ret > 0) {
2299                                 btrfs_release_path(path);
2300                                 metadata = 0;
2301
2302                                 key.objectid = node->bytenr;
2303                                 key.offset = node->num_bytes;
2304                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2305                                 goto again;
2306                         }
2307                 } else {
2308                         err = -EIO;
2309                         goto out;
2310                 }
2311         }
2312
2313         leaf = path->nodes[0];
2314         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2315 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2316         if (item_size < sizeof(*ei)) {
2317                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2318                                              path, (u64)-1, 0);
2319                 if (ret < 0) {
2320                         err = ret;
2321                         goto out;
2322                 }
2323                 leaf = path->nodes[0];
2324                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2325         }
2326 #endif
2327         BUG_ON(item_size < sizeof(*ei));
2328         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2329         __run_delayed_extent_op(extent_op, leaf, ei);
2330
2331         btrfs_mark_buffer_dirty(leaf);
2332 out:
2333         btrfs_free_path(path);
2334         return err;
2335 }
2336
2337 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2338                                 struct btrfs_root *root,
2339                                 struct btrfs_delayed_ref_node *node,
2340                                 struct btrfs_delayed_extent_op *extent_op,
2341                                 int insert_reserved)
2342 {
2343         int ret = 0;
2344         struct btrfs_delayed_tree_ref *ref;
2345         struct btrfs_key ins;
2346         u64 parent = 0;
2347         u64 ref_root = 0;
2348         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2349                                                  SKINNY_METADATA);
2350
2351         ref = btrfs_delayed_node_to_tree_ref(node);
2352         trace_run_delayed_tree_ref(node, ref, node->action);
2353
2354         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2355                 parent = ref->parent;
2356         ref_root = ref->root;
2357
2358         ins.objectid = node->bytenr;
2359         if (skinny_metadata) {
2360                 ins.offset = ref->level;
2361                 ins.type = BTRFS_METADATA_ITEM_KEY;
2362         } else {
2363                 ins.offset = node->num_bytes;
2364                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2365         }
2366
2367         BUG_ON(node->ref_mod != 1);
2368         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2369                 BUG_ON(!extent_op || !extent_op->update_flags);
2370                 ret = alloc_reserved_tree_block(trans, root,
2371                                                 parent, ref_root,
2372                                                 extent_op->flags_to_set,
2373                                                 &extent_op->key,
2374                                                 ref->level, &ins);
2375         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2376                 ret = __btrfs_inc_extent_ref(trans, root, node,
2377                                              parent, ref_root,
2378                                              ref->level, 0, 1,
2379                                              extent_op);
2380         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2381                 ret = __btrfs_free_extent(trans, root, node,
2382                                           parent, ref_root,
2383                                           ref->level, 0, 1, extent_op);
2384         } else {
2385                 BUG();
2386         }
2387         return ret;
2388 }
2389
2390 /* helper function to actually process a single delayed ref entry */
2391 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2392                                struct btrfs_root *root,
2393                                struct btrfs_delayed_ref_node *node,
2394                                struct btrfs_delayed_extent_op *extent_op,
2395                                int insert_reserved)
2396 {
2397         int ret = 0;
2398
2399         if (trans->aborted) {
2400                 if (insert_reserved)
2401                         btrfs_pin_extent(root, node->bytenr,
2402                                          node->num_bytes, 1);
2403                 return 0;
2404         }
2405
2406         if (btrfs_delayed_ref_is_head(node)) {
2407                 struct btrfs_delayed_ref_head *head;
2408                 /*
2409                  * we've hit the end of the chain and we were supposed
2410                  * to insert this extent into the tree.  But, it got
2411                  * deleted before we ever needed to insert it, so all
2412                  * we have to do is clean up the accounting
2413                  */
2414                 BUG_ON(extent_op);
2415                 head = btrfs_delayed_node_to_head(node);
2416                 trace_run_delayed_ref_head(node, head, node->action);
2417
2418                 if (insert_reserved) {
2419                         btrfs_pin_extent(root, node->bytenr,
2420                                          node->num_bytes, 1);
2421                         if (head->is_data) {
2422                                 ret = btrfs_del_csums(trans, root,
2423                                                       node->bytenr,
2424                                                       node->num_bytes);
2425                         }
2426                 }
2427
2428                 /* Also free its reserved qgroup space */
2429                 btrfs_qgroup_free_delayed_ref(root->fs_info,
2430                                               head->qgroup_ref_root,
2431                                               head->qgroup_reserved);
2432                 return ret;
2433         }
2434
2435         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2436             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2437                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2438                                            insert_reserved);
2439         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2440                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2441                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2442                                            insert_reserved);
2443         else
2444                 BUG();
2445         return ret;
2446 }
2447
2448 static inline struct btrfs_delayed_ref_node *
2449 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2450 {
2451         struct btrfs_delayed_ref_node *ref;
2452
2453         if (list_empty(&head->ref_list))
2454                 return NULL;
2455
2456         /*
2457          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2458          * This is to prevent a ref count from going down to zero, which deletes
2459          * the extent item from the extent tree, when there still are references
2460          * to add, which would fail because they would not find the extent item.
2461          */
2462         list_for_each_entry(ref, &head->ref_list, list) {
2463                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2464                         return ref;
2465         }
2466
2467         return list_entry(head->ref_list.next, struct btrfs_delayed_ref_node,
2468                           list);
2469 }
2470
2471 /*
2472  * Returns 0 on success or if called with an already aborted transaction.
2473  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2474  */
2475 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2476                                              struct btrfs_root *root,
2477                                              unsigned long nr)
2478 {
2479         struct btrfs_delayed_ref_root *delayed_refs;
2480         struct btrfs_delayed_ref_node *ref;
2481         struct btrfs_delayed_ref_head *locked_ref = NULL;
2482         struct btrfs_delayed_extent_op *extent_op;
2483         struct btrfs_fs_info *fs_info = root->fs_info;
2484         ktime_t start = ktime_get();
2485         int ret;
2486         unsigned long count = 0;
2487         unsigned long actual_count = 0;
2488         int must_insert_reserved = 0;
2489
2490         delayed_refs = &trans->transaction->delayed_refs;
2491         while (1) {
2492                 if (!locked_ref) {
2493                         if (count >= nr)
2494                                 break;
2495
2496                         spin_lock(&delayed_refs->lock);
2497                         locked_ref = btrfs_select_ref_head(trans);
2498                         if (!locked_ref) {
2499                                 spin_unlock(&delayed_refs->lock);
2500                                 break;
2501                         }
2502
2503                         /* grab the lock that says we are going to process
2504                          * all the refs for this head */
2505                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2506                         spin_unlock(&delayed_refs->lock);
2507                         /*
2508                          * we may have dropped the spin lock to get the head
2509                          * mutex lock, and that might have given someone else
2510                          * time to free the head.  If that's true, it has been
2511                          * removed from our list and we can move on.
2512                          */
2513                         if (ret == -EAGAIN) {
2514                                 locked_ref = NULL;
2515                                 count++;
2516                                 continue;
2517                         }
2518                 }
2519
2520                 /*
2521                  * We need to try and merge add/drops of the same ref since we
2522                  * can run into issues with relocate dropping the implicit ref
2523                  * and then it being added back again before the drop can
2524                  * finish.  If we merged anything we need to re-loop so we can
2525                  * get a good ref.
2526                  * Or we can get node references of the same type that weren't
2527                  * merged when created due to bumps in the tree mod seq, and
2528                  * we need to merge them to prevent adding an inline extent
2529                  * backref before dropping it (triggering a BUG_ON at
2530                  * insert_inline_extent_backref()).
2531                  */
2532                 spin_lock(&locked_ref->lock);
2533                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2534                                          locked_ref);
2535
2536                 /*
2537                  * locked_ref is the head node, so we have to go one
2538                  * node back for any delayed ref updates
2539                  */
2540                 ref = select_delayed_ref(locked_ref);
2541
2542                 if (ref && ref->seq &&
2543                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2544                         spin_unlock(&locked_ref->lock);
2545                         btrfs_delayed_ref_unlock(locked_ref);
2546                         spin_lock(&delayed_refs->lock);
2547                         locked_ref->processing = 0;
2548                         delayed_refs->num_heads_ready++;
2549                         spin_unlock(&delayed_refs->lock);
2550                         locked_ref = NULL;
2551                         cond_resched();
2552                         count++;
2553                         continue;
2554                 }
2555
2556                 /*
2557                  * record the must insert reserved flag before we
2558                  * drop the spin lock.
2559                  */
2560                 must_insert_reserved = locked_ref->must_insert_reserved;
2561                 locked_ref->must_insert_reserved = 0;
2562
2563                 extent_op = locked_ref->extent_op;
2564                 locked_ref->extent_op = NULL;
2565
2566                 if (!ref) {
2567
2568
2569                         /* All delayed refs have been processed, Go ahead
2570                          * and send the head node to run_one_delayed_ref,
2571                          * so that any accounting fixes can happen
2572                          */
2573                         ref = &locked_ref->node;
2574
2575                         if (extent_op && must_insert_reserved) {
2576                                 btrfs_free_delayed_extent_op(extent_op);
2577                                 extent_op = NULL;
2578                         }
2579
2580                         if (extent_op) {
2581                                 spin_unlock(&locked_ref->lock);
2582                                 ret = run_delayed_extent_op(trans, root,
2583                                                             ref, extent_op);
2584                                 btrfs_free_delayed_extent_op(extent_op);
2585
2586                                 if (ret) {
2587                                         /*
2588                                          * Need to reset must_insert_reserved if
2589                                          * there was an error so the abort stuff
2590                                          * can cleanup the reserved space
2591                                          * properly.
2592                                          */
2593                                         if (must_insert_reserved)
2594                                                 locked_ref->must_insert_reserved = 1;
2595                                         locked_ref->processing = 0;
2596                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2597                                         btrfs_delayed_ref_unlock(locked_ref);
2598                                         return ret;
2599                                 }
2600                                 continue;
2601                         }
2602
2603                         /*
2604                          * Need to drop our head ref lock and re-acquire the
2605                          * delayed ref lock and then re-check to make sure
2606                          * nobody got added.
2607                          */
2608                         spin_unlock(&locked_ref->lock);
2609                         spin_lock(&delayed_refs->lock);
2610                         spin_lock(&locked_ref->lock);
2611                         if (!list_empty(&locked_ref->ref_list) ||
2612                             locked_ref->extent_op) {
2613                                 spin_unlock(&locked_ref->lock);
2614                                 spin_unlock(&delayed_refs->lock);
2615                                 continue;
2616                         }
2617                         ref->in_tree = 0;
2618                         delayed_refs->num_heads--;
2619                         rb_erase(&locked_ref->href_node,
2620                                  &delayed_refs->href_root);
2621                         spin_unlock(&delayed_refs->lock);
2622                 } else {
2623                         actual_count++;
2624                         ref->in_tree = 0;
2625                         list_del(&ref->list);
2626                 }
2627                 atomic_dec(&delayed_refs->num_entries);
2628
2629                 if (!btrfs_delayed_ref_is_head(ref)) {
2630                         /*
2631                          * when we play the delayed ref, also correct the
2632                          * ref_mod on head
2633                          */
2634                         switch (ref->action) {
2635                         case BTRFS_ADD_DELAYED_REF:
2636                         case BTRFS_ADD_DELAYED_EXTENT:
2637                                 locked_ref->node.ref_mod -= ref->ref_mod;
2638                                 break;
2639                         case BTRFS_DROP_DELAYED_REF:
2640                                 locked_ref->node.ref_mod += ref->ref_mod;
2641                                 break;
2642                         default:
2643                                 WARN_ON(1);
2644                         }
2645                 }
2646                 spin_unlock(&locked_ref->lock);
2647
2648                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2649                                           must_insert_reserved);
2650
2651                 btrfs_free_delayed_extent_op(extent_op);
2652                 if (ret) {
2653                         locked_ref->processing = 0;
2654                         btrfs_delayed_ref_unlock(locked_ref);
2655                         btrfs_put_delayed_ref(ref);
2656                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2657                         return ret;
2658                 }
2659
2660                 /*
2661                  * If this node is a head, that means all the refs in this head
2662                  * have been dealt with, and we will pick the next head to deal
2663                  * with, so we must unlock the head and drop it from the cluster
2664                  * list before we release it.
2665                  */
2666                 if (btrfs_delayed_ref_is_head(ref)) {
2667                         if (locked_ref->is_data &&
2668                             locked_ref->total_ref_mod < 0) {
2669                                 spin_lock(&delayed_refs->lock);
2670                                 delayed_refs->pending_csums -= ref->num_bytes;
2671                                 spin_unlock(&delayed_refs->lock);
2672                         }
2673                         btrfs_delayed_ref_unlock(locked_ref);
2674                         locked_ref = NULL;
2675                 }
2676                 btrfs_put_delayed_ref(ref);
2677                 count++;
2678                 cond_resched();
2679         }
2680
2681         /*
2682          * We don't want to include ref heads since we can have empty ref heads
2683          * and those will drastically skew our runtime down since we just do
2684          * accounting, no actual extent tree updates.
2685          */
2686         if (actual_count > 0) {
2687                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2688                 u64 avg;
2689
2690                 /*
2691                  * We weigh the current average higher than our current runtime
2692                  * to avoid large swings in the average.
2693                  */
2694                 spin_lock(&delayed_refs->lock);
2695                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2696                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2697                 spin_unlock(&delayed_refs->lock);
2698         }
2699         return 0;
2700 }
2701
2702 #ifdef SCRAMBLE_DELAYED_REFS
2703 /*
2704  * Normally delayed refs get processed in ascending bytenr order. This
2705  * correlates in most cases to the order added. To expose dependencies on this
2706  * order, we start to process the tree in the middle instead of the beginning
2707  */
2708 static u64 find_middle(struct rb_root *root)
2709 {
2710         struct rb_node *n = root->rb_node;
2711         struct btrfs_delayed_ref_node *entry;
2712         int alt = 1;
2713         u64 middle;
2714         u64 first = 0, last = 0;
2715
2716         n = rb_first(root);
2717         if (n) {
2718                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2719                 first = entry->bytenr;
2720         }
2721         n = rb_last(root);
2722         if (n) {
2723                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2724                 last = entry->bytenr;
2725         }
2726         n = root->rb_node;
2727
2728         while (n) {
2729                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2730                 WARN_ON(!entry->in_tree);
2731
2732                 middle = entry->bytenr;
2733
2734                 if (alt)
2735                         n = n->rb_left;
2736                 else
2737                         n = n->rb_right;
2738
2739                 alt = 1 - alt;
2740         }
2741         return middle;
2742 }
2743 #endif
2744
2745 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2746 {
2747         u64 num_bytes;
2748
2749         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2750                              sizeof(struct btrfs_extent_inline_ref));
2751         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2752                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2753
2754         /*
2755          * We don't ever fill up leaves all the way so multiply by 2 just to be
2756          * closer to what we're really going to want to use.
2757          */
2758         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2759 }
2760
2761 /*
2762  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2763  * would require to store the csums for that many bytes.
2764  */
2765 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2766 {
2767         u64 csum_size;
2768         u64 num_csums_per_leaf;
2769         u64 num_csums;
2770
2771         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
2772         num_csums_per_leaf = div64_u64(csum_size,
2773                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2774         num_csums = div64_u64(csum_bytes, root->sectorsize);
2775         num_csums += num_csums_per_leaf - 1;
2776         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2777         return num_csums;
2778 }
2779
2780 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2781                                        struct btrfs_root *root)
2782 {
2783         struct btrfs_block_rsv *global_rsv;
2784         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2785         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2786         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2787         u64 num_bytes, num_dirty_bgs_bytes;
2788         int ret = 0;
2789
2790         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2791         num_heads = heads_to_leaves(root, num_heads);
2792         if (num_heads > 1)
2793                 num_bytes += (num_heads - 1) * root->nodesize;
2794         num_bytes <<= 1;
2795         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2796         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2797                                                              num_dirty_bgs);
2798         global_rsv = &root->fs_info->global_block_rsv;
2799
2800         /*
2801          * If we can't allocate any more chunks lets make sure we have _lots_ of
2802          * wiggle room since running delayed refs can create more delayed refs.
2803          */
2804         if (global_rsv->space_info->full) {
2805                 num_dirty_bgs_bytes <<= 1;
2806                 num_bytes <<= 1;
2807         }
2808
2809         spin_lock(&global_rsv->lock);
2810         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2811                 ret = 1;
2812         spin_unlock(&global_rsv->lock);
2813         return ret;
2814 }
2815
2816 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2817                                        struct btrfs_root *root)
2818 {
2819         struct btrfs_fs_info *fs_info = root->fs_info;
2820         u64 num_entries =
2821                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2822         u64 avg_runtime;
2823         u64 val;
2824
2825         smp_mb();
2826         avg_runtime = fs_info->avg_delayed_ref_runtime;
2827         val = num_entries * avg_runtime;
2828         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2829                 return 1;
2830         if (val >= NSEC_PER_SEC / 2)
2831                 return 2;
2832
2833         return btrfs_check_space_for_delayed_refs(trans, root);
2834 }
2835
2836 struct async_delayed_refs {
2837         struct btrfs_root *root;
2838         int count;
2839         int error;
2840         int sync;
2841         struct completion wait;
2842         struct btrfs_work work;
2843 };
2844
2845 static void delayed_ref_async_start(struct btrfs_work *work)
2846 {
2847         struct async_delayed_refs *async;
2848         struct btrfs_trans_handle *trans;
2849         int ret;
2850
2851         async = container_of(work, struct async_delayed_refs, work);
2852
2853         trans = btrfs_join_transaction(async->root);
2854         if (IS_ERR(trans)) {
2855                 async->error = PTR_ERR(trans);
2856                 goto done;
2857         }
2858
2859         /*
2860          * trans->sync means that when we call end_transaction, we won't
2861          * wait on delayed refs
2862          */
2863         trans->sync = true;
2864         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2865         if (ret)
2866                 async->error = ret;
2867
2868         ret = btrfs_end_transaction(trans, async->root);
2869         if (ret && !async->error)
2870                 async->error = ret;
2871 done:
2872         if (async->sync)
2873                 complete(&async->wait);
2874         else
2875                 kfree(async);
2876 }
2877
2878 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2879                                  unsigned long count, int wait)
2880 {
2881         struct async_delayed_refs *async;
2882         int ret;
2883
2884         async = kmalloc(sizeof(*async), GFP_NOFS);
2885         if (!async)
2886                 return -ENOMEM;
2887
2888         async->root = root->fs_info->tree_root;
2889         async->count = count;
2890         async->error = 0;
2891         if (wait)
2892                 async->sync = 1;
2893         else
2894                 async->sync = 0;
2895         init_completion(&async->wait);
2896
2897         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2898                         delayed_ref_async_start, NULL, NULL);
2899
2900         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2901
2902         if (wait) {
2903                 wait_for_completion(&async->wait);
2904                 ret = async->error;
2905                 kfree(async);
2906                 return ret;
2907         }
2908         return 0;
2909 }
2910
2911 /*
2912  * this starts processing the delayed reference count updates and
2913  * extent insertions we have queued up so far.  count can be
2914  * 0, which means to process everything in the tree at the start
2915  * of the run (but not newly added entries), or it can be some target
2916  * number you'd like to process.
2917  *
2918  * Returns 0 on success or if called with an aborted transaction
2919  * Returns <0 on error and aborts the transaction
2920  */
2921 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2922                            struct btrfs_root *root, unsigned long count)
2923 {
2924         struct rb_node *node;
2925         struct btrfs_delayed_ref_root *delayed_refs;
2926         struct btrfs_delayed_ref_head *head;
2927         int ret;
2928         int run_all = count == (unsigned long)-1;
2929         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2930
2931         /* We'll clean this up in btrfs_cleanup_transaction */
2932         if (trans->aborted)
2933                 return 0;
2934
2935         if (root->fs_info->creating_free_space_tree)
2936                 return 0;
2937
2938         if (root == root->fs_info->extent_root)
2939                 root = root->fs_info->tree_root;
2940
2941         delayed_refs = &trans->transaction->delayed_refs;
2942         if (count == 0)
2943                 count = atomic_read(&delayed_refs->num_entries) * 2;
2944
2945 again:
2946 #ifdef SCRAMBLE_DELAYED_REFS
2947         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2948 #endif
2949         trans->can_flush_pending_bgs = false;
2950         ret = __btrfs_run_delayed_refs(trans, root, count);
2951         if (ret < 0) {
2952                 btrfs_abort_transaction(trans, root, ret);
2953                 return ret;
2954         }
2955
2956         if (run_all) {
2957                 if (!list_empty(&trans->new_bgs))
2958                         btrfs_create_pending_block_groups(trans, root);
2959
2960                 spin_lock(&delayed_refs->lock);
2961                 node = rb_first(&delayed_refs->href_root);
2962                 if (!node) {
2963                         spin_unlock(&delayed_refs->lock);
2964                         goto out;
2965                 }
2966                 count = (unsigned long)-1;
2967
2968                 while (node) {
2969                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2970                                         href_node);
2971                         if (btrfs_delayed_ref_is_head(&head->node)) {
2972                                 struct btrfs_delayed_ref_node *ref;
2973
2974                                 ref = &head->node;
2975                                 atomic_inc(&ref->refs);
2976
2977                                 spin_unlock(&delayed_refs->lock);
2978                                 /*
2979                                  * Mutex was contended, block until it's
2980                                  * released and try again
2981                                  */
2982                                 mutex_lock(&head->mutex);
2983                                 mutex_unlock(&head->mutex);
2984
2985                                 btrfs_put_delayed_ref(ref);
2986                                 cond_resched();
2987                                 goto again;
2988                         } else {
2989                                 WARN_ON(1);
2990                         }
2991                         node = rb_next(node);
2992                 }
2993                 spin_unlock(&delayed_refs->lock);
2994                 cond_resched();
2995                 goto again;
2996         }
2997 out:
2998         assert_qgroups_uptodate(trans);
2999         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3000         return 0;
3001 }
3002
3003 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3004                                 struct btrfs_root *root,
3005                                 u64 bytenr, u64 num_bytes, u64 flags,
3006                                 int level, int is_data)
3007 {
3008         struct btrfs_delayed_extent_op *extent_op;
3009         int ret;
3010
3011         extent_op = btrfs_alloc_delayed_extent_op();
3012         if (!extent_op)
3013                 return -ENOMEM;
3014
3015         extent_op->flags_to_set = flags;
3016         extent_op->update_flags = true;
3017         extent_op->update_key = false;
3018         extent_op->is_data = is_data ? true : false;
3019         extent_op->level = level;
3020
3021         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
3022                                           num_bytes, extent_op);
3023         if (ret)
3024                 btrfs_free_delayed_extent_op(extent_op);
3025         return ret;
3026 }
3027
3028 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
3029                                       struct btrfs_root *root,
3030                                       struct btrfs_path *path,
3031                                       u64 objectid, u64 offset, u64 bytenr)
3032 {
3033         struct btrfs_delayed_ref_head *head;
3034         struct btrfs_delayed_ref_node *ref;
3035         struct btrfs_delayed_data_ref *data_ref;
3036         struct btrfs_delayed_ref_root *delayed_refs;
3037         int ret = 0;
3038
3039         delayed_refs = &trans->transaction->delayed_refs;
3040         spin_lock(&delayed_refs->lock);
3041         head = btrfs_find_delayed_ref_head(trans, bytenr);
3042         if (!head) {
3043                 spin_unlock(&delayed_refs->lock);
3044                 return 0;
3045         }
3046
3047         if (!mutex_trylock(&head->mutex)) {
3048                 atomic_inc(&head->node.refs);
3049                 spin_unlock(&delayed_refs->lock);
3050
3051                 btrfs_release_path(path);
3052
3053                 /*
3054                  * Mutex was contended, block until it's released and let
3055                  * caller try again
3056                  */
3057                 mutex_lock(&head->mutex);
3058                 mutex_unlock(&head->mutex);
3059                 btrfs_put_delayed_ref(&head->node);
3060                 return -EAGAIN;
3061         }
3062         spin_unlock(&delayed_refs->lock);
3063
3064         spin_lock(&head->lock);
3065         list_for_each_entry(ref, &head->ref_list, list) {
3066                 /* If it's a shared ref we know a cross reference exists */
3067                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3068                         ret = 1;
3069                         break;
3070                 }
3071
3072                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3073
3074                 /*
3075                  * If our ref doesn't match the one we're currently looking at
3076                  * then we have a cross reference.
3077                  */
3078                 if (data_ref->root != root->root_key.objectid ||
3079                     data_ref->objectid != objectid ||
3080                     data_ref->offset != offset) {
3081                         ret = 1;
3082                         break;
3083                 }
3084         }
3085         spin_unlock(&head->lock);
3086         mutex_unlock(&head->mutex);
3087         return ret;
3088 }
3089
3090 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
3091                                         struct btrfs_root *root,
3092                                         struct btrfs_path *path,
3093                                         u64 objectid, u64 offset, u64 bytenr)
3094 {
3095         struct btrfs_root *extent_root = root->fs_info->extent_root;
3096         struct extent_buffer *leaf;
3097         struct btrfs_extent_data_ref *ref;
3098         struct btrfs_extent_inline_ref *iref;
3099         struct btrfs_extent_item *ei;
3100         struct btrfs_key key;
3101         u32 item_size;
3102         int ret;
3103
3104         key.objectid = bytenr;
3105         key.offset = (u64)-1;
3106         key.type = BTRFS_EXTENT_ITEM_KEY;
3107
3108         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3109         if (ret < 0)
3110                 goto out;
3111         BUG_ON(ret == 0); /* Corruption */
3112
3113         ret = -ENOENT;
3114         if (path->slots[0] == 0)
3115                 goto out;
3116
3117         path->slots[0]--;
3118         leaf = path->nodes[0];
3119         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3120
3121         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3122                 goto out;
3123
3124         ret = 1;
3125         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3126 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3127         if (item_size < sizeof(*ei)) {
3128                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3129                 goto out;
3130         }
3131 #endif
3132         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3133
3134         if (item_size != sizeof(*ei) +
3135             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3136                 goto out;
3137
3138         if (btrfs_extent_generation(leaf, ei) <=
3139             btrfs_root_last_snapshot(&root->root_item))
3140                 goto out;
3141
3142         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3143         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3144             BTRFS_EXTENT_DATA_REF_KEY)
3145                 goto out;
3146
3147         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3148         if (btrfs_extent_refs(leaf, ei) !=
3149             btrfs_extent_data_ref_count(leaf, ref) ||
3150             btrfs_extent_data_ref_root(leaf, ref) !=
3151             root->root_key.objectid ||
3152             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3153             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3154                 goto out;
3155
3156         ret = 0;
3157 out:
3158         return ret;
3159 }
3160
3161 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3162                           struct btrfs_root *root,
3163                           u64 objectid, u64 offset, u64 bytenr)
3164 {
3165         struct btrfs_path *path;
3166         int ret;
3167         int ret2;
3168
3169         path = btrfs_alloc_path();
3170         if (!path)
3171                 return -ENOENT;
3172
3173         do {
3174                 ret = check_committed_ref(trans, root, path, objectid,
3175                                           offset, bytenr);
3176                 if (ret && ret != -ENOENT)
3177                         goto out;
3178
3179                 ret2 = check_delayed_ref(trans, root, path, objectid,
3180                                          offset, bytenr);
3181         } while (ret2 == -EAGAIN);
3182
3183         if (ret2 && ret2 != -ENOENT) {
3184                 ret = ret2;
3185                 goto out;
3186         }
3187
3188         if (ret != -ENOENT || ret2 != -ENOENT)
3189                 ret = 0;
3190 out:
3191         btrfs_free_path(path);
3192         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3193                 WARN_ON(ret > 0);
3194         return ret;
3195 }
3196
3197 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3198                            struct btrfs_root *root,
3199                            struct extent_buffer *buf,
3200                            int full_backref, int inc)
3201 {
3202         u64 bytenr;
3203         u64 num_bytes;
3204         u64 parent;
3205         u64 ref_root;
3206         u32 nritems;
3207         struct btrfs_key key;
3208         struct btrfs_file_extent_item *fi;
3209         int i;
3210         int level;
3211         int ret = 0;
3212         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3213                             u64, u64, u64, u64, u64, u64);
3214
3215
3216         if (btrfs_test_is_dummy_root(root))
3217                 return 0;
3218
3219         ref_root = btrfs_header_owner(buf);
3220         nritems = btrfs_header_nritems(buf);
3221         level = btrfs_header_level(buf);
3222
3223         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3224                 return 0;
3225
3226         if (inc)
3227                 process_func = btrfs_inc_extent_ref;
3228         else
3229                 process_func = btrfs_free_extent;
3230
3231         if (full_backref)
3232                 parent = buf->start;
3233         else
3234                 parent = 0;
3235
3236         for (i = 0; i < nritems; i++) {
3237                 if (level == 0) {
3238                         btrfs_item_key_to_cpu(buf, &key, i);
3239                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3240                                 continue;
3241                         fi = btrfs_item_ptr(buf, i,
3242                                             struct btrfs_file_extent_item);
3243                         if (btrfs_file_extent_type(buf, fi) ==
3244                             BTRFS_FILE_EXTENT_INLINE)
3245                                 continue;
3246                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3247                         if (bytenr == 0)
3248                                 continue;
3249
3250                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3251                         key.offset -= btrfs_file_extent_offset(buf, fi);
3252                         ret = process_func(trans, root, bytenr, num_bytes,
3253                                            parent, ref_root, key.objectid,
3254                                            key.offset);
3255                         if (ret)
3256                                 goto fail;
3257                 } else {
3258                         bytenr = btrfs_node_blockptr(buf, i);
3259                         num_bytes = root->nodesize;
3260                         ret = process_func(trans, root, bytenr, num_bytes,
3261                                            parent, ref_root, level - 1, 0);
3262                         if (ret)
3263                                 goto fail;
3264                 }
3265         }
3266         return 0;
3267 fail:
3268         return ret;
3269 }
3270
3271 int btrfs_inc_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, 1);
3275 }
3276
3277 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3278                   struct extent_buffer *buf, int full_backref)
3279 {
3280         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3281 }
3282
3283 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3284                                  struct btrfs_root *root,
3285                                  struct btrfs_path *path,
3286                                  struct btrfs_block_group_cache *cache)
3287 {
3288         int ret;
3289         struct btrfs_root *extent_root = root->fs_info->extent_root;
3290         unsigned long bi;
3291         struct extent_buffer *leaf;
3292
3293         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3294         if (ret) {
3295                 if (ret > 0)
3296                         ret = -ENOENT;
3297                 goto fail;
3298         }
3299
3300         leaf = path->nodes[0];
3301         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3302         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3303         btrfs_mark_buffer_dirty(leaf);
3304 fail:
3305         btrfs_release_path(path);
3306         return ret;
3307
3308 }
3309
3310 static struct btrfs_block_group_cache *
3311 next_block_group(struct btrfs_root *root,
3312                  struct btrfs_block_group_cache *cache)
3313 {
3314         struct rb_node *node;
3315
3316         spin_lock(&root->fs_info->block_group_cache_lock);
3317
3318         /* If our block group was removed, we need a full search. */
3319         if (RB_EMPTY_NODE(&cache->cache_node)) {
3320                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3321
3322                 spin_unlock(&root->fs_info->block_group_cache_lock);
3323                 btrfs_put_block_group(cache);
3324                 cache = btrfs_lookup_first_block_group(root->fs_info,
3325                                                        next_bytenr);
3326                 return cache;
3327         }
3328         node = rb_next(&cache->cache_node);
3329         btrfs_put_block_group(cache);
3330         if (node) {
3331                 cache = rb_entry(node, struct btrfs_block_group_cache,
3332                                  cache_node);
3333                 btrfs_get_block_group(cache);
3334         } else
3335                 cache = NULL;
3336         spin_unlock(&root->fs_info->block_group_cache_lock);
3337         return cache;
3338 }
3339
3340 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3341                             struct btrfs_trans_handle *trans,
3342                             struct btrfs_path *path)
3343 {
3344         struct btrfs_root *root = block_group->fs_info->tree_root;
3345         struct inode *inode = NULL;
3346         u64 alloc_hint = 0;
3347         int dcs = BTRFS_DC_ERROR;
3348         u64 num_pages = 0;
3349         int retries = 0;
3350         int ret = 0;
3351
3352         /*
3353          * If this block group is smaller than 100 megs don't bother caching the
3354          * block group.
3355          */
3356         if (block_group->key.offset < (100 * SZ_1M)) {
3357                 spin_lock(&block_group->lock);
3358                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3359                 spin_unlock(&block_group->lock);
3360                 return 0;
3361         }
3362
3363         if (trans->aborted)
3364                 return 0;
3365 again:
3366         inode = lookup_free_space_inode(root, block_group, path);
3367         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3368                 ret = PTR_ERR(inode);
3369                 btrfs_release_path(path);
3370                 goto out;
3371         }
3372
3373         if (IS_ERR(inode)) {
3374                 BUG_ON(retries);
3375                 retries++;
3376
3377                 if (block_group->ro)
3378                         goto out_free;
3379
3380                 ret = create_free_space_inode(root, trans, block_group, path);
3381                 if (ret)
3382                         goto out_free;
3383                 goto again;
3384         }
3385
3386         /* We've already setup this transaction, go ahead and exit */
3387         if (block_group->cache_generation == trans->transid &&
3388             i_size_read(inode)) {
3389                 dcs = BTRFS_DC_SETUP;
3390                 goto out_put;
3391         }
3392
3393         /*
3394          * We want to set the generation to 0, that way if anything goes wrong
3395          * from here on out we know not to trust this cache when we load up next
3396          * time.
3397          */
3398         BTRFS_I(inode)->generation = 0;
3399         ret = btrfs_update_inode(trans, root, inode);
3400         if (ret) {
3401                 /*
3402                  * So theoretically we could recover from this, simply set the
3403                  * super cache generation to 0 so we know to invalidate the
3404                  * cache, but then we'd have to keep track of the block groups
3405                  * that fail this way so we know we _have_ to reset this cache
3406                  * before the next commit or risk reading stale cache.  So to
3407                  * limit our exposure to horrible edge cases lets just abort the
3408                  * transaction, this only happens in really bad situations
3409                  * anyway.
3410                  */
3411                 btrfs_abort_transaction(trans, root, ret);
3412                 goto out_put;
3413         }
3414         WARN_ON(ret);
3415
3416         if (i_size_read(inode) > 0) {
3417                 ret = btrfs_check_trunc_cache_free_space(root,
3418                                         &root->fs_info->global_block_rsv);
3419                 if (ret)
3420                         goto out_put;
3421
3422                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3423                 if (ret)
3424                         goto out_put;
3425         }
3426
3427         spin_lock(&block_group->lock);
3428         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3429             !btrfs_test_opt(root, SPACE_CACHE)) {
3430                 /*
3431                  * don't bother trying to write stuff out _if_
3432                  * a) we're not cached,
3433                  * b) we're with nospace_cache mount option.
3434                  */
3435                 dcs = BTRFS_DC_WRITTEN;
3436                 spin_unlock(&block_group->lock);
3437                 goto out_put;
3438         }
3439         spin_unlock(&block_group->lock);
3440
3441         /*
3442          * We hit an ENOSPC when setting up the cache in this transaction, just
3443          * skip doing the setup, we've already cleared the cache so we're safe.
3444          */
3445         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3446                 ret = -ENOSPC;
3447                 goto out_put;
3448         }
3449
3450         /*
3451          * Try to preallocate enough space based on how big the block group is.
3452          * Keep in mind this has to include any pinned space which could end up
3453          * taking up quite a bit since it's not folded into the other space
3454          * cache.
3455          */
3456         num_pages = div_u64(block_group->key.offset, SZ_256M);
3457         if (!num_pages)
3458                 num_pages = 1;
3459
3460         num_pages *= 16;
3461         num_pages *= PAGE_SIZE;
3462
3463         ret = btrfs_check_data_free_space(inode, 0, num_pages);
3464         if (ret)
3465                 goto out_put;
3466
3467         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3468                                               num_pages, num_pages,
3469                                               &alloc_hint);
3470         /*
3471          * Our cache requires contiguous chunks so that we don't modify a bunch
3472          * of metadata or split extents when writing the cache out, which means
3473          * we can enospc if we are heavily fragmented in addition to just normal
3474          * out of space conditions.  So if we hit this just skip setting up any
3475          * other block groups for this transaction, maybe we'll unpin enough
3476          * space the next time around.
3477          */
3478         if (!ret)
3479                 dcs = BTRFS_DC_SETUP;
3480         else if (ret == -ENOSPC)
3481                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3482         btrfs_free_reserved_data_space(inode, 0, num_pages);
3483
3484 out_put:
3485         iput(inode);
3486 out_free:
3487         btrfs_release_path(path);
3488 out:
3489         spin_lock(&block_group->lock);
3490         if (!ret && dcs == BTRFS_DC_SETUP)
3491                 block_group->cache_generation = trans->transid;
3492         block_group->disk_cache_state = dcs;
3493         spin_unlock(&block_group->lock);
3494
3495         return ret;
3496 }
3497
3498 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3499                             struct btrfs_root *root)
3500 {
3501         struct btrfs_block_group_cache *cache, *tmp;
3502         struct btrfs_transaction *cur_trans = trans->transaction;
3503         struct btrfs_path *path;
3504
3505         if (list_empty(&cur_trans->dirty_bgs) ||
3506             !btrfs_test_opt(root, SPACE_CACHE))
3507                 return 0;
3508
3509         path = btrfs_alloc_path();
3510         if (!path)
3511                 return -ENOMEM;
3512
3513         /* Could add new block groups, use _safe just in case */
3514         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3515                                  dirty_list) {
3516                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3517                         cache_save_setup(cache, trans, path);
3518         }
3519
3520         btrfs_free_path(path);
3521         return 0;
3522 }
3523
3524 /*
3525  * transaction commit does final block group cache writeback during a
3526  * critical section where nothing is allowed to change the FS.  This is
3527  * required in order for the cache to actually match the block group,
3528  * but can introduce a lot of latency into the commit.
3529  *
3530  * So, btrfs_start_dirty_block_groups is here to kick off block group
3531  * cache IO.  There's a chance we'll have to redo some of it if the
3532  * block group changes again during the commit, but it greatly reduces
3533  * the commit latency by getting rid of the easy block groups while
3534  * we're still allowing others to join the commit.
3535  */
3536 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3537                                    struct btrfs_root *root)
3538 {
3539         struct btrfs_block_group_cache *cache;
3540         struct btrfs_transaction *cur_trans = trans->transaction;
3541         int ret = 0;
3542         int should_put;
3543         struct btrfs_path *path = NULL;
3544         LIST_HEAD(dirty);
3545         struct list_head *io = &cur_trans->io_bgs;
3546         int num_started = 0;
3547         int loops = 0;
3548
3549         spin_lock(&cur_trans->dirty_bgs_lock);
3550         if (list_empty(&cur_trans->dirty_bgs)) {
3551                 spin_unlock(&cur_trans->dirty_bgs_lock);
3552                 return 0;
3553         }
3554         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3555         spin_unlock(&cur_trans->dirty_bgs_lock);
3556
3557 again:
3558         /*
3559          * make sure all the block groups on our dirty list actually
3560          * exist
3561          */
3562         btrfs_create_pending_block_groups(trans, root);
3563
3564         if (!path) {
3565                 path = btrfs_alloc_path();
3566                 if (!path)
3567                         return -ENOMEM;
3568         }
3569
3570         /*
3571          * cache_write_mutex is here only to save us from balance or automatic
3572          * removal of empty block groups deleting this block group while we are
3573          * writing out the cache
3574          */
3575         mutex_lock(&trans->transaction->cache_write_mutex);
3576         while (!list_empty(&dirty)) {
3577                 cache = list_first_entry(&dirty,
3578                                          struct btrfs_block_group_cache,
3579                                          dirty_list);
3580                 /*
3581                  * this can happen if something re-dirties a block
3582                  * group that is already under IO.  Just wait for it to
3583                  * finish and then do it all again
3584                  */
3585                 if (!list_empty(&cache->io_list)) {
3586                         list_del_init(&cache->io_list);
3587                         btrfs_wait_cache_io(root, trans, cache,
3588                                             &cache->io_ctl, path,
3589                                             cache->key.objectid);
3590                         btrfs_put_block_group(cache);
3591                 }
3592
3593
3594                 /*
3595                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3596                  * if it should update the cache_state.  Don't delete
3597                  * until after we wait.
3598                  *
3599                  * Since we're not running in the commit critical section
3600                  * we need the dirty_bgs_lock to protect from update_block_group
3601                  */
3602                 spin_lock(&cur_trans->dirty_bgs_lock);
3603                 list_del_init(&cache->dirty_list);
3604                 spin_unlock(&cur_trans->dirty_bgs_lock);
3605
3606                 should_put = 1;
3607
3608                 cache_save_setup(cache, trans, path);
3609
3610                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3611                         cache->io_ctl.inode = NULL;
3612                         ret = btrfs_write_out_cache(root, trans, cache, path);
3613                         if (ret == 0 && cache->io_ctl.inode) {
3614                                 num_started++;
3615                                 should_put = 0;
3616
3617                                 /*
3618                                  * the cache_write_mutex is protecting
3619                                  * the io_list
3620                                  */
3621                                 list_add_tail(&cache->io_list, io);
3622                         } else {
3623                                 /*
3624                                  * if we failed to write the cache, the
3625                                  * generation will be bad and life goes on
3626                                  */
3627                                 ret = 0;
3628                         }
3629                 }
3630                 if (!ret) {
3631                         ret = write_one_cache_group(trans, root, path, cache);
3632                         /*
3633                          * Our block group might still be attached to the list
3634                          * of new block groups in the transaction handle of some
3635                          * other task (struct btrfs_trans_handle->new_bgs). This
3636                          * means its block group item isn't yet in the extent
3637                          * tree. If this happens ignore the error, as we will
3638                          * try again later in the critical section of the
3639                          * transaction commit.
3640                          */
3641                         if (ret == -ENOENT) {
3642                                 ret = 0;
3643                                 spin_lock(&cur_trans->dirty_bgs_lock);
3644                                 if (list_empty(&cache->dirty_list)) {
3645                                         list_add_tail(&cache->dirty_list,
3646                                                       &cur_trans->dirty_bgs);
3647                                         btrfs_get_block_group(cache);
3648                                 }
3649                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3650                         } else if (ret) {
3651                                 btrfs_abort_transaction(trans, root, ret);
3652                         }
3653                 }
3654
3655                 /* if its not on the io list, we need to put the block group */
3656                 if (should_put)
3657                         btrfs_put_block_group(cache);
3658
3659                 if (ret)
3660                         break;
3661
3662                 /*
3663                  * Avoid blocking other tasks for too long. It might even save
3664                  * us from writing caches for block groups that are going to be
3665                  * removed.
3666                  */
3667                 mutex_unlock(&trans->transaction->cache_write_mutex);
3668                 mutex_lock(&trans->transaction->cache_write_mutex);
3669         }
3670         mutex_unlock(&trans->transaction->cache_write_mutex);
3671
3672         /*
3673          * go through delayed refs for all the stuff we've just kicked off
3674          * and then loop back (just once)
3675          */
3676         ret = btrfs_run_delayed_refs(trans, root, 0);
3677         if (!ret && loops == 0) {
3678                 loops++;
3679                 spin_lock(&cur_trans->dirty_bgs_lock);
3680                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3681                 /*
3682                  * dirty_bgs_lock protects us from concurrent block group
3683                  * deletes too (not just cache_write_mutex).
3684                  */
3685                 if (!list_empty(&dirty)) {
3686                         spin_unlock(&cur_trans->dirty_bgs_lock);
3687                         goto again;
3688                 }
3689                 spin_unlock(&cur_trans->dirty_bgs_lock);
3690         }
3691
3692         btrfs_free_path(path);
3693         return ret;
3694 }
3695
3696 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3697                                    struct btrfs_root *root)
3698 {
3699         struct btrfs_block_group_cache *cache;
3700         struct btrfs_transaction *cur_trans = trans->transaction;
3701         int ret = 0;
3702         int should_put;
3703         struct btrfs_path *path;
3704         struct list_head *io = &cur_trans->io_bgs;
3705         int num_started = 0;
3706
3707         path = btrfs_alloc_path();
3708         if (!path)
3709                 return -ENOMEM;
3710
3711         /*
3712          * Even though we are in the critical section of the transaction commit,
3713          * we can still have concurrent tasks adding elements to this
3714          * transaction's list of dirty block groups. These tasks correspond to
3715          * endio free space workers started when writeback finishes for a
3716          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3717          * allocate new block groups as a result of COWing nodes of the root
3718          * tree when updating the free space inode. The writeback for the space
3719          * caches is triggered by an earlier call to
3720          * btrfs_start_dirty_block_groups() and iterations of the following
3721          * loop.
3722          * Also we want to do the cache_save_setup first and then run the
3723          * delayed refs to make sure we have the best chance at doing this all
3724          * in one shot.
3725          */
3726         spin_lock(&cur_trans->dirty_bgs_lock);
3727         while (!list_empty(&cur_trans->dirty_bgs)) {
3728                 cache = list_first_entry(&cur_trans->dirty_bgs,
3729                                          struct btrfs_block_group_cache,
3730                                          dirty_list);
3731
3732                 /*
3733                  * this can happen if cache_save_setup re-dirties a block
3734                  * group that is already under IO.  Just wait for it to
3735                  * finish and then do it all again
3736                  */
3737                 if (!list_empty(&cache->io_list)) {
3738                         spin_unlock(&cur_trans->dirty_bgs_lock);
3739                         list_del_init(&cache->io_list);
3740                         btrfs_wait_cache_io(root, trans, cache,
3741                                             &cache->io_ctl, path,
3742                                             cache->key.objectid);
3743                         btrfs_put_block_group(cache);
3744                         spin_lock(&cur_trans->dirty_bgs_lock);
3745                 }
3746
3747                 /*
3748                  * don't remove from the dirty list until after we've waited
3749                  * on any pending IO
3750                  */
3751                 list_del_init(&cache->dirty_list);
3752                 spin_unlock(&cur_trans->dirty_bgs_lock);
3753                 should_put = 1;
3754
3755                 cache_save_setup(cache, trans, path);
3756
3757                 if (!ret)
3758                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3759
3760                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3761                         cache->io_ctl.inode = NULL;
3762                         ret = btrfs_write_out_cache(root, trans, cache, path);
3763                         if (ret == 0 && cache->io_ctl.inode) {
3764                                 num_started++;
3765                                 should_put = 0;
3766                                 list_add_tail(&cache->io_list, io);
3767                         } else {
3768                                 /*
3769                                  * if we failed to write the cache, the
3770                                  * generation will be bad and life goes on
3771                                  */
3772                                 ret = 0;
3773                         }
3774                 }
3775                 if (!ret) {
3776                         ret = write_one_cache_group(trans, root, path, cache);
3777                         /*
3778                          * One of the free space endio workers might have
3779                          * created a new block group while updating a free space
3780                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3781                          * and hasn't released its transaction handle yet, in
3782                          * which case the new block group is still attached to
3783                          * its transaction handle and its creation has not
3784                          * finished yet (no block group item in the extent tree
3785                          * yet, etc). If this is the case, wait for all free
3786                          * space endio workers to finish and retry. This is a
3787                          * a very rare case so no need for a more efficient and
3788                          * complex approach.
3789                          */
3790                         if (ret == -ENOENT) {
3791                                 wait_event(cur_trans->writer_wait,
3792                                    atomic_read(&cur_trans->num_writers) == 1);
3793                                 ret = write_one_cache_group(trans, root, path,
3794                                                             cache);
3795                         }
3796                         if (ret)
3797                                 btrfs_abort_transaction(trans, root, ret);
3798                 }
3799
3800                 /* if its not on the io list, we need to put the block group */
3801                 if (should_put)
3802                         btrfs_put_block_group(cache);
3803                 spin_lock(&cur_trans->dirty_bgs_lock);
3804         }
3805         spin_unlock(&cur_trans->dirty_bgs_lock);
3806
3807         while (!list_empty(io)) {
3808                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3809                                          io_list);
3810                 list_del_init(&cache->io_list);
3811                 btrfs_wait_cache_io(root, trans, cache,
3812                                     &cache->io_ctl, path, cache->key.objectid);
3813                 btrfs_put_block_group(cache);
3814         }
3815
3816         btrfs_free_path(path);
3817         return ret;
3818 }
3819
3820 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3821 {
3822         struct btrfs_block_group_cache *block_group;
3823         int readonly = 0;
3824
3825         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3826         if (!block_group || block_group->ro)
3827                 readonly = 1;
3828         if (block_group)
3829                 btrfs_put_block_group(block_group);
3830         return readonly;
3831 }
3832
3833 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3834 {
3835         struct btrfs_block_group_cache *bg;
3836         bool ret = true;
3837
3838         bg = btrfs_lookup_block_group(fs_info, bytenr);
3839         if (!bg)
3840                 return false;
3841
3842         spin_lock(&bg->lock);
3843         if (bg->ro)
3844                 ret = false;
3845         else
3846                 atomic_inc(&bg->nocow_writers);
3847         spin_unlock(&bg->lock);
3848
3849         /* no put on block group, done by btrfs_dec_nocow_writers */
3850         if (!ret)
3851                 btrfs_put_block_group(bg);
3852
3853         return ret;
3854
3855 }
3856
3857 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3858 {
3859         struct btrfs_block_group_cache *bg;
3860
3861         bg = btrfs_lookup_block_group(fs_info, bytenr);
3862         ASSERT(bg);
3863         if (atomic_dec_and_test(&bg->nocow_writers))
3864                 wake_up_atomic_t(&bg->nocow_writers);
3865         /*
3866          * Once for our lookup and once for the lookup done by a previous call
3867          * to btrfs_inc_nocow_writers()
3868          */
3869         btrfs_put_block_group(bg);
3870         btrfs_put_block_group(bg);
3871 }
3872
3873 static int btrfs_wait_nocow_writers_atomic_t(atomic_t *a)
3874 {
3875         schedule();
3876         return 0;
3877 }
3878
3879 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3880 {
3881         wait_on_atomic_t(&bg->nocow_writers,
3882                          btrfs_wait_nocow_writers_atomic_t,
3883                          TASK_UNINTERRUPTIBLE);
3884 }
3885
3886 static const char *alloc_name(u64 flags)
3887 {
3888         switch (flags) {
3889         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3890                 return "mixed";
3891         case BTRFS_BLOCK_GROUP_METADATA:
3892                 return "metadata";
3893         case BTRFS_BLOCK_GROUP_DATA:
3894                 return "data";
3895         case BTRFS_BLOCK_GROUP_SYSTEM:
3896                 return "system";
3897         default:
3898                 WARN_ON(1);
3899                 return "invalid-combination";
3900         };
3901 }
3902
3903 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3904                              u64 total_bytes, u64 bytes_used,
3905                              struct btrfs_space_info **space_info)
3906 {
3907         struct btrfs_space_info *found;
3908         int i;
3909         int factor;
3910         int ret;
3911
3912         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3913                      BTRFS_BLOCK_GROUP_RAID10))
3914                 factor = 2;
3915         else
3916                 factor = 1;
3917
3918         found = __find_space_info(info, flags);
3919         if (found) {
3920                 spin_lock(&found->lock);
3921                 found->total_bytes += total_bytes;
3922                 found->disk_total += total_bytes * factor;
3923                 found->bytes_used += bytes_used;
3924                 found->disk_used += bytes_used * factor;
3925                 if (total_bytes > 0)
3926                         found->full = 0;
3927                 spin_unlock(&found->lock);
3928                 *space_info = found;
3929                 return 0;
3930         }
3931         found = kzalloc(sizeof(*found), GFP_NOFS);
3932         if (!found)
3933                 return -ENOMEM;
3934
3935         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3936         if (ret) {
3937                 kfree(found);
3938                 return ret;
3939         }
3940
3941         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3942                 INIT_LIST_HEAD(&found->block_groups[i]);
3943         init_rwsem(&found->groups_sem);
3944         spin_lock_init(&found->lock);
3945         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3946         found->total_bytes = total_bytes;
3947         found->disk_total = total_bytes * factor;
3948         found->bytes_used = bytes_used;
3949         found->disk_used = bytes_used * factor;
3950         found->bytes_pinned = 0;
3951         found->bytes_reserved = 0;
3952         found->bytes_readonly = 0;
3953         found->bytes_may_use = 0;
3954         found->full = 0;
3955         found->max_extent_size = 0;
3956         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3957         found->chunk_alloc = 0;
3958         found->flush = 0;
3959         init_waitqueue_head(&found->wait);
3960         INIT_LIST_HEAD(&found->ro_bgs);
3961
3962         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3963                                     info->space_info_kobj, "%s",
3964                                     alloc_name(found->flags));
3965         if (ret) {
3966                 kfree(found);
3967                 return ret;
3968         }
3969
3970         *space_info = found;
3971         list_add_rcu(&found->list, &info->space_info);
3972         if (flags & BTRFS_BLOCK_GROUP_DATA)
3973                 info->data_sinfo = found;
3974
3975         return ret;
3976 }
3977
3978 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3979 {
3980         u64 extra_flags = chunk_to_extended(flags) &
3981                                 BTRFS_EXTENDED_PROFILE_MASK;
3982
3983         write_seqlock(&fs_info->profiles_lock);
3984         if (flags & BTRFS_BLOCK_GROUP_DATA)
3985                 fs_info->avail_data_alloc_bits |= extra_flags;
3986         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3987                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3988         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3989                 fs_info->avail_system_alloc_bits |= extra_flags;
3990         write_sequnlock(&fs_info->profiles_lock);
3991 }
3992
3993 /*
3994  * returns target flags in extended format or 0 if restripe for this
3995  * chunk_type is not in progress
3996  *
3997  * should be called with either volume_mutex or balance_lock held
3998  */
3999 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4000 {
4001         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4002         u64 target = 0;
4003
4004         if (!bctl)
4005                 return 0;
4006
4007         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4008             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4009                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4010         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4011                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4012                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4013         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4014                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4015                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4016         }
4017
4018         return target;
4019 }
4020
4021 /*
4022  * @flags: available profiles in extended format (see ctree.h)
4023  *
4024  * Returns reduced profile in chunk format.  If profile changing is in
4025  * progress (either running or paused) picks the target profile (if it's
4026  * already available), otherwise falls back to plain reducing.
4027  */
4028 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
4029 {
4030         u64 num_devices = root->fs_info->fs_devices->rw_devices;
4031         u64 target;
4032         u64 raid_type;
4033         u64 allowed = 0;
4034
4035         /*
4036          * see if restripe for this chunk_type is in progress, if so
4037          * try to reduce to the target profile
4038          */
4039         spin_lock(&root->fs_info->balance_lock);
4040         target = get_restripe_target(root->fs_info, flags);
4041         if (target) {
4042                 /* pick target profile only if it's already available */
4043                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4044                         spin_unlock(&root->fs_info->balance_lock);
4045                         return extended_to_chunk(target);
4046                 }
4047         }
4048         spin_unlock(&root->fs_info->balance_lock);
4049
4050         /* First, mask out the RAID levels which aren't possible */
4051         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4052                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4053                         allowed |= btrfs_raid_group[raid_type];
4054         }
4055         allowed &= flags;
4056
4057         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4058                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4059         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4060                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4061         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4062                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4063         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4064                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4065         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4066                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4067
4068         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4069
4070         return extended_to_chunk(flags | allowed);
4071 }
4072
4073 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
4074 {
4075         unsigned seq;
4076         u64 flags;
4077
4078         do {
4079                 flags = orig_flags;
4080                 seq = read_seqbegin(&root->fs_info->profiles_lock);
4081
4082                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4083                         flags |= root->fs_info->avail_data_alloc_bits;
4084                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4085                         flags |= root->fs_info->avail_system_alloc_bits;
4086                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4087                         flags |= root->fs_info->avail_metadata_alloc_bits;
4088         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
4089
4090         return btrfs_reduce_alloc_profile(root, flags);
4091 }
4092
4093 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
4094 {
4095         u64 flags;
4096         u64 ret;
4097
4098         if (data)
4099                 flags = BTRFS_BLOCK_GROUP_DATA;
4100         else if (root == root->fs_info->chunk_root)
4101                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4102         else
4103                 flags = BTRFS_BLOCK_GROUP_METADATA;
4104
4105         ret = get_alloc_profile(root, flags);
4106         return ret;
4107 }
4108
4109 int btrfs_alloc_data_chunk_ondemand(struct inode *inode, u64 bytes)
4110 {
4111         struct btrfs_space_info *data_sinfo;
4112         struct btrfs_root *root = BTRFS_I(inode)->root;
4113         struct btrfs_fs_info *fs_info = root->fs_info;
4114         u64 used;
4115         int ret = 0;
4116         int need_commit = 2;
4117         int have_pinned_space;
4118
4119         /* make sure bytes are sectorsize aligned */
4120         bytes = ALIGN(bytes, root->sectorsize);
4121
4122         if (btrfs_is_free_space_inode(inode)) {
4123                 need_commit = 0;
4124                 ASSERT(current->journal_info);
4125         }
4126
4127         data_sinfo = fs_info->data_sinfo;
4128         if (!data_sinfo)
4129                 goto alloc;
4130
4131 again:
4132         /* make sure we have enough space to handle the data first */
4133         spin_lock(&data_sinfo->lock);
4134         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
4135                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
4136                 data_sinfo->bytes_may_use;
4137
4138         if (used + bytes > data_sinfo->total_bytes) {
4139                 struct btrfs_trans_handle *trans;
4140
4141                 /*
4142                  * if we don't have enough free bytes in this space then we need
4143                  * to alloc a new chunk.
4144                  */
4145                 if (!data_sinfo->full) {
4146                         u64 alloc_target;
4147
4148                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4149                         spin_unlock(&data_sinfo->lock);
4150 alloc:
4151                         alloc_target = btrfs_get_alloc_profile(root, 1);
4152                         /*
4153                          * It is ugly that we don't call nolock join
4154                          * transaction for the free space inode case here.
4155                          * But it is safe because we only do the data space
4156                          * reservation for the free space cache in the
4157                          * transaction context, the common join transaction
4158                          * just increase the counter of the current transaction
4159                          * handler, doesn't try to acquire the trans_lock of
4160                          * the fs.
4161                          */
4162                         trans = btrfs_join_transaction(root);
4163                         if (IS_ERR(trans))
4164                                 return PTR_ERR(trans);
4165
4166                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4167                                              alloc_target,
4168                                              CHUNK_ALLOC_NO_FORCE);
4169                         btrfs_end_transaction(trans, root);
4170                         if (ret < 0) {
4171                                 if (ret != -ENOSPC)
4172                                         return ret;
4173                                 else {
4174                                         have_pinned_space = 1;
4175                                         goto commit_trans;
4176                                 }
4177                         }
4178
4179                         if (!data_sinfo)
4180                                 data_sinfo = fs_info->data_sinfo;
4181
4182                         goto again;
4183                 }
4184
4185                 /*
4186                  * If we don't have enough pinned space to deal with this
4187                  * allocation, and no removed chunk in current transaction,
4188                  * don't bother committing the transaction.
4189                  */
4190                 have_pinned_space = percpu_counter_compare(
4191                         &data_sinfo->total_bytes_pinned,
4192                         used + bytes - data_sinfo->total_bytes);
4193                 spin_unlock(&data_sinfo->lock);
4194
4195                 /* commit the current transaction and try again */
4196 commit_trans:
4197                 if (need_commit &&
4198                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
4199                         need_commit--;
4200
4201                         if (need_commit > 0) {
4202                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4203                                 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
4204                         }
4205
4206                         trans = btrfs_join_transaction(root);
4207                         if (IS_ERR(trans))
4208                                 return PTR_ERR(trans);
4209                         if (have_pinned_space >= 0 ||
4210                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4211                                      &trans->transaction->flags) ||
4212                             need_commit > 0) {
4213                                 ret = btrfs_commit_transaction(trans, root);
4214                                 if (ret)
4215                                         return ret;
4216                                 /*
4217                                  * The cleaner kthread might still be doing iput
4218                                  * operations. Wait for it to finish so that
4219                                  * more space is released.
4220                                  */
4221                                 mutex_lock(&root->fs_info->cleaner_delayed_iput_mutex);
4222                                 mutex_unlock(&root->fs_info->cleaner_delayed_iput_mutex);
4223                                 goto again;
4224                         } else {
4225                                 btrfs_end_transaction(trans, root);
4226                         }
4227                 }
4228
4229                 trace_btrfs_space_reservation(root->fs_info,
4230                                               "space_info:enospc",
4231                                               data_sinfo->flags, bytes, 1);
4232                 return -ENOSPC;
4233         }
4234         data_sinfo->bytes_may_use += bytes;
4235         trace_btrfs_space_reservation(root->fs_info, "space_info",
4236                                       data_sinfo->flags, bytes, 1);
4237         spin_unlock(&data_sinfo->lock);
4238
4239         return ret;
4240 }
4241
4242 /*
4243  * New check_data_free_space() with ability for precious data reservation
4244  * Will replace old btrfs_check_data_free_space(), but for patch split,
4245  * add a new function first and then replace it.
4246  */
4247 int btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4248 {
4249         struct btrfs_root *root = BTRFS_I(inode)->root;
4250         int ret;
4251
4252         /* align the range */
4253         len = round_up(start + len, root->sectorsize) -
4254               round_down(start, root->sectorsize);
4255         start = round_down(start, root->sectorsize);
4256
4257         ret = btrfs_alloc_data_chunk_ondemand(inode, len);
4258         if (ret < 0)
4259                 return ret;
4260
4261         /*
4262          * Use new btrfs_qgroup_reserve_data to reserve precious data space
4263          *
4264          * TODO: Find a good method to avoid reserve data space for NOCOW
4265          * range, but don't impact performance on quota disable case.
4266          */
4267         ret = btrfs_qgroup_reserve_data(inode, start, len);
4268         return ret;
4269 }
4270
4271 /*
4272  * Called if we need to clear a data reservation for this inode
4273  * Normally in a error case.
4274  *
4275  * This one will *NOT* use accurate qgroup reserved space API, just for case
4276  * which we can't sleep and is sure it won't affect qgroup reserved space.
4277  * Like clear_bit_hook().
4278  */
4279 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4280                                             u64 len)
4281 {
4282         struct btrfs_root *root = BTRFS_I(inode)->root;
4283         struct btrfs_space_info *data_sinfo;
4284
4285         /* Make sure the range is aligned to sectorsize */
4286         len = round_up(start + len, root->sectorsize) -
4287               round_down(start, root->sectorsize);
4288         start = round_down(start, root->sectorsize);
4289
4290         data_sinfo = root->fs_info->data_sinfo;
4291         spin_lock(&data_sinfo->lock);
4292         if (WARN_ON(data_sinfo->bytes_may_use < len))
4293                 data_sinfo->bytes_may_use = 0;
4294         else
4295                 data_sinfo->bytes_may_use -= len;
4296         trace_btrfs_space_reservation(root->fs_info, "space_info",
4297                                       data_sinfo->flags, len, 0);
4298         spin_unlock(&data_sinfo->lock);
4299 }
4300
4301 /*
4302  * Called if we need to clear a data reservation for this inode
4303  * Normally in a error case.
4304  *
4305  * This one will handle the per-inode data rsv map for accurate reserved
4306  * space framework.
4307  */
4308 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4309 {
4310         btrfs_free_reserved_data_space_noquota(inode, start, len);
4311         btrfs_qgroup_free_data(inode, start, len);
4312 }
4313
4314 static void force_metadata_allocation(struct btrfs_fs_info *info)
4315 {
4316         struct list_head *head = &info->space_info;
4317         struct btrfs_space_info *found;
4318
4319         rcu_read_lock();
4320         list_for_each_entry_rcu(found, head, list) {
4321                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4322                         found->force_alloc = CHUNK_ALLOC_FORCE;
4323         }
4324         rcu_read_unlock();
4325 }
4326
4327 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4328 {
4329         return (global->size << 1);
4330 }
4331
4332 static int should_alloc_chunk(struct btrfs_root *root,
4333                               struct btrfs_space_info *sinfo, int force)
4334 {
4335         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4336         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4337         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4338         u64 thresh;
4339
4340         if (force == CHUNK_ALLOC_FORCE)
4341                 return 1;
4342
4343         /*
4344          * We need to take into account the global rsv because for all intents
4345          * and purposes it's used space.  Don't worry about locking the
4346          * global_rsv, it doesn't change except when the transaction commits.
4347          */
4348         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4349                 num_allocated += calc_global_rsv_need_space(global_rsv);
4350
4351         /*
4352          * in limited mode, we want to have some free space up to
4353          * about 1% of the FS size.
4354          */
4355         if (force == CHUNK_ALLOC_LIMITED) {
4356                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4357                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4358
4359                 if (num_bytes - num_allocated < thresh)
4360                         return 1;
4361         }
4362
4363         if (num_allocated + SZ_2M < div_factor(num_bytes, 8))
4364                 return 0;
4365         return 1;
4366 }
4367
4368 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4369 {
4370         u64 num_dev;
4371
4372         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4373                     BTRFS_BLOCK_GROUP_RAID0 |
4374                     BTRFS_BLOCK_GROUP_RAID5 |
4375                     BTRFS_BLOCK_GROUP_RAID6))
4376                 num_dev = root->fs_info->fs_devices->rw_devices;
4377         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4378                 num_dev = 2;
4379         else
4380                 num_dev = 1;    /* DUP or single */
4381
4382         return num_dev;
4383 }
4384
4385 /*
4386  * If @is_allocation is true, reserve space in the system space info necessary
4387  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4388  * removing a chunk.
4389  */
4390 void check_system_chunk(struct btrfs_trans_handle *trans,
4391                         struct btrfs_root *root,
4392                         u64 type)
4393 {
4394         struct btrfs_space_info *info;
4395         u64 left;
4396         u64 thresh;
4397         int ret = 0;
4398         u64 num_devs;
4399
4400         /*
4401          * Needed because we can end up allocating a system chunk and for an
4402          * atomic and race free space reservation in the chunk block reserve.
4403          */
4404         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4405
4406         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4407         spin_lock(&info->lock);
4408         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4409                 info->bytes_reserved - info->bytes_readonly -
4410                 info->bytes_may_use;
4411         spin_unlock(&info->lock);
4412
4413         num_devs = get_profile_num_devs(root, type);
4414
4415         /* num_devs device items to update and 1 chunk item to add or remove */
4416         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4417                 btrfs_calc_trans_metadata_size(root, 1);
4418
4419         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4420                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4421                         left, thresh, type);
4422                 dump_space_info(info, 0, 0);
4423         }
4424
4425         if (left < thresh) {
4426                 u64 flags;
4427
4428                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4429                 /*
4430                  * Ignore failure to create system chunk. We might end up not
4431                  * needing it, as we might not need to COW all nodes/leafs from
4432                  * the paths we visit in the chunk tree (they were already COWed
4433                  * or created in the current transaction for example).
4434                  */
4435                 ret = btrfs_alloc_chunk(trans, root, flags);
4436         }
4437
4438         if (!ret) {
4439                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4440                                           &root->fs_info->chunk_block_rsv,
4441                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4442                 if (!ret)
4443                         trans->chunk_bytes_reserved += thresh;
4444         }
4445 }
4446
4447 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4448                           struct btrfs_root *extent_root, u64 flags, int force)
4449 {
4450         struct btrfs_space_info *space_info;
4451         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4452         int wait_for_alloc = 0;
4453         int ret = 0;
4454
4455         /* Don't re-enter if we're already allocating a chunk */
4456         if (trans->allocating_chunk)
4457                 return -ENOSPC;
4458
4459         space_info = __find_space_info(extent_root->fs_info, flags);
4460         if (!space_info) {
4461                 ret = update_space_info(extent_root->fs_info, flags,
4462                                         0, 0, &space_info);
4463                 BUG_ON(ret); /* -ENOMEM */
4464         }
4465         BUG_ON(!space_info); /* Logic error */
4466
4467 again:
4468         spin_lock(&space_info->lock);
4469         if (force < space_info->force_alloc)
4470                 force = space_info->force_alloc;
4471         if (space_info->full) {
4472                 if (should_alloc_chunk(extent_root, space_info, force))
4473                         ret = -ENOSPC;
4474                 else
4475                         ret = 0;
4476                 spin_unlock(&space_info->lock);
4477                 return ret;
4478         }
4479
4480         if (!should_alloc_chunk(extent_root, space_info, force)) {
4481                 spin_unlock(&space_info->lock);
4482                 return 0;
4483         } else if (space_info->chunk_alloc) {
4484                 wait_for_alloc = 1;
4485         } else {
4486                 space_info->chunk_alloc = 1;
4487         }
4488
4489         spin_unlock(&space_info->lock);
4490
4491         mutex_lock(&fs_info->chunk_mutex);
4492
4493         /*
4494          * The chunk_mutex is held throughout the entirety of a chunk
4495          * allocation, so once we've acquired the chunk_mutex we know that the
4496          * other guy is done and we need to recheck and see if we should
4497          * allocate.
4498          */
4499         if (wait_for_alloc) {
4500                 mutex_unlock(&fs_info->chunk_mutex);
4501                 wait_for_alloc = 0;
4502                 goto again;
4503         }
4504
4505         trans->allocating_chunk = true;
4506
4507         /*
4508          * If we have mixed data/metadata chunks we want to make sure we keep
4509          * allocating mixed chunks instead of individual chunks.
4510          */
4511         if (btrfs_mixed_space_info(space_info))
4512                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4513
4514         /*
4515          * if we're doing a data chunk, go ahead and make sure that
4516          * we keep a reasonable number of metadata chunks allocated in the
4517          * FS as well.
4518          */
4519         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4520                 fs_info->data_chunk_allocations++;
4521                 if (!(fs_info->data_chunk_allocations %
4522                       fs_info->metadata_ratio))
4523                         force_metadata_allocation(fs_info);
4524         }
4525
4526         /*
4527          * Check if we have enough space in SYSTEM chunk because we may need
4528          * to update devices.
4529          */
4530         check_system_chunk(trans, extent_root, flags);
4531
4532         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4533         trans->allocating_chunk = false;
4534
4535         spin_lock(&space_info->lock);
4536         if (ret < 0 && ret != -ENOSPC)
4537                 goto out;
4538         if (ret)
4539                 space_info->full = 1;
4540         else
4541                 ret = 1;
4542
4543         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4544 out:
4545         space_info->chunk_alloc = 0;
4546         spin_unlock(&space_info->lock);
4547         mutex_unlock(&fs_info->chunk_mutex);
4548         /*
4549          * When we allocate a new chunk we reserve space in the chunk block
4550          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4551          * add new nodes/leafs to it if we end up needing to do it when
4552          * inserting the chunk item and updating device items as part of the
4553          * second phase of chunk allocation, performed by
4554          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4555          * large number of new block groups to create in our transaction
4556          * handle's new_bgs list to avoid exhausting the chunk block reserve
4557          * in extreme cases - like having a single transaction create many new
4558          * block groups when starting to write out the free space caches of all
4559          * the block groups that were made dirty during the lifetime of the
4560          * transaction.
4561          */
4562         if (trans->can_flush_pending_bgs &&
4563             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4564                 btrfs_create_pending_block_groups(trans, trans->root);
4565                 btrfs_trans_release_chunk_metadata(trans);
4566         }
4567         return ret;
4568 }
4569
4570 static int can_overcommit(struct btrfs_root *root,
4571                           struct btrfs_space_info *space_info, u64 bytes,
4572                           enum btrfs_reserve_flush_enum flush)
4573 {
4574         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4575         u64 profile = btrfs_get_alloc_profile(root, 0);
4576         u64 space_size;
4577         u64 avail;
4578         u64 used;
4579
4580         used = space_info->bytes_used + space_info->bytes_reserved +
4581                 space_info->bytes_pinned + space_info->bytes_readonly;
4582
4583         /*
4584          * We only want to allow over committing if we have lots of actual space
4585          * free, but if we don't have enough space to handle the global reserve
4586          * space then we could end up having a real enospc problem when trying
4587          * to allocate a chunk or some other such important allocation.
4588          */
4589         spin_lock(&global_rsv->lock);
4590         space_size = calc_global_rsv_need_space(global_rsv);
4591         spin_unlock(&global_rsv->lock);
4592         if (used + space_size >= space_info->total_bytes)
4593                 return 0;
4594
4595         used += space_info->bytes_may_use;
4596
4597         spin_lock(&root->fs_info->free_chunk_lock);
4598         avail = root->fs_info->free_chunk_space;
4599         spin_unlock(&root->fs_info->free_chunk_lock);
4600
4601         /*
4602          * If we have dup, raid1 or raid10 then only half of the free
4603          * space is actually useable.  For raid56, the space info used
4604          * doesn't include the parity drive, so we don't have to
4605          * change the math
4606          */
4607         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4608                        BTRFS_BLOCK_GROUP_RAID1 |
4609                        BTRFS_BLOCK_GROUP_RAID10))
4610                 avail >>= 1;
4611
4612         /*
4613          * If we aren't flushing all things, let us overcommit up to
4614          * 1/2th of the space. If we can flush, don't let us overcommit
4615          * too much, let it overcommit up to 1/8 of the space.
4616          */
4617         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4618                 avail >>= 3;
4619         else
4620                 avail >>= 1;
4621
4622         if (used + bytes < space_info->total_bytes + avail)
4623                 return 1;
4624         return 0;
4625 }
4626
4627 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4628                                          unsigned long nr_pages, int nr_items)
4629 {
4630         struct super_block *sb = root->fs_info->sb;
4631
4632         if (down_read_trylock(&sb->s_umount)) {
4633                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4634                 up_read(&sb->s_umount);
4635         } else {
4636                 /*
4637                  * We needn't worry the filesystem going from r/w to r/o though
4638                  * we don't acquire ->s_umount mutex, because the filesystem
4639                  * should guarantee the delalloc inodes list be empty after
4640                  * the filesystem is readonly(all dirty pages are written to
4641                  * the disk).
4642                  */
4643                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4644                 if (!current->journal_info)
4645                         btrfs_wait_ordered_roots(root->fs_info, nr_items,
4646                                                  0, (u64)-1);
4647         }
4648 }
4649
4650 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4651 {
4652         u64 bytes;
4653         int nr;
4654
4655         bytes = btrfs_calc_trans_metadata_size(root, 1);
4656         nr = (int)div64_u64(to_reclaim, bytes);
4657         if (!nr)
4658                 nr = 1;
4659         return nr;
4660 }
4661
4662 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4663
4664 /*
4665  * shrink metadata reservation for delalloc
4666  */
4667 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4668                             bool wait_ordered)
4669 {
4670         struct btrfs_block_rsv *block_rsv;
4671         struct btrfs_space_info *space_info;
4672         struct btrfs_trans_handle *trans;
4673         u64 delalloc_bytes;
4674         u64 max_reclaim;
4675         long time_left;
4676         unsigned long nr_pages;
4677         int loops;
4678         int items;
4679         enum btrfs_reserve_flush_enum flush;
4680
4681         /* Calc the number of the pages we need flush for space reservation */
4682         items = calc_reclaim_items_nr(root, to_reclaim);
4683         to_reclaim = (u64)items * EXTENT_SIZE_PER_ITEM;
4684
4685         trans = (struct btrfs_trans_handle *)current->journal_info;
4686         block_rsv = &root->fs_info->delalloc_block_rsv;
4687         space_info = block_rsv->space_info;
4688
4689         delalloc_bytes = percpu_counter_sum_positive(
4690                                                 &root->fs_info->delalloc_bytes);
4691         if (delalloc_bytes == 0) {
4692                 if (trans)
4693                         return;
4694                 if (wait_ordered)
4695                         btrfs_wait_ordered_roots(root->fs_info, items,
4696                                                  0, (u64)-1);
4697                 return;
4698         }
4699
4700         loops = 0;
4701         while (delalloc_bytes && loops < 3) {
4702                 max_reclaim = min(delalloc_bytes, to_reclaim);
4703                 nr_pages = max_reclaim >> PAGE_SHIFT;
4704                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4705                 /*
4706                  * We need to wait for the async pages to actually start before
4707                  * we do anything.
4708                  */
4709                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4710                 if (!max_reclaim)
4711                         goto skip_async;
4712
4713                 if (max_reclaim <= nr_pages)
4714                         max_reclaim = 0;
4715                 else
4716                         max_reclaim -= nr_pages;
4717
4718                 wait_event(root->fs_info->async_submit_wait,
4719                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4720                            (int)max_reclaim);
4721 skip_async:
4722                 if (!trans)
4723                         flush = BTRFS_RESERVE_FLUSH_ALL;
4724                 else
4725                         flush = BTRFS_RESERVE_NO_FLUSH;
4726                 spin_lock(&space_info->lock);
4727                 if (can_overcommit(root, space_info, orig, flush)) {
4728                         spin_unlock(&space_info->lock);
4729                         break;
4730                 }
4731                 spin_unlock(&space_info->lock);
4732
4733                 loops++;
4734                 if (wait_ordered && !trans) {
4735                         btrfs_wait_ordered_roots(root->fs_info, items,
4736                                                  0, (u64)-1);
4737                 } else {
4738                         time_left = schedule_timeout_killable(1);
4739                         if (time_left)
4740                                 break;
4741                 }
4742                 delalloc_bytes = percpu_counter_sum_positive(
4743                                                 &root->fs_info->delalloc_bytes);
4744         }
4745 }
4746
4747 /**
4748  * maybe_commit_transaction - possibly commit the transaction if its ok to
4749  * @root - the root we're allocating for
4750  * @bytes - the number of bytes we want to reserve
4751  * @force - force the commit
4752  *
4753  * This will check to make sure that committing the transaction will actually
4754  * get us somewhere and then commit the transaction if it does.  Otherwise it
4755  * will return -ENOSPC.
4756  */
4757 static int may_commit_transaction(struct btrfs_root *root,
4758                                   struct btrfs_space_info *space_info,
4759                                   u64 bytes, int force)
4760 {
4761         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4762         struct btrfs_trans_handle *trans;
4763
4764         trans = (struct btrfs_trans_handle *)current->journal_info;
4765         if (trans)
4766                 return -EAGAIN;
4767
4768         if (force)
4769                 goto commit;
4770
4771         /* See if there is enough pinned space to make this reservation */
4772         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4773                                    bytes) >= 0)
4774                 goto commit;
4775
4776         /*
4777          * See if there is some space in the delayed insertion reservation for
4778          * this reservation.
4779          */
4780         if (space_info != delayed_rsv->space_info)
4781                 return -ENOSPC;
4782
4783         spin_lock(&delayed_rsv->lock);
4784         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4785                                    bytes - delayed_rsv->size) >= 0) {
4786                 spin_unlock(&delayed_rsv->lock);
4787                 return -ENOSPC;
4788         }
4789         spin_unlock(&delayed_rsv->lock);
4790
4791 commit:
4792         trans = btrfs_join_transaction(root);
4793         if (IS_ERR(trans))
4794                 return -ENOSPC;
4795
4796         return btrfs_commit_transaction(trans, root);
4797 }
4798
4799 enum flush_state {
4800         FLUSH_DELAYED_ITEMS_NR  =       1,
4801         FLUSH_DELAYED_ITEMS     =       2,
4802         FLUSH_DELALLOC          =       3,
4803         FLUSH_DELALLOC_WAIT     =       4,
4804         ALLOC_CHUNK             =       5,
4805         COMMIT_TRANS            =       6,
4806 };
4807
4808 static int flush_space(struct btrfs_root *root,
4809                        struct btrfs_space_info *space_info, u64 num_bytes,
4810                        u64 orig_bytes, int state)
4811 {
4812         struct btrfs_trans_handle *trans;
4813         int nr;
4814         int ret = 0;
4815
4816         switch (state) {
4817         case FLUSH_DELAYED_ITEMS_NR:
4818         case FLUSH_DELAYED_ITEMS:
4819                 if (state == FLUSH_DELAYED_ITEMS_NR)
4820                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4821                 else
4822                         nr = -1;
4823
4824                 trans = btrfs_join_transaction(root);
4825                 if (IS_ERR(trans)) {
4826                         ret = PTR_ERR(trans);
4827                         break;
4828                 }
4829                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4830                 btrfs_end_transaction(trans, root);
4831                 break;
4832         case FLUSH_DELALLOC:
4833         case FLUSH_DELALLOC_WAIT:
4834                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4835                                 state == FLUSH_DELALLOC_WAIT);
4836                 break;
4837         case ALLOC_CHUNK:
4838                 trans = btrfs_join_transaction(root);
4839                 if (IS_ERR(trans)) {
4840                         ret = PTR_ERR(trans);
4841                         break;
4842                 }
4843                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4844                                      btrfs_get_alloc_profile(root, 0),
4845                                      CHUNK_ALLOC_NO_FORCE);
4846                 btrfs_end_transaction(trans, root);
4847                 if (ret == -ENOSPC)
4848                         ret = 0;
4849                 break;
4850         case COMMIT_TRANS:
4851                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4852                 break;
4853         default:
4854                 ret = -ENOSPC;
4855                 break;
4856         }
4857
4858         return ret;
4859 }
4860
4861 static inline u64
4862 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4863                                  struct btrfs_space_info *space_info)
4864 {
4865         u64 used;
4866         u64 expected;
4867         u64 to_reclaim;
4868
4869         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4870         spin_lock(&space_info->lock);
4871         if (can_overcommit(root, space_info, to_reclaim,
4872                            BTRFS_RESERVE_FLUSH_ALL)) {
4873                 to_reclaim = 0;
4874                 goto out;
4875         }
4876
4877         used = space_info->bytes_used + space_info->bytes_reserved +
4878                space_info->bytes_pinned + space_info->bytes_readonly +
4879                space_info->bytes_may_use;
4880         if (can_overcommit(root, space_info, SZ_1M, BTRFS_RESERVE_FLUSH_ALL))
4881                 expected = div_factor_fine(space_info->total_bytes, 95);
4882         else
4883                 expected = div_factor_fine(space_info->total_bytes, 90);
4884
4885         if (used > expected)
4886                 to_reclaim = used - expected;
4887         else
4888                 to_reclaim = 0;
4889         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4890                                      space_info->bytes_reserved);
4891 out:
4892         spin_unlock(&space_info->lock);
4893
4894         return to_reclaim;
4895 }
4896
4897 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4898                                         struct btrfs_fs_info *fs_info, u64 used)
4899 {
4900         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4901
4902         /* If we're just plain full then async reclaim just slows us down. */
4903         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4904                 return 0;
4905
4906         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4907                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4908 }
4909
4910 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4911                                        struct btrfs_fs_info *fs_info,
4912                                        int flush_state)
4913 {
4914         u64 used;
4915
4916         spin_lock(&space_info->lock);
4917         /*
4918          * We run out of space and have not got any free space via flush_space,
4919          * so don't bother doing async reclaim.
4920          */
4921         if (flush_state > COMMIT_TRANS && space_info->full) {
4922                 spin_unlock(&space_info->lock);
4923                 return 0;
4924         }
4925
4926         used = space_info->bytes_used + space_info->bytes_reserved +
4927                space_info->bytes_pinned + space_info->bytes_readonly +
4928                space_info->bytes_may_use;
4929         if (need_do_async_reclaim(space_info, fs_info, used)) {
4930                 spin_unlock(&space_info->lock);
4931                 return 1;
4932         }
4933         spin_unlock(&space_info->lock);
4934
4935         return 0;
4936 }
4937
4938 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4939 {
4940         struct btrfs_fs_info *fs_info;
4941         struct btrfs_space_info *space_info;
4942         u64 to_reclaim;
4943         int flush_state;
4944
4945         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4946         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4947
4948         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4949                                                       space_info);
4950         if (!to_reclaim)
4951                 return;
4952
4953         flush_state = FLUSH_DELAYED_ITEMS_NR;
4954         do {
4955                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4956                             to_reclaim, flush_state);
4957                 flush_state++;
4958                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4959                                                  flush_state))
4960                         return;
4961         } while (flush_state < COMMIT_TRANS);
4962 }
4963
4964 void btrfs_init_async_reclaim_work(struct work_struct *work)
4965 {
4966         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4967 }
4968
4969 /**
4970  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4971  * @root - the root we're allocating for
4972  * @block_rsv - the block_rsv we're allocating for
4973  * @orig_bytes - the number of bytes we want
4974  * @flush - whether or not we can flush to make our reservation
4975  *
4976  * This will reserve orig_bytes number of bytes from the space info associated
4977  * with the block_rsv.  If there is not enough space it will make an attempt to
4978  * flush out space to make room.  It will do this by flushing delalloc if
4979  * possible or committing the transaction.  If flush is 0 then no attempts to
4980  * regain reservations will be made and this will fail if there is not enough
4981  * space already.
4982  */
4983 static int reserve_metadata_bytes(struct btrfs_root *root,
4984                                   struct btrfs_block_rsv *block_rsv,
4985                                   u64 orig_bytes,
4986                                   enum btrfs_reserve_flush_enum flush)
4987 {
4988         struct btrfs_space_info *space_info = block_rsv->space_info;
4989         u64 used;
4990         u64 num_bytes = orig_bytes;
4991         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4992         int ret = 0;
4993         bool flushing = false;
4994
4995 again:
4996         ret = 0;
4997         spin_lock(&space_info->lock);
4998         /*
4999          * We only want to wait if somebody other than us is flushing and we
5000          * are actually allowed to flush all things.
5001          */
5002         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
5003                space_info->flush) {
5004                 spin_unlock(&space_info->lock);
5005                 /*
5006                  * If we have a trans handle we can't wait because the flusher
5007                  * may have to commit the transaction, which would mean we would
5008                  * deadlock since we are waiting for the flusher to finish, but
5009                  * hold the current transaction open.
5010                  */
5011                 if (current->journal_info)
5012                         return -EAGAIN;
5013                 ret = wait_event_killable(space_info->wait, !space_info->flush);
5014                 /* Must have been killed, return */
5015                 if (ret)
5016                         return -EINTR;
5017
5018                 spin_lock(&space_info->lock);
5019         }
5020
5021         ret = -ENOSPC;
5022         used = space_info->bytes_used + space_info->bytes_reserved +
5023                 space_info->bytes_pinned + space_info->bytes_readonly +
5024                 space_info->bytes_may_use;
5025
5026         /*
5027          * The idea here is that we've not already over-reserved the block group
5028          * then we can go ahead and save our reservation first and then start
5029          * flushing if we need to.  Otherwise if we've already overcommitted
5030          * lets start flushing stuff first and then come back and try to make
5031          * our reservation.
5032          */
5033         if (used <= space_info->total_bytes) {
5034                 if (used + orig_bytes <= space_info->total_bytes) {
5035                         space_info->bytes_may_use += orig_bytes;
5036                         trace_btrfs_space_reservation(root->fs_info,
5037                                 "space_info", space_info->flags, orig_bytes, 1);
5038                         ret = 0;
5039                 } else {
5040                         /*
5041                          * Ok set num_bytes to orig_bytes since we aren't
5042                          * overocmmitted, this way we only try and reclaim what
5043                          * we need.
5044                          */
5045                         num_bytes = orig_bytes;
5046                 }
5047         } else {
5048                 /*
5049                  * Ok we're over committed, set num_bytes to the overcommitted
5050                  * amount plus the amount of bytes that we need for this
5051                  * reservation.
5052                  */
5053                 num_bytes = used - space_info->total_bytes +
5054                         (orig_bytes * 2);
5055         }
5056
5057         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
5058                 space_info->bytes_may_use += orig_bytes;
5059                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5060                                               space_info->flags, orig_bytes,
5061                                               1);
5062                 ret = 0;
5063         }
5064
5065         /*
5066          * Couldn't make our reservation, save our place so while we're trying
5067          * to reclaim space we can actually use it instead of somebody else
5068          * stealing it from us.
5069          *
5070          * We make the other tasks wait for the flush only when we can flush
5071          * all things.
5072          */
5073         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5074                 flushing = true;
5075                 space_info->flush = 1;
5076         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5077                 used += orig_bytes;
5078                 /*
5079                  * We will do the space reservation dance during log replay,
5080                  * which means we won't have fs_info->fs_root set, so don't do
5081                  * the async reclaim as we will panic.
5082                  */
5083                 if (!root->fs_info->log_root_recovering &&
5084                     need_do_async_reclaim(space_info, root->fs_info, used) &&
5085                     !work_busy(&root->fs_info->async_reclaim_work))
5086                         queue_work(system_unbound_wq,
5087                                    &root->fs_info->async_reclaim_work);
5088         }
5089         spin_unlock(&space_info->lock);
5090
5091         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5092                 goto out;
5093
5094         ret = flush_space(root, space_info, num_bytes, orig_bytes,
5095                           flush_state);
5096         flush_state++;
5097
5098         /*
5099          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
5100          * would happen. So skip delalloc flush.
5101          */
5102         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
5103             (flush_state == FLUSH_DELALLOC ||
5104              flush_state == FLUSH_DELALLOC_WAIT))
5105                 flush_state = ALLOC_CHUNK;
5106
5107         if (!ret)
5108                 goto again;
5109         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
5110                  flush_state < COMMIT_TRANS)
5111                 goto again;
5112         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
5113                  flush_state <= COMMIT_TRANS)
5114                 goto again;
5115
5116 out:
5117         if (ret == -ENOSPC &&
5118             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5119                 struct btrfs_block_rsv *global_rsv =
5120                         &root->fs_info->global_block_rsv;
5121
5122                 if (block_rsv != global_rsv &&
5123                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5124                         ret = 0;
5125         }
5126         if (ret == -ENOSPC)
5127                 trace_btrfs_space_reservation(root->fs_info,
5128                                               "space_info:enospc",
5129                                               space_info->flags, orig_bytes, 1);
5130         if (flushing) {
5131                 spin_lock(&space_info->lock);
5132                 space_info->flush = 0;
5133                 wake_up_all(&space_info->wait);
5134                 spin_unlock(&space_info->lock);
5135         }
5136         return ret;
5137 }
5138
5139 static struct btrfs_block_rsv *get_block_rsv(
5140                                         const struct btrfs_trans_handle *trans,
5141                                         const struct btrfs_root *root)
5142 {
5143         struct btrfs_block_rsv *block_rsv = NULL;
5144
5145         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5146             (root == root->fs_info->csum_root && trans->adding_csums) ||
5147              (root == root->fs_info->uuid_root))
5148                 block_rsv = trans->block_rsv;
5149
5150         if (!block_rsv)
5151                 block_rsv = root->block_rsv;
5152
5153         if (!block_rsv)
5154                 block_rsv = &root->fs_info->empty_block_rsv;
5155
5156         return block_rsv;
5157 }
5158
5159 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5160                                u64 num_bytes)
5161 {
5162         int ret = -ENOSPC;
5163         spin_lock(&block_rsv->lock);
5164         if (block_rsv->reserved >= num_bytes) {
5165                 block_rsv->reserved -= num_bytes;
5166                 if (block_rsv->reserved < block_rsv->size)
5167                         block_rsv->full = 0;
5168                 ret = 0;
5169         }
5170         spin_unlock(&block_rsv->lock);
5171         return ret;
5172 }
5173
5174 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5175                                 u64 num_bytes, int update_size)
5176 {
5177         spin_lock(&block_rsv->lock);
5178         block_rsv->reserved += num_bytes;
5179         if (update_size)
5180                 block_rsv->size += num_bytes;
5181         else if (block_rsv->reserved >= block_rsv->size)
5182                 block_rsv->full = 1;
5183         spin_unlock(&block_rsv->lock);
5184 }
5185
5186 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5187                              struct btrfs_block_rsv *dest, u64 num_bytes,
5188                              int min_factor)
5189 {
5190         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5191         u64 min_bytes;
5192
5193         if (global_rsv->space_info != dest->space_info)
5194                 return -ENOSPC;
5195
5196         spin_lock(&global_rsv->lock);
5197         min_bytes = div_factor(global_rsv->size, min_factor);
5198         if (global_rsv->reserved < min_bytes + num_bytes) {
5199                 spin_unlock(&global_rsv->lock);
5200                 return -ENOSPC;
5201         }
5202         global_rsv->reserved -= num_bytes;
5203         if (global_rsv->reserved < global_rsv->size)
5204                 global_rsv->full = 0;
5205         spin_unlock(&global_rsv->lock);
5206
5207         block_rsv_add_bytes(dest, num_bytes, 1);
5208         return 0;
5209 }
5210
5211 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5212                                     struct btrfs_block_rsv *block_rsv,
5213                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5214 {
5215         struct btrfs_space_info *space_info = block_rsv->space_info;
5216
5217         spin_lock(&block_rsv->lock);
5218         if (num_bytes == (u64)-1)
5219                 num_bytes = block_rsv->size;
5220         block_rsv->size -= num_bytes;
5221         if (block_rsv->reserved >= block_rsv->size) {
5222                 num_bytes = block_rsv->reserved - block_rsv->size;
5223                 block_rsv->reserved = block_rsv->size;
5224                 block_rsv->full = 1;
5225         } else {
5226                 num_bytes = 0;
5227         }
5228         spin_unlock(&block_rsv->lock);
5229
5230         if (num_bytes > 0) {
5231                 if (dest) {
5232                         spin_lock(&dest->lock);
5233                         if (!dest->full) {
5234                                 u64 bytes_to_add;
5235
5236                                 bytes_to_add = dest->size - dest->reserved;
5237                                 bytes_to_add = min(num_bytes, bytes_to_add);
5238                                 dest->reserved += bytes_to_add;
5239                                 if (dest->reserved >= dest->size)
5240                                         dest->full = 1;
5241                                 num_bytes -= bytes_to_add;
5242                         }
5243                         spin_unlock(&dest->lock);
5244                 }
5245                 if (num_bytes) {
5246                         spin_lock(&space_info->lock);
5247                         space_info->bytes_may_use -= num_bytes;
5248                         trace_btrfs_space_reservation(fs_info, "space_info",
5249                                         space_info->flags, num_bytes, 0);
5250                         spin_unlock(&space_info->lock);
5251                 }
5252         }
5253 }
5254
5255 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
5256                                    struct btrfs_block_rsv *dst, u64 num_bytes)
5257 {
5258         int ret;
5259
5260         ret = block_rsv_use_bytes(src, num_bytes);
5261         if (ret)
5262                 return ret;
5263
5264         block_rsv_add_bytes(dst, num_bytes, 1);
5265         return 0;
5266 }
5267
5268 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5269 {
5270         memset(rsv, 0, sizeof(*rsv));
5271         spin_lock_init(&rsv->lock);
5272         rsv->type = type;
5273 }
5274
5275 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5276                                               unsigned short type)
5277 {
5278         struct btrfs_block_rsv *block_rsv;
5279         struct btrfs_fs_info *fs_info = root->fs_info;
5280
5281         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5282         if (!block_rsv)
5283                 return NULL;
5284
5285         btrfs_init_block_rsv(block_rsv, type);
5286         block_rsv->space_info = __find_space_info(fs_info,
5287                                                   BTRFS_BLOCK_GROUP_METADATA);
5288         return block_rsv;
5289 }
5290
5291 void btrfs_free_block_rsv(struct btrfs_root *root,
5292                           struct btrfs_block_rsv *rsv)
5293 {
5294         if (!rsv)
5295                 return;
5296         btrfs_block_rsv_release(root, rsv, (u64)-1);
5297         kfree(rsv);
5298 }
5299
5300 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5301 {
5302         kfree(rsv);
5303 }
5304
5305 int btrfs_block_rsv_add(struct btrfs_root *root,
5306                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5307                         enum btrfs_reserve_flush_enum flush)
5308 {
5309         int ret;
5310
5311         if (num_bytes == 0)
5312                 return 0;
5313
5314         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5315         if (!ret) {
5316                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5317                 return 0;
5318         }
5319
5320         return ret;
5321 }
5322
5323 int btrfs_block_rsv_check(struct btrfs_root *root,
5324                           struct btrfs_block_rsv *block_rsv, int min_factor)
5325 {
5326         u64 num_bytes = 0;
5327         int ret = -ENOSPC;
5328
5329         if (!block_rsv)
5330                 return 0;
5331
5332         spin_lock(&block_rsv->lock);
5333         num_bytes = div_factor(block_rsv->size, min_factor);
5334         if (block_rsv->reserved >= num_bytes)
5335                 ret = 0;
5336         spin_unlock(&block_rsv->lock);
5337
5338         return ret;
5339 }
5340
5341 int btrfs_block_rsv_refill(struct btrfs_root *root,
5342                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5343                            enum btrfs_reserve_flush_enum flush)
5344 {
5345         u64 num_bytes = 0;
5346         int ret = -ENOSPC;
5347
5348         if (!block_rsv)
5349                 return 0;
5350
5351         spin_lock(&block_rsv->lock);
5352         num_bytes = min_reserved;
5353         if (block_rsv->reserved >= num_bytes)
5354                 ret = 0;
5355         else
5356                 num_bytes -= block_rsv->reserved;
5357         spin_unlock(&block_rsv->lock);
5358
5359         if (!ret)
5360                 return 0;
5361
5362         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5363         if (!ret) {
5364                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5365                 return 0;
5366         }
5367
5368         return ret;
5369 }
5370
5371 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5372                             struct btrfs_block_rsv *dst_rsv,
5373                             u64 num_bytes)
5374 {
5375         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5376 }
5377
5378 void btrfs_block_rsv_release(struct btrfs_root *root,
5379                              struct btrfs_block_rsv *block_rsv,
5380                              u64 num_bytes)
5381 {
5382         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5383         if (global_rsv == block_rsv ||
5384             block_rsv->space_info != global_rsv->space_info)
5385                 global_rsv = NULL;
5386         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5387                                 num_bytes);
5388 }
5389
5390 /*
5391  * helper to calculate size of global block reservation.
5392  * the desired value is sum of space used by extent tree,
5393  * checksum tree and root tree
5394  */
5395 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5396 {
5397         struct btrfs_space_info *sinfo;
5398         u64 num_bytes;
5399         u64 meta_used;
5400         u64 data_used;
5401         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5402
5403         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5404         spin_lock(&sinfo->lock);
5405         data_used = sinfo->bytes_used;
5406         spin_unlock(&sinfo->lock);
5407
5408         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5409         spin_lock(&sinfo->lock);
5410         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5411                 data_used = 0;
5412         meta_used = sinfo->bytes_used;
5413         spin_unlock(&sinfo->lock);
5414
5415         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5416                     csum_size * 2;
5417         num_bytes += div_u64(data_used + meta_used, 50);
5418
5419         if (num_bytes * 3 > meta_used)
5420                 num_bytes = div_u64(meta_used, 3);
5421
5422         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5423 }
5424
5425 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5426 {
5427         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5428         struct btrfs_space_info *sinfo = block_rsv->space_info;
5429         u64 num_bytes;
5430
5431         num_bytes = calc_global_metadata_size(fs_info);
5432
5433         spin_lock(&sinfo->lock);
5434         spin_lock(&block_rsv->lock);
5435
5436         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5437
5438         if (block_rsv->reserved < block_rsv->size) {
5439                 num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5440                         sinfo->bytes_reserved + sinfo->bytes_readonly +
5441                         sinfo->bytes_may_use;
5442                 if (sinfo->total_bytes > num_bytes) {
5443                         num_bytes = sinfo->total_bytes - num_bytes;
5444                         num_bytes = min(num_bytes,
5445                                         block_rsv->size - block_rsv->reserved);
5446                         block_rsv->reserved += num_bytes;
5447                         sinfo->bytes_may_use += num_bytes;
5448                         trace_btrfs_space_reservation(fs_info, "space_info",
5449                                                       sinfo->flags, num_bytes,
5450                                                       1);
5451                 }
5452         } else if (block_rsv->reserved > block_rsv->size) {
5453                 num_bytes = block_rsv->reserved - block_rsv->size;
5454                 sinfo->bytes_may_use -= num_bytes;
5455                 trace_btrfs_space_reservation(fs_info, "space_info",
5456                                       sinfo->flags, num_bytes, 0);
5457                 block_rsv->reserved = block_rsv->size;
5458         }
5459
5460         if (block_rsv->reserved == block_rsv->size)
5461                 block_rsv->full = 1;
5462         else
5463                 block_rsv->full = 0;
5464
5465         spin_unlock(&block_rsv->lock);
5466         spin_unlock(&sinfo->lock);
5467 }
5468
5469 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5470 {
5471         struct btrfs_space_info *space_info;
5472
5473         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5474         fs_info->chunk_block_rsv.space_info = space_info;
5475
5476         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5477         fs_info->global_block_rsv.space_info = space_info;
5478         fs_info->delalloc_block_rsv.space_info = space_info;
5479         fs_info->trans_block_rsv.space_info = space_info;
5480         fs_info->empty_block_rsv.space_info = space_info;
5481         fs_info->delayed_block_rsv.space_info = space_info;
5482
5483         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5484         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5485         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5486         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5487         if (fs_info->quota_root)
5488                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5489         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5490
5491         update_global_block_rsv(fs_info);
5492 }
5493
5494 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5495 {
5496         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5497                                 (u64)-1);
5498         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5499         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5500         WARN_ON(fs_info->trans_block_rsv.size > 0);
5501         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5502         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5503         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5504         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5505         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5506 }
5507
5508 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5509                                   struct btrfs_root *root)
5510 {
5511         if (!trans->block_rsv)
5512                 return;
5513
5514         if (!trans->bytes_reserved)
5515                 return;
5516
5517         trace_btrfs_space_reservation(root->fs_info, "transaction",
5518                                       trans->transid, trans->bytes_reserved, 0);
5519         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5520         trans->bytes_reserved = 0;
5521 }
5522
5523 /*
5524  * To be called after all the new block groups attached to the transaction
5525  * handle have been created (btrfs_create_pending_block_groups()).
5526  */
5527 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5528 {
5529         struct btrfs_fs_info *fs_info = trans->root->fs_info;
5530
5531         if (!trans->chunk_bytes_reserved)
5532                 return;
5533
5534         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5535
5536         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5537                                 trans->chunk_bytes_reserved);
5538         trans->chunk_bytes_reserved = 0;
5539 }
5540
5541 /* Can only return 0 or -ENOSPC */
5542 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5543                                   struct inode *inode)
5544 {
5545         struct btrfs_root *root = BTRFS_I(inode)->root;
5546         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5547         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5548
5549         /*
5550          * We need to hold space in order to delete our orphan item once we've
5551          * added it, so this takes the reservation so we can release it later
5552          * when we are truly done with the orphan item.
5553          */
5554         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5555         trace_btrfs_space_reservation(root->fs_info, "orphan",
5556                                       btrfs_ino(inode), num_bytes, 1);
5557         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5558 }
5559
5560 void btrfs_orphan_release_metadata(struct inode *inode)
5561 {
5562         struct btrfs_root *root = BTRFS_I(inode)->root;
5563         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5564         trace_btrfs_space_reservation(root->fs_info, "orphan",
5565                                       btrfs_ino(inode), num_bytes, 0);
5566         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5567 }
5568
5569 /*
5570  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5571  * root: the root of the parent directory
5572  * rsv: block reservation
5573  * items: the number of items that we need do reservation
5574  * qgroup_reserved: used to return the reserved size in qgroup
5575  *
5576  * This function is used to reserve the space for snapshot/subvolume
5577  * creation and deletion. Those operations are different with the
5578  * common file/directory operations, they change two fs/file trees
5579  * and root tree, the number of items that the qgroup reserves is
5580  * different with the free space reservation. So we can not use
5581  * the space reservation mechanism in start_transaction().
5582  */
5583 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5584                                      struct btrfs_block_rsv *rsv,
5585                                      int items,
5586                                      u64 *qgroup_reserved,
5587                                      bool use_global_rsv)
5588 {
5589         u64 num_bytes;
5590         int ret;
5591         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5592
5593         if (root->fs_info->quota_enabled) {
5594                 /* One for parent inode, two for dir entries */
5595                 num_bytes = 3 * root->nodesize;
5596                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5597                 if (ret)
5598                         return ret;
5599         } else {
5600                 num_bytes = 0;
5601         }
5602
5603         *qgroup_reserved = num_bytes;
5604
5605         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5606         rsv->space_info = __find_space_info(root->fs_info,
5607                                             BTRFS_BLOCK_GROUP_METADATA);
5608         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5609                                   BTRFS_RESERVE_FLUSH_ALL);
5610
5611         if (ret == -ENOSPC && use_global_rsv)
5612                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5613
5614         if (ret && *qgroup_reserved)
5615                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5616
5617         return ret;
5618 }
5619
5620 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5621                                       struct btrfs_block_rsv *rsv,
5622                                       u64 qgroup_reserved)
5623 {
5624         btrfs_block_rsv_release(root, rsv, (u64)-1);
5625 }
5626
5627 /**
5628  * drop_outstanding_extent - drop an outstanding extent
5629  * @inode: the inode we're dropping the extent for
5630  * @num_bytes: the number of bytes we're releasing.
5631  *
5632  * This is called when we are freeing up an outstanding extent, either called
5633  * after an error or after an extent is written.  This will return the number of
5634  * reserved extents that need to be freed.  This must be called with
5635  * BTRFS_I(inode)->lock held.
5636  */
5637 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5638 {
5639         unsigned drop_inode_space = 0;
5640         unsigned dropped_extents = 0;
5641         unsigned num_extents = 0;
5642
5643         num_extents = (unsigned)div64_u64(num_bytes +
5644                                           BTRFS_MAX_EXTENT_SIZE - 1,
5645                                           BTRFS_MAX_EXTENT_SIZE);
5646         ASSERT(num_extents);
5647         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5648         BTRFS_I(inode)->outstanding_extents -= num_extents;
5649
5650         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5651             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5652                                &BTRFS_I(inode)->runtime_flags))
5653                 drop_inode_space = 1;
5654
5655         /*
5656          * If we have more or the same amount of outstanding extents than we have
5657          * reserved then we need to leave the reserved extents count alone.
5658          */
5659         if (BTRFS_I(inode)->outstanding_extents >=
5660             BTRFS_I(inode)->reserved_extents)
5661                 return drop_inode_space;
5662
5663         dropped_extents = BTRFS_I(inode)->reserved_extents -
5664                 BTRFS_I(inode)->outstanding_extents;
5665         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5666         return dropped_extents + drop_inode_space;
5667 }
5668
5669 /**
5670  * calc_csum_metadata_size - return the amount of metadata space that must be
5671  *      reserved/freed for the given bytes.
5672  * @inode: the inode we're manipulating
5673  * @num_bytes: the number of bytes in question
5674  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5675  *
5676  * This adjusts the number of csum_bytes in the inode and then returns the
5677  * correct amount of metadata that must either be reserved or freed.  We
5678  * calculate how many checksums we can fit into one leaf and then divide the
5679  * number of bytes that will need to be checksumed by this value to figure out
5680  * how many checksums will be required.  If we are adding bytes then the number
5681  * may go up and we will return the number of additional bytes that must be
5682  * reserved.  If it is going down we will return the number of bytes that must
5683  * be freed.
5684  *
5685  * This must be called with BTRFS_I(inode)->lock held.
5686  */
5687 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5688                                    int reserve)
5689 {
5690         struct btrfs_root *root = BTRFS_I(inode)->root;
5691         u64 old_csums, num_csums;
5692
5693         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5694             BTRFS_I(inode)->csum_bytes == 0)
5695                 return 0;
5696
5697         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5698         if (reserve)
5699                 BTRFS_I(inode)->csum_bytes += num_bytes;
5700         else
5701                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5702         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5703
5704         /* No change, no need to reserve more */
5705         if (old_csums == num_csums)
5706                 return 0;
5707
5708         if (reserve)
5709                 return btrfs_calc_trans_metadata_size(root,
5710                                                       num_csums - old_csums);
5711
5712         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5713 }
5714
5715 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5716 {
5717         struct btrfs_root *root = BTRFS_I(inode)->root;
5718         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5719         u64 to_reserve = 0;
5720         u64 csum_bytes;
5721         unsigned nr_extents = 0;
5722         int extra_reserve = 0;
5723         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5724         int ret = 0;
5725         bool delalloc_lock = true;
5726         u64 to_free = 0;
5727         unsigned dropped;
5728
5729         /* If we are a free space inode we need to not flush since we will be in
5730          * the middle of a transaction commit.  We also don't need the delalloc
5731          * mutex since we won't race with anybody.  We need this mostly to make
5732          * lockdep shut its filthy mouth.
5733          */
5734         if (btrfs_is_free_space_inode(inode)) {
5735                 flush = BTRFS_RESERVE_NO_FLUSH;
5736                 delalloc_lock = false;
5737         }
5738
5739         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5740             btrfs_transaction_in_commit(root->fs_info))
5741                 schedule_timeout(1);
5742
5743         if (delalloc_lock)
5744                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5745
5746         num_bytes = ALIGN(num_bytes, root->sectorsize);
5747
5748         spin_lock(&BTRFS_I(inode)->lock);
5749         nr_extents = (unsigned)div64_u64(num_bytes +
5750                                          BTRFS_MAX_EXTENT_SIZE - 1,
5751                                          BTRFS_MAX_EXTENT_SIZE);
5752         BTRFS_I(inode)->outstanding_extents += nr_extents;
5753         nr_extents = 0;
5754
5755         if (BTRFS_I(inode)->outstanding_extents >
5756             BTRFS_I(inode)->reserved_extents)
5757                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5758                         BTRFS_I(inode)->reserved_extents;
5759
5760         /*
5761          * Add an item to reserve for updating the inode when we complete the
5762          * delalloc io.
5763          */
5764         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5765                       &BTRFS_I(inode)->runtime_flags)) {
5766                 nr_extents++;
5767                 extra_reserve = 1;
5768         }
5769
5770         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5771         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5772         csum_bytes = BTRFS_I(inode)->csum_bytes;
5773         spin_unlock(&BTRFS_I(inode)->lock);
5774
5775         if (root->fs_info->quota_enabled) {
5776                 ret = btrfs_qgroup_reserve_meta(root,
5777                                 nr_extents * root->nodesize);
5778                 if (ret)
5779                         goto out_fail;
5780         }
5781
5782         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5783         if (unlikely(ret)) {
5784                 btrfs_qgroup_free_meta(root, nr_extents * root->nodesize);
5785                 goto out_fail;
5786         }
5787
5788         spin_lock(&BTRFS_I(inode)->lock);
5789         if (extra_reserve) {
5790                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5791                         &BTRFS_I(inode)->runtime_flags);
5792                 nr_extents--;
5793         }
5794         BTRFS_I(inode)->reserved_extents += nr_extents;
5795         spin_unlock(&BTRFS_I(inode)->lock);
5796
5797         if (delalloc_lock)
5798                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5799
5800         if (to_reserve)
5801                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5802                                               btrfs_ino(inode), to_reserve, 1);
5803         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5804
5805         return 0;
5806
5807 out_fail:
5808         spin_lock(&BTRFS_I(inode)->lock);
5809         dropped = drop_outstanding_extent(inode, num_bytes);
5810         /*
5811          * If the inodes csum_bytes is the same as the original
5812          * csum_bytes then we know we haven't raced with any free()ers
5813          * so we can just reduce our inodes csum bytes and carry on.
5814          */
5815         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5816                 calc_csum_metadata_size(inode, num_bytes, 0);
5817         } else {
5818                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5819                 u64 bytes;
5820
5821                 /*
5822                  * This is tricky, but first we need to figure out how much we
5823                  * freed from any free-ers that occurred during this
5824                  * reservation, so we reset ->csum_bytes to the csum_bytes
5825                  * before we dropped our lock, and then call the free for the
5826                  * number of bytes that were freed while we were trying our
5827                  * reservation.
5828                  */
5829                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5830                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5831                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5832
5833
5834                 /*
5835                  * Now we need to see how much we would have freed had we not
5836                  * been making this reservation and our ->csum_bytes were not
5837                  * artificially inflated.
5838                  */
5839                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5840                 bytes = csum_bytes - orig_csum_bytes;
5841                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5842
5843                 /*
5844                  * Now reset ->csum_bytes to what it should be.  If bytes is
5845                  * more than to_free then we would have freed more space had we
5846                  * not had an artificially high ->csum_bytes, so we need to free
5847                  * the remainder.  If bytes is the same or less then we don't
5848                  * need to do anything, the other free-ers did the correct
5849                  * thing.
5850                  */
5851                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5852                 if (bytes > to_free)
5853                         to_free = bytes - to_free;
5854                 else
5855                         to_free = 0;
5856         }
5857         spin_unlock(&BTRFS_I(inode)->lock);
5858         if (dropped)
5859                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5860
5861         if (to_free) {
5862                 btrfs_block_rsv_release(root, block_rsv, to_free);
5863                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5864                                               btrfs_ino(inode), to_free, 0);
5865         }
5866         if (delalloc_lock)
5867                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5868         return ret;
5869 }
5870
5871 /**
5872  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5873  * @inode: the inode to release the reservation for
5874  * @num_bytes: the number of bytes we're releasing
5875  *
5876  * This will release the metadata reservation for an inode.  This can be called
5877  * once we complete IO for a given set of bytes to release their metadata
5878  * reservations.
5879  */
5880 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5881 {
5882         struct btrfs_root *root = BTRFS_I(inode)->root;
5883         u64 to_free = 0;
5884         unsigned dropped;
5885
5886         num_bytes = ALIGN(num_bytes, root->sectorsize);
5887         spin_lock(&BTRFS_I(inode)->lock);
5888         dropped = drop_outstanding_extent(inode, num_bytes);
5889
5890         if (num_bytes)
5891                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5892         spin_unlock(&BTRFS_I(inode)->lock);
5893         if (dropped > 0)
5894                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5895
5896         if (btrfs_test_is_dummy_root(root))
5897                 return;
5898
5899         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5900                                       btrfs_ino(inode), to_free, 0);
5901
5902         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5903                                 to_free);
5904 }
5905
5906 /**
5907  * btrfs_delalloc_reserve_space - reserve data and metadata space for
5908  * delalloc
5909  * @inode: inode we're writing to
5910  * @start: start range we are writing to
5911  * @len: how long the range we are writing to
5912  *
5913  * TODO: This function will finally replace old btrfs_delalloc_reserve_space()
5914  *
5915  * This will do the following things
5916  *
5917  * o reserve space in data space info for num bytes
5918  *   and reserve precious corresponding qgroup space
5919  *   (Done in check_data_free_space)
5920  *
5921  * o reserve space for metadata space, based on the number of outstanding
5922  *   extents and how much csums will be needed
5923  *   also reserve metadata space in a per root over-reserve method.
5924  * o add to the inodes->delalloc_bytes
5925  * o add it to the fs_info's delalloc inodes list.
5926  *   (Above 3 all done in delalloc_reserve_metadata)
5927  *
5928  * Return 0 for success
5929  * Return <0 for error(-ENOSPC or -EQUOT)
5930  */
5931 int btrfs_delalloc_reserve_space(struct inode *inode, u64 start, u64 len)
5932 {
5933         int ret;
5934
5935         ret = btrfs_check_data_free_space(inode, start, len);
5936         if (ret < 0)
5937                 return ret;
5938         ret = btrfs_delalloc_reserve_metadata(inode, len);
5939         if (ret < 0)
5940                 btrfs_free_reserved_data_space(inode, start, len);
5941         return ret;
5942 }
5943
5944 /**
5945  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5946  * @inode: inode we're releasing space for
5947  * @start: start position of the space already reserved
5948  * @len: the len of the space already reserved
5949  *
5950  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5951  * called in the case that we don't need the metadata AND data reservations
5952  * anymore.  So if there is an error or we insert an inline extent.
5953  *
5954  * This function will release the metadata space that was not used and will
5955  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5956  * list if there are no delalloc bytes left.
5957  * Also it will handle the qgroup reserved space.
5958  */
5959 void btrfs_delalloc_release_space(struct inode *inode, u64 start, u64 len)
5960 {
5961         btrfs_delalloc_release_metadata(inode, len);
5962         btrfs_free_reserved_data_space(inode, start, len);
5963 }
5964
5965 static int update_block_group(struct btrfs_trans_handle *trans,
5966                               struct btrfs_root *root, u64 bytenr,
5967                               u64 num_bytes, int alloc)
5968 {
5969         struct btrfs_block_group_cache *cache = NULL;
5970         struct btrfs_fs_info *info = root->fs_info;
5971         u64 total = num_bytes;
5972         u64 old_val;
5973         u64 byte_in_group;
5974         int factor;
5975
5976         /* block accounting for super block */
5977         spin_lock(&info->delalloc_root_lock);
5978         old_val = btrfs_super_bytes_used(info->super_copy);
5979         if (alloc)
5980                 old_val += num_bytes;
5981         else
5982                 old_val -= num_bytes;
5983         btrfs_set_super_bytes_used(info->super_copy, old_val);
5984         spin_unlock(&info->delalloc_root_lock);
5985
5986         while (total) {
5987                 cache = btrfs_lookup_block_group(info, bytenr);
5988                 if (!cache)
5989                         return -ENOENT;
5990                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5991                                     BTRFS_BLOCK_GROUP_RAID1 |
5992                                     BTRFS_BLOCK_GROUP_RAID10))
5993                         factor = 2;
5994                 else
5995                         factor = 1;
5996                 /*
5997                  * If this block group has free space cache written out, we
5998                  * need to make sure to load it if we are removing space.  This
5999                  * is because we need the unpinning stage to actually add the
6000                  * space back to the block group, otherwise we will leak space.
6001                  */
6002                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6003                         cache_block_group(cache, 1);
6004
6005                 byte_in_group = bytenr - cache->key.objectid;
6006                 WARN_ON(byte_in_group > cache->key.offset);
6007
6008                 spin_lock(&cache->space_info->lock);
6009                 spin_lock(&cache->lock);
6010
6011                 if (btrfs_test_opt(root, SPACE_CACHE) &&
6012                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6013                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6014
6015                 old_val = btrfs_block_group_used(&cache->item);
6016                 num_bytes = min(total, cache->key.offset - byte_in_group);
6017                 if (alloc) {
6018                         old_val += num_bytes;
6019                         btrfs_set_block_group_used(&cache->item, old_val);
6020                         cache->reserved -= num_bytes;
6021                         cache->space_info->bytes_reserved -= num_bytes;
6022                         cache->space_info->bytes_used += num_bytes;
6023                         cache->space_info->disk_used += num_bytes * factor;
6024                         spin_unlock(&cache->lock);
6025                         spin_unlock(&cache->space_info->lock);
6026                 } else {
6027                         old_val -= num_bytes;
6028                         btrfs_set_block_group_used(&cache->item, old_val);
6029                         cache->pinned += num_bytes;
6030                         cache->space_info->bytes_pinned += num_bytes;
6031                         cache->space_info->bytes_used -= num_bytes;
6032                         cache->space_info->disk_used -= num_bytes * factor;
6033                         spin_unlock(&cache->lock);
6034                         spin_unlock(&cache->space_info->lock);
6035
6036                         set_extent_dirty(info->pinned_extents,
6037                                          bytenr, bytenr + num_bytes - 1,
6038                                          GFP_NOFS | __GFP_NOFAIL);
6039                 }
6040
6041                 spin_lock(&trans->transaction->dirty_bgs_lock);
6042                 if (list_empty(&cache->dirty_list)) {
6043                         list_add_tail(&cache->dirty_list,
6044                                       &trans->transaction->dirty_bgs);
6045                                 trans->transaction->num_dirty_bgs++;
6046                         btrfs_get_block_group(cache);
6047                 }
6048                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6049
6050                 /*
6051                  * No longer have used bytes in this block group, queue it for
6052                  * deletion. We do this after adding the block group to the
6053                  * dirty list to avoid races between cleaner kthread and space
6054                  * cache writeout.
6055                  */
6056                 if (!alloc && old_val == 0) {
6057                         spin_lock(&info->unused_bgs_lock);
6058                         if (list_empty(&cache->bg_list)) {
6059                                 btrfs_get_block_group(cache);
6060                                 list_add_tail(&cache->bg_list,
6061                                               &info->unused_bgs);
6062                         }
6063                         spin_unlock(&info->unused_bgs_lock);
6064                 }
6065
6066                 btrfs_put_block_group(cache);
6067                 total -= num_bytes;
6068                 bytenr += num_bytes;
6069         }
6070         return 0;
6071 }
6072
6073 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
6074 {
6075         struct btrfs_block_group_cache *cache;
6076         u64 bytenr;
6077
6078         spin_lock(&root->fs_info->block_group_cache_lock);
6079         bytenr = root->fs_info->first_logical_byte;
6080         spin_unlock(&root->fs_info->block_group_cache_lock);
6081
6082         if (bytenr < (u64)-1)
6083                 return bytenr;
6084
6085         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
6086         if (!cache)
6087                 return 0;
6088
6089         bytenr = cache->key.objectid;
6090         btrfs_put_block_group(cache);
6091
6092         return bytenr;
6093 }
6094
6095 static int pin_down_extent(struct btrfs_root *root,
6096                            struct btrfs_block_group_cache *cache,
6097                            u64 bytenr, u64 num_bytes, int reserved)
6098 {
6099         spin_lock(&cache->space_info->lock);
6100         spin_lock(&cache->lock);
6101         cache->pinned += num_bytes;
6102         cache->space_info->bytes_pinned += num_bytes;
6103         if (reserved) {
6104                 cache->reserved -= num_bytes;
6105                 cache->space_info->bytes_reserved -= num_bytes;
6106         }
6107         spin_unlock(&cache->lock);
6108         spin_unlock(&cache->space_info->lock);
6109
6110         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
6111                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6112         if (reserved)
6113                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
6114         return 0;
6115 }
6116
6117 /*
6118  * this function must be called within transaction
6119  */
6120 int btrfs_pin_extent(struct btrfs_root *root,
6121                      u64 bytenr, u64 num_bytes, int reserved)
6122 {
6123         struct btrfs_block_group_cache *cache;
6124
6125         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6126         BUG_ON(!cache); /* Logic error */
6127
6128         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
6129
6130         btrfs_put_block_group(cache);
6131         return 0;
6132 }
6133
6134 /*
6135  * this function must be called within transaction
6136  */
6137 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
6138                                     u64 bytenr, u64 num_bytes)
6139 {
6140         struct btrfs_block_group_cache *cache;
6141         int ret;
6142
6143         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6144         if (!cache)
6145                 return -EINVAL;
6146
6147         /*
6148          * pull in the free space cache (if any) so that our pin
6149          * removes the free space from the cache.  We have load_only set
6150          * to one because the slow code to read in the free extents does check
6151          * the pinned extents.
6152          */
6153         cache_block_group(cache, 1);
6154
6155         pin_down_extent(root, cache, bytenr, num_bytes, 0);
6156
6157         /* remove us from the free space cache (if we're there at all) */
6158         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6159         btrfs_put_block_group(cache);
6160         return ret;
6161 }
6162
6163 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
6164 {
6165         int ret;
6166         struct btrfs_block_group_cache *block_group;
6167         struct btrfs_caching_control *caching_ctl;
6168
6169         block_group = btrfs_lookup_block_group(root->fs_info, start);
6170         if (!block_group)
6171                 return -EINVAL;
6172
6173         cache_block_group(block_group, 0);
6174         caching_ctl = get_caching_control(block_group);
6175
6176         if (!caching_ctl) {
6177                 /* Logic error */
6178                 BUG_ON(!block_group_cache_done(block_group));
6179                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6180         } else {
6181                 mutex_lock(&caching_ctl->mutex);
6182
6183                 if (start >= caching_ctl->progress) {
6184                         ret = add_excluded_extent(root, start, num_bytes);
6185                 } else if (start + num_bytes <= caching_ctl->progress) {
6186                         ret = btrfs_remove_free_space(block_group,
6187                                                       start, num_bytes);
6188                 } else {
6189                         num_bytes = caching_ctl->progress - start;
6190                         ret = btrfs_remove_free_space(block_group,
6191                                                       start, num_bytes);
6192                         if (ret)
6193                                 goto out_lock;
6194
6195                         num_bytes = (start + num_bytes) -
6196                                 caching_ctl->progress;
6197                         start = caching_ctl->progress;
6198                         ret = add_excluded_extent(root, start, num_bytes);
6199                 }
6200 out_lock:
6201                 mutex_unlock(&caching_ctl->mutex);
6202                 put_caching_control(caching_ctl);
6203         }
6204         btrfs_put_block_group(block_group);
6205         return ret;
6206 }
6207
6208 int btrfs_exclude_logged_extents(struct btrfs_root *log,
6209                                  struct extent_buffer *eb)
6210 {
6211         struct btrfs_file_extent_item *item;
6212         struct btrfs_key key;
6213         int found_type;
6214         int i;
6215
6216         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
6217                 return 0;
6218
6219         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6220                 btrfs_item_key_to_cpu(eb, &key, i);
6221                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6222                         continue;
6223                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6224                 found_type = btrfs_file_extent_type(eb, item);
6225                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6226                         continue;
6227                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6228                         continue;
6229                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6230                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6231                 __exclude_logged_extent(log, key.objectid, key.offset);
6232         }
6233
6234         return 0;
6235 }
6236
6237 static void
6238 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6239 {
6240         atomic_inc(&bg->reservations);
6241 }
6242
6243 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6244                                         const u64 start)
6245 {
6246         struct btrfs_block_group_cache *bg;
6247
6248         bg = btrfs_lookup_block_group(fs_info, start);
6249         ASSERT(bg);
6250         if (atomic_dec_and_test(&bg->reservations))
6251                 wake_up_atomic_t(&bg->reservations);
6252         btrfs_put_block_group(bg);
6253 }
6254
6255 static int btrfs_wait_bg_reservations_atomic_t(atomic_t *a)
6256 {
6257         schedule();
6258         return 0;
6259 }
6260
6261 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6262 {
6263         struct btrfs_space_info *space_info = bg->space_info;
6264
6265         ASSERT(bg->ro);
6266
6267         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6268                 return;
6269
6270         /*
6271          * Our block group is read only but before we set it to read only,
6272          * some task might have had allocated an extent from it already, but it
6273          * has not yet created a respective ordered extent (and added it to a
6274          * root's list of ordered extents).
6275          * Therefore wait for any task currently allocating extents, since the
6276          * block group's reservations counter is incremented while a read lock
6277          * on the groups' semaphore is held and decremented after releasing
6278          * the read access on that semaphore and creating the ordered extent.
6279          */
6280         down_write(&space_info->groups_sem);
6281         up_write(&space_info->groups_sem);
6282
6283         wait_on_atomic_t(&bg->reservations,
6284                          btrfs_wait_bg_reservations_atomic_t,
6285                          TASK_UNINTERRUPTIBLE);
6286 }
6287
6288 /**
6289  * btrfs_update_reserved_bytes - update the block_group and space info counters
6290  * @cache:      The cache we are manipulating
6291  * @num_bytes:  The number of bytes in question
6292  * @reserve:    One of the reservation enums
6293  * @delalloc:   The blocks are allocated for the delalloc write
6294  *
6295  * This is called by the allocator when it reserves space, or by somebody who is
6296  * freeing space that was never actually used on disk.  For example if you
6297  * reserve some space for a new leaf in transaction A and before transaction A
6298  * commits you free that leaf, you call this with reserve set to 0 in order to
6299  * clear the reservation.
6300  *
6301  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
6302  * ENOSPC accounting.  For data we handle the reservation through clearing the
6303  * delalloc bits in the io_tree.  We have to do this since we could end up
6304  * allocating less disk space for the amount of data we have reserved in the
6305  * case of compression.
6306  *
6307  * If this is a reservation and the block group has become read only we cannot
6308  * make the reservation and return -EAGAIN, otherwise this function always
6309  * succeeds.
6310  */
6311 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
6312                                        u64 num_bytes, int reserve, int delalloc)
6313 {
6314         struct btrfs_space_info *space_info = cache->space_info;
6315         int ret = 0;
6316
6317         spin_lock(&space_info->lock);
6318         spin_lock(&cache->lock);
6319         if (reserve != RESERVE_FREE) {
6320                 if (cache->ro) {
6321                         ret = -EAGAIN;
6322                 } else {
6323                         cache->reserved += num_bytes;
6324                         space_info->bytes_reserved += num_bytes;
6325                         if (reserve == RESERVE_ALLOC) {
6326                                 trace_btrfs_space_reservation(cache->fs_info,
6327                                                 "space_info", space_info->flags,
6328                                                 num_bytes, 0);
6329                                 space_info->bytes_may_use -= num_bytes;
6330                         }
6331
6332                         if (delalloc)
6333                                 cache->delalloc_bytes += num_bytes;
6334                 }
6335         } else {
6336                 if (cache->ro)
6337                         space_info->bytes_readonly += num_bytes;
6338                 cache->reserved -= num_bytes;
6339                 space_info->bytes_reserved -= num_bytes;
6340
6341                 if (delalloc)
6342                         cache->delalloc_bytes -= num_bytes;
6343         }
6344         spin_unlock(&cache->lock);
6345         spin_unlock(&space_info->lock);
6346         return ret;
6347 }
6348
6349 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6350                                 struct btrfs_root *root)
6351 {
6352         struct btrfs_fs_info *fs_info = root->fs_info;
6353         struct btrfs_caching_control *next;
6354         struct btrfs_caching_control *caching_ctl;
6355         struct btrfs_block_group_cache *cache;
6356
6357         down_write(&fs_info->commit_root_sem);
6358
6359         list_for_each_entry_safe(caching_ctl, next,
6360                                  &fs_info->caching_block_groups, list) {
6361                 cache = caching_ctl->block_group;
6362                 if (block_group_cache_done(cache)) {
6363                         cache->last_byte_to_unpin = (u64)-1;
6364                         list_del_init(&caching_ctl->list);
6365                         put_caching_control(caching_ctl);
6366                 } else {
6367                         cache->last_byte_to_unpin = caching_ctl->progress;
6368                 }
6369         }
6370
6371         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6372                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6373         else
6374                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6375
6376         up_write(&fs_info->commit_root_sem);
6377
6378         update_global_block_rsv(fs_info);
6379 }
6380
6381 /*
6382  * Returns the free cluster for the given space info and sets empty_cluster to
6383  * what it should be based on the mount options.
6384  */
6385 static struct btrfs_free_cluster *
6386 fetch_cluster_info(struct btrfs_root *root, struct btrfs_space_info *space_info,
6387                    u64 *empty_cluster)
6388 {
6389         struct btrfs_free_cluster *ret = NULL;
6390         bool ssd = btrfs_test_opt(root, SSD);
6391
6392         *empty_cluster = 0;
6393         if (btrfs_mixed_space_info(space_info))
6394                 return ret;
6395
6396         if (ssd)
6397                 *empty_cluster = SZ_2M;
6398         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6399                 ret = &root->fs_info->meta_alloc_cluster;
6400                 if (!ssd)
6401                         *empty_cluster = SZ_64K;
6402         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) && ssd) {
6403                 ret = &root->fs_info->data_alloc_cluster;
6404         }
6405
6406         return ret;
6407 }
6408
6409 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
6410                               const bool return_free_space)
6411 {
6412         struct btrfs_fs_info *fs_info = root->fs_info;
6413         struct btrfs_block_group_cache *cache = NULL;
6414         struct btrfs_space_info *space_info;
6415         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6416         struct btrfs_free_cluster *cluster = NULL;
6417         u64 len;
6418         u64 total_unpinned = 0;
6419         u64 empty_cluster = 0;
6420         bool readonly;
6421
6422         while (start <= end) {
6423                 readonly = false;
6424                 if (!cache ||
6425                     start >= cache->key.objectid + cache->key.offset) {
6426                         if (cache)
6427                                 btrfs_put_block_group(cache);
6428                         total_unpinned = 0;
6429                         cache = btrfs_lookup_block_group(fs_info, start);
6430                         BUG_ON(!cache); /* Logic error */
6431
6432                         cluster = fetch_cluster_info(root,
6433                                                      cache->space_info,
6434                                                      &empty_cluster);
6435                         empty_cluster <<= 1;
6436                 }
6437
6438                 len = cache->key.objectid + cache->key.offset - start;
6439                 len = min(len, end + 1 - start);
6440
6441                 if (start < cache->last_byte_to_unpin) {
6442                         len = min(len, cache->last_byte_to_unpin - start);
6443                         if (return_free_space)
6444                                 btrfs_add_free_space(cache, start, len);
6445                 }
6446
6447                 start += len;
6448                 total_unpinned += len;
6449                 space_info = cache->space_info;
6450
6451                 /*
6452                  * If this space cluster has been marked as fragmented and we've
6453                  * unpinned enough in this block group to potentially allow a
6454                  * cluster to be created inside of it go ahead and clear the
6455                  * fragmented check.
6456                  */
6457                 if (cluster && cluster->fragmented &&
6458                     total_unpinned > empty_cluster) {
6459                         spin_lock(&cluster->lock);
6460                         cluster->fragmented = 0;
6461                         spin_unlock(&cluster->lock);
6462                 }
6463
6464                 spin_lock(&space_info->lock);
6465                 spin_lock(&cache->lock);
6466                 cache->pinned -= len;
6467                 space_info->bytes_pinned -= len;
6468                 space_info->max_extent_size = 0;
6469                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6470                 if (cache->ro) {
6471                         space_info->bytes_readonly += len;
6472                         readonly = true;
6473                 }
6474                 spin_unlock(&cache->lock);
6475                 if (!readonly && global_rsv->space_info == space_info) {
6476                         spin_lock(&global_rsv->lock);
6477                         if (!global_rsv->full) {
6478                                 len = min(len, global_rsv->size -
6479                                           global_rsv->reserved);
6480                                 global_rsv->reserved += len;
6481                                 space_info->bytes_may_use += len;
6482                                 if (global_rsv->reserved >= global_rsv->size)
6483                                         global_rsv->full = 1;
6484                         }
6485                         spin_unlock(&global_rsv->lock);
6486                 }
6487                 spin_unlock(&space_info->lock);
6488         }
6489
6490         if (cache)
6491                 btrfs_put_block_group(cache);
6492         return 0;
6493 }
6494
6495 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6496                                struct btrfs_root *root)
6497 {
6498         struct btrfs_fs_info *fs_info = root->fs_info;
6499         struct btrfs_block_group_cache *block_group, *tmp;
6500         struct list_head *deleted_bgs;
6501         struct extent_io_tree *unpin;
6502         u64 start;
6503         u64 end;
6504         int ret;
6505
6506         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6507                 unpin = &fs_info->freed_extents[1];
6508         else
6509                 unpin = &fs_info->freed_extents[0];
6510
6511         while (!trans->aborted) {
6512                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6513                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6514                                             EXTENT_DIRTY, NULL);
6515                 if (ret) {
6516                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6517                         break;
6518                 }
6519
6520                 if (btrfs_test_opt(root, DISCARD))
6521                         ret = btrfs_discard_extent(root, start,
6522                                                    end + 1 - start, NULL);
6523
6524                 clear_extent_dirty(unpin, start, end);
6525                 unpin_extent_range(root, start, end, true);
6526                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6527                 cond_resched();
6528         }
6529
6530         /*
6531          * Transaction is finished.  We don't need the lock anymore.  We
6532          * do need to clean up the block groups in case of a transaction
6533          * abort.
6534          */
6535         deleted_bgs = &trans->transaction->deleted_bgs;
6536         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6537                 u64 trimmed = 0;
6538
6539                 ret = -EROFS;
6540                 if (!trans->aborted)
6541                         ret = btrfs_discard_extent(root,
6542                                                    block_group->key.objectid,
6543                                                    block_group->key.offset,
6544                                                    &trimmed);
6545
6546                 list_del_init(&block_group->bg_list);
6547                 btrfs_put_block_group_trimming(block_group);
6548                 btrfs_put_block_group(block_group);
6549
6550                 if (ret) {
6551                         const char *errstr = btrfs_decode_error(ret);
6552                         btrfs_warn(fs_info,
6553                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6554                                    ret, errstr);
6555                 }
6556         }
6557
6558         return 0;
6559 }
6560
6561 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6562                              u64 owner, u64 root_objectid)
6563 {
6564         struct btrfs_space_info *space_info;
6565         u64 flags;
6566
6567         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6568                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6569                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6570                 else
6571                         flags = BTRFS_BLOCK_GROUP_METADATA;
6572         } else {
6573                 flags = BTRFS_BLOCK_GROUP_DATA;
6574         }
6575
6576         space_info = __find_space_info(fs_info, flags);
6577         BUG_ON(!space_info); /* Logic bug */
6578         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6579 }
6580
6581
6582 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6583                                 struct btrfs_root *root,
6584                                 struct btrfs_delayed_ref_node *node, u64 parent,
6585                                 u64 root_objectid, u64 owner_objectid,
6586                                 u64 owner_offset, int refs_to_drop,
6587                                 struct btrfs_delayed_extent_op *extent_op)
6588 {
6589         struct btrfs_key key;
6590         struct btrfs_path *path;
6591         struct btrfs_fs_info *info = root->fs_info;
6592         struct btrfs_root *extent_root = info->extent_root;
6593         struct extent_buffer *leaf;
6594         struct btrfs_extent_item *ei;
6595         struct btrfs_extent_inline_ref *iref;
6596         int ret;
6597         int is_data;
6598         int extent_slot = 0;
6599         int found_extent = 0;
6600         int num_to_del = 1;
6601         u32 item_size;
6602         u64 refs;
6603         u64 bytenr = node->bytenr;
6604         u64 num_bytes = node->num_bytes;
6605         int last_ref = 0;
6606         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6607                                                  SKINNY_METADATA);
6608
6609         path = btrfs_alloc_path();
6610         if (!path)
6611                 return -ENOMEM;
6612
6613         path->reada = READA_FORWARD;
6614         path->leave_spinning = 1;
6615
6616         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6617         BUG_ON(!is_data && refs_to_drop != 1);
6618
6619         if (is_data)
6620                 skinny_metadata = 0;
6621
6622         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6623                                     bytenr, num_bytes, parent,
6624                                     root_objectid, owner_objectid,
6625                                     owner_offset);
6626         if (ret == 0) {
6627                 extent_slot = path->slots[0];
6628                 while (extent_slot >= 0) {
6629                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6630                                               extent_slot);
6631                         if (key.objectid != bytenr)
6632                                 break;
6633                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6634                             key.offset == num_bytes) {
6635                                 found_extent = 1;
6636                                 break;
6637                         }
6638                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6639                             key.offset == owner_objectid) {
6640                                 found_extent = 1;
6641                                 break;
6642                         }
6643                         if (path->slots[0] - extent_slot > 5)
6644                                 break;
6645                         extent_slot--;
6646                 }
6647 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6648                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6649                 if (found_extent && item_size < sizeof(*ei))
6650                         found_extent = 0;
6651 #endif
6652                 if (!found_extent) {
6653                         BUG_ON(iref);
6654                         ret = remove_extent_backref(trans, extent_root, path,
6655                                                     NULL, refs_to_drop,
6656                                                     is_data, &last_ref);
6657                         if (ret) {
6658                                 btrfs_abort_transaction(trans, extent_root, ret);
6659                                 goto out;
6660                         }
6661                         btrfs_release_path(path);
6662                         path->leave_spinning = 1;
6663
6664                         key.objectid = bytenr;
6665                         key.type = BTRFS_EXTENT_ITEM_KEY;
6666                         key.offset = num_bytes;
6667
6668                         if (!is_data && skinny_metadata) {
6669                                 key.type = BTRFS_METADATA_ITEM_KEY;
6670                                 key.offset = owner_objectid;
6671                         }
6672
6673                         ret = btrfs_search_slot(trans, extent_root,
6674                                                 &key, path, -1, 1);
6675                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6676                                 /*
6677                                  * Couldn't find our skinny metadata item,
6678                                  * see if we have ye olde extent item.
6679                                  */
6680                                 path->slots[0]--;
6681                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6682                                                       path->slots[0]);
6683                                 if (key.objectid == bytenr &&
6684                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6685                                     key.offset == num_bytes)
6686                                         ret = 0;
6687                         }
6688
6689                         if (ret > 0 && skinny_metadata) {
6690                                 skinny_metadata = false;
6691                                 key.objectid = bytenr;
6692                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6693                                 key.offset = num_bytes;
6694                                 btrfs_release_path(path);
6695                                 ret = btrfs_search_slot(trans, extent_root,
6696                                                         &key, path, -1, 1);
6697                         }
6698
6699                         if (ret) {
6700                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6701                                         ret, bytenr);
6702                                 if (ret > 0)
6703                                         btrfs_print_leaf(extent_root,
6704                                                          path->nodes[0]);
6705                         }
6706                         if (ret < 0) {
6707                                 btrfs_abort_transaction(trans, extent_root, ret);
6708                                 goto out;
6709                         }
6710                         extent_slot = path->slots[0];
6711                 }
6712         } else if (WARN_ON(ret == -ENOENT)) {
6713                 btrfs_print_leaf(extent_root, path->nodes[0]);
6714                 btrfs_err(info,
6715                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6716                         bytenr, parent, root_objectid, owner_objectid,
6717                         owner_offset);
6718                 btrfs_abort_transaction(trans, extent_root, ret);
6719                 goto out;
6720         } else {
6721                 btrfs_abort_transaction(trans, extent_root, ret);
6722                 goto out;
6723         }
6724
6725         leaf = path->nodes[0];
6726         item_size = btrfs_item_size_nr(leaf, extent_slot);
6727 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6728         if (item_size < sizeof(*ei)) {
6729                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6730                 ret = convert_extent_item_v0(trans, extent_root, path,
6731                                              owner_objectid, 0);
6732                 if (ret < 0) {
6733                         btrfs_abort_transaction(trans, extent_root, ret);
6734                         goto out;
6735                 }
6736
6737                 btrfs_release_path(path);
6738                 path->leave_spinning = 1;
6739
6740                 key.objectid = bytenr;
6741                 key.type = BTRFS_EXTENT_ITEM_KEY;
6742                 key.offset = num_bytes;
6743
6744                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6745                                         -1, 1);
6746                 if (ret) {
6747                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6748                                 ret, bytenr);
6749                         btrfs_print_leaf(extent_root, path->nodes[0]);
6750                 }
6751                 if (ret < 0) {
6752                         btrfs_abort_transaction(trans, extent_root, ret);
6753                         goto out;
6754                 }
6755
6756                 extent_slot = path->slots[0];
6757                 leaf = path->nodes[0];
6758                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6759         }
6760 #endif
6761         BUG_ON(item_size < sizeof(*ei));
6762         ei = btrfs_item_ptr(leaf, extent_slot,
6763                             struct btrfs_extent_item);
6764         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6765             key.type == BTRFS_EXTENT_ITEM_KEY) {
6766                 struct btrfs_tree_block_info *bi;
6767                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6768                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6769                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6770         }
6771
6772         refs = btrfs_extent_refs(leaf, ei);
6773         if (refs < refs_to_drop) {
6774                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6775                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6776                 ret = -EINVAL;
6777                 btrfs_abort_transaction(trans, extent_root, ret);
6778                 goto out;
6779         }
6780         refs -= refs_to_drop;
6781
6782         if (refs > 0) {
6783                 if (extent_op)
6784                         __run_delayed_extent_op(extent_op, leaf, ei);
6785                 /*
6786                  * In the case of inline back ref, reference count will
6787                  * be updated by remove_extent_backref
6788                  */
6789                 if (iref) {
6790                         BUG_ON(!found_extent);
6791                 } else {
6792                         btrfs_set_extent_refs(leaf, ei, refs);
6793                         btrfs_mark_buffer_dirty(leaf);
6794                 }
6795                 if (found_extent) {
6796                         ret = remove_extent_backref(trans, extent_root, path,
6797                                                     iref, refs_to_drop,
6798                                                     is_data, &last_ref);
6799                         if (ret) {
6800                                 btrfs_abort_transaction(trans, extent_root, ret);
6801                                 goto out;
6802                         }
6803                 }
6804                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6805                                  root_objectid);
6806         } else {
6807                 if (found_extent) {
6808                         BUG_ON(is_data && refs_to_drop !=
6809                                extent_data_ref_count(path, iref));
6810                         if (iref) {
6811                                 BUG_ON(path->slots[0] != extent_slot);
6812                         } else {
6813                                 BUG_ON(path->slots[0] != extent_slot + 1);
6814                                 path->slots[0] = extent_slot;
6815                                 num_to_del = 2;
6816                         }
6817                 }
6818
6819                 last_ref = 1;
6820                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6821                                       num_to_del);
6822                 if (ret) {
6823                         btrfs_abort_transaction(trans, extent_root, ret);
6824                         goto out;
6825                 }
6826                 btrfs_release_path(path);
6827
6828                 if (is_data) {
6829                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6830                         if (ret) {
6831                                 btrfs_abort_transaction(trans, extent_root, ret);
6832                                 goto out;
6833                         }
6834                 }
6835
6836                 ret = add_to_free_space_tree(trans, root->fs_info, bytenr,
6837                                              num_bytes);
6838                 if (ret) {
6839                         btrfs_abort_transaction(trans, extent_root, ret);
6840                         goto out;
6841                 }
6842
6843                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6844                 if (ret) {
6845                         btrfs_abort_transaction(trans, extent_root, ret);
6846                         goto out;
6847                 }
6848         }
6849         btrfs_release_path(path);
6850
6851 out:
6852         btrfs_free_path(path);
6853         return ret;
6854 }
6855
6856 /*
6857  * when we free an block, it is possible (and likely) that we free the last
6858  * delayed ref for that extent as well.  This searches the delayed ref tree for
6859  * a given extent, and if there are no other delayed refs to be processed, it
6860  * removes it from the tree.
6861  */
6862 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6863                                       struct btrfs_root *root, u64 bytenr)
6864 {
6865         struct btrfs_delayed_ref_head *head;
6866         struct btrfs_delayed_ref_root *delayed_refs;
6867         int ret = 0;
6868
6869         delayed_refs = &trans->transaction->delayed_refs;
6870         spin_lock(&delayed_refs->lock);
6871         head = btrfs_find_delayed_ref_head(trans, bytenr);
6872         if (!head)
6873                 goto out_delayed_unlock;
6874
6875         spin_lock(&head->lock);
6876         if (!list_empty(&head->ref_list))
6877                 goto out;
6878
6879         if (head->extent_op) {
6880                 if (!head->must_insert_reserved)
6881                         goto out;
6882                 btrfs_free_delayed_extent_op(head->extent_op);
6883                 head->extent_op = NULL;
6884         }
6885
6886         /*
6887          * waiting for the lock here would deadlock.  If someone else has it
6888          * locked they are already in the process of dropping it anyway
6889          */
6890         if (!mutex_trylock(&head->mutex))
6891                 goto out;
6892
6893         /*
6894          * at this point we have a head with no other entries.  Go
6895          * ahead and process it.
6896          */
6897         head->node.in_tree = 0;
6898         rb_erase(&head->href_node, &delayed_refs->href_root);
6899
6900         atomic_dec(&delayed_refs->num_entries);
6901
6902         /*
6903          * we don't take a ref on the node because we're removing it from the
6904          * tree, so we just steal the ref the tree was holding.
6905          */
6906         delayed_refs->num_heads--;
6907         if (head->processing == 0)
6908                 delayed_refs->num_heads_ready--;
6909         head->processing = 0;
6910         spin_unlock(&head->lock);
6911         spin_unlock(&delayed_refs->lock);
6912
6913         BUG_ON(head->extent_op);
6914         if (head->must_insert_reserved)
6915                 ret = 1;
6916
6917         mutex_unlock(&head->mutex);
6918         btrfs_put_delayed_ref(&head->node);
6919         return ret;
6920 out:
6921         spin_unlock(&head->lock);
6922
6923 out_delayed_unlock:
6924         spin_unlock(&delayed_refs->lock);
6925         return 0;
6926 }
6927
6928 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6929                            struct btrfs_root *root,
6930                            struct extent_buffer *buf,
6931                            u64 parent, int last_ref)
6932 {
6933         int pin = 1;
6934         int ret;
6935
6936         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6937                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6938                                         buf->start, buf->len,
6939                                         parent, root->root_key.objectid,
6940                                         btrfs_header_level(buf),
6941                                         BTRFS_DROP_DELAYED_REF, NULL);
6942                 BUG_ON(ret); /* -ENOMEM */
6943         }
6944
6945         if (!last_ref)
6946                 return;
6947
6948         if (btrfs_header_generation(buf) == trans->transid) {
6949                 struct btrfs_block_group_cache *cache;
6950
6951                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6952                         ret = check_ref_cleanup(trans, root, buf->start);
6953                         if (!ret)
6954                                 goto out;
6955                 }
6956
6957                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6958
6959                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6960                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6961                         btrfs_put_block_group(cache);
6962                         goto out;
6963                 }
6964
6965                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6966
6967                 btrfs_add_free_space(cache, buf->start, buf->len);
6968                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6969                 btrfs_put_block_group(cache);
6970                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6971                 pin = 0;
6972         }
6973 out:
6974         if (pin)
6975                 add_pinned_bytes(root->fs_info, buf->len,
6976                                  btrfs_header_level(buf),
6977                                  root->root_key.objectid);
6978
6979         /*
6980          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6981          * anymore.
6982          */
6983         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6984 }
6985
6986 /* Can return -ENOMEM */
6987 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6988                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6989                       u64 owner, u64 offset)
6990 {
6991         int ret;
6992         struct btrfs_fs_info *fs_info = root->fs_info;
6993
6994         if (btrfs_test_is_dummy_root(root))
6995                 return 0;
6996
6997         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6998
6999         /*
7000          * tree log blocks never actually go into the extent allocation
7001          * tree, just update pinning info and exit early.
7002          */
7003         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7004                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7005                 /* unlocks the pinned mutex */
7006                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
7007                 ret = 0;
7008         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7009                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7010                                         num_bytes,
7011                                         parent, root_objectid, (int)owner,
7012                                         BTRFS_DROP_DELAYED_REF, NULL);
7013         } else {
7014                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7015                                                 num_bytes,
7016                                                 parent, root_objectid, owner,
7017                                                 offset, 0,
7018                                                 BTRFS_DROP_DELAYED_REF, NULL);
7019         }
7020         return ret;
7021 }
7022
7023 /*
7024  * when we wait for progress in the block group caching, its because
7025  * our allocation attempt failed at least once.  So, we must sleep
7026  * and let some progress happen before we try again.
7027  *
7028  * This function will sleep at least once waiting for new free space to
7029  * show up, and then it will check the block group free space numbers
7030  * for our min num_bytes.  Another option is to have it go ahead
7031  * and look in the rbtree for a free extent of a given size, but this
7032  * is a good start.
7033  *
7034  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7035  * any of the information in this block group.
7036  */
7037 static noinline void
7038 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7039                                 u64 num_bytes)
7040 {
7041         struct btrfs_caching_control *caching_ctl;
7042
7043         caching_ctl = get_caching_control(cache);
7044         if (!caching_ctl)
7045                 return;
7046
7047         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7048                    (cache->free_space_ctl->free_space >= num_bytes));
7049
7050         put_caching_control(caching_ctl);
7051 }
7052
7053 static noinline int
7054 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7055 {
7056         struct btrfs_caching_control *caching_ctl;
7057         int ret = 0;
7058
7059         caching_ctl = get_caching_control(cache);
7060         if (!caching_ctl)
7061                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7062
7063         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7064         if (cache->cached == BTRFS_CACHE_ERROR)
7065                 ret = -EIO;
7066         put_caching_control(caching_ctl);
7067         return ret;
7068 }
7069
7070 int __get_raid_index(u64 flags)
7071 {
7072         if (flags & BTRFS_BLOCK_GROUP_RAID10)
7073                 return BTRFS_RAID_RAID10;
7074         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
7075                 return BTRFS_RAID_RAID1;
7076         else if (flags & BTRFS_BLOCK_GROUP_DUP)
7077                 return BTRFS_RAID_DUP;
7078         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
7079                 return BTRFS_RAID_RAID0;
7080         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
7081                 return BTRFS_RAID_RAID5;
7082         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
7083                 return BTRFS_RAID_RAID6;
7084
7085         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
7086 }
7087
7088 int get_block_group_index(struct btrfs_block_group_cache *cache)
7089 {
7090         return __get_raid_index(cache->flags);
7091 }
7092
7093 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7094         [BTRFS_RAID_RAID10]     = "raid10",
7095         [BTRFS_RAID_RAID1]      = "raid1",
7096         [BTRFS_RAID_DUP]        = "dup",
7097         [BTRFS_RAID_RAID0]      = "raid0",
7098         [BTRFS_RAID_SINGLE]     = "single",
7099         [BTRFS_RAID_RAID5]      = "raid5",
7100         [BTRFS_RAID_RAID6]      = "raid6",
7101 };
7102
7103 static const char *get_raid_name(enum btrfs_raid_types type)
7104 {
7105         if (type >= BTRFS_NR_RAID_TYPES)
7106                 return NULL;
7107
7108         return btrfs_raid_type_names[type];
7109 }
7110
7111 enum btrfs_loop_type {
7112         LOOP_CACHING_NOWAIT = 0,
7113         LOOP_CACHING_WAIT = 1,
7114         LOOP_ALLOC_CHUNK = 2,
7115         LOOP_NO_EMPTY_SIZE = 3,
7116 };
7117
7118 static inline void
7119 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7120                        int delalloc)
7121 {
7122         if (delalloc)
7123                 down_read(&cache->data_rwsem);
7124 }
7125
7126 static inline void
7127 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7128                        int delalloc)
7129 {
7130         btrfs_get_block_group(cache);
7131         if (delalloc)
7132                 down_read(&cache->data_rwsem);
7133 }
7134
7135 static struct btrfs_block_group_cache *
7136 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7137                    struct btrfs_free_cluster *cluster,
7138                    int delalloc)
7139 {
7140         struct btrfs_block_group_cache *used_bg = NULL;
7141
7142         spin_lock(&cluster->refill_lock);
7143         while (1) {
7144                 used_bg = cluster->block_group;
7145                 if (!used_bg)
7146                         return NULL;
7147
7148                 if (used_bg == block_group)
7149                         return used_bg;
7150
7151                 btrfs_get_block_group(used_bg);
7152
7153                 if (!delalloc)
7154                         return used_bg;
7155
7156                 if (down_read_trylock(&used_bg->data_rwsem))
7157                         return used_bg;
7158
7159                 spin_unlock(&cluster->refill_lock);
7160
7161                 down_read(&used_bg->data_rwsem);
7162
7163                 spin_lock(&cluster->refill_lock);
7164                 if (used_bg == cluster->block_group)
7165                         return used_bg;
7166
7167                 up_read(&used_bg->data_rwsem);
7168                 btrfs_put_block_group(used_bg);
7169         }
7170 }
7171
7172 static inline void
7173 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7174                          int delalloc)
7175 {
7176         if (delalloc)
7177                 up_read(&cache->data_rwsem);
7178         btrfs_put_block_group(cache);
7179 }
7180
7181 /*
7182  * walks the btree of allocated extents and find a hole of a given size.
7183  * The key ins is changed to record the hole:
7184  * ins->objectid == start position
7185  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7186  * ins->offset == the size of the hole.
7187  * Any available blocks before search_start are skipped.
7188  *
7189  * If there is no suitable free space, we will record the max size of
7190  * the free space extent currently.
7191  */
7192 static noinline int find_free_extent(struct btrfs_root *orig_root,
7193                                      u64 num_bytes, u64 empty_size,
7194                                      u64 hint_byte, struct btrfs_key *ins,
7195                                      u64 flags, int delalloc)
7196 {
7197         int ret = 0;
7198         struct btrfs_root *root = orig_root->fs_info->extent_root;
7199         struct btrfs_free_cluster *last_ptr = NULL;
7200         struct btrfs_block_group_cache *block_group = NULL;
7201         u64 search_start = 0;
7202         u64 max_extent_size = 0;
7203         u64 empty_cluster = 0;
7204         struct btrfs_space_info *space_info;
7205         int loop = 0;
7206         int index = __get_raid_index(flags);
7207         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
7208                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
7209         bool failed_cluster_refill = false;
7210         bool failed_alloc = false;
7211         bool use_cluster = true;
7212         bool have_caching_bg = false;
7213         bool orig_have_caching_bg = false;
7214         bool full_search = false;
7215
7216         WARN_ON(num_bytes < root->sectorsize);
7217         ins->type = BTRFS_EXTENT_ITEM_KEY;
7218         ins->objectid = 0;
7219         ins->offset = 0;
7220
7221         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
7222
7223         space_info = __find_space_info(root->fs_info, flags);
7224         if (!space_info) {
7225                 btrfs_err(root->fs_info, "No space info for %llu", flags);
7226                 return -ENOSPC;
7227         }
7228
7229         /*
7230          * If our free space is heavily fragmented we may not be able to make
7231          * big contiguous allocations, so instead of doing the expensive search
7232          * for free space, simply return ENOSPC with our max_extent_size so we
7233          * can go ahead and search for a more manageable chunk.
7234          *
7235          * If our max_extent_size is large enough for our allocation simply
7236          * disable clustering since we will likely not be able to find enough
7237          * space to create a cluster and induce latency trying.
7238          */
7239         if (unlikely(space_info->max_extent_size)) {
7240                 spin_lock(&space_info->lock);
7241                 if (space_info->max_extent_size &&
7242                     num_bytes > space_info->max_extent_size) {
7243                         ins->offset = space_info->max_extent_size;
7244                         spin_unlock(&space_info->lock);
7245                         return -ENOSPC;
7246                 } else if (space_info->max_extent_size) {
7247                         use_cluster = false;
7248                 }
7249                 spin_unlock(&space_info->lock);
7250         }
7251
7252         last_ptr = fetch_cluster_info(orig_root, space_info, &empty_cluster);
7253         if (last_ptr) {
7254                 spin_lock(&last_ptr->lock);
7255                 if (last_ptr->block_group)
7256                         hint_byte = last_ptr->window_start;
7257                 if (last_ptr->fragmented) {
7258                         /*
7259                          * We still set window_start so we can keep track of the
7260                          * last place we found an allocation to try and save
7261                          * some time.
7262                          */
7263                         hint_byte = last_ptr->window_start;
7264                         use_cluster = false;
7265                 }
7266                 spin_unlock(&last_ptr->lock);
7267         }
7268
7269         search_start = max(search_start, first_logical_byte(root, 0));
7270         search_start = max(search_start, hint_byte);
7271         if (search_start == hint_byte) {
7272                 block_group = btrfs_lookup_block_group(root->fs_info,
7273                                                        search_start);
7274                 /*
7275                  * we don't want to use the block group if it doesn't match our
7276                  * allocation bits, or if its not cached.
7277                  *
7278                  * However if we are re-searching with an ideal block group
7279                  * picked out then we don't care that the block group is cached.
7280                  */
7281                 if (block_group && block_group_bits(block_group, flags) &&
7282                     block_group->cached != BTRFS_CACHE_NO) {
7283                         down_read(&space_info->groups_sem);
7284                         if (list_empty(&block_group->list) ||
7285                             block_group->ro) {
7286                                 /*
7287                                  * someone is removing this block group,
7288                                  * we can't jump into the have_block_group
7289                                  * target because our list pointers are not
7290                                  * valid
7291                                  */
7292                                 btrfs_put_block_group(block_group);
7293                                 up_read(&space_info->groups_sem);
7294                         } else {
7295                                 index = get_block_group_index(block_group);
7296                                 btrfs_lock_block_group(block_group, delalloc);
7297                                 goto have_block_group;
7298                         }
7299                 } else if (block_group) {
7300                         btrfs_put_block_group(block_group);
7301                 }
7302         }
7303 search:
7304         have_caching_bg = false;
7305         if (index == 0 || index == __get_raid_index(flags))
7306                 full_search = true;
7307         down_read(&space_info->groups_sem);
7308         list_for_each_entry(block_group, &space_info->block_groups[index],
7309                             list) {
7310                 u64 offset;
7311                 int cached;
7312
7313                 btrfs_grab_block_group(block_group, delalloc);
7314                 search_start = block_group->key.objectid;
7315
7316                 /*
7317                  * this can happen if we end up cycling through all the
7318                  * raid types, but we want to make sure we only allocate
7319                  * for the proper type.
7320                  */
7321                 if (!block_group_bits(block_group, flags)) {
7322                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7323                                 BTRFS_BLOCK_GROUP_RAID1 |
7324                                 BTRFS_BLOCK_GROUP_RAID5 |
7325                                 BTRFS_BLOCK_GROUP_RAID6 |
7326                                 BTRFS_BLOCK_GROUP_RAID10;
7327
7328                         /*
7329                          * if they asked for extra copies and this block group
7330                          * doesn't provide them, bail.  This does allow us to
7331                          * fill raid0 from raid1.
7332                          */
7333                         if ((flags & extra) && !(block_group->flags & extra))
7334                                 goto loop;
7335                 }
7336
7337 have_block_group:
7338                 cached = block_group_cache_done(block_group);
7339                 if (unlikely(!cached)) {
7340                         have_caching_bg = true;
7341                         ret = cache_block_group(block_group, 0);
7342                         BUG_ON(ret < 0);
7343                         ret = 0;
7344                 }
7345
7346                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7347                         goto loop;
7348                 if (unlikely(block_group->ro))
7349                         goto loop;
7350
7351                 /*
7352                  * Ok we want to try and use the cluster allocator, so
7353                  * lets look there
7354                  */
7355                 if (last_ptr && use_cluster) {
7356                         struct btrfs_block_group_cache *used_block_group;
7357                         unsigned long aligned_cluster;
7358                         /*
7359                          * the refill lock keeps out other
7360                          * people trying to start a new cluster
7361                          */
7362                         used_block_group = btrfs_lock_cluster(block_group,
7363                                                               last_ptr,
7364                                                               delalloc);
7365                         if (!used_block_group)
7366                                 goto refill_cluster;
7367
7368                         if (used_block_group != block_group &&
7369                             (used_block_group->ro ||
7370                              !block_group_bits(used_block_group, flags)))
7371                                 goto release_cluster;
7372
7373                         offset = btrfs_alloc_from_cluster(used_block_group,
7374                                                 last_ptr,
7375                                                 num_bytes,
7376                                                 used_block_group->key.objectid,
7377                                                 &max_extent_size);
7378                         if (offset) {
7379                                 /* we have a block, we're done */
7380                                 spin_unlock(&last_ptr->refill_lock);
7381                                 trace_btrfs_reserve_extent_cluster(root,
7382                                                 used_block_group,
7383                                                 search_start, num_bytes);
7384                                 if (used_block_group != block_group) {
7385                                         btrfs_release_block_group(block_group,
7386                                                                   delalloc);
7387                                         block_group = used_block_group;
7388                                 }
7389                                 goto checks;
7390                         }
7391
7392                         WARN_ON(last_ptr->block_group != used_block_group);
7393 release_cluster:
7394                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7395                          * set up a new clusters, so lets just skip it
7396                          * and let the allocator find whatever block
7397                          * it can find.  If we reach this point, we
7398                          * will have tried the cluster allocator
7399                          * plenty of times and not have found
7400                          * anything, so we are likely way too
7401                          * fragmented for the clustering stuff to find
7402                          * anything.
7403                          *
7404                          * However, if the cluster is taken from the
7405                          * current block group, release the cluster
7406                          * first, so that we stand a better chance of
7407                          * succeeding in the unclustered
7408                          * allocation.  */
7409                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7410                             used_block_group != block_group) {
7411                                 spin_unlock(&last_ptr->refill_lock);
7412                                 btrfs_release_block_group(used_block_group,
7413                                                           delalloc);
7414                                 goto unclustered_alloc;
7415                         }
7416
7417                         /*
7418                          * this cluster didn't work out, free it and
7419                          * start over
7420                          */
7421                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7422
7423                         if (used_block_group != block_group)
7424                                 btrfs_release_block_group(used_block_group,
7425                                                           delalloc);
7426 refill_cluster:
7427                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7428                                 spin_unlock(&last_ptr->refill_lock);
7429                                 goto unclustered_alloc;
7430                         }
7431
7432                         aligned_cluster = max_t(unsigned long,
7433                                                 empty_cluster + empty_size,
7434                                               block_group->full_stripe_len);
7435
7436                         /* allocate a cluster in this block group */
7437                         ret = btrfs_find_space_cluster(root, block_group,
7438                                                        last_ptr, search_start,
7439                                                        num_bytes,
7440                                                        aligned_cluster);
7441                         if (ret == 0) {
7442                                 /*
7443                                  * now pull our allocation out of this
7444                                  * cluster
7445                                  */
7446                                 offset = btrfs_alloc_from_cluster(block_group,
7447                                                         last_ptr,
7448                                                         num_bytes,
7449                                                         search_start,
7450                                                         &max_extent_size);
7451                                 if (offset) {
7452                                         /* we found one, proceed */
7453                                         spin_unlock(&last_ptr->refill_lock);
7454                                         trace_btrfs_reserve_extent_cluster(root,
7455                                                 block_group, search_start,
7456                                                 num_bytes);
7457                                         goto checks;
7458                                 }
7459                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7460                                    && !failed_cluster_refill) {
7461                                 spin_unlock(&last_ptr->refill_lock);
7462
7463                                 failed_cluster_refill = true;
7464                                 wait_block_group_cache_progress(block_group,
7465                                        num_bytes + empty_cluster + empty_size);
7466                                 goto have_block_group;
7467                         }
7468
7469                         /*
7470                          * at this point we either didn't find a cluster
7471                          * or we weren't able to allocate a block from our
7472                          * cluster.  Free the cluster we've been trying
7473                          * to use, and go to the next block group
7474                          */
7475                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7476                         spin_unlock(&last_ptr->refill_lock);
7477                         goto loop;
7478                 }
7479
7480 unclustered_alloc:
7481                 /*
7482                  * We are doing an unclustered alloc, set the fragmented flag so
7483                  * we don't bother trying to setup a cluster again until we get
7484                  * more space.
7485                  */
7486                 if (unlikely(last_ptr)) {
7487                         spin_lock(&last_ptr->lock);
7488                         last_ptr->fragmented = 1;
7489                         spin_unlock(&last_ptr->lock);
7490                 }
7491                 spin_lock(&block_group->free_space_ctl->tree_lock);
7492                 if (cached &&
7493                     block_group->free_space_ctl->free_space <
7494                     num_bytes + empty_cluster + empty_size) {
7495                         if (block_group->free_space_ctl->free_space >
7496                             max_extent_size)
7497                                 max_extent_size =
7498                                         block_group->free_space_ctl->free_space;
7499                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7500                         goto loop;
7501                 }
7502                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7503
7504                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7505                                                     num_bytes, empty_size,
7506                                                     &max_extent_size);
7507                 /*
7508                  * If we didn't find a chunk, and we haven't failed on this
7509                  * block group before, and this block group is in the middle of
7510                  * caching and we are ok with waiting, then go ahead and wait
7511                  * for progress to be made, and set failed_alloc to true.
7512                  *
7513                  * If failed_alloc is true then we've already waited on this
7514                  * block group once and should move on to the next block group.
7515                  */
7516                 if (!offset && !failed_alloc && !cached &&
7517                     loop > LOOP_CACHING_NOWAIT) {
7518                         wait_block_group_cache_progress(block_group,
7519                                                 num_bytes + empty_size);
7520                         failed_alloc = true;
7521                         goto have_block_group;
7522                 } else if (!offset) {
7523                         goto loop;
7524                 }
7525 checks:
7526                 search_start = ALIGN(offset, root->stripesize);
7527
7528                 /* move on to the next group */
7529                 if (search_start + num_bytes >
7530                     block_group->key.objectid + block_group->key.offset) {
7531                         btrfs_add_free_space(block_group, offset, num_bytes);
7532                         goto loop;
7533                 }
7534
7535                 if (offset < search_start)
7536                         btrfs_add_free_space(block_group, offset,
7537                                              search_start - offset);
7538                 BUG_ON(offset > search_start);
7539
7540                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7541                                                   alloc_type, delalloc);
7542                 if (ret == -EAGAIN) {
7543                         btrfs_add_free_space(block_group, offset, num_bytes);
7544                         goto loop;
7545                 }
7546                 btrfs_inc_block_group_reservations(block_group);
7547
7548                 /* we are all good, lets return */
7549                 ins->objectid = search_start;
7550                 ins->offset = num_bytes;
7551
7552                 trace_btrfs_reserve_extent(orig_root, block_group,
7553                                            search_start, num_bytes);
7554                 btrfs_release_block_group(block_group, delalloc);
7555                 break;
7556 loop:
7557                 failed_cluster_refill = false;
7558                 failed_alloc = false;
7559                 BUG_ON(index != get_block_group_index(block_group));
7560                 btrfs_release_block_group(block_group, delalloc);
7561         }
7562         up_read(&space_info->groups_sem);
7563
7564         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7565                 && !orig_have_caching_bg)
7566                 orig_have_caching_bg = true;
7567
7568         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7569                 goto search;
7570
7571         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7572                 goto search;
7573
7574         /*
7575          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7576          *                      caching kthreads as we move along
7577          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7578          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7579          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7580          *                      again
7581          */
7582         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7583                 index = 0;
7584                 if (loop == LOOP_CACHING_NOWAIT) {
7585                         /*
7586                          * We want to skip the LOOP_CACHING_WAIT step if we
7587                          * don't have any uncached bgs and we've already done a
7588                          * full search through.
7589                          */
7590                         if (orig_have_caching_bg || !full_search)
7591                                 loop = LOOP_CACHING_WAIT;
7592                         else
7593                                 loop = LOOP_ALLOC_CHUNK;
7594                 } else {
7595                         loop++;
7596                 }
7597
7598                 if (loop == LOOP_ALLOC_CHUNK) {
7599                         struct btrfs_trans_handle *trans;
7600                         int exist = 0;
7601
7602                         trans = current->journal_info;
7603                         if (trans)
7604                                 exist = 1;
7605                         else
7606                                 trans = btrfs_join_transaction(root);
7607
7608                         if (IS_ERR(trans)) {
7609                                 ret = PTR_ERR(trans);
7610                                 goto out;
7611                         }
7612
7613                         ret = do_chunk_alloc(trans, root, flags,
7614                                              CHUNK_ALLOC_FORCE);
7615
7616                         /*
7617                          * If we can't allocate a new chunk we've already looped
7618                          * through at least once, move on to the NO_EMPTY_SIZE
7619                          * case.
7620                          */
7621                         if (ret == -ENOSPC)
7622                                 loop = LOOP_NO_EMPTY_SIZE;
7623
7624                         /*
7625                          * Do not bail out on ENOSPC since we
7626                          * can do more things.
7627                          */
7628                         if (ret < 0 && ret != -ENOSPC)
7629                                 btrfs_abort_transaction(trans,
7630                                                         root, ret);
7631                         else
7632                                 ret = 0;
7633                         if (!exist)
7634                                 btrfs_end_transaction(trans, root);
7635                         if (ret)
7636                                 goto out;
7637                 }
7638
7639                 if (loop == LOOP_NO_EMPTY_SIZE) {
7640                         /*
7641                          * Don't loop again if we already have no empty_size and
7642                          * no empty_cluster.
7643                          */
7644                         if (empty_size == 0 &&
7645                             empty_cluster == 0) {
7646                                 ret = -ENOSPC;
7647                                 goto out;
7648                         }
7649                         empty_size = 0;
7650                         empty_cluster = 0;
7651                 }
7652
7653                 goto search;
7654         } else if (!ins->objectid) {
7655                 ret = -ENOSPC;
7656         } else if (ins->objectid) {
7657                 if (!use_cluster && last_ptr) {
7658                         spin_lock(&last_ptr->lock);
7659                         last_ptr->window_start = ins->objectid;
7660                         spin_unlock(&last_ptr->lock);
7661                 }
7662                 ret = 0;
7663         }
7664 out:
7665         if (ret == -ENOSPC) {
7666                 spin_lock(&space_info->lock);
7667                 space_info->max_extent_size = max_extent_size;
7668                 spin_unlock(&space_info->lock);
7669                 ins->offset = max_extent_size;
7670         }
7671         return ret;
7672 }
7673
7674 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7675                             int dump_block_groups)
7676 {
7677         struct btrfs_block_group_cache *cache;
7678         int index = 0;
7679
7680         spin_lock(&info->lock);
7681         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7682                info->flags,
7683                info->total_bytes - info->bytes_used - info->bytes_pinned -
7684                info->bytes_reserved - info->bytes_readonly,
7685                (info->full) ? "" : "not ");
7686         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7687                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7688                info->total_bytes, info->bytes_used, info->bytes_pinned,
7689                info->bytes_reserved, info->bytes_may_use,
7690                info->bytes_readonly);
7691         spin_unlock(&info->lock);
7692
7693         if (!dump_block_groups)
7694                 return;
7695
7696         down_read(&info->groups_sem);
7697 again:
7698         list_for_each_entry(cache, &info->block_groups[index], list) {
7699                 spin_lock(&cache->lock);
7700                 printk(KERN_INFO "BTRFS: "
7701                            "block group %llu has %llu bytes, "
7702                            "%llu used %llu pinned %llu reserved %s\n",
7703                        cache->key.objectid, cache->key.offset,
7704                        btrfs_block_group_used(&cache->item), cache->pinned,
7705                        cache->reserved, cache->ro ? "[readonly]" : "");
7706                 btrfs_dump_free_space(cache, bytes);
7707                 spin_unlock(&cache->lock);
7708         }
7709         if (++index < BTRFS_NR_RAID_TYPES)
7710                 goto again;
7711         up_read(&info->groups_sem);
7712 }
7713
7714 int btrfs_reserve_extent(struct btrfs_root *root,
7715                          u64 num_bytes, u64 min_alloc_size,
7716                          u64 empty_size, u64 hint_byte,
7717                          struct btrfs_key *ins, int is_data, int delalloc)
7718 {
7719         bool final_tried = num_bytes == min_alloc_size;
7720         u64 flags;
7721         int ret;
7722
7723         flags = btrfs_get_alloc_profile(root, is_data);
7724 again:
7725         WARN_ON(num_bytes < root->sectorsize);
7726         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7727                                flags, delalloc);
7728         if (!ret && !is_data) {
7729                 btrfs_dec_block_group_reservations(root->fs_info,
7730                                                    ins->objectid);
7731         } else if (ret == -ENOSPC) {
7732                 if (!final_tried && ins->offset) {
7733                         num_bytes = min(num_bytes >> 1, ins->offset);
7734                         num_bytes = round_down(num_bytes, root->sectorsize);
7735                         num_bytes = max(num_bytes, min_alloc_size);
7736                         if (num_bytes == min_alloc_size)
7737                                 final_tried = true;
7738                         goto again;
7739                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7740                         struct btrfs_space_info *sinfo;
7741
7742                         sinfo = __find_space_info(root->fs_info, flags);
7743                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7744                                 flags, num_bytes);
7745                         if (sinfo)
7746                                 dump_space_info(sinfo, num_bytes, 1);
7747                 }
7748         }
7749
7750         return ret;
7751 }
7752
7753 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7754                                         u64 start, u64 len,
7755                                         int pin, int delalloc)
7756 {
7757         struct btrfs_block_group_cache *cache;
7758         int ret = 0;
7759
7760         cache = btrfs_lookup_block_group(root->fs_info, start);
7761         if (!cache) {
7762                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7763                         start);
7764                 return -ENOSPC;
7765         }
7766
7767         if (pin)
7768                 pin_down_extent(root, cache, start, len, 1);
7769         else {
7770                 if (btrfs_test_opt(root, DISCARD))
7771                         ret = btrfs_discard_extent(root, start, len, NULL);
7772                 btrfs_add_free_space(cache, start, len);
7773                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7774         }
7775
7776         btrfs_put_block_group(cache);
7777
7778         trace_btrfs_reserved_extent_free(root, start, len);
7779
7780         return ret;
7781 }
7782
7783 int btrfs_free_reserved_extent(struct btrfs_root *root,
7784                                u64 start, u64 len, int delalloc)
7785 {
7786         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7787 }
7788
7789 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7790                                        u64 start, u64 len)
7791 {
7792         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7793 }
7794
7795 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7796                                       struct btrfs_root *root,
7797                                       u64 parent, u64 root_objectid,
7798                                       u64 flags, u64 owner, u64 offset,
7799                                       struct btrfs_key *ins, int ref_mod)
7800 {
7801         int ret;
7802         struct btrfs_fs_info *fs_info = root->fs_info;
7803         struct btrfs_extent_item *extent_item;
7804         struct btrfs_extent_inline_ref *iref;
7805         struct btrfs_path *path;
7806         struct extent_buffer *leaf;
7807         int type;
7808         u32 size;
7809
7810         if (parent > 0)
7811                 type = BTRFS_SHARED_DATA_REF_KEY;
7812         else
7813                 type = BTRFS_EXTENT_DATA_REF_KEY;
7814
7815         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7816
7817         path = btrfs_alloc_path();
7818         if (!path)
7819                 return -ENOMEM;
7820
7821         path->leave_spinning = 1;
7822         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7823                                       ins, size);
7824         if (ret) {
7825                 btrfs_free_path(path);
7826                 return ret;
7827         }
7828
7829         leaf = path->nodes[0];
7830         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7831                                      struct btrfs_extent_item);
7832         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7833         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7834         btrfs_set_extent_flags(leaf, extent_item,
7835                                flags | BTRFS_EXTENT_FLAG_DATA);
7836
7837         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7838         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7839         if (parent > 0) {
7840                 struct btrfs_shared_data_ref *ref;
7841                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7842                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7843                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7844         } else {
7845                 struct btrfs_extent_data_ref *ref;
7846                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7847                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7848                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7849                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7850                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7851         }
7852
7853         btrfs_mark_buffer_dirty(path->nodes[0]);
7854         btrfs_free_path(path);
7855
7856         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
7857                                           ins->offset);
7858         if (ret)
7859                 return ret;
7860
7861         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7862         if (ret) { /* -ENOENT, logic error */
7863                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7864                         ins->objectid, ins->offset);
7865                 BUG();
7866         }
7867         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7868         return ret;
7869 }
7870
7871 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7872                                      struct btrfs_root *root,
7873                                      u64 parent, u64 root_objectid,
7874                                      u64 flags, struct btrfs_disk_key *key,
7875                                      int level, struct btrfs_key *ins)
7876 {
7877         int ret;
7878         struct btrfs_fs_info *fs_info = root->fs_info;
7879         struct btrfs_extent_item *extent_item;
7880         struct btrfs_tree_block_info *block_info;
7881         struct btrfs_extent_inline_ref *iref;
7882         struct btrfs_path *path;
7883         struct extent_buffer *leaf;
7884         u32 size = sizeof(*extent_item) + sizeof(*iref);
7885         u64 num_bytes = ins->offset;
7886         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7887                                                  SKINNY_METADATA);
7888
7889         if (!skinny_metadata)
7890                 size += sizeof(*block_info);
7891
7892         path = btrfs_alloc_path();
7893         if (!path) {
7894                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7895                                                    root->nodesize);
7896                 return -ENOMEM;
7897         }
7898
7899         path->leave_spinning = 1;
7900         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7901                                       ins, size);
7902         if (ret) {
7903                 btrfs_free_path(path);
7904                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7905                                                    root->nodesize);
7906                 return ret;
7907         }
7908
7909         leaf = path->nodes[0];
7910         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7911                                      struct btrfs_extent_item);
7912         btrfs_set_extent_refs(leaf, extent_item, 1);
7913         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7914         btrfs_set_extent_flags(leaf, extent_item,
7915                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7916
7917         if (skinny_metadata) {
7918                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7919                 num_bytes = root->nodesize;
7920         } else {
7921                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7922                 btrfs_set_tree_block_key(leaf, block_info, key);
7923                 btrfs_set_tree_block_level(leaf, block_info, level);
7924                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7925         }
7926
7927         if (parent > 0) {
7928                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7929                 btrfs_set_extent_inline_ref_type(leaf, iref,
7930                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7931                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7932         } else {
7933                 btrfs_set_extent_inline_ref_type(leaf, iref,
7934                                                  BTRFS_TREE_BLOCK_REF_KEY);
7935                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7936         }
7937
7938         btrfs_mark_buffer_dirty(leaf);
7939         btrfs_free_path(path);
7940
7941         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
7942                                           num_bytes);
7943         if (ret)
7944                 return ret;
7945
7946         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7947                                  1);
7948         if (ret) { /* -ENOENT, logic error */
7949                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7950                         ins->objectid, ins->offset);
7951                 BUG();
7952         }
7953
7954         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7955         return ret;
7956 }
7957
7958 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7959                                      struct btrfs_root *root,
7960                                      u64 root_objectid, u64 owner,
7961                                      u64 offset, u64 ram_bytes,
7962                                      struct btrfs_key *ins)
7963 {
7964         int ret;
7965
7966         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7967
7968         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7969                                          ins->offset, 0,
7970                                          root_objectid, owner, offset,
7971                                          ram_bytes, BTRFS_ADD_DELAYED_EXTENT,
7972                                          NULL);
7973         return ret;
7974 }
7975
7976 /*
7977  * this is used by the tree logging recovery code.  It records that
7978  * an extent has been allocated and makes sure to clear the free
7979  * space cache bits as well
7980  */
7981 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7982                                    struct btrfs_root *root,
7983                                    u64 root_objectid, u64 owner, u64 offset,
7984                                    struct btrfs_key *ins)
7985 {
7986         int ret;
7987         struct btrfs_block_group_cache *block_group;
7988
7989         /*
7990          * Mixed block groups will exclude before processing the log so we only
7991          * need to do the exclude dance if this fs isn't mixed.
7992          */
7993         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7994                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7995                 if (ret)
7996                         return ret;
7997         }
7998
7999         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
8000         if (!block_group)
8001                 return -EINVAL;
8002
8003         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
8004                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
8005         BUG_ON(ret); /* logic error */
8006         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
8007                                          0, owner, offset, ins, 1);
8008         btrfs_put_block_group(block_group);
8009         return ret;
8010 }
8011
8012 static struct extent_buffer *
8013 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8014                       u64 bytenr, int level)
8015 {
8016         struct extent_buffer *buf;
8017
8018         buf = btrfs_find_create_tree_block(root, bytenr);
8019         if (IS_ERR(buf))
8020                 return buf;
8021
8022         btrfs_set_header_generation(buf, trans->transid);
8023         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8024         btrfs_tree_lock(buf);
8025         clean_tree_block(trans, root->fs_info, buf);
8026         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8027
8028         btrfs_set_lock_blocking(buf);
8029         set_extent_buffer_uptodate(buf);
8030
8031         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8032                 buf->log_index = root->log_transid % 2;
8033                 /*
8034                  * we allow two log transactions at a time, use different
8035                  * EXENT bit to differentiate dirty pages.
8036                  */
8037                 if (buf->log_index == 0)
8038                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8039                                         buf->start + buf->len - 1, GFP_NOFS);
8040                 else
8041                         set_extent_new(&root->dirty_log_pages, buf->start,
8042                                         buf->start + buf->len - 1);
8043         } else {
8044                 buf->log_index = -1;
8045                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8046                          buf->start + buf->len - 1, GFP_NOFS);
8047         }
8048         trans->dirty = true;
8049         /* this returns a buffer locked for blocking */
8050         return buf;
8051 }
8052
8053 static struct btrfs_block_rsv *
8054 use_block_rsv(struct btrfs_trans_handle *trans,
8055               struct btrfs_root *root, u32 blocksize)
8056 {
8057         struct btrfs_block_rsv *block_rsv;
8058         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
8059         int ret;
8060         bool global_updated = false;
8061
8062         block_rsv = get_block_rsv(trans, root);
8063
8064         if (unlikely(block_rsv->size == 0))
8065                 goto try_reserve;
8066 again:
8067         ret = block_rsv_use_bytes(block_rsv, blocksize);
8068         if (!ret)
8069                 return block_rsv;
8070
8071         if (block_rsv->failfast)
8072                 return ERR_PTR(ret);
8073
8074         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8075                 global_updated = true;
8076                 update_global_block_rsv(root->fs_info);
8077                 goto again;
8078         }
8079
8080         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
8081                 static DEFINE_RATELIMIT_STATE(_rs,
8082                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8083                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8084                 if (__ratelimit(&_rs))
8085                         WARN(1, KERN_DEBUG
8086                                 "BTRFS: block rsv returned %d\n", ret);
8087         }
8088 try_reserve:
8089         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8090                                      BTRFS_RESERVE_NO_FLUSH);
8091         if (!ret)
8092                 return block_rsv;
8093         /*
8094          * If we couldn't reserve metadata bytes try and use some from
8095          * the global reserve if its space type is the same as the global
8096          * reservation.
8097          */
8098         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8099             block_rsv->space_info == global_rsv->space_info) {
8100                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8101                 if (!ret)
8102                         return global_rsv;
8103         }
8104         return ERR_PTR(ret);
8105 }
8106
8107 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8108                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8109 {
8110         block_rsv_add_bytes(block_rsv, blocksize, 0);
8111         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8112 }
8113
8114 /*
8115  * finds a free extent and does all the dirty work required for allocation
8116  * returns the tree buffer or an ERR_PTR on error.
8117  */
8118 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8119                                         struct btrfs_root *root,
8120                                         u64 parent, u64 root_objectid,
8121                                         struct btrfs_disk_key *key, int level,
8122                                         u64 hint, u64 empty_size)
8123 {
8124         struct btrfs_key ins;
8125         struct btrfs_block_rsv *block_rsv;
8126         struct extent_buffer *buf;
8127         struct btrfs_delayed_extent_op *extent_op;
8128         u64 flags = 0;
8129         int ret;
8130         u32 blocksize = root->nodesize;
8131         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8132                                                  SKINNY_METADATA);
8133
8134         if (btrfs_test_is_dummy_root(root)) {
8135                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8136                                             level);
8137                 if (!IS_ERR(buf))
8138                         root->alloc_bytenr += blocksize;
8139                 return buf;
8140         }
8141
8142         block_rsv = use_block_rsv(trans, root, blocksize);
8143         if (IS_ERR(block_rsv))
8144                 return ERR_CAST(block_rsv);
8145
8146         ret = btrfs_reserve_extent(root, blocksize, blocksize,
8147                                    empty_size, hint, &ins, 0, 0);
8148         if (ret)
8149                 goto out_unuse;
8150
8151         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8152         if (IS_ERR(buf)) {
8153                 ret = PTR_ERR(buf);
8154                 goto out_free_reserved;
8155         }
8156
8157         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8158                 if (parent == 0)
8159                         parent = ins.objectid;
8160                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8161         } else
8162                 BUG_ON(parent > 0);
8163
8164         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8165                 extent_op = btrfs_alloc_delayed_extent_op();
8166                 if (!extent_op) {
8167                         ret = -ENOMEM;
8168                         goto out_free_buf;
8169                 }
8170                 if (key)
8171                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8172                 else
8173                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8174                 extent_op->flags_to_set = flags;
8175                 extent_op->update_key = skinny_metadata ? false : true;
8176                 extent_op->update_flags = true;
8177                 extent_op->is_data = false;
8178                 extent_op->level = level;
8179
8180                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
8181                                                  ins.objectid, ins.offset,
8182                                                  parent, root_objectid, level,
8183                                                  BTRFS_ADD_DELAYED_EXTENT,
8184                                                  extent_op);
8185                 if (ret)
8186                         goto out_free_delayed;
8187         }
8188         return buf;
8189
8190 out_free_delayed:
8191         btrfs_free_delayed_extent_op(extent_op);
8192 out_free_buf:
8193         free_extent_buffer(buf);
8194 out_free_reserved:
8195         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
8196 out_unuse:
8197         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
8198         return ERR_PTR(ret);
8199 }
8200
8201 struct walk_control {
8202         u64 refs[BTRFS_MAX_LEVEL];
8203         u64 flags[BTRFS_MAX_LEVEL];
8204         struct btrfs_key update_progress;
8205         int stage;
8206         int level;
8207         int shared_level;
8208         int update_ref;
8209         int keep_locks;
8210         int reada_slot;
8211         int reada_count;
8212         int for_reloc;
8213 };
8214
8215 #define DROP_REFERENCE  1
8216 #define UPDATE_BACKREF  2
8217
8218 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8219                                      struct btrfs_root *root,
8220                                      struct walk_control *wc,
8221                                      struct btrfs_path *path)
8222 {
8223         u64 bytenr;
8224         u64 generation;
8225         u64 refs;
8226         u64 flags;
8227         u32 nritems;
8228         u32 blocksize;
8229         struct btrfs_key key;
8230         struct extent_buffer *eb;
8231         int ret;
8232         int slot;
8233         int nread = 0;
8234
8235         if (path->slots[wc->level] < wc->reada_slot) {
8236                 wc->reada_count = wc->reada_count * 2 / 3;
8237                 wc->reada_count = max(wc->reada_count, 2);
8238         } else {
8239                 wc->reada_count = wc->reada_count * 3 / 2;
8240                 wc->reada_count = min_t(int, wc->reada_count,
8241                                         BTRFS_NODEPTRS_PER_BLOCK(root));
8242         }
8243
8244         eb = path->nodes[wc->level];
8245         nritems = btrfs_header_nritems(eb);
8246         blocksize = root->nodesize;
8247
8248         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8249                 if (nread >= wc->reada_count)
8250                         break;
8251
8252                 cond_resched();
8253                 bytenr = btrfs_node_blockptr(eb, slot);
8254                 generation = btrfs_node_ptr_generation(eb, slot);
8255
8256                 if (slot == path->slots[wc->level])
8257                         goto reada;
8258
8259                 if (wc->stage == UPDATE_BACKREF &&
8260                     generation <= root->root_key.offset)
8261                         continue;
8262
8263                 /* We don't lock the tree block, it's OK to be racy here */
8264                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
8265                                                wc->level - 1, 1, &refs,
8266                                                &flags);
8267                 /* We don't care about errors in readahead. */
8268                 if (ret < 0)
8269                         continue;
8270                 BUG_ON(refs == 0);
8271
8272                 if (wc->stage == DROP_REFERENCE) {
8273                         if (refs == 1)
8274                                 goto reada;
8275
8276                         if (wc->level == 1 &&
8277                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8278                                 continue;
8279                         if (!wc->update_ref ||
8280                             generation <= root->root_key.offset)
8281                                 continue;
8282                         btrfs_node_key_to_cpu(eb, &key, slot);
8283                         ret = btrfs_comp_cpu_keys(&key,
8284                                                   &wc->update_progress);
8285                         if (ret < 0)
8286                                 continue;
8287                 } else {
8288                         if (wc->level == 1 &&
8289                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8290                                 continue;
8291                 }
8292 reada:
8293                 readahead_tree_block(root, bytenr);
8294                 nread++;
8295         }
8296         wc->reada_slot = slot;
8297 }
8298
8299 /*
8300  * These may not be seen by the usual inc/dec ref code so we have to
8301  * add them here.
8302  */
8303 static int record_one_subtree_extent(struct btrfs_trans_handle *trans,
8304                                      struct btrfs_root *root, u64 bytenr,
8305                                      u64 num_bytes)
8306 {
8307         struct btrfs_qgroup_extent_record *qrecord;
8308         struct btrfs_delayed_ref_root *delayed_refs;
8309
8310         qrecord = kmalloc(sizeof(*qrecord), GFP_NOFS);
8311         if (!qrecord)
8312                 return -ENOMEM;
8313
8314         qrecord->bytenr = bytenr;
8315         qrecord->num_bytes = num_bytes;
8316         qrecord->old_roots = NULL;
8317
8318         delayed_refs = &trans->transaction->delayed_refs;
8319         spin_lock(&delayed_refs->lock);
8320         if (btrfs_qgroup_insert_dirty_extent(delayed_refs, qrecord))
8321                 kfree(qrecord);
8322         spin_unlock(&delayed_refs->lock);
8323
8324         return 0;
8325 }
8326
8327 static int account_leaf_items(struct btrfs_trans_handle *trans,
8328                               struct btrfs_root *root,
8329                               struct extent_buffer *eb)
8330 {
8331         int nr = btrfs_header_nritems(eb);
8332         int i, extent_type, ret;
8333         struct btrfs_key key;
8334         struct btrfs_file_extent_item *fi;
8335         u64 bytenr, num_bytes;
8336
8337         /* We can be called directly from walk_up_proc() */
8338         if (!root->fs_info->quota_enabled)
8339                 return 0;
8340
8341         for (i = 0; i < nr; i++) {
8342                 btrfs_item_key_to_cpu(eb, &key, i);
8343
8344                 if (key.type != BTRFS_EXTENT_DATA_KEY)
8345                         continue;
8346
8347                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
8348                 /* filter out non qgroup-accountable extents  */
8349                 extent_type = btrfs_file_extent_type(eb, fi);
8350
8351                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
8352                         continue;
8353
8354                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
8355                 if (!bytenr)
8356                         continue;
8357
8358                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
8359
8360                 ret = record_one_subtree_extent(trans, root, bytenr, num_bytes);
8361                 if (ret)
8362                         return ret;
8363         }
8364         return 0;
8365 }
8366
8367 /*
8368  * Walk up the tree from the bottom, freeing leaves and any interior
8369  * nodes which have had all slots visited. If a node (leaf or
8370  * interior) is freed, the node above it will have it's slot
8371  * incremented. The root node will never be freed.
8372  *
8373  * At the end of this function, we should have a path which has all
8374  * slots incremented to the next position for a search. If we need to
8375  * read a new node it will be NULL and the node above it will have the
8376  * correct slot selected for a later read.
8377  *
8378  * If we increment the root nodes slot counter past the number of
8379  * elements, 1 is returned to signal completion of the search.
8380  */
8381 static int adjust_slots_upwards(struct btrfs_root *root,
8382                                 struct btrfs_path *path, int root_level)
8383 {
8384         int level = 0;
8385         int nr, slot;
8386         struct extent_buffer *eb;
8387
8388         if (root_level == 0)
8389                 return 1;
8390
8391         while (level <= root_level) {
8392                 eb = path->nodes[level];
8393                 nr = btrfs_header_nritems(eb);
8394                 path->slots[level]++;
8395                 slot = path->slots[level];
8396                 if (slot >= nr || level == 0) {
8397                         /*
8398                          * Don't free the root -  we will detect this
8399                          * condition after our loop and return a
8400                          * positive value for caller to stop walking the tree.
8401                          */
8402                         if (level != root_level) {
8403                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8404                                 path->locks[level] = 0;
8405
8406                                 free_extent_buffer(eb);
8407                                 path->nodes[level] = NULL;
8408                                 path->slots[level] = 0;
8409                         }
8410                 } else {
8411                         /*
8412                          * We have a valid slot to walk back down
8413                          * from. Stop here so caller can process these
8414                          * new nodes.
8415                          */
8416                         break;
8417                 }
8418
8419                 level++;
8420         }
8421
8422         eb = path->nodes[root_level];
8423         if (path->slots[root_level] >= btrfs_header_nritems(eb))
8424                 return 1;
8425
8426         return 0;
8427 }
8428
8429 /*
8430  * root_eb is the subtree root and is locked before this function is called.
8431  */
8432 static int account_shared_subtree(struct btrfs_trans_handle *trans,
8433                                   struct btrfs_root *root,
8434                                   struct extent_buffer *root_eb,
8435                                   u64 root_gen,
8436                                   int root_level)
8437 {
8438         int ret = 0;
8439         int level;
8440         struct extent_buffer *eb = root_eb;
8441         struct btrfs_path *path = NULL;
8442
8443         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
8444         BUG_ON(root_eb == NULL);
8445
8446         if (!root->fs_info->quota_enabled)
8447                 return 0;
8448
8449         if (!extent_buffer_uptodate(root_eb)) {
8450                 ret = btrfs_read_buffer(root_eb, root_gen);
8451                 if (ret)
8452                         goto out;
8453         }
8454
8455         if (root_level == 0) {
8456                 ret = account_leaf_items(trans, root, root_eb);
8457                 goto out;
8458         }
8459
8460         path = btrfs_alloc_path();
8461         if (!path)
8462                 return -ENOMEM;
8463
8464         /*
8465          * Walk down the tree.  Missing extent blocks are filled in as
8466          * we go. Metadata is accounted every time we read a new
8467          * extent block.
8468          *
8469          * When we reach a leaf, we account for file extent items in it,
8470          * walk back up the tree (adjusting slot pointers as we go)
8471          * and restart the search process.
8472          */
8473         extent_buffer_get(root_eb); /* For path */
8474         path->nodes[root_level] = root_eb;
8475         path->slots[root_level] = 0;
8476         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
8477 walk_down:
8478         level = root_level;
8479         while (level >= 0) {
8480                 if (path->nodes[level] == NULL) {
8481                         int parent_slot;
8482                         u64 child_gen;
8483                         u64 child_bytenr;
8484
8485                         /* We need to get child blockptr/gen from
8486                          * parent before we can read it. */
8487                         eb = path->nodes[level + 1];
8488                         parent_slot = path->slots[level + 1];
8489                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8490                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8491
8492                         eb = read_tree_block(root, child_bytenr, child_gen);
8493                         if (IS_ERR(eb)) {
8494                                 ret = PTR_ERR(eb);
8495                                 goto out;
8496                         } else if (!extent_buffer_uptodate(eb)) {
8497                                 free_extent_buffer(eb);
8498                                 ret = -EIO;
8499                                 goto out;
8500                         }
8501
8502                         path->nodes[level] = eb;
8503                         path->slots[level] = 0;
8504
8505                         btrfs_tree_read_lock(eb);
8506                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8507                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8508
8509                         ret = record_one_subtree_extent(trans, root, child_bytenr,
8510                                                         root->nodesize);
8511                         if (ret)
8512                                 goto out;
8513                 }
8514
8515                 if (level == 0) {
8516                         ret = account_leaf_items(trans, root, path->nodes[level]);
8517                         if (ret)
8518                                 goto out;
8519
8520                         /* Nonzero return here means we completed our search */
8521                         ret = adjust_slots_upwards(root, path, root_level);
8522                         if (ret)
8523                                 break;
8524
8525                         /* Restart search with new slots */
8526                         goto walk_down;
8527                 }
8528
8529                 level--;
8530         }
8531
8532         ret = 0;
8533 out:
8534         btrfs_free_path(path);
8535
8536         return ret;
8537 }
8538
8539 /*
8540  * helper to process tree block while walking down the tree.
8541  *
8542  * when wc->stage == UPDATE_BACKREF, this function updates
8543  * back refs for pointers in the block.
8544  *
8545  * NOTE: return value 1 means we should stop walking down.
8546  */
8547 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8548                                    struct btrfs_root *root,
8549                                    struct btrfs_path *path,
8550                                    struct walk_control *wc, int lookup_info)
8551 {
8552         int level = wc->level;
8553         struct extent_buffer *eb = path->nodes[level];
8554         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8555         int ret;
8556
8557         if (wc->stage == UPDATE_BACKREF &&
8558             btrfs_header_owner(eb) != root->root_key.objectid)
8559                 return 1;
8560
8561         /*
8562          * when reference count of tree block is 1, it won't increase
8563          * again. once full backref flag is set, we never clear it.
8564          */
8565         if (lookup_info &&
8566             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8567              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8568                 BUG_ON(!path->locks[level]);
8569                 ret = btrfs_lookup_extent_info(trans, root,
8570                                                eb->start, level, 1,
8571                                                &wc->refs[level],
8572                                                &wc->flags[level]);
8573                 BUG_ON(ret == -ENOMEM);
8574                 if (ret)
8575                         return ret;
8576                 BUG_ON(wc->refs[level] == 0);
8577         }
8578
8579         if (wc->stage == DROP_REFERENCE) {
8580                 if (wc->refs[level] > 1)
8581                         return 1;
8582
8583                 if (path->locks[level] && !wc->keep_locks) {
8584                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8585                         path->locks[level] = 0;
8586                 }
8587                 return 0;
8588         }
8589
8590         /* wc->stage == UPDATE_BACKREF */
8591         if (!(wc->flags[level] & flag)) {
8592                 BUG_ON(!path->locks[level]);
8593                 ret = btrfs_inc_ref(trans, root, eb, 1);
8594                 BUG_ON(ret); /* -ENOMEM */
8595                 ret = btrfs_dec_ref(trans, root, eb, 0);
8596                 BUG_ON(ret); /* -ENOMEM */
8597                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8598                                                   eb->len, flag,
8599                                                   btrfs_header_level(eb), 0);
8600                 BUG_ON(ret); /* -ENOMEM */
8601                 wc->flags[level] |= flag;
8602         }
8603
8604         /*
8605          * the block is shared by multiple trees, so it's not good to
8606          * keep the tree lock
8607          */
8608         if (path->locks[level] && level > 0) {
8609                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8610                 path->locks[level] = 0;
8611         }
8612         return 0;
8613 }
8614
8615 /*
8616  * helper to process tree block pointer.
8617  *
8618  * when wc->stage == DROP_REFERENCE, this function checks
8619  * reference count of the block pointed to. if the block
8620  * is shared and we need update back refs for the subtree
8621  * rooted at the block, this function changes wc->stage to
8622  * UPDATE_BACKREF. if the block is shared and there is no
8623  * need to update back, this function drops the reference
8624  * to the block.
8625  *
8626  * NOTE: return value 1 means we should stop walking down.
8627  */
8628 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8629                                  struct btrfs_root *root,
8630                                  struct btrfs_path *path,
8631                                  struct walk_control *wc, int *lookup_info)
8632 {
8633         u64 bytenr;
8634         u64 generation;
8635         u64 parent;
8636         u32 blocksize;
8637         struct btrfs_key key;
8638         struct extent_buffer *next;
8639         int level = wc->level;
8640         int reada = 0;
8641         int ret = 0;
8642         bool need_account = false;
8643
8644         generation = btrfs_node_ptr_generation(path->nodes[level],
8645                                                path->slots[level]);
8646         /*
8647          * if the lower level block was created before the snapshot
8648          * was created, we know there is no need to update back refs
8649          * for the subtree
8650          */
8651         if (wc->stage == UPDATE_BACKREF &&
8652             generation <= root->root_key.offset) {
8653                 *lookup_info = 1;
8654                 return 1;
8655         }
8656
8657         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8658         blocksize = root->nodesize;
8659
8660         next = btrfs_find_tree_block(root->fs_info, bytenr);
8661         if (!next) {
8662                 next = btrfs_find_create_tree_block(root, bytenr);
8663                 if (IS_ERR(next))
8664                         return PTR_ERR(next);
8665
8666                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8667                                                level - 1);
8668                 reada = 1;
8669         }
8670         btrfs_tree_lock(next);
8671         btrfs_set_lock_blocking(next);
8672
8673         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8674                                        &wc->refs[level - 1],
8675                                        &wc->flags[level - 1]);
8676         if (ret < 0) {
8677                 btrfs_tree_unlock(next);
8678                 return ret;
8679         }
8680
8681         if (unlikely(wc->refs[level - 1] == 0)) {
8682                 btrfs_err(root->fs_info, "Missing references.");
8683                 BUG();
8684         }
8685         *lookup_info = 0;
8686
8687         if (wc->stage == DROP_REFERENCE) {
8688                 if (wc->refs[level - 1] > 1) {
8689                         need_account = true;
8690                         if (level == 1 &&
8691                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8692                                 goto skip;
8693
8694                         if (!wc->update_ref ||
8695                             generation <= root->root_key.offset)
8696                                 goto skip;
8697
8698                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8699                                               path->slots[level]);
8700                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8701                         if (ret < 0)
8702                                 goto skip;
8703
8704                         wc->stage = UPDATE_BACKREF;
8705                         wc->shared_level = level - 1;
8706                 }
8707         } else {
8708                 if (level == 1 &&
8709                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8710                         goto skip;
8711         }
8712
8713         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8714                 btrfs_tree_unlock(next);
8715                 free_extent_buffer(next);
8716                 next = NULL;
8717                 *lookup_info = 1;
8718         }
8719
8720         if (!next) {
8721                 if (reada && level == 1)
8722                         reada_walk_down(trans, root, wc, path);
8723                 next = read_tree_block(root, bytenr, generation);
8724                 if (IS_ERR(next)) {
8725                         return PTR_ERR(next);
8726                 } else if (!extent_buffer_uptodate(next)) {
8727                         free_extent_buffer(next);
8728                         return -EIO;
8729                 }
8730                 btrfs_tree_lock(next);
8731                 btrfs_set_lock_blocking(next);
8732         }
8733
8734         level--;
8735         BUG_ON(level != btrfs_header_level(next));
8736         path->nodes[level] = next;
8737         path->slots[level] = 0;
8738         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8739         wc->level = level;
8740         if (wc->level == 1)
8741                 wc->reada_slot = 0;
8742         return 0;
8743 skip:
8744         wc->refs[level - 1] = 0;
8745         wc->flags[level - 1] = 0;
8746         if (wc->stage == DROP_REFERENCE) {
8747                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8748                         parent = path->nodes[level]->start;
8749                 } else {
8750                         BUG_ON(root->root_key.objectid !=
8751                                btrfs_header_owner(path->nodes[level]));
8752                         parent = 0;
8753                 }
8754
8755                 if (need_account) {
8756                         ret = account_shared_subtree(trans, root, next,
8757                                                      generation, level - 1);
8758                         if (ret) {
8759                                 btrfs_err_rl(root->fs_info,
8760                                         "Error "
8761                                         "%d accounting shared subtree. Quota "
8762                                         "is out of sync, rescan required.",
8763                                         ret);
8764                         }
8765                 }
8766                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8767                                 root->root_key.objectid, level - 1, 0);
8768                 BUG_ON(ret); /* -ENOMEM */
8769         }
8770         btrfs_tree_unlock(next);
8771         free_extent_buffer(next);
8772         *lookup_info = 1;
8773         return 1;
8774 }
8775
8776 /*
8777  * helper to process tree block while walking up the tree.
8778  *
8779  * when wc->stage == DROP_REFERENCE, this function drops
8780  * reference count on the block.
8781  *
8782  * when wc->stage == UPDATE_BACKREF, this function changes
8783  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8784  * to UPDATE_BACKREF previously while processing the block.
8785  *
8786  * NOTE: return value 1 means we should stop walking up.
8787  */
8788 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8789                                  struct btrfs_root *root,
8790                                  struct btrfs_path *path,
8791                                  struct walk_control *wc)
8792 {
8793         int ret;
8794         int level = wc->level;
8795         struct extent_buffer *eb = path->nodes[level];
8796         u64 parent = 0;
8797
8798         if (wc->stage == UPDATE_BACKREF) {
8799                 BUG_ON(wc->shared_level < level);
8800                 if (level < wc->shared_level)
8801                         goto out;
8802
8803                 ret = find_next_key(path, level + 1, &wc->update_progress);
8804                 if (ret > 0)
8805                         wc->update_ref = 0;
8806
8807                 wc->stage = DROP_REFERENCE;
8808                 wc->shared_level = -1;
8809                 path->slots[level] = 0;
8810
8811                 /*
8812                  * check reference count again if the block isn't locked.
8813                  * we should start walking down the tree again if reference
8814                  * count is one.
8815                  */
8816                 if (!path->locks[level]) {
8817                         BUG_ON(level == 0);
8818                         btrfs_tree_lock(eb);
8819                         btrfs_set_lock_blocking(eb);
8820                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8821
8822                         ret = btrfs_lookup_extent_info(trans, root,
8823                                                        eb->start, level, 1,
8824                                                        &wc->refs[level],
8825                                                        &wc->flags[level]);
8826                         if (ret < 0) {
8827                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8828                                 path->locks[level] = 0;
8829                                 return ret;
8830                         }
8831                         BUG_ON(wc->refs[level] == 0);
8832                         if (wc->refs[level] == 1) {
8833                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8834                                 path->locks[level] = 0;
8835                                 return 1;
8836                         }
8837                 }
8838         }
8839
8840         /* wc->stage == DROP_REFERENCE */
8841         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8842
8843         if (wc->refs[level] == 1) {
8844                 if (level == 0) {
8845                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8846                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8847                         else
8848                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8849                         BUG_ON(ret); /* -ENOMEM */
8850                         ret = account_leaf_items(trans, root, eb);
8851                         if (ret) {
8852                                 btrfs_err_rl(root->fs_info,
8853                                         "error "
8854                                         "%d accounting leaf items. Quota "
8855                                         "is out of sync, rescan required.",
8856                                         ret);
8857                         }
8858                 }
8859                 /* make block locked assertion in clean_tree_block happy */
8860                 if (!path->locks[level] &&
8861                     btrfs_header_generation(eb) == trans->transid) {
8862                         btrfs_tree_lock(eb);
8863                         btrfs_set_lock_blocking(eb);
8864                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8865                 }
8866                 clean_tree_block(trans, root->fs_info, eb);
8867         }
8868
8869         if (eb == root->node) {
8870                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8871                         parent = eb->start;
8872                 else
8873                         BUG_ON(root->root_key.objectid !=
8874                                btrfs_header_owner(eb));
8875         } else {
8876                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8877                         parent = path->nodes[level + 1]->start;
8878                 else
8879                         BUG_ON(root->root_key.objectid !=
8880                                btrfs_header_owner(path->nodes[level + 1]));
8881         }
8882
8883         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8884 out:
8885         wc->refs[level] = 0;
8886         wc->flags[level] = 0;
8887         return 0;
8888 }
8889
8890 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8891                                    struct btrfs_root *root,
8892                                    struct btrfs_path *path,
8893                                    struct walk_control *wc)
8894 {
8895         int level = wc->level;
8896         int lookup_info = 1;
8897         int ret;
8898
8899         while (level >= 0) {
8900                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8901                 if (ret > 0)
8902                         break;
8903
8904                 if (level == 0)
8905                         break;
8906
8907                 if (path->slots[level] >=
8908                     btrfs_header_nritems(path->nodes[level]))
8909                         break;
8910
8911                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8912                 if (ret > 0) {
8913                         path->slots[level]++;
8914                         continue;
8915                 } else if (ret < 0)
8916                         return ret;
8917                 level = wc->level;
8918         }
8919         return 0;
8920 }
8921
8922 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8923                                  struct btrfs_root *root,
8924                                  struct btrfs_path *path,
8925                                  struct walk_control *wc, int max_level)
8926 {
8927         int level = wc->level;
8928         int ret;
8929
8930         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8931         while (level < max_level && path->nodes[level]) {
8932                 wc->level = level;
8933                 if (path->slots[level] + 1 <
8934                     btrfs_header_nritems(path->nodes[level])) {
8935                         path->slots[level]++;
8936                         return 0;
8937                 } else {
8938                         ret = walk_up_proc(trans, root, path, wc);
8939                         if (ret > 0)
8940                                 return 0;
8941
8942                         if (path->locks[level]) {
8943                                 btrfs_tree_unlock_rw(path->nodes[level],
8944                                                      path->locks[level]);
8945                                 path->locks[level] = 0;
8946                         }
8947                         free_extent_buffer(path->nodes[level]);
8948                         path->nodes[level] = NULL;
8949                         level++;
8950                 }
8951         }
8952         return 1;
8953 }
8954
8955 /*
8956  * drop a subvolume tree.
8957  *
8958  * this function traverses the tree freeing any blocks that only
8959  * referenced by the tree.
8960  *
8961  * when a shared tree block is found. this function decreases its
8962  * reference count by one. if update_ref is true, this function
8963  * also make sure backrefs for the shared block and all lower level
8964  * blocks are properly updated.
8965  *
8966  * If called with for_reloc == 0, may exit early with -EAGAIN
8967  */
8968 int btrfs_drop_snapshot(struct btrfs_root *root,
8969                          struct btrfs_block_rsv *block_rsv, int update_ref,
8970                          int for_reloc)
8971 {
8972         struct btrfs_path *path;
8973         struct btrfs_trans_handle *trans;
8974         struct btrfs_root *tree_root = root->fs_info->tree_root;
8975         struct btrfs_root_item *root_item = &root->root_item;
8976         struct walk_control *wc;
8977         struct btrfs_key key;
8978         int err = 0;
8979         int ret;
8980         int level;
8981         bool root_dropped = false;
8982
8983         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8984
8985         path = btrfs_alloc_path();
8986         if (!path) {
8987                 err = -ENOMEM;
8988                 goto out;
8989         }
8990
8991         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8992         if (!wc) {
8993                 btrfs_free_path(path);
8994                 err = -ENOMEM;
8995                 goto out;
8996         }
8997
8998         trans = btrfs_start_transaction(tree_root, 0);
8999         if (IS_ERR(trans)) {
9000                 err = PTR_ERR(trans);
9001                 goto out_free;
9002         }
9003
9004         if (block_rsv)
9005                 trans->block_rsv = block_rsv;
9006
9007         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9008                 level = btrfs_header_level(root->node);
9009                 path->nodes[level] = btrfs_lock_root_node(root);
9010                 btrfs_set_lock_blocking(path->nodes[level]);
9011                 path->slots[level] = 0;
9012                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9013                 memset(&wc->update_progress, 0,
9014                        sizeof(wc->update_progress));
9015         } else {
9016                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9017                 memcpy(&wc->update_progress, &key,
9018                        sizeof(wc->update_progress));
9019
9020                 level = root_item->drop_level;
9021                 BUG_ON(level == 0);
9022                 path->lowest_level = level;
9023                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9024                 path->lowest_level = 0;
9025                 if (ret < 0) {
9026                         err = ret;
9027                         goto out_end_trans;
9028                 }
9029                 WARN_ON(ret > 0);
9030
9031                 /*
9032                  * unlock our path, this is safe because only this
9033                  * function is allowed to delete this snapshot
9034                  */
9035                 btrfs_unlock_up_safe(path, 0);
9036
9037                 level = btrfs_header_level(root->node);
9038                 while (1) {
9039                         btrfs_tree_lock(path->nodes[level]);
9040                         btrfs_set_lock_blocking(path->nodes[level]);
9041                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9042
9043                         ret = btrfs_lookup_extent_info(trans, root,
9044                                                 path->nodes[level]->start,
9045                                                 level, 1, &wc->refs[level],
9046                                                 &wc->flags[level]);
9047                         if (ret < 0) {
9048                                 err = ret;
9049                                 goto out_end_trans;
9050                         }
9051                         BUG_ON(wc->refs[level] == 0);
9052
9053                         if (level == root_item->drop_level)
9054                                 break;
9055
9056                         btrfs_tree_unlock(path->nodes[level]);
9057                         path->locks[level] = 0;
9058                         WARN_ON(wc->refs[level] != 1);
9059                         level--;
9060                 }
9061         }
9062
9063         wc->level = level;
9064         wc->shared_level = -1;
9065         wc->stage = DROP_REFERENCE;
9066         wc->update_ref = update_ref;
9067         wc->keep_locks = 0;
9068         wc->for_reloc = for_reloc;
9069         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9070
9071         while (1) {
9072
9073                 ret = walk_down_tree(trans, root, path, wc);
9074                 if (ret < 0) {
9075                         err = ret;
9076                         break;
9077                 }
9078
9079                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9080                 if (ret < 0) {
9081                         err = ret;
9082                         break;
9083                 }
9084
9085                 if (ret > 0) {
9086                         BUG_ON(wc->stage != DROP_REFERENCE);
9087                         break;
9088                 }
9089
9090                 if (wc->stage == DROP_REFERENCE) {
9091                         level = wc->level;
9092                         btrfs_node_key(path->nodes[level],
9093                                        &root_item->drop_progress,
9094                                        path->slots[level]);
9095                         root_item->drop_level = level;
9096                 }
9097
9098                 BUG_ON(wc->level == 0);
9099                 if (btrfs_should_end_transaction(trans, tree_root) ||
9100                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
9101                         ret = btrfs_update_root(trans, tree_root,
9102                                                 &root->root_key,
9103                                                 root_item);
9104                         if (ret) {
9105                                 btrfs_abort_transaction(trans, tree_root, ret);
9106                                 err = ret;
9107                                 goto out_end_trans;
9108                         }
9109
9110                         btrfs_end_transaction_throttle(trans, tree_root);
9111                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
9112                                 pr_debug("BTRFS: drop snapshot early exit\n");
9113                                 err = -EAGAIN;
9114                                 goto out_free;
9115                         }
9116
9117                         trans = btrfs_start_transaction(tree_root, 0);
9118                         if (IS_ERR(trans)) {
9119                                 err = PTR_ERR(trans);
9120                                 goto out_free;
9121                         }
9122                         if (block_rsv)
9123                                 trans->block_rsv = block_rsv;
9124                 }
9125         }
9126         btrfs_release_path(path);
9127         if (err)
9128                 goto out_end_trans;
9129
9130         ret = btrfs_del_root(trans, tree_root, &root->root_key);
9131         if (ret) {
9132                 btrfs_abort_transaction(trans, tree_root, ret);
9133                 goto out_end_trans;
9134         }
9135
9136         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9137                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9138                                       NULL, NULL);
9139                 if (ret < 0) {
9140                         btrfs_abort_transaction(trans, tree_root, ret);
9141                         err = ret;
9142                         goto out_end_trans;
9143                 } else if (ret > 0) {
9144                         /* if we fail to delete the orphan item this time
9145                          * around, it'll get picked up the next time.
9146                          *
9147                          * The most common failure here is just -ENOENT.
9148                          */
9149                         btrfs_del_orphan_item(trans, tree_root,
9150                                               root->root_key.objectid);
9151                 }
9152         }
9153
9154         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9155                 btrfs_add_dropped_root(trans, root);
9156         } else {
9157                 free_extent_buffer(root->node);
9158                 free_extent_buffer(root->commit_root);
9159                 btrfs_put_fs_root(root);
9160         }
9161         root_dropped = true;
9162 out_end_trans:
9163         btrfs_end_transaction_throttle(trans, tree_root);
9164 out_free:
9165         kfree(wc);
9166         btrfs_free_path(path);
9167 out:
9168         /*
9169          * So if we need to stop dropping the snapshot for whatever reason we
9170          * need to make sure to add it back to the dead root list so that we
9171          * keep trying to do the work later.  This also cleans up roots if we
9172          * don't have it in the radix (like when we recover after a power fail
9173          * or unmount) so we don't leak memory.
9174          */
9175         if (!for_reloc && root_dropped == false)
9176                 btrfs_add_dead_root(root);
9177         if (err && err != -EAGAIN)
9178                 btrfs_handle_fs_error(root->fs_info, err, NULL);
9179         return err;
9180 }
9181
9182 /*
9183  * drop subtree rooted at tree block 'node'.
9184  *
9185  * NOTE: this function will unlock and release tree block 'node'
9186  * only used by relocation code
9187  */
9188 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9189                         struct btrfs_root *root,
9190                         struct extent_buffer *node,
9191                         struct extent_buffer *parent)
9192 {
9193         struct btrfs_path *path;
9194         struct walk_control *wc;
9195         int level;
9196         int parent_level;
9197         int ret = 0;
9198         int wret;
9199
9200         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9201
9202         path = btrfs_alloc_path();
9203         if (!path)
9204                 return -ENOMEM;
9205
9206         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9207         if (!wc) {
9208                 btrfs_free_path(path);
9209                 return -ENOMEM;
9210         }
9211
9212         btrfs_assert_tree_locked(parent);
9213         parent_level = btrfs_header_level(parent);
9214         extent_buffer_get(parent);
9215         path->nodes[parent_level] = parent;
9216         path->slots[parent_level] = btrfs_header_nritems(parent);
9217
9218         btrfs_assert_tree_locked(node);
9219         level = btrfs_header_level(node);
9220         path->nodes[level] = node;
9221         path->slots[level] = 0;
9222         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9223
9224         wc->refs[parent_level] = 1;
9225         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9226         wc->level = level;
9227         wc->shared_level = -1;
9228         wc->stage = DROP_REFERENCE;
9229         wc->update_ref = 0;
9230         wc->keep_locks = 1;
9231         wc->for_reloc = 1;
9232         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9233
9234         while (1) {
9235                 wret = walk_down_tree(trans, root, path, wc);
9236                 if (wret < 0) {
9237                         ret = wret;
9238                         break;
9239                 }
9240
9241                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9242                 if (wret < 0)
9243                         ret = wret;
9244                 if (wret != 0)
9245                         break;
9246         }
9247
9248         kfree(wc);
9249         btrfs_free_path(path);
9250         return ret;
9251 }
9252
9253 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
9254 {
9255         u64 num_devices;
9256         u64 stripped;
9257
9258         /*
9259          * if restripe for this chunk_type is on pick target profile and
9260          * return, otherwise do the usual balance
9261          */
9262         stripped = get_restripe_target(root->fs_info, flags);
9263         if (stripped)
9264                 return extended_to_chunk(stripped);
9265
9266         num_devices = root->fs_info->fs_devices->rw_devices;
9267
9268         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9269                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9270                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9271
9272         if (num_devices == 1) {
9273                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9274                 stripped = flags & ~stripped;
9275
9276                 /* turn raid0 into single device chunks */
9277                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9278                         return stripped;
9279
9280                 /* turn mirroring into duplication */
9281                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9282                              BTRFS_BLOCK_GROUP_RAID10))
9283                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9284         } else {
9285                 /* they already had raid on here, just return */
9286                 if (flags & stripped)
9287                         return flags;
9288
9289                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9290                 stripped = flags & ~stripped;
9291
9292                 /* switch duplicated blocks with raid1 */
9293                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9294                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9295
9296                 /* this is drive concat, leave it alone */
9297         }
9298
9299         return flags;
9300 }
9301
9302 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9303 {
9304         struct btrfs_space_info *sinfo = cache->space_info;
9305         u64 num_bytes;
9306         u64 min_allocable_bytes;
9307         int ret = -ENOSPC;
9308
9309         /*
9310          * We need some metadata space and system metadata space for
9311          * allocating chunks in some corner cases until we force to set
9312          * it to be readonly.
9313          */
9314         if ((sinfo->flags &
9315              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9316             !force)
9317                 min_allocable_bytes = SZ_1M;
9318         else
9319                 min_allocable_bytes = 0;
9320
9321         spin_lock(&sinfo->lock);
9322         spin_lock(&cache->lock);
9323
9324         if (cache->ro) {
9325                 cache->ro++;
9326                 ret = 0;
9327                 goto out;
9328         }
9329
9330         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9331                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9332
9333         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
9334             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
9335             min_allocable_bytes <= sinfo->total_bytes) {
9336                 sinfo->bytes_readonly += num_bytes;
9337                 cache->ro++;
9338                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9339                 ret = 0;
9340         }
9341 out:
9342         spin_unlock(&cache->lock);
9343         spin_unlock(&sinfo->lock);
9344         return ret;
9345 }
9346
9347 int btrfs_inc_block_group_ro(struct btrfs_root *root,
9348                              struct btrfs_block_group_cache *cache)
9349
9350 {
9351         struct btrfs_trans_handle *trans;
9352         u64 alloc_flags;
9353         int ret;
9354
9355 again:
9356         trans = btrfs_join_transaction(root);
9357         if (IS_ERR(trans))
9358                 return PTR_ERR(trans);
9359
9360         /*
9361          * we're not allowed to set block groups readonly after the dirty
9362          * block groups cache has started writing.  If it already started,
9363          * back off and let this transaction commit
9364          */
9365         mutex_lock(&root->fs_info->ro_block_group_mutex);
9366         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9367                 u64 transid = trans->transid;
9368
9369                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
9370                 btrfs_end_transaction(trans, root);
9371
9372                 ret = btrfs_wait_for_commit(root, transid);
9373                 if (ret)
9374                         return ret;
9375                 goto again;
9376         }
9377
9378         /*
9379          * if we are changing raid levels, try to allocate a corresponding
9380          * block group with the new raid level.
9381          */
9382         alloc_flags = update_block_group_flags(root, cache->flags);
9383         if (alloc_flags != cache->flags) {
9384                 ret = do_chunk_alloc(trans, root, alloc_flags,
9385                                      CHUNK_ALLOC_FORCE);
9386                 /*
9387                  * ENOSPC is allowed here, we may have enough space
9388                  * already allocated at the new raid level to
9389                  * carry on
9390                  */
9391                 if (ret == -ENOSPC)
9392                         ret = 0;
9393                 if (ret < 0)
9394                         goto out;
9395         }
9396
9397         ret = inc_block_group_ro(cache, 0);
9398         if (!ret)
9399                 goto out;
9400         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
9401         ret = do_chunk_alloc(trans, root, alloc_flags,
9402                              CHUNK_ALLOC_FORCE);
9403         if (ret < 0)
9404                 goto out;
9405         ret = inc_block_group_ro(cache, 0);
9406 out:
9407         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9408                 alloc_flags = update_block_group_flags(root, cache->flags);
9409                 lock_chunks(root->fs_info->chunk_root);
9410                 check_system_chunk(trans, root, alloc_flags);
9411                 unlock_chunks(root->fs_info->chunk_root);
9412         }
9413         mutex_unlock(&root->fs_info->ro_block_group_mutex);
9414
9415         btrfs_end_transaction(trans, root);
9416         return ret;
9417 }
9418
9419 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9420                             struct btrfs_root *root, u64 type)
9421 {
9422         u64 alloc_flags = get_alloc_profile(root, type);
9423         return do_chunk_alloc(trans, root, alloc_flags,
9424                               CHUNK_ALLOC_FORCE);
9425 }
9426
9427 /*
9428  * helper to account the unused space of all the readonly block group in the
9429  * space_info. takes mirrors into account.
9430  */
9431 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9432 {
9433         struct btrfs_block_group_cache *block_group;
9434         u64 free_bytes = 0;
9435         int factor;
9436
9437         /* It's df, we don't care if it's racy */
9438         if (list_empty(&sinfo->ro_bgs))
9439                 return 0;
9440
9441         spin_lock(&sinfo->lock);
9442         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9443                 spin_lock(&block_group->lock);
9444
9445                 if (!block_group->ro) {
9446                         spin_unlock(&block_group->lock);
9447                         continue;
9448                 }
9449
9450                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9451                                           BTRFS_BLOCK_GROUP_RAID10 |
9452                                           BTRFS_BLOCK_GROUP_DUP))
9453                         factor = 2;
9454                 else
9455                         factor = 1;
9456
9457                 free_bytes += (block_group->key.offset -
9458                                btrfs_block_group_used(&block_group->item)) *
9459                                factor;
9460
9461                 spin_unlock(&block_group->lock);
9462         }
9463         spin_unlock(&sinfo->lock);
9464
9465         return free_bytes;
9466 }
9467
9468 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9469                               struct btrfs_block_group_cache *cache)
9470 {
9471         struct btrfs_space_info *sinfo = cache->space_info;
9472         u64 num_bytes;
9473
9474         BUG_ON(!cache->ro);
9475
9476         spin_lock(&sinfo->lock);
9477         spin_lock(&cache->lock);
9478         if (!--cache->ro) {
9479                 num_bytes = cache->key.offset - cache->reserved -
9480                             cache->pinned - cache->bytes_super -
9481                             btrfs_block_group_used(&cache->item);
9482                 sinfo->bytes_readonly -= num_bytes;
9483                 list_del_init(&cache->ro_list);
9484         }
9485         spin_unlock(&cache->lock);
9486         spin_unlock(&sinfo->lock);
9487 }
9488
9489 /*
9490  * checks to see if its even possible to relocate this block group.
9491  *
9492  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9493  * ok to go ahead and try.
9494  */
9495 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9496 {
9497         struct btrfs_block_group_cache *block_group;
9498         struct btrfs_space_info *space_info;
9499         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9500         struct btrfs_device *device;
9501         struct btrfs_trans_handle *trans;
9502         u64 min_free;
9503         u64 dev_min = 1;
9504         u64 dev_nr = 0;
9505         u64 target;
9506         int debug;
9507         int index;
9508         int full = 0;
9509         int ret = 0;
9510
9511         debug = btrfs_test_opt(root, ENOSPC_DEBUG);
9512
9513         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9514
9515         /* odd, couldn't find the block group, leave it alone */
9516         if (!block_group) {
9517                 if (debug)
9518                         btrfs_warn(root->fs_info,
9519                                    "can't find block group for bytenr %llu",
9520                                    bytenr);
9521                 return -1;
9522         }
9523
9524         min_free = btrfs_block_group_used(&block_group->item);
9525
9526         /* no bytes used, we're good */
9527         if (!min_free)
9528                 goto out;
9529
9530         space_info = block_group->space_info;
9531         spin_lock(&space_info->lock);
9532
9533         full = space_info->full;
9534
9535         /*
9536          * if this is the last block group we have in this space, we can't
9537          * relocate it unless we're able to allocate a new chunk below.
9538          *
9539          * Otherwise, we need to make sure we have room in the space to handle
9540          * all of the extents from this block group.  If we can, we're good
9541          */
9542         if ((space_info->total_bytes != block_group->key.offset) &&
9543             (space_info->bytes_used + space_info->bytes_reserved +
9544              space_info->bytes_pinned + space_info->bytes_readonly +
9545              min_free < space_info->total_bytes)) {
9546                 spin_unlock(&space_info->lock);
9547                 goto out;
9548         }
9549         spin_unlock(&space_info->lock);
9550
9551         /*
9552          * ok we don't have enough space, but maybe we have free space on our
9553          * devices to allocate new chunks for relocation, so loop through our
9554          * alloc devices and guess if we have enough space.  if this block
9555          * group is going to be restriped, run checks against the target
9556          * profile instead of the current one.
9557          */
9558         ret = -1;
9559
9560         /*
9561          * index:
9562          *      0: raid10
9563          *      1: raid1
9564          *      2: dup
9565          *      3: raid0
9566          *      4: single
9567          */
9568         target = get_restripe_target(root->fs_info, block_group->flags);
9569         if (target) {
9570                 index = __get_raid_index(extended_to_chunk(target));
9571         } else {
9572                 /*
9573                  * this is just a balance, so if we were marked as full
9574                  * we know there is no space for a new chunk
9575                  */
9576                 if (full) {
9577                         if (debug)
9578                                 btrfs_warn(root->fs_info,
9579                                         "no space to alloc new chunk for block group %llu",
9580                                         block_group->key.objectid);
9581                         goto out;
9582                 }
9583
9584                 index = get_block_group_index(block_group);
9585         }
9586
9587         if (index == BTRFS_RAID_RAID10) {
9588                 dev_min = 4;
9589                 /* Divide by 2 */
9590                 min_free >>= 1;
9591         } else if (index == BTRFS_RAID_RAID1) {
9592                 dev_min = 2;
9593         } else if (index == BTRFS_RAID_DUP) {
9594                 /* Multiply by 2 */
9595                 min_free <<= 1;
9596         } else if (index == BTRFS_RAID_RAID0) {
9597                 dev_min = fs_devices->rw_devices;
9598                 min_free = div64_u64(min_free, dev_min);
9599         }
9600
9601         /* We need to do this so that we can look at pending chunks */
9602         trans = btrfs_join_transaction(root);
9603         if (IS_ERR(trans)) {
9604                 ret = PTR_ERR(trans);
9605                 goto out;
9606         }
9607
9608         mutex_lock(&root->fs_info->chunk_mutex);
9609         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9610                 u64 dev_offset;
9611
9612                 /*
9613                  * check to make sure we can actually find a chunk with enough
9614                  * space to fit our block group in.
9615                  */
9616                 if (device->total_bytes > device->bytes_used + min_free &&
9617                     !device->is_tgtdev_for_dev_replace) {
9618                         ret = find_free_dev_extent(trans, device, min_free,
9619                                                    &dev_offset, NULL);
9620                         if (!ret)
9621                                 dev_nr++;
9622
9623                         if (dev_nr >= dev_min)
9624                                 break;
9625
9626                         ret = -1;
9627                 }
9628         }
9629         if (debug && ret == -1)
9630                 btrfs_warn(root->fs_info,
9631                         "no space to allocate a new chunk for block group %llu",
9632                         block_group->key.objectid);
9633         mutex_unlock(&root->fs_info->chunk_mutex);
9634         btrfs_end_transaction(trans, root);
9635 out:
9636         btrfs_put_block_group(block_group);
9637         return ret;
9638 }
9639
9640 static int find_first_block_group(struct btrfs_root *root,
9641                 struct btrfs_path *path, struct btrfs_key *key)
9642 {
9643         int ret = 0;
9644         struct btrfs_key found_key;
9645         struct extent_buffer *leaf;
9646         int slot;
9647
9648         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9649         if (ret < 0)
9650                 goto out;
9651
9652         while (1) {
9653                 slot = path->slots[0];
9654                 leaf = path->nodes[0];
9655                 if (slot >= btrfs_header_nritems(leaf)) {
9656                         ret = btrfs_next_leaf(root, path);
9657                         if (ret == 0)
9658                                 continue;
9659                         if (ret < 0)
9660                                 goto out;
9661                         break;
9662                 }
9663                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9664
9665                 if (found_key.objectid >= key->objectid &&
9666                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9667                         ret = 0;
9668                         goto out;
9669                 }
9670                 path->slots[0]++;
9671         }
9672 out:
9673         return ret;
9674 }
9675
9676 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9677 {
9678         struct btrfs_block_group_cache *block_group;
9679         u64 last = 0;
9680
9681         while (1) {
9682                 struct inode *inode;
9683
9684                 block_group = btrfs_lookup_first_block_group(info, last);
9685                 while (block_group) {
9686                         spin_lock(&block_group->lock);
9687                         if (block_group->iref)
9688                                 break;
9689                         spin_unlock(&block_group->lock);
9690                         block_group = next_block_group(info->tree_root,
9691                                                        block_group);
9692                 }
9693                 if (!block_group) {
9694                         if (last == 0)
9695                                 break;
9696                         last = 0;
9697                         continue;
9698                 }
9699
9700                 inode = block_group->inode;
9701                 block_group->iref = 0;
9702                 block_group->inode = NULL;
9703                 spin_unlock(&block_group->lock);
9704                 iput(inode);
9705                 last = block_group->key.objectid + block_group->key.offset;
9706                 btrfs_put_block_group(block_group);
9707         }
9708 }
9709
9710 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9711 {
9712         struct btrfs_block_group_cache *block_group;
9713         struct btrfs_space_info *space_info;
9714         struct btrfs_caching_control *caching_ctl;
9715         struct rb_node *n;
9716
9717         down_write(&info->commit_root_sem);
9718         while (!list_empty(&info->caching_block_groups)) {
9719                 caching_ctl = list_entry(info->caching_block_groups.next,
9720                                          struct btrfs_caching_control, list);
9721                 list_del(&caching_ctl->list);
9722                 put_caching_control(caching_ctl);
9723         }
9724         up_write(&info->commit_root_sem);
9725
9726         spin_lock(&info->unused_bgs_lock);
9727         while (!list_empty(&info->unused_bgs)) {
9728                 block_group = list_first_entry(&info->unused_bgs,
9729                                                struct btrfs_block_group_cache,
9730                                                bg_list);
9731                 list_del_init(&block_group->bg_list);
9732                 btrfs_put_block_group(block_group);
9733         }
9734         spin_unlock(&info->unused_bgs_lock);
9735
9736         spin_lock(&info->block_group_cache_lock);
9737         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9738                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9739                                        cache_node);
9740                 rb_erase(&block_group->cache_node,
9741                          &info->block_group_cache_tree);
9742                 RB_CLEAR_NODE(&block_group->cache_node);
9743                 spin_unlock(&info->block_group_cache_lock);
9744
9745                 down_write(&block_group->space_info->groups_sem);
9746                 list_del(&block_group->list);
9747                 up_write(&block_group->space_info->groups_sem);
9748
9749                 if (block_group->cached == BTRFS_CACHE_STARTED)
9750                         wait_block_group_cache_done(block_group);
9751
9752                 /*
9753                  * We haven't cached this block group, which means we could
9754                  * possibly have excluded extents on this block group.
9755                  */
9756                 if (block_group->cached == BTRFS_CACHE_NO ||
9757                     block_group->cached == BTRFS_CACHE_ERROR)
9758                         free_excluded_extents(info->extent_root, block_group);
9759
9760                 btrfs_remove_free_space_cache(block_group);
9761                 btrfs_put_block_group(block_group);
9762
9763                 spin_lock(&info->block_group_cache_lock);
9764         }
9765         spin_unlock(&info->block_group_cache_lock);
9766
9767         /* now that all the block groups are freed, go through and
9768          * free all the space_info structs.  This is only called during
9769          * the final stages of unmount, and so we know nobody is
9770          * using them.  We call synchronize_rcu() once before we start,
9771          * just to be on the safe side.
9772          */
9773         synchronize_rcu();
9774
9775         release_global_block_rsv(info);
9776
9777         while (!list_empty(&info->space_info)) {
9778                 int i;
9779
9780                 space_info = list_entry(info->space_info.next,
9781                                         struct btrfs_space_info,
9782                                         list);
9783                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9784                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9785                             space_info->bytes_reserved > 0 ||
9786                             space_info->bytes_may_use > 0)) {
9787                                 dump_space_info(space_info, 0, 0);
9788                         }
9789                 }
9790                 list_del(&space_info->list);
9791                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9792                         struct kobject *kobj;
9793                         kobj = space_info->block_group_kobjs[i];
9794                         space_info->block_group_kobjs[i] = NULL;
9795                         if (kobj) {
9796                                 kobject_del(kobj);
9797                                 kobject_put(kobj);
9798                         }
9799                 }
9800                 kobject_del(&space_info->kobj);
9801                 kobject_put(&space_info->kobj);
9802         }
9803         return 0;
9804 }
9805
9806 static void __link_block_group(struct btrfs_space_info *space_info,
9807                                struct btrfs_block_group_cache *cache)
9808 {
9809         int index = get_block_group_index(cache);
9810         bool first = false;
9811
9812         down_write(&space_info->groups_sem);
9813         if (list_empty(&space_info->block_groups[index]))
9814                 first = true;
9815         list_add_tail(&cache->list, &space_info->block_groups[index]);
9816         up_write(&space_info->groups_sem);
9817
9818         if (first) {
9819                 struct raid_kobject *rkobj;
9820                 int ret;
9821
9822                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9823                 if (!rkobj)
9824                         goto out_err;
9825                 rkobj->raid_type = index;
9826                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9827                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9828                                   "%s", get_raid_name(index));
9829                 if (ret) {
9830                         kobject_put(&rkobj->kobj);
9831                         goto out_err;
9832                 }
9833                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9834         }
9835
9836         return;
9837 out_err:
9838         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9839 }
9840
9841 static struct btrfs_block_group_cache *
9842 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9843 {
9844         struct btrfs_block_group_cache *cache;
9845
9846         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9847         if (!cache)
9848                 return NULL;
9849
9850         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9851                                         GFP_NOFS);
9852         if (!cache->free_space_ctl) {
9853                 kfree(cache);
9854                 return NULL;
9855         }
9856
9857         cache->key.objectid = start;
9858         cache->key.offset = size;
9859         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9860
9861         cache->sectorsize = root->sectorsize;
9862         cache->fs_info = root->fs_info;
9863         cache->full_stripe_len = btrfs_full_stripe_len(root,
9864                                                &root->fs_info->mapping_tree,
9865                                                start);
9866         set_free_space_tree_thresholds(cache);
9867
9868         atomic_set(&cache->count, 1);
9869         spin_lock_init(&cache->lock);
9870         init_rwsem(&cache->data_rwsem);
9871         INIT_LIST_HEAD(&cache->list);
9872         INIT_LIST_HEAD(&cache->cluster_list);
9873         INIT_LIST_HEAD(&cache->bg_list);
9874         INIT_LIST_HEAD(&cache->ro_list);
9875         INIT_LIST_HEAD(&cache->dirty_list);
9876         INIT_LIST_HEAD(&cache->io_list);
9877         btrfs_init_free_space_ctl(cache);
9878         atomic_set(&cache->trimming, 0);
9879         mutex_init(&cache->free_space_lock);
9880
9881         return cache;
9882 }
9883
9884 int btrfs_read_block_groups(struct btrfs_root *root)
9885 {
9886         struct btrfs_path *path;
9887         int ret;
9888         struct btrfs_block_group_cache *cache;
9889         struct btrfs_fs_info *info = root->fs_info;
9890         struct btrfs_space_info *space_info;
9891         struct btrfs_key key;
9892         struct btrfs_key found_key;
9893         struct extent_buffer *leaf;
9894         int need_clear = 0;
9895         u64 cache_gen;
9896
9897         root = info->extent_root;
9898         key.objectid = 0;
9899         key.offset = 0;
9900         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9901         path = btrfs_alloc_path();
9902         if (!path)
9903                 return -ENOMEM;
9904         path->reada = READA_FORWARD;
9905
9906         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9907         if (btrfs_test_opt(root, SPACE_CACHE) &&
9908             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9909                 need_clear = 1;
9910         if (btrfs_test_opt(root, CLEAR_CACHE))
9911                 need_clear = 1;
9912
9913         while (1) {
9914                 ret = find_first_block_group(root, path, &key);
9915                 if (ret > 0)
9916                         break;
9917                 if (ret != 0)
9918                         goto error;
9919
9920                 leaf = path->nodes[0];
9921                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9922
9923                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9924                                                        found_key.offset);
9925                 if (!cache) {
9926                         ret = -ENOMEM;
9927                         goto error;
9928                 }
9929
9930                 if (need_clear) {
9931                         /*
9932                          * When we mount with old space cache, we need to
9933                          * set BTRFS_DC_CLEAR and set dirty flag.
9934                          *
9935                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9936                          *    truncate the old free space cache inode and
9937                          *    setup a new one.
9938                          * b) Setting 'dirty flag' makes sure that we flush
9939                          *    the new space cache info onto disk.
9940                          */
9941                         if (btrfs_test_opt(root, SPACE_CACHE))
9942                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9943                 }
9944
9945                 read_extent_buffer(leaf, &cache->item,
9946                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9947                                    sizeof(cache->item));
9948                 cache->flags = btrfs_block_group_flags(&cache->item);
9949
9950                 key.objectid = found_key.objectid + found_key.offset;
9951                 btrfs_release_path(path);
9952
9953                 /*
9954                  * We need to exclude the super stripes now so that the space
9955                  * info has super bytes accounted for, otherwise we'll think
9956                  * we have more space than we actually do.
9957                  */
9958                 ret = exclude_super_stripes(root, cache);
9959                 if (ret) {
9960                         /*
9961                          * We may have excluded something, so call this just in
9962                          * case.
9963                          */
9964                         free_excluded_extents(root, cache);
9965                         btrfs_put_block_group(cache);
9966                         goto error;
9967                 }
9968
9969                 /*
9970                  * check for two cases, either we are full, and therefore
9971                  * don't need to bother with the caching work since we won't
9972                  * find any space, or we are empty, and we can just add all
9973                  * the space in and be done with it.  This saves us _alot_ of
9974                  * time, particularly in the full case.
9975                  */
9976                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9977                         cache->last_byte_to_unpin = (u64)-1;
9978                         cache->cached = BTRFS_CACHE_FINISHED;
9979                         free_excluded_extents(root, cache);
9980                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9981                         cache->last_byte_to_unpin = (u64)-1;
9982                         cache->cached = BTRFS_CACHE_FINISHED;
9983                         add_new_free_space(cache, root->fs_info,
9984                                            found_key.objectid,
9985                                            found_key.objectid +
9986                                            found_key.offset);
9987                         free_excluded_extents(root, cache);
9988                 }
9989
9990                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9991                 if (ret) {
9992                         btrfs_remove_free_space_cache(cache);
9993                         btrfs_put_block_group(cache);
9994                         goto error;
9995                 }
9996
9997                 ret = update_space_info(info, cache->flags, found_key.offset,
9998                                         btrfs_block_group_used(&cache->item),
9999                                         &space_info);
10000                 if (ret) {
10001                         btrfs_remove_free_space_cache(cache);
10002                         spin_lock(&info->block_group_cache_lock);
10003                         rb_erase(&cache->cache_node,
10004                                  &info->block_group_cache_tree);
10005                         RB_CLEAR_NODE(&cache->cache_node);
10006                         spin_unlock(&info->block_group_cache_lock);
10007                         btrfs_put_block_group(cache);
10008                         goto error;
10009                 }
10010
10011                 cache->space_info = space_info;
10012                 spin_lock(&cache->space_info->lock);
10013                 cache->space_info->bytes_readonly += cache->bytes_super;
10014                 spin_unlock(&cache->space_info->lock);
10015
10016                 __link_block_group(space_info, cache);
10017
10018                 set_avail_alloc_bits(root->fs_info, cache->flags);
10019                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
10020                         inc_block_group_ro(cache, 1);
10021                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10022                         spin_lock(&info->unused_bgs_lock);
10023                         /* Should always be true but just in case. */
10024                         if (list_empty(&cache->bg_list)) {
10025                                 btrfs_get_block_group(cache);
10026                                 list_add_tail(&cache->bg_list,
10027                                               &info->unused_bgs);
10028                         }
10029                         spin_unlock(&info->unused_bgs_lock);
10030                 }
10031         }
10032
10033         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
10034                 if (!(get_alloc_profile(root, space_info->flags) &
10035                       (BTRFS_BLOCK_GROUP_RAID10 |
10036                        BTRFS_BLOCK_GROUP_RAID1 |
10037                        BTRFS_BLOCK_GROUP_RAID5 |
10038                        BTRFS_BLOCK_GROUP_RAID6 |
10039                        BTRFS_BLOCK_GROUP_DUP)))
10040                         continue;
10041                 /*
10042                  * avoid allocating from un-mirrored block group if there are
10043                  * mirrored block groups.
10044                  */
10045                 list_for_each_entry(cache,
10046                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10047                                 list)
10048                         inc_block_group_ro(cache, 1);
10049                 list_for_each_entry(cache,
10050                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10051                                 list)
10052                         inc_block_group_ro(cache, 1);
10053         }
10054
10055         init_global_block_rsv(info);
10056         ret = 0;
10057 error:
10058         btrfs_free_path(path);
10059         return ret;
10060 }
10061
10062 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
10063                                        struct btrfs_root *root)
10064 {
10065         struct btrfs_block_group_cache *block_group, *tmp;
10066         struct btrfs_root *extent_root = root->fs_info->extent_root;
10067         struct btrfs_block_group_item item;
10068         struct btrfs_key key;
10069         int ret = 0;
10070         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10071
10072         trans->can_flush_pending_bgs = false;
10073         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10074                 if (ret)
10075                         goto next;
10076
10077                 spin_lock(&block_group->lock);
10078                 memcpy(&item, &block_group->item, sizeof(item));
10079                 memcpy(&key, &block_group->key, sizeof(key));
10080                 spin_unlock(&block_group->lock);
10081
10082                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10083                                         sizeof(item));
10084                 if (ret)
10085                         btrfs_abort_transaction(trans, extent_root, ret);
10086                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
10087                                                key.objectid, key.offset);
10088                 if (ret)
10089                         btrfs_abort_transaction(trans, extent_root, ret);
10090                 add_block_group_free_space(trans, root->fs_info, block_group);
10091                 /* already aborted the transaction if it failed. */
10092 next:
10093                 list_del_init(&block_group->bg_list);
10094         }
10095         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10096 }
10097
10098 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10099                            struct btrfs_root *root, u64 bytes_used,
10100                            u64 type, u64 chunk_objectid, u64 chunk_offset,
10101                            u64 size)
10102 {
10103         int ret;
10104         struct btrfs_root *extent_root;
10105         struct btrfs_block_group_cache *cache;
10106
10107         extent_root = root->fs_info->extent_root;
10108
10109         btrfs_set_log_full_commit(root->fs_info, trans);
10110
10111         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
10112         if (!cache)
10113                 return -ENOMEM;
10114
10115         btrfs_set_block_group_used(&cache->item, bytes_used);
10116         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
10117         btrfs_set_block_group_flags(&cache->item, type);
10118
10119         cache->flags = type;
10120         cache->last_byte_to_unpin = (u64)-1;
10121         cache->cached = BTRFS_CACHE_FINISHED;
10122         cache->needs_free_space = 1;
10123         ret = exclude_super_stripes(root, cache);
10124         if (ret) {
10125                 /*
10126                  * We may have excluded something, so call this just in
10127                  * case.
10128                  */
10129                 free_excluded_extents(root, cache);
10130                 btrfs_put_block_group(cache);
10131                 return ret;
10132         }
10133
10134         add_new_free_space(cache, root->fs_info, chunk_offset,
10135                            chunk_offset + size);
10136
10137         free_excluded_extents(root, cache);
10138
10139 #ifdef CONFIG_BTRFS_DEBUG
10140         if (btrfs_should_fragment_free_space(root, cache)) {
10141                 u64 new_bytes_used = size - bytes_used;
10142
10143                 bytes_used += new_bytes_used >> 1;
10144                 fragment_free_space(root, cache);
10145         }
10146 #endif
10147         /*
10148          * Call to ensure the corresponding space_info object is created and
10149          * assigned to our block group, but don't update its counters just yet.
10150          * We want our bg to be added to the rbtree with its ->space_info set.
10151          */
10152         ret = update_space_info(root->fs_info, cache->flags, 0, 0,
10153                                 &cache->space_info);
10154         if (ret) {
10155                 btrfs_remove_free_space_cache(cache);
10156                 btrfs_put_block_group(cache);
10157                 return ret;
10158         }
10159
10160         ret = btrfs_add_block_group_cache(root->fs_info, cache);
10161         if (ret) {
10162                 btrfs_remove_free_space_cache(cache);
10163                 btrfs_put_block_group(cache);
10164                 return ret;
10165         }
10166
10167         /*
10168          * Now that our block group has its ->space_info set and is inserted in
10169          * the rbtree, update the space info's counters.
10170          */
10171         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
10172                                 &cache->space_info);
10173         if (ret) {
10174                 btrfs_remove_free_space_cache(cache);
10175                 spin_lock(&root->fs_info->block_group_cache_lock);
10176                 rb_erase(&cache->cache_node,
10177                          &root->fs_info->block_group_cache_tree);
10178                 RB_CLEAR_NODE(&cache->cache_node);
10179                 spin_unlock(&root->fs_info->block_group_cache_lock);
10180                 btrfs_put_block_group(cache);
10181                 return ret;
10182         }
10183         update_global_block_rsv(root->fs_info);
10184
10185         spin_lock(&cache->space_info->lock);
10186         cache->space_info->bytes_readonly += cache->bytes_super;
10187         spin_unlock(&cache->space_info->lock);
10188
10189         __link_block_group(cache->space_info, cache);
10190
10191         list_add_tail(&cache->bg_list, &trans->new_bgs);
10192
10193         set_avail_alloc_bits(extent_root->fs_info, type);
10194
10195         return 0;
10196 }
10197
10198 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10199 {
10200         u64 extra_flags = chunk_to_extended(flags) &
10201                                 BTRFS_EXTENDED_PROFILE_MASK;
10202
10203         write_seqlock(&fs_info->profiles_lock);
10204         if (flags & BTRFS_BLOCK_GROUP_DATA)
10205                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10206         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10207                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10208         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10209                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10210         write_sequnlock(&fs_info->profiles_lock);
10211 }
10212
10213 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10214                              struct btrfs_root *root, u64 group_start,
10215                              struct extent_map *em)
10216 {
10217         struct btrfs_path *path;
10218         struct btrfs_block_group_cache *block_group;
10219         struct btrfs_free_cluster *cluster;
10220         struct btrfs_root *tree_root = root->fs_info->tree_root;
10221         struct btrfs_key key;
10222         struct inode *inode;
10223         struct kobject *kobj = NULL;
10224         int ret;
10225         int index;
10226         int factor;
10227         struct btrfs_caching_control *caching_ctl = NULL;
10228         bool remove_em;
10229
10230         root = root->fs_info->extent_root;
10231
10232         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
10233         BUG_ON(!block_group);
10234         BUG_ON(!block_group->ro);
10235
10236         /*
10237          * Free the reserved super bytes from this block group before
10238          * remove it.
10239          */
10240         free_excluded_extents(root, block_group);
10241
10242         memcpy(&key, &block_group->key, sizeof(key));
10243         index = get_block_group_index(block_group);
10244         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10245                                   BTRFS_BLOCK_GROUP_RAID1 |
10246                                   BTRFS_BLOCK_GROUP_RAID10))
10247                 factor = 2;
10248         else
10249                 factor = 1;
10250
10251         /* make sure this block group isn't part of an allocation cluster */
10252         cluster = &root->fs_info->data_alloc_cluster;
10253         spin_lock(&cluster->refill_lock);
10254         btrfs_return_cluster_to_free_space(block_group, cluster);
10255         spin_unlock(&cluster->refill_lock);
10256
10257         /*
10258          * make sure this block group isn't part of a metadata
10259          * allocation cluster
10260          */
10261         cluster = &root->fs_info->meta_alloc_cluster;
10262         spin_lock(&cluster->refill_lock);
10263         btrfs_return_cluster_to_free_space(block_group, cluster);
10264         spin_unlock(&cluster->refill_lock);
10265
10266         path = btrfs_alloc_path();
10267         if (!path) {
10268                 ret = -ENOMEM;
10269                 goto out;
10270         }
10271
10272         /*
10273          * get the inode first so any iput calls done for the io_list
10274          * aren't the final iput (no unlinks allowed now)
10275          */
10276         inode = lookup_free_space_inode(tree_root, block_group, path);
10277
10278         mutex_lock(&trans->transaction->cache_write_mutex);
10279         /*
10280          * make sure our free spache cache IO is done before remove the
10281          * free space inode
10282          */
10283         spin_lock(&trans->transaction->dirty_bgs_lock);
10284         if (!list_empty(&block_group->io_list)) {
10285                 list_del_init(&block_group->io_list);
10286
10287                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10288
10289                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10290                 btrfs_wait_cache_io(root, trans, block_group,
10291                                     &block_group->io_ctl, path,
10292                                     block_group->key.objectid);
10293                 btrfs_put_block_group(block_group);
10294                 spin_lock(&trans->transaction->dirty_bgs_lock);
10295         }
10296
10297         if (!list_empty(&block_group->dirty_list)) {
10298                 list_del_init(&block_group->dirty_list);
10299                 btrfs_put_block_group(block_group);
10300         }
10301         spin_unlock(&trans->transaction->dirty_bgs_lock);
10302         mutex_unlock(&trans->transaction->cache_write_mutex);
10303
10304         if (!IS_ERR(inode)) {
10305                 ret = btrfs_orphan_add(trans, inode);
10306                 if (ret) {
10307                         btrfs_add_delayed_iput(inode);
10308                         goto out;
10309                 }
10310                 clear_nlink(inode);
10311                 /* One for the block groups ref */
10312                 spin_lock(&block_group->lock);
10313                 if (block_group->iref) {
10314                         block_group->iref = 0;
10315                         block_group->inode = NULL;
10316                         spin_unlock(&block_group->lock);
10317                         iput(inode);
10318                 } else {
10319                         spin_unlock(&block_group->lock);
10320                 }
10321                 /* One for our lookup ref */
10322                 btrfs_add_delayed_iput(inode);
10323         }
10324
10325         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10326         key.offset = block_group->key.objectid;
10327         key.type = 0;
10328
10329         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10330         if (ret < 0)
10331                 goto out;
10332         if (ret > 0)
10333                 btrfs_release_path(path);
10334         if (ret == 0) {
10335                 ret = btrfs_del_item(trans, tree_root, path);
10336                 if (ret)
10337                         goto out;
10338                 btrfs_release_path(path);
10339         }
10340
10341         spin_lock(&root->fs_info->block_group_cache_lock);
10342         rb_erase(&block_group->cache_node,
10343                  &root->fs_info->block_group_cache_tree);
10344         RB_CLEAR_NODE(&block_group->cache_node);
10345
10346         if (root->fs_info->first_logical_byte == block_group->key.objectid)
10347                 root->fs_info->first_logical_byte = (u64)-1;
10348         spin_unlock(&root->fs_info->block_group_cache_lock);
10349
10350         down_write(&block_group->space_info->groups_sem);
10351         /*
10352          * we must use list_del_init so people can check to see if they
10353          * are still on the list after taking the semaphore
10354          */
10355         list_del_init(&block_group->list);
10356         if (list_empty(&block_group->space_info->block_groups[index])) {
10357                 kobj = block_group->space_info->block_group_kobjs[index];
10358                 block_group->space_info->block_group_kobjs[index] = NULL;
10359                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
10360         }
10361         up_write(&block_group->space_info->groups_sem);
10362         if (kobj) {
10363                 kobject_del(kobj);
10364                 kobject_put(kobj);
10365         }
10366
10367         if (block_group->has_caching_ctl)
10368                 caching_ctl = get_caching_control(block_group);
10369         if (block_group->cached == BTRFS_CACHE_STARTED)
10370                 wait_block_group_cache_done(block_group);
10371         if (block_group->has_caching_ctl) {
10372                 down_write(&root->fs_info->commit_root_sem);
10373                 if (!caching_ctl) {
10374                         struct btrfs_caching_control *ctl;
10375
10376                         list_for_each_entry(ctl,
10377                                     &root->fs_info->caching_block_groups, list)
10378                                 if (ctl->block_group == block_group) {
10379                                         caching_ctl = ctl;
10380                                         atomic_inc(&caching_ctl->count);
10381                                         break;
10382                                 }
10383                 }
10384                 if (caching_ctl)
10385                         list_del_init(&caching_ctl->list);
10386                 up_write(&root->fs_info->commit_root_sem);
10387                 if (caching_ctl) {
10388                         /* Once for the caching bgs list and once for us. */
10389                         put_caching_control(caching_ctl);
10390                         put_caching_control(caching_ctl);
10391                 }
10392         }
10393
10394         spin_lock(&trans->transaction->dirty_bgs_lock);
10395         if (!list_empty(&block_group->dirty_list)) {
10396                 WARN_ON(1);
10397         }
10398         if (!list_empty(&block_group->io_list)) {
10399                 WARN_ON(1);
10400         }
10401         spin_unlock(&trans->transaction->dirty_bgs_lock);
10402         btrfs_remove_free_space_cache(block_group);
10403
10404         spin_lock(&block_group->space_info->lock);
10405         list_del_init(&block_group->ro_list);
10406
10407         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
10408                 WARN_ON(block_group->space_info->total_bytes
10409                         < block_group->key.offset);
10410                 WARN_ON(block_group->space_info->bytes_readonly
10411                         < block_group->key.offset);
10412                 WARN_ON(block_group->space_info->disk_total
10413                         < block_group->key.offset * factor);
10414         }
10415         block_group->space_info->total_bytes -= block_group->key.offset;
10416         block_group->space_info->bytes_readonly -= block_group->key.offset;
10417         block_group->space_info->disk_total -= block_group->key.offset * factor;
10418
10419         spin_unlock(&block_group->space_info->lock);
10420
10421         memcpy(&key, &block_group->key, sizeof(key));
10422
10423         lock_chunks(root);
10424         if (!list_empty(&em->list)) {
10425                 /* We're in the transaction->pending_chunks list. */
10426                 free_extent_map(em);
10427         }
10428         spin_lock(&block_group->lock);
10429         block_group->removed = 1;
10430         /*
10431          * At this point trimming can't start on this block group, because we
10432          * removed the block group from the tree fs_info->block_group_cache_tree
10433          * so no one can't find it anymore and even if someone already got this
10434          * block group before we removed it from the rbtree, they have already
10435          * incremented block_group->trimming - if they didn't, they won't find
10436          * any free space entries because we already removed them all when we
10437          * called btrfs_remove_free_space_cache().
10438          *
10439          * And we must not remove the extent map from the fs_info->mapping_tree
10440          * to prevent the same logical address range and physical device space
10441          * ranges from being reused for a new block group. This is because our
10442          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10443          * completely transactionless, so while it is trimming a range the
10444          * currently running transaction might finish and a new one start,
10445          * allowing for new block groups to be created that can reuse the same
10446          * physical device locations unless we take this special care.
10447          *
10448          * There may also be an implicit trim operation if the file system
10449          * is mounted with -odiscard. The same protections must remain
10450          * in place until the extents have been discarded completely when
10451          * the transaction commit has completed.
10452          */
10453         remove_em = (atomic_read(&block_group->trimming) == 0);
10454         /*
10455          * Make sure a trimmer task always sees the em in the pinned_chunks list
10456          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10457          * before checking block_group->removed).
10458          */
10459         if (!remove_em) {
10460                 /*
10461                  * Our em might be in trans->transaction->pending_chunks which
10462                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10463                  * and so is the fs_info->pinned_chunks list.
10464                  *
10465                  * So at this point we must be holding the chunk_mutex to avoid
10466                  * any races with chunk allocation (more specifically at
10467                  * volumes.c:contains_pending_extent()), to ensure it always
10468                  * sees the em, either in the pending_chunks list or in the
10469                  * pinned_chunks list.
10470                  */
10471                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10472         }
10473         spin_unlock(&block_group->lock);
10474
10475         if (remove_em) {
10476                 struct extent_map_tree *em_tree;
10477
10478                 em_tree = &root->fs_info->mapping_tree.map_tree;
10479                 write_lock(&em_tree->lock);
10480                 /*
10481                  * The em might be in the pending_chunks list, so make sure the
10482                  * chunk mutex is locked, since remove_extent_mapping() will
10483                  * delete us from that list.
10484                  */
10485                 remove_extent_mapping(em_tree, em);
10486                 write_unlock(&em_tree->lock);
10487                 /* once for the tree */
10488                 free_extent_map(em);
10489         }
10490
10491         unlock_chunks(root);
10492
10493         ret = remove_block_group_free_space(trans, root->fs_info, block_group);
10494         if (ret)
10495                 goto out;
10496
10497         btrfs_put_block_group(block_group);
10498         btrfs_put_block_group(block_group);
10499
10500         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10501         if (ret > 0)
10502                 ret = -EIO;
10503         if (ret < 0)
10504                 goto out;
10505
10506         ret = btrfs_del_item(trans, root, path);
10507 out:
10508         btrfs_free_path(path);
10509         return ret;
10510 }
10511
10512 struct btrfs_trans_handle *
10513 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10514                                      const u64 chunk_offset)
10515 {
10516         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10517         struct extent_map *em;
10518         struct map_lookup *map;
10519         unsigned int num_items;
10520
10521         read_lock(&em_tree->lock);
10522         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10523         read_unlock(&em_tree->lock);
10524         ASSERT(em && em->start == chunk_offset);
10525
10526         /*
10527          * We need to reserve 3 + N units from the metadata space info in order
10528          * to remove a block group (done at btrfs_remove_chunk() and at
10529          * btrfs_remove_block_group()), which are used for:
10530          *
10531          * 1 unit for adding the free space inode's orphan (located in the tree
10532          * of tree roots).
10533          * 1 unit for deleting the block group item (located in the extent
10534          * tree).
10535          * 1 unit for deleting the free space item (located in tree of tree
10536          * roots).
10537          * N units for deleting N device extent items corresponding to each
10538          * stripe (located in the device tree).
10539          *
10540          * In order to remove a block group we also need to reserve units in the
10541          * system space info in order to update the chunk tree (update one or
10542          * more device items and remove one chunk item), but this is done at
10543          * btrfs_remove_chunk() through a call to check_system_chunk().
10544          */
10545         map = em->map_lookup;
10546         num_items = 3 + map->num_stripes;
10547         free_extent_map(em);
10548
10549         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10550                                                            num_items, 1);
10551 }
10552
10553 /*
10554  * Process the unused_bgs list and remove any that don't have any allocated
10555  * space inside of them.
10556  */
10557 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10558 {
10559         struct btrfs_block_group_cache *block_group;
10560         struct btrfs_space_info *space_info;
10561         struct btrfs_root *root = fs_info->extent_root;
10562         struct btrfs_trans_handle *trans;
10563         int ret = 0;
10564
10565         if (!fs_info->open)
10566                 return;
10567
10568         spin_lock(&fs_info->unused_bgs_lock);
10569         while (!list_empty(&fs_info->unused_bgs)) {
10570                 u64 start, end;
10571                 int trimming;
10572
10573                 block_group = list_first_entry(&fs_info->unused_bgs,
10574                                                struct btrfs_block_group_cache,
10575                                                bg_list);
10576                 list_del_init(&block_group->bg_list);
10577
10578                 space_info = block_group->space_info;
10579
10580                 if (ret || btrfs_mixed_space_info(space_info)) {
10581                         btrfs_put_block_group(block_group);
10582                         continue;
10583                 }
10584                 spin_unlock(&fs_info->unused_bgs_lock);
10585
10586                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10587
10588                 /* Don't want to race with allocators so take the groups_sem */
10589                 down_write(&space_info->groups_sem);
10590                 spin_lock(&block_group->lock);
10591                 if (block_group->reserved ||
10592                     btrfs_block_group_used(&block_group->item) ||
10593                     block_group->ro ||
10594                     list_is_singular(&block_group->list)) {
10595                         /*
10596                          * We want to bail if we made new allocations or have
10597                          * outstanding allocations in this block group.  We do
10598                          * the ro check in case balance is currently acting on
10599                          * this block group.
10600                          */
10601                         spin_unlock(&block_group->lock);
10602                         up_write(&space_info->groups_sem);
10603                         goto next;
10604                 }
10605                 spin_unlock(&block_group->lock);
10606
10607                 /* We don't want to force the issue, only flip if it's ok. */
10608                 ret = inc_block_group_ro(block_group, 0);
10609                 up_write(&space_info->groups_sem);
10610                 if (ret < 0) {
10611                         ret = 0;
10612                         goto next;
10613                 }
10614
10615                 /*
10616                  * Want to do this before we do anything else so we can recover
10617                  * properly if we fail to join the transaction.
10618                  */
10619                 trans = btrfs_start_trans_remove_block_group(fs_info,
10620                                                      block_group->key.objectid);
10621                 if (IS_ERR(trans)) {
10622                         btrfs_dec_block_group_ro(root, block_group);
10623                         ret = PTR_ERR(trans);
10624                         goto next;
10625                 }
10626
10627                 /*
10628                  * We could have pending pinned extents for this block group,
10629                  * just delete them, we don't care about them anymore.
10630                  */
10631                 start = block_group->key.objectid;
10632                 end = start + block_group->key.offset - 1;
10633                 /*
10634                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10635                  * btrfs_finish_extent_commit(). If we are at transaction N,
10636                  * another task might be running finish_extent_commit() for the
10637                  * previous transaction N - 1, and have seen a range belonging
10638                  * to the block group in freed_extents[] before we were able to
10639                  * clear the whole block group range from freed_extents[]. This
10640                  * means that task can lookup for the block group after we
10641                  * unpinned it from freed_extents[] and removed it, leading to
10642                  * a BUG_ON() at btrfs_unpin_extent_range().
10643                  */
10644                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10645                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10646                                   EXTENT_DIRTY);
10647                 if (ret) {
10648                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10649                         btrfs_dec_block_group_ro(root, block_group);
10650                         goto end_trans;
10651                 }
10652                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10653                                   EXTENT_DIRTY);
10654                 if (ret) {
10655                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10656                         btrfs_dec_block_group_ro(root, block_group);
10657                         goto end_trans;
10658                 }
10659                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10660
10661                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10662                 spin_lock(&space_info->lock);
10663                 spin_lock(&block_group->lock);
10664
10665                 space_info->bytes_pinned -= block_group->pinned;
10666                 space_info->bytes_readonly += block_group->pinned;
10667                 percpu_counter_add(&space_info->total_bytes_pinned,
10668                                    -block_group->pinned);
10669                 block_group->pinned = 0;
10670
10671                 spin_unlock(&block_group->lock);
10672                 spin_unlock(&space_info->lock);
10673
10674                 /* DISCARD can flip during remount */
10675                 trimming = btrfs_test_opt(root, DISCARD);
10676
10677                 /* Implicit trim during transaction commit. */
10678                 if (trimming)
10679                         btrfs_get_block_group_trimming(block_group);
10680
10681                 /*
10682                  * Btrfs_remove_chunk will abort the transaction if things go
10683                  * horribly wrong.
10684                  */
10685                 ret = btrfs_remove_chunk(trans, root,
10686                                          block_group->key.objectid);
10687
10688                 if (ret) {
10689                         if (trimming)
10690                                 btrfs_put_block_group_trimming(block_group);
10691                         goto end_trans;
10692                 }
10693
10694                 /*
10695                  * If we're not mounted with -odiscard, we can just forget
10696                  * about this block group. Otherwise we'll need to wait
10697                  * until transaction commit to do the actual discard.
10698                  */
10699                 if (trimming) {
10700                         spin_lock(&fs_info->unused_bgs_lock);
10701                         /*
10702                          * A concurrent scrub might have added us to the list
10703                          * fs_info->unused_bgs, so use a list_move operation
10704                          * to add the block group to the deleted_bgs list.
10705                          */
10706                         list_move(&block_group->bg_list,
10707                                   &trans->transaction->deleted_bgs);
10708                         spin_unlock(&fs_info->unused_bgs_lock);
10709                         btrfs_get_block_group(block_group);
10710                 }
10711 end_trans:
10712                 btrfs_end_transaction(trans, root);
10713 next:
10714                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10715                 btrfs_put_block_group(block_group);
10716                 spin_lock(&fs_info->unused_bgs_lock);
10717         }
10718         spin_unlock(&fs_info->unused_bgs_lock);
10719 }
10720
10721 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10722 {
10723         struct btrfs_space_info *space_info;
10724         struct btrfs_super_block *disk_super;
10725         u64 features;
10726         u64 flags;
10727         int mixed = 0;
10728         int ret;
10729
10730         disk_super = fs_info->super_copy;
10731         if (!btrfs_super_root(disk_super))
10732                 return -EINVAL;
10733
10734         features = btrfs_super_incompat_flags(disk_super);
10735         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10736                 mixed = 1;
10737
10738         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10739         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10740         if (ret)
10741                 goto out;
10742
10743         if (mixed) {
10744                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10745                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10746         } else {
10747                 flags = BTRFS_BLOCK_GROUP_METADATA;
10748                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10749                 if (ret)
10750                         goto out;
10751
10752                 flags = BTRFS_BLOCK_GROUP_DATA;
10753                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10754         }
10755 out:
10756         return ret;
10757 }
10758
10759 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10760 {
10761         return unpin_extent_range(root, start, end, false);
10762 }
10763
10764 /*
10765  * It used to be that old block groups would be left around forever.
10766  * Iterating over them would be enough to trim unused space.  Since we
10767  * now automatically remove them, we also need to iterate over unallocated
10768  * space.
10769  *
10770  * We don't want a transaction for this since the discard may take a
10771  * substantial amount of time.  We don't require that a transaction be
10772  * running, but we do need to take a running transaction into account
10773  * to ensure that we're not discarding chunks that were released in
10774  * the current transaction.
10775  *
10776  * Holding the chunks lock will prevent other threads from allocating
10777  * or releasing chunks, but it won't prevent a running transaction
10778  * from committing and releasing the memory that the pending chunks
10779  * list head uses.  For that, we need to take a reference to the
10780  * transaction.
10781  */
10782 static int btrfs_trim_free_extents(struct btrfs_device *device,
10783                                    u64 minlen, u64 *trimmed)
10784 {
10785         u64 start = 0, len = 0;
10786         int ret;
10787
10788         *trimmed = 0;
10789
10790         /* Not writeable = nothing to do. */
10791         if (!device->writeable)
10792                 return 0;
10793
10794         /* No free space = nothing to do. */
10795         if (device->total_bytes <= device->bytes_used)
10796                 return 0;
10797
10798         ret = 0;
10799
10800         while (1) {
10801                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
10802                 struct btrfs_transaction *trans;
10803                 u64 bytes;
10804
10805                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10806                 if (ret)
10807                         return ret;
10808
10809                 down_read(&fs_info->commit_root_sem);
10810
10811                 spin_lock(&fs_info->trans_lock);
10812                 trans = fs_info->running_transaction;
10813                 if (trans)
10814                         atomic_inc(&trans->use_count);
10815                 spin_unlock(&fs_info->trans_lock);
10816
10817                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10818                                                  &start, &len);
10819                 if (trans)
10820                         btrfs_put_transaction(trans);
10821
10822                 if (ret) {
10823                         up_read(&fs_info->commit_root_sem);
10824                         mutex_unlock(&fs_info->chunk_mutex);
10825                         if (ret == -ENOSPC)
10826                                 ret = 0;
10827                         break;
10828                 }
10829
10830                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10831                 up_read(&fs_info->commit_root_sem);
10832                 mutex_unlock(&fs_info->chunk_mutex);
10833
10834                 if (ret)
10835                         break;
10836
10837                 start += len;
10838                 *trimmed += bytes;
10839
10840                 if (fatal_signal_pending(current)) {
10841                         ret = -ERESTARTSYS;
10842                         break;
10843                 }
10844
10845                 cond_resched();
10846         }
10847
10848         return ret;
10849 }
10850
10851 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10852 {
10853         struct btrfs_fs_info *fs_info = root->fs_info;
10854         struct btrfs_block_group_cache *cache = NULL;
10855         struct btrfs_device *device;
10856         struct list_head *devices;
10857         u64 group_trimmed;
10858         u64 start;
10859         u64 end;
10860         u64 trimmed = 0;
10861         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10862         int ret = 0;
10863
10864         /*
10865          * try to trim all FS space, our block group may start from non-zero.
10866          */
10867         if (range->len == total_bytes)
10868                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10869         else
10870                 cache = btrfs_lookup_block_group(fs_info, range->start);
10871
10872         while (cache) {
10873                 if (cache->key.objectid >= (range->start + range->len)) {
10874                         btrfs_put_block_group(cache);
10875                         break;
10876                 }
10877
10878                 start = max(range->start, cache->key.objectid);
10879                 end = min(range->start + range->len,
10880                                 cache->key.objectid + cache->key.offset);
10881
10882                 if (end - start >= range->minlen) {
10883                         if (!block_group_cache_done(cache)) {
10884                                 ret = cache_block_group(cache, 0);
10885                                 if (ret) {
10886                                         btrfs_put_block_group(cache);
10887                                         break;
10888                                 }
10889                                 ret = wait_block_group_cache_done(cache);
10890                                 if (ret) {
10891                                         btrfs_put_block_group(cache);
10892                                         break;
10893                                 }
10894                         }
10895                         ret = btrfs_trim_block_group(cache,
10896                                                      &group_trimmed,
10897                                                      start,
10898                                                      end,
10899                                                      range->minlen);
10900
10901                         trimmed += group_trimmed;
10902                         if (ret) {
10903                                 btrfs_put_block_group(cache);
10904                                 break;
10905                         }
10906                 }
10907
10908                 cache = next_block_group(fs_info->tree_root, cache);
10909         }
10910
10911         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
10912         devices = &root->fs_info->fs_devices->alloc_list;
10913         list_for_each_entry(device, devices, dev_alloc_list) {
10914                 ret = btrfs_trim_free_extents(device, range->minlen,
10915                                               &group_trimmed);
10916                 if (ret)
10917                         break;
10918
10919                 trimmed += group_trimmed;
10920         }
10921         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
10922
10923         range->len = trimmed;
10924         return ret;
10925 }
10926
10927 /*
10928  * btrfs_{start,end}_write_no_snapshoting() are similar to
10929  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10930  * data into the page cache through nocow before the subvolume is snapshoted,
10931  * but flush the data into disk after the snapshot creation, or to prevent
10932  * operations while snapshoting is ongoing and that cause the snapshot to be
10933  * inconsistent (writes followed by expanding truncates for example).
10934  */
10935 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10936 {
10937         percpu_counter_dec(&root->subv_writers->counter);
10938         /*
10939          * Make sure counter is updated before we wake up waiters.
10940          */
10941         smp_mb();
10942         if (waitqueue_active(&root->subv_writers->wait))
10943                 wake_up(&root->subv_writers->wait);
10944 }
10945
10946 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10947 {
10948         if (atomic_read(&root->will_be_snapshoted))
10949                 return 0;
10950
10951         percpu_counter_inc(&root->subv_writers->counter);
10952         /*
10953          * Make sure counter is updated before we check for snapshot creation.
10954          */
10955         smp_mb();
10956         if (atomic_read(&root->will_be_snapshoted)) {
10957                 btrfs_end_write_no_snapshoting(root);
10958                 return 0;
10959         }
10960         return 1;
10961 }
10962
10963 static int wait_snapshoting_atomic_t(atomic_t *a)
10964 {
10965         schedule();
10966         return 0;
10967 }
10968
10969 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
10970 {
10971         while (true) {
10972                 int ret;
10973
10974                 ret = btrfs_start_write_no_snapshoting(root);
10975                 if (ret)
10976                         break;
10977                 wait_on_atomic_t(&root->will_be_snapshoted,
10978                                  wait_snapshoting_atomic_t,
10979                                  TASK_UNINTERRUPTIBLE);
10980         }
10981 }