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         u64 transid;
2839         int count;
2840         int error;
2841         int sync;
2842         struct completion wait;
2843         struct btrfs_work work;
2844 };
2845
2846 static void delayed_ref_async_start(struct btrfs_work *work)
2847 {
2848         struct async_delayed_refs *async;
2849         struct btrfs_trans_handle *trans;
2850         int ret;
2851
2852         async = container_of(work, struct async_delayed_refs, work);
2853
2854         /* if the commit is already started, we don't need to wait here */
2855         if (btrfs_transaction_blocked(async->root->fs_info))
2856                 goto done;
2857
2858         trans = btrfs_join_transaction(async->root);
2859         if (IS_ERR(trans)) {
2860                 async->error = PTR_ERR(trans);
2861                 goto done;
2862         }
2863
2864         /*
2865          * trans->sync means that when we call end_transaction, we won't
2866          * wait on delayed refs
2867          */
2868         trans->sync = true;
2869
2870         /* Don't bother flushing if we got into a different transaction */
2871         if (trans->transid > async->transid)
2872                 goto end;
2873
2874         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2875         if (ret)
2876                 async->error = ret;
2877 end:
2878         ret = btrfs_end_transaction(trans, async->root);
2879         if (ret && !async->error)
2880                 async->error = ret;
2881 done:
2882         if (async->sync)
2883                 complete(&async->wait);
2884         else
2885                 kfree(async);
2886 }
2887
2888 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2889                                  unsigned long count, u64 transid, int wait)
2890 {
2891         struct async_delayed_refs *async;
2892         int ret;
2893
2894         async = kmalloc(sizeof(*async), GFP_NOFS);
2895         if (!async)
2896                 return -ENOMEM;
2897
2898         async->root = root->fs_info->tree_root;
2899         async->count = count;
2900         async->error = 0;
2901         async->transid = transid;
2902         if (wait)
2903                 async->sync = 1;
2904         else
2905                 async->sync = 0;
2906         init_completion(&async->wait);
2907
2908         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2909                         delayed_ref_async_start, NULL, NULL);
2910
2911         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2912
2913         if (wait) {
2914                 wait_for_completion(&async->wait);
2915                 ret = async->error;
2916                 kfree(async);
2917                 return ret;
2918         }
2919         return 0;
2920 }
2921
2922 /*
2923  * this starts processing the delayed reference count updates and
2924  * extent insertions we have queued up so far.  count can be
2925  * 0, which means to process everything in the tree at the start
2926  * of the run (but not newly added entries), or it can be some target
2927  * number you'd like to process.
2928  *
2929  * Returns 0 on success or if called with an aborted transaction
2930  * Returns <0 on error and aborts the transaction
2931  */
2932 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2933                            struct btrfs_root *root, unsigned long count)
2934 {
2935         struct rb_node *node;
2936         struct btrfs_delayed_ref_root *delayed_refs;
2937         struct btrfs_delayed_ref_head *head;
2938         int ret;
2939         int run_all = count == (unsigned long)-1;
2940         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2941
2942         /* We'll clean this up in btrfs_cleanup_transaction */
2943         if (trans->aborted)
2944                 return 0;
2945
2946         if (root->fs_info->creating_free_space_tree)
2947                 return 0;
2948
2949         if (root == root->fs_info->extent_root)
2950                 root = root->fs_info->tree_root;
2951
2952         delayed_refs = &trans->transaction->delayed_refs;
2953         if (count == 0)
2954                 count = atomic_read(&delayed_refs->num_entries) * 2;
2955
2956 again:
2957 #ifdef SCRAMBLE_DELAYED_REFS
2958         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2959 #endif
2960         trans->can_flush_pending_bgs = false;
2961         ret = __btrfs_run_delayed_refs(trans, root, count);
2962         if (ret < 0) {
2963                 btrfs_abort_transaction(trans, root, ret);
2964                 return ret;
2965         }
2966
2967         if (run_all) {
2968                 if (!list_empty(&trans->new_bgs))
2969                         btrfs_create_pending_block_groups(trans, root);
2970
2971                 spin_lock(&delayed_refs->lock);
2972                 node = rb_first(&delayed_refs->href_root);
2973                 if (!node) {
2974                         spin_unlock(&delayed_refs->lock);
2975                         goto out;
2976                 }
2977                 count = (unsigned long)-1;
2978
2979                 while (node) {
2980                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2981                                         href_node);
2982                         if (btrfs_delayed_ref_is_head(&head->node)) {
2983                                 struct btrfs_delayed_ref_node *ref;
2984
2985                                 ref = &head->node;
2986                                 atomic_inc(&ref->refs);
2987
2988                                 spin_unlock(&delayed_refs->lock);
2989                                 /*
2990                                  * Mutex was contended, block until it's
2991                                  * released and try again
2992                                  */
2993                                 mutex_lock(&head->mutex);
2994                                 mutex_unlock(&head->mutex);
2995
2996                                 btrfs_put_delayed_ref(ref);
2997                                 cond_resched();
2998                                 goto again;
2999                         } else {
3000                                 WARN_ON(1);
3001                         }
3002                         node = rb_next(node);
3003                 }
3004                 spin_unlock(&delayed_refs->lock);
3005                 cond_resched();
3006                 goto again;
3007         }
3008 out:
3009         assert_qgroups_uptodate(trans);
3010         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3011         return 0;
3012 }
3013
3014 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3015                                 struct btrfs_root *root,
3016                                 u64 bytenr, u64 num_bytes, u64 flags,
3017                                 int level, int is_data)
3018 {
3019         struct btrfs_delayed_extent_op *extent_op;
3020         int ret;
3021
3022         extent_op = btrfs_alloc_delayed_extent_op();
3023         if (!extent_op)
3024                 return -ENOMEM;
3025
3026         extent_op->flags_to_set = flags;
3027         extent_op->update_flags = true;
3028         extent_op->update_key = false;
3029         extent_op->is_data = is_data ? true : false;
3030         extent_op->level = level;
3031
3032         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
3033                                           num_bytes, extent_op);
3034         if (ret)
3035                 btrfs_free_delayed_extent_op(extent_op);
3036         return ret;
3037 }
3038
3039 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
3040                                       struct btrfs_root *root,
3041                                       struct btrfs_path *path,
3042                                       u64 objectid, u64 offset, u64 bytenr)
3043 {
3044         struct btrfs_delayed_ref_head *head;
3045         struct btrfs_delayed_ref_node *ref;
3046         struct btrfs_delayed_data_ref *data_ref;
3047         struct btrfs_delayed_ref_root *delayed_refs;
3048         int ret = 0;
3049
3050         delayed_refs = &trans->transaction->delayed_refs;
3051         spin_lock(&delayed_refs->lock);
3052         head = btrfs_find_delayed_ref_head(trans, bytenr);
3053         if (!head) {
3054                 spin_unlock(&delayed_refs->lock);
3055                 return 0;
3056         }
3057
3058         if (!mutex_trylock(&head->mutex)) {
3059                 atomic_inc(&head->node.refs);
3060                 spin_unlock(&delayed_refs->lock);
3061
3062                 btrfs_release_path(path);
3063
3064                 /*
3065                  * Mutex was contended, block until it's released and let
3066                  * caller try again
3067                  */
3068                 mutex_lock(&head->mutex);
3069                 mutex_unlock(&head->mutex);
3070                 btrfs_put_delayed_ref(&head->node);
3071                 return -EAGAIN;
3072         }
3073         spin_unlock(&delayed_refs->lock);
3074
3075         spin_lock(&head->lock);
3076         list_for_each_entry(ref, &head->ref_list, list) {
3077                 /* If it's a shared ref we know a cross reference exists */
3078                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3079                         ret = 1;
3080                         break;
3081                 }
3082
3083                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3084
3085                 /*
3086                  * If our ref doesn't match the one we're currently looking at
3087                  * then we have a cross reference.
3088                  */
3089                 if (data_ref->root != root->root_key.objectid ||
3090                     data_ref->objectid != objectid ||
3091                     data_ref->offset != offset) {
3092                         ret = 1;
3093                         break;
3094                 }
3095         }
3096         spin_unlock(&head->lock);
3097         mutex_unlock(&head->mutex);
3098         return ret;
3099 }
3100
3101 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
3102                                         struct btrfs_root *root,
3103                                         struct btrfs_path *path,
3104                                         u64 objectid, u64 offset, u64 bytenr)
3105 {
3106         struct btrfs_root *extent_root = root->fs_info->extent_root;
3107         struct extent_buffer *leaf;
3108         struct btrfs_extent_data_ref *ref;
3109         struct btrfs_extent_inline_ref *iref;
3110         struct btrfs_extent_item *ei;
3111         struct btrfs_key key;
3112         u32 item_size;
3113         int ret;
3114
3115         key.objectid = bytenr;
3116         key.offset = (u64)-1;
3117         key.type = BTRFS_EXTENT_ITEM_KEY;
3118
3119         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3120         if (ret < 0)
3121                 goto out;
3122         BUG_ON(ret == 0); /* Corruption */
3123
3124         ret = -ENOENT;
3125         if (path->slots[0] == 0)
3126                 goto out;
3127
3128         path->slots[0]--;
3129         leaf = path->nodes[0];
3130         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3131
3132         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3133                 goto out;
3134
3135         ret = 1;
3136         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3137 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3138         if (item_size < sizeof(*ei)) {
3139                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3140                 goto out;
3141         }
3142 #endif
3143         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3144
3145         if (item_size != sizeof(*ei) +
3146             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3147                 goto out;
3148
3149         if (btrfs_extent_generation(leaf, ei) <=
3150             btrfs_root_last_snapshot(&root->root_item))
3151                 goto out;
3152
3153         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3154         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3155             BTRFS_EXTENT_DATA_REF_KEY)
3156                 goto out;
3157
3158         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3159         if (btrfs_extent_refs(leaf, ei) !=
3160             btrfs_extent_data_ref_count(leaf, ref) ||
3161             btrfs_extent_data_ref_root(leaf, ref) !=
3162             root->root_key.objectid ||
3163             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3164             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3165                 goto out;
3166
3167         ret = 0;
3168 out:
3169         return ret;
3170 }
3171
3172 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3173                           struct btrfs_root *root,
3174                           u64 objectid, u64 offset, u64 bytenr)
3175 {
3176         struct btrfs_path *path;
3177         int ret;
3178         int ret2;
3179
3180         path = btrfs_alloc_path();
3181         if (!path)
3182                 return -ENOENT;
3183
3184         do {
3185                 ret = check_committed_ref(trans, root, path, objectid,
3186                                           offset, bytenr);
3187                 if (ret && ret != -ENOENT)
3188                         goto out;
3189
3190                 ret2 = check_delayed_ref(trans, root, path, objectid,
3191                                          offset, bytenr);
3192         } while (ret2 == -EAGAIN);
3193
3194         if (ret2 && ret2 != -ENOENT) {
3195                 ret = ret2;
3196                 goto out;
3197         }
3198
3199         if (ret != -ENOENT || ret2 != -ENOENT)
3200                 ret = 0;
3201 out:
3202         btrfs_free_path(path);
3203         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3204                 WARN_ON(ret > 0);
3205         return ret;
3206 }
3207
3208 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3209                            struct btrfs_root *root,
3210                            struct extent_buffer *buf,
3211                            int full_backref, int inc)
3212 {
3213         u64 bytenr;
3214         u64 num_bytes;
3215         u64 parent;
3216         u64 ref_root;
3217         u32 nritems;
3218         struct btrfs_key key;
3219         struct btrfs_file_extent_item *fi;
3220         int i;
3221         int level;
3222         int ret = 0;
3223         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3224                             u64, u64, u64, u64, u64, u64);
3225
3226
3227         if (btrfs_test_is_dummy_root(root))
3228                 return 0;
3229
3230         ref_root = btrfs_header_owner(buf);
3231         nritems = btrfs_header_nritems(buf);
3232         level = btrfs_header_level(buf);
3233
3234         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3235                 return 0;
3236
3237         if (inc)
3238                 process_func = btrfs_inc_extent_ref;
3239         else
3240                 process_func = btrfs_free_extent;
3241
3242         if (full_backref)
3243                 parent = buf->start;
3244         else
3245                 parent = 0;
3246
3247         for (i = 0; i < nritems; i++) {
3248                 if (level == 0) {
3249                         btrfs_item_key_to_cpu(buf, &key, i);
3250                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3251                                 continue;
3252                         fi = btrfs_item_ptr(buf, i,
3253                                             struct btrfs_file_extent_item);
3254                         if (btrfs_file_extent_type(buf, fi) ==
3255                             BTRFS_FILE_EXTENT_INLINE)
3256                                 continue;
3257                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3258                         if (bytenr == 0)
3259                                 continue;
3260
3261                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3262                         key.offset -= btrfs_file_extent_offset(buf, fi);
3263                         ret = process_func(trans, root, bytenr, num_bytes,
3264                                            parent, ref_root, key.objectid,
3265                                            key.offset);
3266                         if (ret)
3267                                 goto fail;
3268                 } else {
3269                         bytenr = btrfs_node_blockptr(buf, i);
3270                         num_bytes = root->nodesize;
3271                         ret = process_func(trans, root, bytenr, num_bytes,
3272                                            parent, ref_root, level - 1, 0);
3273                         if (ret)
3274                                 goto fail;
3275                 }
3276         }
3277         return 0;
3278 fail:
3279         return ret;
3280 }
3281
3282 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3283                   struct extent_buffer *buf, int full_backref)
3284 {
3285         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3286 }
3287
3288 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3289                   struct extent_buffer *buf, int full_backref)
3290 {
3291         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3292 }
3293
3294 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3295                                  struct btrfs_root *root,
3296                                  struct btrfs_path *path,
3297                                  struct btrfs_block_group_cache *cache)
3298 {
3299         int ret;
3300         struct btrfs_root *extent_root = root->fs_info->extent_root;
3301         unsigned long bi;
3302         struct extent_buffer *leaf;
3303
3304         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3305         if (ret) {
3306                 if (ret > 0)
3307                         ret = -ENOENT;
3308                 goto fail;
3309         }
3310
3311         leaf = path->nodes[0];
3312         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3313         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3314         btrfs_mark_buffer_dirty(leaf);
3315 fail:
3316         btrfs_release_path(path);
3317         return ret;
3318
3319 }
3320
3321 static struct btrfs_block_group_cache *
3322 next_block_group(struct btrfs_root *root,
3323                  struct btrfs_block_group_cache *cache)
3324 {
3325         struct rb_node *node;
3326
3327         spin_lock(&root->fs_info->block_group_cache_lock);
3328
3329         /* If our block group was removed, we need a full search. */
3330         if (RB_EMPTY_NODE(&cache->cache_node)) {
3331                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3332
3333                 spin_unlock(&root->fs_info->block_group_cache_lock);
3334                 btrfs_put_block_group(cache);
3335                 cache = btrfs_lookup_first_block_group(root->fs_info,
3336                                                        next_bytenr);
3337                 return cache;
3338         }
3339         node = rb_next(&cache->cache_node);
3340         btrfs_put_block_group(cache);
3341         if (node) {
3342                 cache = rb_entry(node, struct btrfs_block_group_cache,
3343                                  cache_node);
3344                 btrfs_get_block_group(cache);
3345         } else
3346                 cache = NULL;
3347         spin_unlock(&root->fs_info->block_group_cache_lock);
3348         return cache;
3349 }
3350
3351 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3352                             struct btrfs_trans_handle *trans,
3353                             struct btrfs_path *path)
3354 {
3355         struct btrfs_root *root = block_group->fs_info->tree_root;
3356         struct inode *inode = NULL;
3357         u64 alloc_hint = 0;
3358         int dcs = BTRFS_DC_ERROR;
3359         u64 num_pages = 0;
3360         int retries = 0;
3361         int ret = 0;
3362
3363         /*
3364          * If this block group is smaller than 100 megs don't bother caching the
3365          * block group.
3366          */
3367         if (block_group->key.offset < (100 * SZ_1M)) {
3368                 spin_lock(&block_group->lock);
3369                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3370                 spin_unlock(&block_group->lock);
3371                 return 0;
3372         }
3373
3374         if (trans->aborted)
3375                 return 0;
3376 again:
3377         inode = lookup_free_space_inode(root, block_group, path);
3378         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3379                 ret = PTR_ERR(inode);
3380                 btrfs_release_path(path);
3381                 goto out;
3382         }
3383
3384         if (IS_ERR(inode)) {
3385                 BUG_ON(retries);
3386                 retries++;
3387
3388                 if (block_group->ro)
3389                         goto out_free;
3390
3391                 ret = create_free_space_inode(root, trans, block_group, path);
3392                 if (ret)
3393                         goto out_free;
3394                 goto again;
3395         }
3396
3397         /* We've already setup this transaction, go ahead and exit */
3398         if (block_group->cache_generation == trans->transid &&
3399             i_size_read(inode)) {
3400                 dcs = BTRFS_DC_SETUP;
3401                 goto out_put;
3402         }
3403
3404         /*
3405          * We want to set the generation to 0, that way if anything goes wrong
3406          * from here on out we know not to trust this cache when we load up next
3407          * time.
3408          */
3409         BTRFS_I(inode)->generation = 0;
3410         ret = btrfs_update_inode(trans, root, inode);
3411         if (ret) {
3412                 /*
3413                  * So theoretically we could recover from this, simply set the
3414                  * super cache generation to 0 so we know to invalidate the
3415                  * cache, but then we'd have to keep track of the block groups
3416                  * that fail this way so we know we _have_ to reset this cache
3417                  * before the next commit or risk reading stale cache.  So to
3418                  * limit our exposure to horrible edge cases lets just abort the
3419                  * transaction, this only happens in really bad situations
3420                  * anyway.
3421                  */
3422                 btrfs_abort_transaction(trans, root, ret);
3423                 goto out_put;
3424         }
3425         WARN_ON(ret);
3426
3427         if (i_size_read(inode) > 0) {
3428                 ret = btrfs_check_trunc_cache_free_space(root,
3429                                         &root->fs_info->global_block_rsv);
3430                 if (ret)
3431                         goto out_put;
3432
3433                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3434                 if (ret)
3435                         goto out_put;
3436         }
3437
3438         spin_lock(&block_group->lock);
3439         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3440             !btrfs_test_opt(root, SPACE_CACHE)) {
3441                 /*
3442                  * don't bother trying to write stuff out _if_
3443                  * a) we're not cached,
3444                  * b) we're with nospace_cache mount option.
3445                  */
3446                 dcs = BTRFS_DC_WRITTEN;
3447                 spin_unlock(&block_group->lock);
3448                 goto out_put;
3449         }
3450         spin_unlock(&block_group->lock);
3451
3452         /*
3453          * We hit an ENOSPC when setting up the cache in this transaction, just
3454          * skip doing the setup, we've already cleared the cache so we're safe.
3455          */
3456         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3457                 ret = -ENOSPC;
3458                 goto out_put;
3459         }
3460
3461         /*
3462          * Try to preallocate enough space based on how big the block group is.
3463          * Keep in mind this has to include any pinned space which could end up
3464          * taking up quite a bit since it's not folded into the other space
3465          * cache.
3466          */
3467         num_pages = div_u64(block_group->key.offset, SZ_256M);
3468         if (!num_pages)
3469                 num_pages = 1;
3470
3471         num_pages *= 16;
3472         num_pages *= PAGE_SIZE;
3473
3474         ret = btrfs_check_data_free_space(inode, 0, num_pages);
3475         if (ret)
3476                 goto out_put;
3477
3478         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3479                                               num_pages, num_pages,
3480                                               &alloc_hint);
3481         /*
3482          * Our cache requires contiguous chunks so that we don't modify a bunch
3483          * of metadata or split extents when writing the cache out, which means
3484          * we can enospc if we are heavily fragmented in addition to just normal
3485          * out of space conditions.  So if we hit this just skip setting up any
3486          * other block groups for this transaction, maybe we'll unpin enough
3487          * space the next time around.
3488          */
3489         if (!ret)
3490                 dcs = BTRFS_DC_SETUP;
3491         else if (ret == -ENOSPC)
3492                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3493         btrfs_free_reserved_data_space(inode, 0, num_pages);
3494
3495 out_put:
3496         iput(inode);
3497 out_free:
3498         btrfs_release_path(path);
3499 out:
3500         spin_lock(&block_group->lock);
3501         if (!ret && dcs == BTRFS_DC_SETUP)
3502                 block_group->cache_generation = trans->transid;
3503         block_group->disk_cache_state = dcs;
3504         spin_unlock(&block_group->lock);
3505
3506         return ret;
3507 }
3508
3509 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3510                             struct btrfs_root *root)
3511 {
3512         struct btrfs_block_group_cache *cache, *tmp;
3513         struct btrfs_transaction *cur_trans = trans->transaction;
3514         struct btrfs_path *path;
3515
3516         if (list_empty(&cur_trans->dirty_bgs) ||
3517             !btrfs_test_opt(root, SPACE_CACHE))
3518                 return 0;
3519
3520         path = btrfs_alloc_path();
3521         if (!path)
3522                 return -ENOMEM;
3523
3524         /* Could add new block groups, use _safe just in case */
3525         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3526                                  dirty_list) {
3527                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3528                         cache_save_setup(cache, trans, path);
3529         }
3530
3531         btrfs_free_path(path);
3532         return 0;
3533 }
3534
3535 /*
3536  * transaction commit does final block group cache writeback during a
3537  * critical section where nothing is allowed to change the FS.  This is
3538  * required in order for the cache to actually match the block group,
3539  * but can introduce a lot of latency into the commit.
3540  *
3541  * So, btrfs_start_dirty_block_groups is here to kick off block group
3542  * cache IO.  There's a chance we'll have to redo some of it if the
3543  * block group changes again during the commit, but it greatly reduces
3544  * the commit latency by getting rid of the easy block groups while
3545  * we're still allowing others to join the commit.
3546  */
3547 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3548                                    struct btrfs_root *root)
3549 {
3550         struct btrfs_block_group_cache *cache;
3551         struct btrfs_transaction *cur_trans = trans->transaction;
3552         int ret = 0;
3553         int should_put;
3554         struct btrfs_path *path = NULL;
3555         LIST_HEAD(dirty);
3556         struct list_head *io = &cur_trans->io_bgs;
3557         int num_started = 0;
3558         int loops = 0;
3559
3560         spin_lock(&cur_trans->dirty_bgs_lock);
3561         if (list_empty(&cur_trans->dirty_bgs)) {
3562                 spin_unlock(&cur_trans->dirty_bgs_lock);
3563                 return 0;
3564         }
3565         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3566         spin_unlock(&cur_trans->dirty_bgs_lock);
3567
3568 again:
3569         /*
3570          * make sure all the block groups on our dirty list actually
3571          * exist
3572          */
3573         btrfs_create_pending_block_groups(trans, root);
3574
3575         if (!path) {
3576                 path = btrfs_alloc_path();
3577                 if (!path)
3578                         return -ENOMEM;
3579         }
3580
3581         /*
3582          * cache_write_mutex is here only to save us from balance or automatic
3583          * removal of empty block groups deleting this block group while we are
3584          * writing out the cache
3585          */
3586         mutex_lock(&trans->transaction->cache_write_mutex);
3587         while (!list_empty(&dirty)) {
3588                 cache = list_first_entry(&dirty,
3589                                          struct btrfs_block_group_cache,
3590                                          dirty_list);
3591                 /*
3592                  * this can happen if something re-dirties a block
3593                  * group that is already under IO.  Just wait for it to
3594                  * finish and then do it all again
3595                  */
3596                 if (!list_empty(&cache->io_list)) {
3597                         list_del_init(&cache->io_list);
3598                         btrfs_wait_cache_io(root, trans, cache,
3599                                             &cache->io_ctl, path,
3600                                             cache->key.objectid);
3601                         btrfs_put_block_group(cache);
3602                 }
3603
3604
3605                 /*
3606                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3607                  * if it should update the cache_state.  Don't delete
3608                  * until after we wait.
3609                  *
3610                  * Since we're not running in the commit critical section
3611                  * we need the dirty_bgs_lock to protect from update_block_group
3612                  */
3613                 spin_lock(&cur_trans->dirty_bgs_lock);
3614                 list_del_init(&cache->dirty_list);
3615                 spin_unlock(&cur_trans->dirty_bgs_lock);
3616
3617                 should_put = 1;
3618
3619                 cache_save_setup(cache, trans, path);
3620
3621                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3622                         cache->io_ctl.inode = NULL;
3623                         ret = btrfs_write_out_cache(root, trans, cache, path);
3624                         if (ret == 0 && cache->io_ctl.inode) {
3625                                 num_started++;
3626                                 should_put = 0;
3627
3628                                 /*
3629                                  * the cache_write_mutex is protecting
3630                                  * the io_list
3631                                  */
3632                                 list_add_tail(&cache->io_list, io);
3633                         } else {
3634                                 /*
3635                                  * if we failed to write the cache, the
3636                                  * generation will be bad and life goes on
3637                                  */
3638                                 ret = 0;
3639                         }
3640                 }
3641                 if (!ret) {
3642                         ret = write_one_cache_group(trans, root, path, cache);
3643                         /*
3644                          * Our block group might still be attached to the list
3645                          * of new block groups in the transaction handle of some
3646                          * other task (struct btrfs_trans_handle->new_bgs). This
3647                          * means its block group item isn't yet in the extent
3648                          * tree. If this happens ignore the error, as we will
3649                          * try again later in the critical section of the
3650                          * transaction commit.
3651                          */
3652                         if (ret == -ENOENT) {
3653                                 ret = 0;
3654                                 spin_lock(&cur_trans->dirty_bgs_lock);
3655                                 if (list_empty(&cache->dirty_list)) {
3656                                         list_add_tail(&cache->dirty_list,
3657                                                       &cur_trans->dirty_bgs);
3658                                         btrfs_get_block_group(cache);
3659                                 }
3660                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3661                         } else if (ret) {
3662                                 btrfs_abort_transaction(trans, root, ret);
3663                         }
3664                 }
3665
3666                 /* if its not on the io list, we need to put the block group */
3667                 if (should_put)
3668                         btrfs_put_block_group(cache);
3669
3670                 if (ret)
3671                         break;
3672
3673                 /*
3674                  * Avoid blocking other tasks for too long. It might even save
3675                  * us from writing caches for block groups that are going to be
3676                  * removed.
3677                  */
3678                 mutex_unlock(&trans->transaction->cache_write_mutex);
3679                 mutex_lock(&trans->transaction->cache_write_mutex);
3680         }
3681         mutex_unlock(&trans->transaction->cache_write_mutex);
3682
3683         /*
3684          * go through delayed refs for all the stuff we've just kicked off
3685          * and then loop back (just once)
3686          */
3687         ret = btrfs_run_delayed_refs(trans, root, 0);
3688         if (!ret && loops == 0) {
3689                 loops++;
3690                 spin_lock(&cur_trans->dirty_bgs_lock);
3691                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3692                 /*
3693                  * dirty_bgs_lock protects us from concurrent block group
3694                  * deletes too (not just cache_write_mutex).
3695                  */
3696                 if (!list_empty(&dirty)) {
3697                         spin_unlock(&cur_trans->dirty_bgs_lock);
3698                         goto again;
3699                 }
3700                 spin_unlock(&cur_trans->dirty_bgs_lock);
3701         }
3702
3703         btrfs_free_path(path);
3704         return ret;
3705 }
3706
3707 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3708                                    struct btrfs_root *root)
3709 {
3710         struct btrfs_block_group_cache *cache;
3711         struct btrfs_transaction *cur_trans = trans->transaction;
3712         int ret = 0;
3713         int should_put;
3714         struct btrfs_path *path;
3715         struct list_head *io = &cur_trans->io_bgs;
3716         int num_started = 0;
3717
3718         path = btrfs_alloc_path();
3719         if (!path)
3720                 return -ENOMEM;
3721
3722         /*
3723          * Even though we are in the critical section of the transaction commit,
3724          * we can still have concurrent tasks adding elements to this
3725          * transaction's list of dirty block groups. These tasks correspond to
3726          * endio free space workers started when writeback finishes for a
3727          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3728          * allocate new block groups as a result of COWing nodes of the root
3729          * tree when updating the free space inode. The writeback for the space
3730          * caches is triggered by an earlier call to
3731          * btrfs_start_dirty_block_groups() and iterations of the following
3732          * loop.
3733          * Also we want to do the cache_save_setup first and then run the
3734          * delayed refs to make sure we have the best chance at doing this all
3735          * in one shot.
3736          */
3737         spin_lock(&cur_trans->dirty_bgs_lock);
3738         while (!list_empty(&cur_trans->dirty_bgs)) {
3739                 cache = list_first_entry(&cur_trans->dirty_bgs,
3740                                          struct btrfs_block_group_cache,
3741                                          dirty_list);
3742
3743                 /*
3744                  * this can happen if cache_save_setup re-dirties a block
3745                  * group that is already under IO.  Just wait for it to
3746                  * finish and then do it all again
3747                  */
3748                 if (!list_empty(&cache->io_list)) {
3749                         spin_unlock(&cur_trans->dirty_bgs_lock);
3750                         list_del_init(&cache->io_list);
3751                         btrfs_wait_cache_io(root, trans, cache,
3752                                             &cache->io_ctl, path,
3753                                             cache->key.objectid);
3754                         btrfs_put_block_group(cache);
3755                         spin_lock(&cur_trans->dirty_bgs_lock);
3756                 }
3757
3758                 /*
3759                  * don't remove from the dirty list until after we've waited
3760                  * on any pending IO
3761                  */
3762                 list_del_init(&cache->dirty_list);
3763                 spin_unlock(&cur_trans->dirty_bgs_lock);
3764                 should_put = 1;
3765
3766                 cache_save_setup(cache, trans, path);
3767
3768                 if (!ret)
3769                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3770
3771                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3772                         cache->io_ctl.inode = NULL;
3773                         ret = btrfs_write_out_cache(root, trans, cache, path);
3774                         if (ret == 0 && cache->io_ctl.inode) {
3775                                 num_started++;
3776                                 should_put = 0;
3777                                 list_add_tail(&cache->io_list, io);
3778                         } else {
3779                                 /*
3780                                  * if we failed to write the cache, the
3781                                  * generation will be bad and life goes on
3782                                  */
3783                                 ret = 0;
3784                         }
3785                 }
3786                 if (!ret) {
3787                         ret = write_one_cache_group(trans, root, path, cache);
3788                         /*
3789                          * One of the free space endio workers might have
3790                          * created a new block group while updating a free space
3791                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3792                          * and hasn't released its transaction handle yet, in
3793                          * which case the new block group is still attached to
3794                          * its transaction handle and its creation has not
3795                          * finished yet (no block group item in the extent tree
3796                          * yet, etc). If this is the case, wait for all free
3797                          * space endio workers to finish and retry. This is a
3798                          * a very rare case so no need for a more efficient and
3799                          * complex approach.
3800                          */
3801                         if (ret == -ENOENT) {
3802                                 wait_event(cur_trans->writer_wait,
3803                                    atomic_read(&cur_trans->num_writers) == 1);
3804                                 ret = write_one_cache_group(trans, root, path,
3805                                                             cache);
3806                         }
3807                         if (ret)
3808                                 btrfs_abort_transaction(trans, root, ret);
3809                 }
3810
3811                 /* if its not on the io list, we need to put the block group */
3812                 if (should_put)
3813                         btrfs_put_block_group(cache);
3814                 spin_lock(&cur_trans->dirty_bgs_lock);
3815         }
3816         spin_unlock(&cur_trans->dirty_bgs_lock);
3817
3818         while (!list_empty(io)) {
3819                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3820                                          io_list);
3821                 list_del_init(&cache->io_list);
3822                 btrfs_wait_cache_io(root, trans, cache,
3823                                     &cache->io_ctl, path, cache->key.objectid);
3824                 btrfs_put_block_group(cache);
3825         }
3826
3827         btrfs_free_path(path);
3828         return ret;
3829 }
3830
3831 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3832 {
3833         struct btrfs_block_group_cache *block_group;
3834         int readonly = 0;
3835
3836         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3837         if (!block_group || block_group->ro)
3838                 readonly = 1;
3839         if (block_group)
3840                 btrfs_put_block_group(block_group);
3841         return readonly;
3842 }
3843
3844 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3845 {
3846         struct btrfs_block_group_cache *bg;
3847         bool ret = true;
3848
3849         bg = btrfs_lookup_block_group(fs_info, bytenr);
3850         if (!bg)
3851                 return false;
3852
3853         spin_lock(&bg->lock);
3854         if (bg->ro)
3855                 ret = false;
3856         else
3857                 atomic_inc(&bg->nocow_writers);
3858         spin_unlock(&bg->lock);
3859
3860         /* no put on block group, done by btrfs_dec_nocow_writers */
3861         if (!ret)
3862                 btrfs_put_block_group(bg);
3863
3864         return ret;
3865
3866 }
3867
3868 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3869 {
3870         struct btrfs_block_group_cache *bg;
3871
3872         bg = btrfs_lookup_block_group(fs_info, bytenr);
3873         ASSERT(bg);
3874         if (atomic_dec_and_test(&bg->nocow_writers))
3875                 wake_up_atomic_t(&bg->nocow_writers);
3876         /*
3877          * Once for our lookup and once for the lookup done by a previous call
3878          * to btrfs_inc_nocow_writers()
3879          */
3880         btrfs_put_block_group(bg);
3881         btrfs_put_block_group(bg);
3882 }
3883
3884 static int btrfs_wait_nocow_writers_atomic_t(atomic_t *a)
3885 {
3886         schedule();
3887         return 0;
3888 }
3889
3890 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3891 {
3892         wait_on_atomic_t(&bg->nocow_writers,
3893                          btrfs_wait_nocow_writers_atomic_t,
3894                          TASK_UNINTERRUPTIBLE);
3895 }
3896
3897 static const char *alloc_name(u64 flags)
3898 {
3899         switch (flags) {
3900         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3901                 return "mixed";
3902         case BTRFS_BLOCK_GROUP_METADATA:
3903                 return "metadata";
3904         case BTRFS_BLOCK_GROUP_DATA:
3905                 return "data";
3906         case BTRFS_BLOCK_GROUP_SYSTEM:
3907                 return "system";
3908         default:
3909                 WARN_ON(1);
3910                 return "invalid-combination";
3911         };
3912 }
3913
3914 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3915                              u64 total_bytes, u64 bytes_used,
3916                              u64 bytes_readonly,
3917                              struct btrfs_space_info **space_info)
3918 {
3919         struct btrfs_space_info *found;
3920         int i;
3921         int factor;
3922         int ret;
3923
3924         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3925                      BTRFS_BLOCK_GROUP_RAID10))
3926                 factor = 2;
3927         else
3928                 factor = 1;
3929
3930         found = __find_space_info(info, flags);
3931         if (found) {
3932                 spin_lock(&found->lock);
3933                 found->total_bytes += total_bytes;
3934                 found->disk_total += total_bytes * factor;
3935                 found->bytes_used += bytes_used;
3936                 found->disk_used += bytes_used * factor;
3937                 found->bytes_readonly += bytes_readonly;
3938                 if (total_bytes > 0)
3939                         found->full = 0;
3940                 spin_unlock(&found->lock);
3941                 *space_info = found;
3942                 return 0;
3943         }
3944         found = kzalloc(sizeof(*found), GFP_NOFS);
3945         if (!found)
3946                 return -ENOMEM;
3947
3948         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3949         if (ret) {
3950                 kfree(found);
3951                 return ret;
3952         }
3953
3954         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3955                 INIT_LIST_HEAD(&found->block_groups[i]);
3956         init_rwsem(&found->groups_sem);
3957         spin_lock_init(&found->lock);
3958         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3959         found->total_bytes = total_bytes;
3960         found->disk_total = total_bytes * factor;
3961         found->bytes_used = bytes_used;
3962         found->disk_used = bytes_used * factor;
3963         found->bytes_pinned = 0;
3964         found->bytes_reserved = 0;
3965         found->bytes_readonly = bytes_readonly;
3966         found->bytes_may_use = 0;
3967         found->full = 0;
3968         found->max_extent_size = 0;
3969         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3970         found->chunk_alloc = 0;
3971         found->flush = 0;
3972         init_waitqueue_head(&found->wait);
3973         INIT_LIST_HEAD(&found->ro_bgs);
3974
3975         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3976                                     info->space_info_kobj, "%s",
3977                                     alloc_name(found->flags));
3978         if (ret) {
3979                 kfree(found);
3980                 return ret;
3981         }
3982
3983         *space_info = found;
3984         list_add_rcu(&found->list, &info->space_info);
3985         if (flags & BTRFS_BLOCK_GROUP_DATA)
3986                 info->data_sinfo = found;
3987
3988         return ret;
3989 }
3990
3991 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3992 {
3993         u64 extra_flags = chunk_to_extended(flags) &
3994                                 BTRFS_EXTENDED_PROFILE_MASK;
3995
3996         write_seqlock(&fs_info->profiles_lock);
3997         if (flags & BTRFS_BLOCK_GROUP_DATA)
3998                 fs_info->avail_data_alloc_bits |= extra_flags;
3999         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4000                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4001         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4002                 fs_info->avail_system_alloc_bits |= extra_flags;
4003         write_sequnlock(&fs_info->profiles_lock);
4004 }
4005
4006 /*
4007  * returns target flags in extended format or 0 if restripe for this
4008  * chunk_type is not in progress
4009  *
4010  * should be called with either volume_mutex or balance_lock held
4011  */
4012 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4013 {
4014         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4015         u64 target = 0;
4016
4017         if (!bctl)
4018                 return 0;
4019
4020         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4021             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4022                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4023         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4024                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4025                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4026         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4027                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4028                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4029         }
4030
4031         return target;
4032 }
4033
4034 /*
4035  * @flags: available profiles in extended format (see ctree.h)
4036  *
4037  * Returns reduced profile in chunk format.  If profile changing is in
4038  * progress (either running or paused) picks the target profile (if it's
4039  * already available), otherwise falls back to plain reducing.
4040  */
4041 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
4042 {
4043         u64 num_devices = root->fs_info->fs_devices->rw_devices;
4044         u64 target;
4045         u64 raid_type;
4046         u64 allowed = 0;
4047
4048         /*
4049          * see if restripe for this chunk_type is in progress, if so
4050          * try to reduce to the target profile
4051          */
4052         spin_lock(&root->fs_info->balance_lock);
4053         target = get_restripe_target(root->fs_info, flags);
4054         if (target) {
4055                 /* pick target profile only if it's already available */
4056                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4057                         spin_unlock(&root->fs_info->balance_lock);
4058                         return extended_to_chunk(target);
4059                 }
4060         }
4061         spin_unlock(&root->fs_info->balance_lock);
4062
4063         /* First, mask out the RAID levels which aren't possible */
4064         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4065                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4066                         allowed |= btrfs_raid_group[raid_type];
4067         }
4068         allowed &= flags;
4069
4070         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4071                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4072         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4073                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4074         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4075                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4076         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4077                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4078         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4079                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4080
4081         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4082
4083         return extended_to_chunk(flags | allowed);
4084 }
4085
4086 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
4087 {
4088         unsigned seq;
4089         u64 flags;
4090
4091         do {
4092                 flags = orig_flags;
4093                 seq = read_seqbegin(&root->fs_info->profiles_lock);
4094
4095                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4096                         flags |= root->fs_info->avail_data_alloc_bits;
4097                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4098                         flags |= root->fs_info->avail_system_alloc_bits;
4099                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4100                         flags |= root->fs_info->avail_metadata_alloc_bits;
4101         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
4102
4103         return btrfs_reduce_alloc_profile(root, flags);
4104 }
4105
4106 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
4107 {
4108         u64 flags;
4109         u64 ret;
4110
4111         if (data)
4112                 flags = BTRFS_BLOCK_GROUP_DATA;
4113         else if (root == root->fs_info->chunk_root)
4114                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4115         else
4116                 flags = BTRFS_BLOCK_GROUP_METADATA;
4117
4118         ret = get_alloc_profile(root, flags);
4119         return ret;
4120 }
4121
4122 int btrfs_alloc_data_chunk_ondemand(struct inode *inode, u64 bytes)
4123 {
4124         struct btrfs_space_info *data_sinfo;
4125         struct btrfs_root *root = BTRFS_I(inode)->root;
4126         struct btrfs_fs_info *fs_info = root->fs_info;
4127         u64 used;
4128         int ret = 0;
4129         int need_commit = 2;
4130         int have_pinned_space;
4131
4132         /* make sure bytes are sectorsize aligned */
4133         bytes = ALIGN(bytes, root->sectorsize);
4134
4135         if (btrfs_is_free_space_inode(inode)) {
4136                 need_commit = 0;
4137                 ASSERT(current->journal_info);
4138         }
4139
4140         data_sinfo = fs_info->data_sinfo;
4141         if (!data_sinfo)
4142                 goto alloc;
4143
4144 again:
4145         /* make sure we have enough space to handle the data first */
4146         spin_lock(&data_sinfo->lock);
4147         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
4148                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
4149                 data_sinfo->bytes_may_use;
4150
4151         if (used + bytes > data_sinfo->total_bytes) {
4152                 struct btrfs_trans_handle *trans;
4153
4154                 /*
4155                  * if we don't have enough free bytes in this space then we need
4156                  * to alloc a new chunk.
4157                  */
4158                 if (!data_sinfo->full) {
4159                         u64 alloc_target;
4160
4161                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4162                         spin_unlock(&data_sinfo->lock);
4163 alloc:
4164                         alloc_target = btrfs_get_alloc_profile(root, 1);
4165                         /*
4166                          * It is ugly that we don't call nolock join
4167                          * transaction for the free space inode case here.
4168                          * But it is safe because we only do the data space
4169                          * reservation for the free space cache in the
4170                          * transaction context, the common join transaction
4171                          * just increase the counter of the current transaction
4172                          * handler, doesn't try to acquire the trans_lock of
4173                          * the fs.
4174                          */
4175                         trans = btrfs_join_transaction(root);
4176                         if (IS_ERR(trans))
4177                                 return PTR_ERR(trans);
4178
4179                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4180                                              alloc_target,
4181                                              CHUNK_ALLOC_NO_FORCE);
4182                         btrfs_end_transaction(trans, root);
4183                         if (ret < 0) {
4184                                 if (ret != -ENOSPC)
4185                                         return ret;
4186                                 else {
4187                                         have_pinned_space = 1;
4188                                         goto commit_trans;
4189                                 }
4190                         }
4191
4192                         if (!data_sinfo)
4193                                 data_sinfo = fs_info->data_sinfo;
4194
4195                         goto again;
4196                 }
4197
4198                 /*
4199                  * If we don't have enough pinned space to deal with this
4200                  * allocation, and no removed chunk in current transaction,
4201                  * don't bother committing the transaction.
4202                  */
4203                 have_pinned_space = percpu_counter_compare(
4204                         &data_sinfo->total_bytes_pinned,
4205                         used + bytes - data_sinfo->total_bytes);
4206                 spin_unlock(&data_sinfo->lock);
4207
4208                 /* commit the current transaction and try again */
4209 commit_trans:
4210                 if (need_commit &&
4211                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
4212                         need_commit--;
4213
4214                         if (need_commit > 0) {
4215                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4216                                 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
4217                         }
4218
4219                         trans = btrfs_join_transaction(root);
4220                         if (IS_ERR(trans))
4221                                 return PTR_ERR(trans);
4222                         if (have_pinned_space >= 0 ||
4223                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4224                                      &trans->transaction->flags) ||
4225                             need_commit > 0) {
4226                                 ret = btrfs_commit_transaction(trans, root);
4227                                 if (ret)
4228                                         return ret;
4229                                 /*
4230                                  * The cleaner kthread might still be doing iput
4231                                  * operations. Wait for it to finish so that
4232                                  * more space is released.
4233                                  */
4234                                 mutex_lock(&root->fs_info->cleaner_delayed_iput_mutex);
4235                                 mutex_unlock(&root->fs_info->cleaner_delayed_iput_mutex);
4236                                 goto again;
4237                         } else {
4238                                 btrfs_end_transaction(trans, root);
4239                         }
4240                 }
4241
4242                 trace_btrfs_space_reservation(root->fs_info,
4243                                               "space_info:enospc",
4244                                               data_sinfo->flags, bytes, 1);
4245                 return -ENOSPC;
4246         }
4247         data_sinfo->bytes_may_use += bytes;
4248         trace_btrfs_space_reservation(root->fs_info, "space_info",
4249                                       data_sinfo->flags, bytes, 1);
4250         spin_unlock(&data_sinfo->lock);
4251
4252         return ret;
4253 }
4254
4255 /*
4256  * New check_data_free_space() with ability for precious data reservation
4257  * Will replace old btrfs_check_data_free_space(), but for patch split,
4258  * add a new function first and then replace it.
4259  */
4260 int btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4261 {
4262         struct btrfs_root *root = BTRFS_I(inode)->root;
4263         int ret;
4264
4265         /* align the range */
4266         len = round_up(start + len, root->sectorsize) -
4267               round_down(start, root->sectorsize);
4268         start = round_down(start, root->sectorsize);
4269
4270         ret = btrfs_alloc_data_chunk_ondemand(inode, len);
4271         if (ret < 0)
4272                 return ret;
4273
4274         /*
4275          * Use new btrfs_qgroup_reserve_data to reserve precious data space
4276          *
4277          * TODO: Find a good method to avoid reserve data space for NOCOW
4278          * range, but don't impact performance on quota disable case.
4279          */
4280         ret = btrfs_qgroup_reserve_data(inode, start, len);
4281         return ret;
4282 }
4283
4284 /*
4285  * Called if we need to clear a data reservation for this inode
4286  * Normally in a error case.
4287  *
4288  * This one will *NOT* use accurate qgroup reserved space API, just for case
4289  * which we can't sleep and is sure it won't affect qgroup reserved space.
4290  * Like clear_bit_hook().
4291  */
4292 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4293                                             u64 len)
4294 {
4295         struct btrfs_root *root = BTRFS_I(inode)->root;
4296         struct btrfs_space_info *data_sinfo;
4297
4298         /* Make sure the range is aligned to sectorsize */
4299         len = round_up(start + len, root->sectorsize) -
4300               round_down(start, root->sectorsize);
4301         start = round_down(start, root->sectorsize);
4302
4303         data_sinfo = root->fs_info->data_sinfo;
4304         spin_lock(&data_sinfo->lock);
4305         if (WARN_ON(data_sinfo->bytes_may_use < len))
4306                 data_sinfo->bytes_may_use = 0;
4307         else
4308                 data_sinfo->bytes_may_use -= len;
4309         trace_btrfs_space_reservation(root->fs_info, "space_info",
4310                                       data_sinfo->flags, len, 0);
4311         spin_unlock(&data_sinfo->lock);
4312 }
4313
4314 /*
4315  * Called if we need to clear a data reservation for this inode
4316  * Normally in a error case.
4317  *
4318  * This one will handle the per-inode data rsv map for accurate reserved
4319  * space framework.
4320  */
4321 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4322 {
4323         btrfs_free_reserved_data_space_noquota(inode, start, len);
4324         btrfs_qgroup_free_data(inode, start, len);
4325 }
4326
4327 static void force_metadata_allocation(struct btrfs_fs_info *info)
4328 {
4329         struct list_head *head = &info->space_info;
4330         struct btrfs_space_info *found;
4331
4332         rcu_read_lock();
4333         list_for_each_entry_rcu(found, head, list) {
4334                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4335                         found->force_alloc = CHUNK_ALLOC_FORCE;
4336         }
4337         rcu_read_unlock();
4338 }
4339
4340 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4341 {
4342         return (global->size << 1);
4343 }
4344
4345 static int should_alloc_chunk(struct btrfs_root *root,
4346                               struct btrfs_space_info *sinfo, int force)
4347 {
4348         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4349         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4350         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4351         u64 thresh;
4352
4353         if (force == CHUNK_ALLOC_FORCE)
4354                 return 1;
4355
4356         /*
4357          * We need to take into account the global rsv because for all intents
4358          * and purposes it's used space.  Don't worry about locking the
4359          * global_rsv, it doesn't change except when the transaction commits.
4360          */
4361         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4362                 num_allocated += calc_global_rsv_need_space(global_rsv);
4363
4364         /*
4365          * in limited mode, we want to have some free space up to
4366          * about 1% of the FS size.
4367          */
4368         if (force == CHUNK_ALLOC_LIMITED) {
4369                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4370                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4371
4372                 if (num_bytes - num_allocated < thresh)
4373                         return 1;
4374         }
4375
4376         if (num_allocated + SZ_2M < div_factor(num_bytes, 8))
4377                 return 0;
4378         return 1;
4379 }
4380
4381 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4382 {
4383         u64 num_dev;
4384
4385         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4386                     BTRFS_BLOCK_GROUP_RAID0 |
4387                     BTRFS_BLOCK_GROUP_RAID5 |
4388                     BTRFS_BLOCK_GROUP_RAID6))
4389                 num_dev = root->fs_info->fs_devices->rw_devices;
4390         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4391                 num_dev = 2;
4392         else
4393                 num_dev = 1;    /* DUP or single */
4394
4395         return num_dev;
4396 }
4397
4398 /*
4399  * If @is_allocation is true, reserve space in the system space info necessary
4400  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4401  * removing a chunk.
4402  */
4403 void check_system_chunk(struct btrfs_trans_handle *trans,
4404                         struct btrfs_root *root,
4405                         u64 type)
4406 {
4407         struct btrfs_space_info *info;
4408         u64 left;
4409         u64 thresh;
4410         int ret = 0;
4411         u64 num_devs;
4412
4413         /*
4414          * Needed because we can end up allocating a system chunk and for an
4415          * atomic and race free space reservation in the chunk block reserve.
4416          */
4417         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4418
4419         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4420         spin_lock(&info->lock);
4421         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4422                 info->bytes_reserved - info->bytes_readonly -
4423                 info->bytes_may_use;
4424         spin_unlock(&info->lock);
4425
4426         num_devs = get_profile_num_devs(root, type);
4427
4428         /* num_devs device items to update and 1 chunk item to add or remove */
4429         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4430                 btrfs_calc_trans_metadata_size(root, 1);
4431
4432         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4433                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4434                         left, thresh, type);
4435                 dump_space_info(info, 0, 0);
4436         }
4437
4438         if (left < thresh) {
4439                 u64 flags;
4440
4441                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4442                 /*
4443                  * Ignore failure to create system chunk. We might end up not
4444                  * needing it, as we might not need to COW all nodes/leafs from
4445                  * the paths we visit in the chunk tree (they were already COWed
4446                  * or created in the current transaction for example).
4447                  */
4448                 ret = btrfs_alloc_chunk(trans, root, flags);
4449         }
4450
4451         if (!ret) {
4452                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4453                                           &root->fs_info->chunk_block_rsv,
4454                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4455                 if (!ret)
4456                         trans->chunk_bytes_reserved += thresh;
4457         }
4458 }
4459
4460 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4461                           struct btrfs_root *extent_root, u64 flags, int force)
4462 {
4463         struct btrfs_space_info *space_info;
4464         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4465         int wait_for_alloc = 0;
4466         int ret = 0;
4467
4468         /* Don't re-enter if we're already allocating a chunk */
4469         if (trans->allocating_chunk)
4470                 return -ENOSPC;
4471
4472         space_info = __find_space_info(extent_root->fs_info, flags);
4473         if (!space_info) {
4474                 ret = update_space_info(extent_root->fs_info, flags,
4475                                         0, 0, 0, &space_info);
4476                 BUG_ON(ret); /* -ENOMEM */
4477         }
4478         BUG_ON(!space_info); /* Logic error */
4479
4480 again:
4481         spin_lock(&space_info->lock);
4482         if (force < space_info->force_alloc)
4483                 force = space_info->force_alloc;
4484         if (space_info->full) {
4485                 if (should_alloc_chunk(extent_root, space_info, force))
4486                         ret = -ENOSPC;
4487                 else
4488                         ret = 0;
4489                 spin_unlock(&space_info->lock);
4490                 return ret;
4491         }
4492
4493         if (!should_alloc_chunk(extent_root, space_info, force)) {
4494                 spin_unlock(&space_info->lock);
4495                 return 0;
4496         } else if (space_info->chunk_alloc) {
4497                 wait_for_alloc = 1;
4498         } else {
4499                 space_info->chunk_alloc = 1;
4500         }
4501
4502         spin_unlock(&space_info->lock);
4503
4504         mutex_lock(&fs_info->chunk_mutex);
4505
4506         /*
4507          * The chunk_mutex is held throughout the entirety of a chunk
4508          * allocation, so once we've acquired the chunk_mutex we know that the
4509          * other guy is done and we need to recheck and see if we should
4510          * allocate.
4511          */
4512         if (wait_for_alloc) {
4513                 mutex_unlock(&fs_info->chunk_mutex);
4514                 wait_for_alloc = 0;
4515                 goto again;
4516         }
4517
4518         trans->allocating_chunk = true;
4519
4520         /*
4521          * If we have mixed data/metadata chunks we want to make sure we keep
4522          * allocating mixed chunks instead of individual chunks.
4523          */
4524         if (btrfs_mixed_space_info(space_info))
4525                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4526
4527         /*
4528          * if we're doing a data chunk, go ahead and make sure that
4529          * we keep a reasonable number of metadata chunks allocated in the
4530          * FS as well.
4531          */
4532         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4533                 fs_info->data_chunk_allocations++;
4534                 if (!(fs_info->data_chunk_allocations %
4535                       fs_info->metadata_ratio))
4536                         force_metadata_allocation(fs_info);
4537         }
4538
4539         /*
4540          * Check if we have enough space in SYSTEM chunk because we may need
4541          * to update devices.
4542          */
4543         check_system_chunk(trans, extent_root, flags);
4544
4545         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4546         trans->allocating_chunk = false;
4547
4548         spin_lock(&space_info->lock);
4549         if (ret < 0 && ret != -ENOSPC)
4550                 goto out;
4551         if (ret)
4552                 space_info->full = 1;
4553         else
4554                 ret = 1;
4555
4556         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4557 out:
4558         space_info->chunk_alloc = 0;
4559         spin_unlock(&space_info->lock);
4560         mutex_unlock(&fs_info->chunk_mutex);
4561         /*
4562          * When we allocate a new chunk we reserve space in the chunk block
4563          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4564          * add new nodes/leafs to it if we end up needing to do it when
4565          * inserting the chunk item and updating device items as part of the
4566          * second phase of chunk allocation, performed by
4567          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4568          * large number of new block groups to create in our transaction
4569          * handle's new_bgs list to avoid exhausting the chunk block reserve
4570          * in extreme cases - like having a single transaction create many new
4571          * block groups when starting to write out the free space caches of all
4572          * the block groups that were made dirty during the lifetime of the
4573          * transaction.
4574          */
4575         if (trans->can_flush_pending_bgs &&
4576             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4577                 btrfs_create_pending_block_groups(trans, trans->root);
4578                 btrfs_trans_release_chunk_metadata(trans);
4579         }
4580         return ret;
4581 }
4582
4583 static int can_overcommit(struct btrfs_root *root,
4584                           struct btrfs_space_info *space_info, u64 bytes,
4585                           enum btrfs_reserve_flush_enum flush)
4586 {
4587         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4588         u64 profile = btrfs_get_alloc_profile(root, 0);
4589         u64 space_size;
4590         u64 avail;
4591         u64 used;
4592
4593         used = space_info->bytes_used + space_info->bytes_reserved +
4594                 space_info->bytes_pinned + space_info->bytes_readonly;
4595
4596         /*
4597          * We only want to allow over committing if we have lots of actual space
4598          * free, but if we don't have enough space to handle the global reserve
4599          * space then we could end up having a real enospc problem when trying
4600          * to allocate a chunk or some other such important allocation.
4601          */
4602         spin_lock(&global_rsv->lock);
4603         space_size = calc_global_rsv_need_space(global_rsv);
4604         spin_unlock(&global_rsv->lock);
4605         if (used + space_size >= space_info->total_bytes)
4606                 return 0;
4607
4608         used += space_info->bytes_may_use;
4609
4610         spin_lock(&root->fs_info->free_chunk_lock);
4611         avail = root->fs_info->free_chunk_space;
4612         spin_unlock(&root->fs_info->free_chunk_lock);
4613
4614         /*
4615          * If we have dup, raid1 or raid10 then only half of the free
4616          * space is actually useable.  For raid56, the space info used
4617          * doesn't include the parity drive, so we don't have to
4618          * change the math
4619          */
4620         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4621                        BTRFS_BLOCK_GROUP_RAID1 |
4622                        BTRFS_BLOCK_GROUP_RAID10))
4623                 avail >>= 1;
4624
4625         /*
4626          * If we aren't flushing all things, let us overcommit up to
4627          * 1/2th of the space. If we can flush, don't let us overcommit
4628          * too much, let it overcommit up to 1/8 of the space.
4629          */
4630         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4631                 avail >>= 3;
4632         else
4633                 avail >>= 1;
4634
4635         if (used + bytes < space_info->total_bytes + avail)
4636                 return 1;
4637         return 0;
4638 }
4639
4640 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4641                                          unsigned long nr_pages, int nr_items)
4642 {
4643         struct super_block *sb = root->fs_info->sb;
4644
4645         if (down_read_trylock(&sb->s_umount)) {
4646                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4647                 up_read(&sb->s_umount);
4648         } else {
4649                 /*
4650                  * We needn't worry the filesystem going from r/w to r/o though
4651                  * we don't acquire ->s_umount mutex, because the filesystem
4652                  * should guarantee the delalloc inodes list be empty after
4653                  * the filesystem is readonly(all dirty pages are written to
4654                  * the disk).
4655                  */
4656                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4657                 if (!current->journal_info)
4658                         btrfs_wait_ordered_roots(root->fs_info, nr_items,
4659                                                  0, (u64)-1);
4660         }
4661 }
4662
4663 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4664 {
4665         u64 bytes;
4666         int nr;
4667
4668         bytes = btrfs_calc_trans_metadata_size(root, 1);
4669         nr = (int)div64_u64(to_reclaim, bytes);
4670         if (!nr)
4671                 nr = 1;
4672         return nr;
4673 }
4674
4675 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4676
4677 /*
4678  * shrink metadata reservation for delalloc
4679  */
4680 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4681                             bool wait_ordered)
4682 {
4683         struct btrfs_block_rsv *block_rsv;
4684         struct btrfs_space_info *space_info;
4685         struct btrfs_trans_handle *trans;
4686         u64 delalloc_bytes;
4687         u64 max_reclaim;
4688         long time_left;
4689         unsigned long nr_pages;
4690         int loops;
4691         int items;
4692         enum btrfs_reserve_flush_enum flush;
4693
4694         /* Calc the number of the pages we need flush for space reservation */
4695         items = calc_reclaim_items_nr(root, to_reclaim);
4696         to_reclaim = (u64)items * EXTENT_SIZE_PER_ITEM;
4697
4698         trans = (struct btrfs_trans_handle *)current->journal_info;
4699         block_rsv = &root->fs_info->delalloc_block_rsv;
4700         space_info = block_rsv->space_info;
4701
4702         delalloc_bytes = percpu_counter_sum_positive(
4703                                                 &root->fs_info->delalloc_bytes);
4704         if (delalloc_bytes == 0) {
4705                 if (trans)
4706                         return;
4707                 if (wait_ordered)
4708                         btrfs_wait_ordered_roots(root->fs_info, items,
4709                                                  0, (u64)-1);
4710                 return;
4711         }
4712
4713         loops = 0;
4714         while (delalloc_bytes && loops < 3) {
4715                 max_reclaim = min(delalloc_bytes, to_reclaim);
4716                 nr_pages = max_reclaim >> PAGE_SHIFT;
4717                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4718                 /*
4719                  * We need to wait for the async pages to actually start before
4720                  * we do anything.
4721                  */
4722                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4723                 if (!max_reclaim)
4724                         goto skip_async;
4725
4726                 if (max_reclaim <= nr_pages)
4727                         max_reclaim = 0;
4728                 else
4729                         max_reclaim -= nr_pages;
4730
4731                 wait_event(root->fs_info->async_submit_wait,
4732                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4733                            (int)max_reclaim);
4734 skip_async:
4735                 if (!trans)
4736                         flush = BTRFS_RESERVE_FLUSH_ALL;
4737                 else
4738                         flush = BTRFS_RESERVE_NO_FLUSH;
4739                 spin_lock(&space_info->lock);
4740                 if (can_overcommit(root, space_info, orig, flush)) {
4741                         spin_unlock(&space_info->lock);
4742                         break;
4743                 }
4744                 spin_unlock(&space_info->lock);
4745
4746                 loops++;
4747                 if (wait_ordered && !trans) {
4748                         btrfs_wait_ordered_roots(root->fs_info, items,
4749                                                  0, (u64)-1);
4750                 } else {
4751                         time_left = schedule_timeout_killable(1);
4752                         if (time_left)
4753                                 break;
4754                 }
4755                 delalloc_bytes = percpu_counter_sum_positive(
4756                                                 &root->fs_info->delalloc_bytes);
4757         }
4758 }
4759
4760 /**
4761  * maybe_commit_transaction - possibly commit the transaction if its ok to
4762  * @root - the root we're allocating for
4763  * @bytes - the number of bytes we want to reserve
4764  * @force - force the commit
4765  *
4766  * This will check to make sure that committing the transaction will actually
4767  * get us somewhere and then commit the transaction if it does.  Otherwise it
4768  * will return -ENOSPC.
4769  */
4770 static int may_commit_transaction(struct btrfs_root *root,
4771                                   struct btrfs_space_info *space_info,
4772                                   u64 bytes, int force)
4773 {
4774         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4775         struct btrfs_trans_handle *trans;
4776
4777         trans = (struct btrfs_trans_handle *)current->journal_info;
4778         if (trans)
4779                 return -EAGAIN;
4780
4781         if (force)
4782                 goto commit;
4783
4784         /* See if there is enough pinned space to make this reservation */
4785         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4786                                    bytes) >= 0)
4787                 goto commit;
4788
4789         /*
4790          * See if there is some space in the delayed insertion reservation for
4791          * this reservation.
4792          */
4793         if (space_info != delayed_rsv->space_info)
4794                 return -ENOSPC;
4795
4796         spin_lock(&delayed_rsv->lock);
4797         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4798                                    bytes - delayed_rsv->size) >= 0) {
4799                 spin_unlock(&delayed_rsv->lock);
4800                 return -ENOSPC;
4801         }
4802         spin_unlock(&delayed_rsv->lock);
4803
4804 commit:
4805         trans = btrfs_join_transaction(root);
4806         if (IS_ERR(trans))
4807                 return -ENOSPC;
4808
4809         return btrfs_commit_transaction(trans, root);
4810 }
4811
4812 enum flush_state {
4813         FLUSH_DELAYED_ITEMS_NR  =       1,
4814         FLUSH_DELAYED_ITEMS     =       2,
4815         FLUSH_DELALLOC          =       3,
4816         FLUSH_DELALLOC_WAIT     =       4,
4817         ALLOC_CHUNK             =       5,
4818         COMMIT_TRANS            =       6,
4819 };
4820
4821 static int flush_space(struct btrfs_root *root,
4822                        struct btrfs_space_info *space_info, u64 num_bytes,
4823                        u64 orig_bytes, int state)
4824 {
4825         struct btrfs_trans_handle *trans;
4826         int nr;
4827         int ret = 0;
4828
4829         switch (state) {
4830         case FLUSH_DELAYED_ITEMS_NR:
4831         case FLUSH_DELAYED_ITEMS:
4832                 if (state == FLUSH_DELAYED_ITEMS_NR)
4833                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4834                 else
4835                         nr = -1;
4836
4837                 trans = btrfs_join_transaction(root);
4838                 if (IS_ERR(trans)) {
4839                         ret = PTR_ERR(trans);
4840                         break;
4841                 }
4842                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4843                 btrfs_end_transaction(trans, root);
4844                 break;
4845         case FLUSH_DELALLOC:
4846         case FLUSH_DELALLOC_WAIT:
4847                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4848                                 state == FLUSH_DELALLOC_WAIT);
4849                 break;
4850         case ALLOC_CHUNK:
4851                 trans = btrfs_join_transaction(root);
4852                 if (IS_ERR(trans)) {
4853                         ret = PTR_ERR(trans);
4854                         break;
4855                 }
4856                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4857                                      btrfs_get_alloc_profile(root, 0),
4858                                      CHUNK_ALLOC_NO_FORCE);
4859                 btrfs_end_transaction(trans, root);
4860                 if (ret == -ENOSPC)
4861                         ret = 0;
4862                 break;
4863         case COMMIT_TRANS:
4864                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4865                 break;
4866         default:
4867                 ret = -ENOSPC;
4868                 break;
4869         }
4870
4871         return ret;
4872 }
4873
4874 static inline u64
4875 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4876                                  struct btrfs_space_info *space_info)
4877 {
4878         u64 used;
4879         u64 expected;
4880         u64 to_reclaim;
4881
4882         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4883         spin_lock(&space_info->lock);
4884         if (can_overcommit(root, space_info, to_reclaim,
4885                            BTRFS_RESERVE_FLUSH_ALL)) {
4886                 to_reclaim = 0;
4887                 goto out;
4888         }
4889
4890         used = space_info->bytes_used + space_info->bytes_reserved +
4891                space_info->bytes_pinned + space_info->bytes_readonly +
4892                space_info->bytes_may_use;
4893         if (can_overcommit(root, space_info, SZ_1M, BTRFS_RESERVE_FLUSH_ALL))
4894                 expected = div_factor_fine(space_info->total_bytes, 95);
4895         else
4896                 expected = div_factor_fine(space_info->total_bytes, 90);
4897
4898         if (used > expected)
4899                 to_reclaim = used - expected;
4900         else
4901                 to_reclaim = 0;
4902         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4903                                      space_info->bytes_reserved);
4904 out:
4905         spin_unlock(&space_info->lock);
4906
4907         return to_reclaim;
4908 }
4909
4910 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4911                                         struct btrfs_fs_info *fs_info, u64 used)
4912 {
4913         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4914
4915         /* If we're just plain full then async reclaim just slows us down. */
4916         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4917                 return 0;
4918
4919         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4920                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4921 }
4922
4923 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4924                                        struct btrfs_fs_info *fs_info,
4925                                        int flush_state)
4926 {
4927         u64 used;
4928
4929         spin_lock(&space_info->lock);
4930         /*
4931          * We run out of space and have not got any free space via flush_space,
4932          * so don't bother doing async reclaim.
4933          */
4934         if (flush_state > COMMIT_TRANS && space_info->full) {
4935                 spin_unlock(&space_info->lock);
4936                 return 0;
4937         }
4938
4939         used = space_info->bytes_used + space_info->bytes_reserved +
4940                space_info->bytes_pinned + space_info->bytes_readonly +
4941                space_info->bytes_may_use;
4942         if (need_do_async_reclaim(space_info, fs_info, used)) {
4943                 spin_unlock(&space_info->lock);
4944                 return 1;
4945         }
4946         spin_unlock(&space_info->lock);
4947
4948         return 0;
4949 }
4950
4951 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4952 {
4953         struct btrfs_fs_info *fs_info;
4954         struct btrfs_space_info *space_info;
4955         u64 to_reclaim;
4956         int flush_state;
4957
4958         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4959         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4960
4961         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4962                                                       space_info);
4963         if (!to_reclaim)
4964                 return;
4965
4966         flush_state = FLUSH_DELAYED_ITEMS_NR;
4967         do {
4968                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4969                             to_reclaim, flush_state);
4970                 flush_state++;
4971                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4972                                                  flush_state))
4973                         return;
4974         } while (flush_state < COMMIT_TRANS);
4975 }
4976
4977 void btrfs_init_async_reclaim_work(struct work_struct *work)
4978 {
4979         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4980 }
4981
4982 /**
4983  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4984  * @root - the root we're allocating for
4985  * @block_rsv - the block_rsv we're allocating for
4986  * @orig_bytes - the number of bytes we want
4987  * @flush - whether or not we can flush to make our reservation
4988  *
4989  * This will reserve orig_bytes number of bytes from the space info associated
4990  * with the block_rsv.  If there is not enough space it will make an attempt to
4991  * flush out space to make room.  It will do this by flushing delalloc if
4992  * possible or committing the transaction.  If flush is 0 then no attempts to
4993  * regain reservations will be made and this will fail if there is not enough
4994  * space already.
4995  */
4996 static int reserve_metadata_bytes(struct btrfs_root *root,
4997                                   struct btrfs_block_rsv *block_rsv,
4998                                   u64 orig_bytes,
4999                                   enum btrfs_reserve_flush_enum flush)
5000 {
5001         struct btrfs_space_info *space_info = block_rsv->space_info;
5002         u64 used;
5003         u64 num_bytes = orig_bytes;
5004         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5005         int ret = 0;
5006         bool flushing = false;
5007
5008 again:
5009         ret = 0;
5010         spin_lock(&space_info->lock);
5011         /*
5012          * We only want to wait if somebody other than us is flushing and we
5013          * are actually allowed to flush all things.
5014          */
5015         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
5016                space_info->flush) {
5017                 spin_unlock(&space_info->lock);
5018                 /*
5019                  * If we have a trans handle we can't wait because the flusher
5020                  * may have to commit the transaction, which would mean we would
5021                  * deadlock since we are waiting for the flusher to finish, but
5022                  * hold the current transaction open.
5023                  */
5024                 if (current->journal_info)
5025                         return -EAGAIN;
5026                 ret = wait_event_killable(space_info->wait, !space_info->flush);
5027                 /* Must have been killed, return */
5028                 if (ret)
5029                         return -EINTR;
5030
5031                 spin_lock(&space_info->lock);
5032         }
5033
5034         ret = -ENOSPC;
5035         used = space_info->bytes_used + space_info->bytes_reserved +
5036                 space_info->bytes_pinned + space_info->bytes_readonly +
5037                 space_info->bytes_may_use;
5038
5039         /*
5040          * The idea here is that we've not already over-reserved the block group
5041          * then we can go ahead and save our reservation first and then start
5042          * flushing if we need to.  Otherwise if we've already overcommitted
5043          * lets start flushing stuff first and then come back and try to make
5044          * our reservation.
5045          */
5046         if (used <= space_info->total_bytes) {
5047                 if (used + orig_bytes <= space_info->total_bytes) {
5048                         space_info->bytes_may_use += orig_bytes;
5049                         trace_btrfs_space_reservation(root->fs_info,
5050                                 "space_info", space_info->flags, orig_bytes, 1);
5051                         ret = 0;
5052                 } else {
5053                         /*
5054                          * Ok set num_bytes to orig_bytes since we aren't
5055                          * overocmmitted, this way we only try and reclaim what
5056                          * we need.
5057                          */
5058                         num_bytes = orig_bytes;
5059                 }
5060         } else {
5061                 /*
5062                  * Ok we're over committed, set num_bytes to the overcommitted
5063                  * amount plus the amount of bytes that we need for this
5064                  * reservation.
5065                  */
5066                 num_bytes = used - space_info->total_bytes +
5067                         (orig_bytes * 2);
5068         }
5069
5070         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
5071                 space_info->bytes_may_use += orig_bytes;
5072                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5073                                               space_info->flags, orig_bytes,
5074                                               1);
5075                 ret = 0;
5076         }
5077
5078         /*
5079          * Couldn't make our reservation, save our place so while we're trying
5080          * to reclaim space we can actually use it instead of somebody else
5081          * stealing it from us.
5082          *
5083          * We make the other tasks wait for the flush only when we can flush
5084          * all things.
5085          */
5086         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5087                 flushing = true;
5088                 space_info->flush = 1;
5089         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5090                 used += orig_bytes;
5091                 /*
5092                  * We will do the space reservation dance during log replay,
5093                  * which means we won't have fs_info->fs_root set, so don't do
5094                  * the async reclaim as we will panic.
5095                  */
5096                 if (!root->fs_info->log_root_recovering &&
5097                     need_do_async_reclaim(space_info, root->fs_info, used) &&
5098                     !work_busy(&root->fs_info->async_reclaim_work))
5099                         queue_work(system_unbound_wq,
5100                                    &root->fs_info->async_reclaim_work);
5101         }
5102         spin_unlock(&space_info->lock);
5103
5104         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5105                 goto out;
5106
5107         ret = flush_space(root, space_info, num_bytes, orig_bytes,
5108                           flush_state);
5109         flush_state++;
5110
5111         /*
5112          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
5113          * would happen. So skip delalloc flush.
5114          */
5115         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
5116             (flush_state == FLUSH_DELALLOC ||
5117              flush_state == FLUSH_DELALLOC_WAIT))
5118                 flush_state = ALLOC_CHUNK;
5119
5120         if (!ret)
5121                 goto again;
5122         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
5123                  flush_state < COMMIT_TRANS)
5124                 goto again;
5125         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
5126                  flush_state <= COMMIT_TRANS)
5127                 goto again;
5128
5129 out:
5130         if (ret == -ENOSPC &&
5131             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5132                 struct btrfs_block_rsv *global_rsv =
5133                         &root->fs_info->global_block_rsv;
5134
5135                 if (block_rsv != global_rsv &&
5136                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5137                         ret = 0;
5138         }
5139         if (ret == -ENOSPC)
5140                 trace_btrfs_space_reservation(root->fs_info,
5141                                               "space_info:enospc",
5142                                               space_info->flags, orig_bytes, 1);
5143         if (flushing) {
5144                 spin_lock(&space_info->lock);
5145                 space_info->flush = 0;
5146                 wake_up_all(&space_info->wait);
5147                 spin_unlock(&space_info->lock);
5148         }
5149         return ret;
5150 }
5151
5152 static struct btrfs_block_rsv *get_block_rsv(
5153                                         const struct btrfs_trans_handle *trans,
5154                                         const struct btrfs_root *root)
5155 {
5156         struct btrfs_block_rsv *block_rsv = NULL;
5157
5158         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5159             (root == root->fs_info->csum_root && trans->adding_csums) ||
5160              (root == root->fs_info->uuid_root))
5161                 block_rsv = trans->block_rsv;
5162
5163         if (!block_rsv)
5164                 block_rsv = root->block_rsv;
5165
5166         if (!block_rsv)
5167                 block_rsv = &root->fs_info->empty_block_rsv;
5168
5169         return block_rsv;
5170 }
5171
5172 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5173                                u64 num_bytes)
5174 {
5175         int ret = -ENOSPC;
5176         spin_lock(&block_rsv->lock);
5177         if (block_rsv->reserved >= num_bytes) {
5178                 block_rsv->reserved -= num_bytes;
5179                 if (block_rsv->reserved < block_rsv->size)
5180                         block_rsv->full = 0;
5181                 ret = 0;
5182         }
5183         spin_unlock(&block_rsv->lock);
5184         return ret;
5185 }
5186
5187 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5188                                 u64 num_bytes, int update_size)
5189 {
5190         spin_lock(&block_rsv->lock);
5191         block_rsv->reserved += num_bytes;
5192         if (update_size)
5193                 block_rsv->size += num_bytes;
5194         else if (block_rsv->reserved >= block_rsv->size)
5195                 block_rsv->full = 1;
5196         spin_unlock(&block_rsv->lock);
5197 }
5198
5199 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5200                              struct btrfs_block_rsv *dest, u64 num_bytes,
5201                              int min_factor)
5202 {
5203         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5204         u64 min_bytes;
5205
5206         if (global_rsv->space_info != dest->space_info)
5207                 return -ENOSPC;
5208
5209         spin_lock(&global_rsv->lock);
5210         min_bytes = div_factor(global_rsv->size, min_factor);
5211         if (global_rsv->reserved < min_bytes + num_bytes) {
5212                 spin_unlock(&global_rsv->lock);
5213                 return -ENOSPC;
5214         }
5215         global_rsv->reserved -= num_bytes;
5216         if (global_rsv->reserved < global_rsv->size)
5217                 global_rsv->full = 0;
5218         spin_unlock(&global_rsv->lock);
5219
5220         block_rsv_add_bytes(dest, num_bytes, 1);
5221         return 0;
5222 }
5223
5224 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5225                                     struct btrfs_block_rsv *block_rsv,
5226                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5227 {
5228         struct btrfs_space_info *space_info = block_rsv->space_info;
5229
5230         spin_lock(&block_rsv->lock);
5231         if (num_bytes == (u64)-1)
5232                 num_bytes = block_rsv->size;
5233         block_rsv->size -= num_bytes;
5234         if (block_rsv->reserved >= block_rsv->size) {
5235                 num_bytes = block_rsv->reserved - block_rsv->size;
5236                 block_rsv->reserved = block_rsv->size;
5237                 block_rsv->full = 1;
5238         } else {
5239                 num_bytes = 0;
5240         }
5241         spin_unlock(&block_rsv->lock);
5242
5243         if (num_bytes > 0) {
5244                 if (dest) {
5245                         spin_lock(&dest->lock);
5246                         if (!dest->full) {
5247                                 u64 bytes_to_add;
5248
5249                                 bytes_to_add = dest->size - dest->reserved;
5250                                 bytes_to_add = min(num_bytes, bytes_to_add);
5251                                 dest->reserved += bytes_to_add;
5252                                 if (dest->reserved >= dest->size)
5253                                         dest->full = 1;
5254                                 num_bytes -= bytes_to_add;
5255                         }
5256                         spin_unlock(&dest->lock);
5257                 }
5258                 if (num_bytes) {
5259                         spin_lock(&space_info->lock);
5260                         space_info->bytes_may_use -= num_bytes;
5261                         trace_btrfs_space_reservation(fs_info, "space_info",
5262                                         space_info->flags, num_bytes, 0);
5263                         spin_unlock(&space_info->lock);
5264                 }
5265         }
5266 }
5267
5268 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5269                             struct btrfs_block_rsv *dst, u64 num_bytes,
5270                             int update_size)
5271 {
5272         int ret;
5273
5274         ret = block_rsv_use_bytes(src, num_bytes);
5275         if (ret)
5276                 return ret;
5277
5278         block_rsv_add_bytes(dst, num_bytes, update_size);
5279         return 0;
5280 }
5281
5282 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5283 {
5284         memset(rsv, 0, sizeof(*rsv));
5285         spin_lock_init(&rsv->lock);
5286         rsv->type = type;
5287 }
5288
5289 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5290                                               unsigned short type)
5291 {
5292         struct btrfs_block_rsv *block_rsv;
5293         struct btrfs_fs_info *fs_info = root->fs_info;
5294
5295         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5296         if (!block_rsv)
5297                 return NULL;
5298
5299         btrfs_init_block_rsv(block_rsv, type);
5300         block_rsv->space_info = __find_space_info(fs_info,
5301                                                   BTRFS_BLOCK_GROUP_METADATA);
5302         return block_rsv;
5303 }
5304
5305 void btrfs_free_block_rsv(struct btrfs_root *root,
5306                           struct btrfs_block_rsv *rsv)
5307 {
5308         if (!rsv)
5309                 return;
5310         btrfs_block_rsv_release(root, rsv, (u64)-1);
5311         kfree(rsv);
5312 }
5313
5314 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5315 {
5316         kfree(rsv);
5317 }
5318
5319 int btrfs_block_rsv_add(struct btrfs_root *root,
5320                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5321                         enum btrfs_reserve_flush_enum flush)
5322 {
5323         int ret;
5324
5325         if (num_bytes == 0)
5326                 return 0;
5327
5328         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5329         if (!ret) {
5330                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5331                 return 0;
5332         }
5333
5334         return ret;
5335 }
5336
5337 int btrfs_block_rsv_check(struct btrfs_root *root,
5338                           struct btrfs_block_rsv *block_rsv, int min_factor)
5339 {
5340         u64 num_bytes = 0;
5341         int ret = -ENOSPC;
5342
5343         if (!block_rsv)
5344                 return 0;
5345
5346         spin_lock(&block_rsv->lock);
5347         num_bytes = div_factor(block_rsv->size, min_factor);
5348         if (block_rsv->reserved >= num_bytes)
5349                 ret = 0;
5350         spin_unlock(&block_rsv->lock);
5351
5352         return ret;
5353 }
5354
5355 int btrfs_block_rsv_refill(struct btrfs_root *root,
5356                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5357                            enum btrfs_reserve_flush_enum flush)
5358 {
5359         u64 num_bytes = 0;
5360         int ret = -ENOSPC;
5361
5362         if (!block_rsv)
5363                 return 0;
5364
5365         spin_lock(&block_rsv->lock);
5366         num_bytes = min_reserved;
5367         if (block_rsv->reserved >= num_bytes)
5368                 ret = 0;
5369         else
5370                 num_bytes -= block_rsv->reserved;
5371         spin_unlock(&block_rsv->lock);
5372
5373         if (!ret)
5374                 return 0;
5375
5376         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5377         if (!ret) {
5378                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5379                 return 0;
5380         }
5381
5382         return ret;
5383 }
5384
5385 void btrfs_block_rsv_release(struct btrfs_root *root,
5386                              struct btrfs_block_rsv *block_rsv,
5387                              u64 num_bytes)
5388 {
5389         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5390         if (global_rsv == block_rsv ||
5391             block_rsv->space_info != global_rsv->space_info)
5392                 global_rsv = NULL;
5393         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5394                                 num_bytes);
5395 }
5396
5397 /*
5398  * helper to calculate size of global block reservation.
5399  * the desired value is sum of space used by extent tree,
5400  * checksum tree and root tree
5401  */
5402 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5403 {
5404         struct btrfs_space_info *sinfo;
5405         u64 num_bytes;
5406         u64 meta_used;
5407         u64 data_used;
5408         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5409
5410         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5411         spin_lock(&sinfo->lock);
5412         data_used = sinfo->bytes_used;
5413         spin_unlock(&sinfo->lock);
5414
5415         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5416         spin_lock(&sinfo->lock);
5417         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5418                 data_used = 0;
5419         meta_used = sinfo->bytes_used;
5420         spin_unlock(&sinfo->lock);
5421
5422         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5423                     csum_size * 2;
5424         num_bytes += div_u64(data_used + meta_used, 50);
5425
5426         if (num_bytes * 3 > meta_used)
5427                 num_bytes = div_u64(meta_used, 3);
5428
5429         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5430 }
5431
5432 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5433 {
5434         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5435         struct btrfs_space_info *sinfo = block_rsv->space_info;
5436         u64 num_bytes;
5437
5438         num_bytes = calc_global_metadata_size(fs_info);
5439
5440         spin_lock(&sinfo->lock);
5441         spin_lock(&block_rsv->lock);
5442
5443         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5444
5445         if (block_rsv->reserved < block_rsv->size) {
5446                 num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5447                         sinfo->bytes_reserved + sinfo->bytes_readonly +
5448                         sinfo->bytes_may_use;
5449                 if (sinfo->total_bytes > num_bytes) {
5450                         num_bytes = sinfo->total_bytes - num_bytes;
5451                         num_bytes = min(num_bytes,
5452                                         block_rsv->size - block_rsv->reserved);
5453                         block_rsv->reserved += num_bytes;
5454                         sinfo->bytes_may_use += num_bytes;
5455                         trace_btrfs_space_reservation(fs_info, "space_info",
5456                                                       sinfo->flags, num_bytes,
5457                                                       1);
5458                 }
5459         } else if (block_rsv->reserved > block_rsv->size) {
5460                 num_bytes = block_rsv->reserved - block_rsv->size;
5461                 sinfo->bytes_may_use -= num_bytes;
5462                 trace_btrfs_space_reservation(fs_info, "space_info",
5463                                       sinfo->flags, num_bytes, 0);
5464                 block_rsv->reserved = block_rsv->size;
5465         }
5466
5467         if (block_rsv->reserved == block_rsv->size)
5468                 block_rsv->full = 1;
5469         else
5470                 block_rsv->full = 0;
5471
5472         spin_unlock(&block_rsv->lock);
5473         spin_unlock(&sinfo->lock);
5474 }
5475
5476 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5477 {
5478         struct btrfs_space_info *space_info;
5479
5480         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5481         fs_info->chunk_block_rsv.space_info = space_info;
5482
5483         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5484         fs_info->global_block_rsv.space_info = space_info;
5485         fs_info->delalloc_block_rsv.space_info = space_info;
5486         fs_info->trans_block_rsv.space_info = space_info;
5487         fs_info->empty_block_rsv.space_info = space_info;
5488         fs_info->delayed_block_rsv.space_info = space_info;
5489
5490         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5491         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5492         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5493         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5494         if (fs_info->quota_root)
5495                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5496         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5497
5498         update_global_block_rsv(fs_info);
5499 }
5500
5501 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5502 {
5503         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5504                                 (u64)-1);
5505         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5506         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5507         WARN_ON(fs_info->trans_block_rsv.size > 0);
5508         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5509         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5510         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5511         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5512         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5513 }
5514
5515 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5516                                   struct btrfs_root *root)
5517 {
5518         if (!trans->block_rsv)
5519                 return;
5520
5521         if (!trans->bytes_reserved)
5522                 return;
5523
5524         trace_btrfs_space_reservation(root->fs_info, "transaction",
5525                                       trans->transid, trans->bytes_reserved, 0);
5526         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5527         trans->bytes_reserved = 0;
5528 }
5529
5530 /*
5531  * To be called after all the new block groups attached to the transaction
5532  * handle have been created (btrfs_create_pending_block_groups()).
5533  */
5534 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5535 {
5536         struct btrfs_fs_info *fs_info = trans->root->fs_info;
5537
5538         if (!trans->chunk_bytes_reserved)
5539                 return;
5540
5541         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5542
5543         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5544                                 trans->chunk_bytes_reserved);
5545         trans->chunk_bytes_reserved = 0;
5546 }
5547
5548 /* Can only return 0 or -ENOSPC */
5549 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5550                                   struct inode *inode)
5551 {
5552         struct btrfs_root *root = BTRFS_I(inode)->root;
5553         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5554         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5555
5556         /*
5557          * We need to hold space in order to delete our orphan item once we've
5558          * added it, so this takes the reservation so we can release it later
5559          * when we are truly done with the orphan item.
5560          */
5561         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5562         trace_btrfs_space_reservation(root->fs_info, "orphan",
5563                                       btrfs_ino(inode), num_bytes, 1);
5564         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5565 }
5566
5567 void btrfs_orphan_release_metadata(struct inode *inode)
5568 {
5569         struct btrfs_root *root = BTRFS_I(inode)->root;
5570         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5571         trace_btrfs_space_reservation(root->fs_info, "orphan",
5572                                       btrfs_ino(inode), num_bytes, 0);
5573         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5574 }
5575
5576 /*
5577  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5578  * root: the root of the parent directory
5579  * rsv: block reservation
5580  * items: the number of items that we need do reservation
5581  * qgroup_reserved: used to return the reserved size in qgroup
5582  *
5583  * This function is used to reserve the space for snapshot/subvolume
5584  * creation and deletion. Those operations are different with the
5585  * common file/directory operations, they change two fs/file trees
5586  * and root tree, the number of items that the qgroup reserves is
5587  * different with the free space reservation. So we can not use
5588  * the space reservation mechanism in start_transaction().
5589  */
5590 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5591                                      struct btrfs_block_rsv *rsv,
5592                                      int items,
5593                                      u64 *qgroup_reserved,
5594                                      bool use_global_rsv)
5595 {
5596         u64 num_bytes;
5597         int ret;
5598         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5599
5600         if (root->fs_info->quota_enabled) {
5601                 /* One for parent inode, two for dir entries */
5602                 num_bytes = 3 * root->nodesize;
5603                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5604                 if (ret)
5605                         return ret;
5606         } else {
5607                 num_bytes = 0;
5608         }
5609
5610         *qgroup_reserved = num_bytes;
5611
5612         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5613         rsv->space_info = __find_space_info(root->fs_info,
5614                                             BTRFS_BLOCK_GROUP_METADATA);
5615         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5616                                   BTRFS_RESERVE_FLUSH_ALL);
5617
5618         if (ret == -ENOSPC && use_global_rsv)
5619                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5620
5621         if (ret && *qgroup_reserved)
5622                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5623
5624         return ret;
5625 }
5626
5627 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5628                                       struct btrfs_block_rsv *rsv,
5629                                       u64 qgroup_reserved)
5630 {
5631         btrfs_block_rsv_release(root, rsv, (u64)-1);
5632 }
5633
5634 /**
5635  * drop_outstanding_extent - drop an outstanding extent
5636  * @inode: the inode we're dropping the extent for
5637  * @num_bytes: the number of bytes we're releasing.
5638  *
5639  * This is called when we are freeing up an outstanding extent, either called
5640  * after an error or after an extent is written.  This will return the number of
5641  * reserved extents that need to be freed.  This must be called with
5642  * BTRFS_I(inode)->lock held.
5643  */
5644 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5645 {
5646         unsigned drop_inode_space = 0;
5647         unsigned dropped_extents = 0;
5648         unsigned num_extents = 0;
5649
5650         num_extents = (unsigned)div64_u64(num_bytes +
5651                                           BTRFS_MAX_EXTENT_SIZE - 1,
5652                                           BTRFS_MAX_EXTENT_SIZE);
5653         ASSERT(num_extents);
5654         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5655         BTRFS_I(inode)->outstanding_extents -= num_extents;
5656
5657         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5658             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5659                                &BTRFS_I(inode)->runtime_flags))
5660                 drop_inode_space = 1;
5661
5662         /*
5663          * If we have more or the same amount of outstanding extents than we have
5664          * reserved then we need to leave the reserved extents count alone.
5665          */
5666         if (BTRFS_I(inode)->outstanding_extents >=
5667             BTRFS_I(inode)->reserved_extents)
5668                 return drop_inode_space;
5669
5670         dropped_extents = BTRFS_I(inode)->reserved_extents -
5671                 BTRFS_I(inode)->outstanding_extents;
5672         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5673         return dropped_extents + drop_inode_space;
5674 }
5675
5676 /**
5677  * calc_csum_metadata_size - return the amount of metadata space that must be
5678  *      reserved/freed for the given bytes.
5679  * @inode: the inode we're manipulating
5680  * @num_bytes: the number of bytes in question
5681  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5682  *
5683  * This adjusts the number of csum_bytes in the inode and then returns the
5684  * correct amount of metadata that must either be reserved or freed.  We
5685  * calculate how many checksums we can fit into one leaf and then divide the
5686  * number of bytes that will need to be checksumed by this value to figure out
5687  * how many checksums will be required.  If we are adding bytes then the number
5688  * may go up and we will return the number of additional bytes that must be
5689  * reserved.  If it is going down we will return the number of bytes that must
5690  * be freed.
5691  *
5692  * This must be called with BTRFS_I(inode)->lock held.
5693  */
5694 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5695                                    int reserve)
5696 {
5697         struct btrfs_root *root = BTRFS_I(inode)->root;
5698         u64 old_csums, num_csums;
5699
5700         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5701             BTRFS_I(inode)->csum_bytes == 0)
5702                 return 0;
5703
5704         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5705         if (reserve)
5706                 BTRFS_I(inode)->csum_bytes += num_bytes;
5707         else
5708                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5709         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5710
5711         /* No change, no need to reserve more */
5712         if (old_csums == num_csums)
5713                 return 0;
5714
5715         if (reserve)
5716                 return btrfs_calc_trans_metadata_size(root,
5717                                                       num_csums - old_csums);
5718
5719         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5720 }
5721
5722 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5723 {
5724         struct btrfs_root *root = BTRFS_I(inode)->root;
5725         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5726         u64 to_reserve = 0;
5727         u64 csum_bytes;
5728         unsigned nr_extents = 0;
5729         int extra_reserve = 0;
5730         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5731         int ret = 0;
5732         bool delalloc_lock = true;
5733         u64 to_free = 0;
5734         unsigned dropped;
5735
5736         /* If we are a free space inode we need to not flush since we will be in
5737          * the middle of a transaction commit.  We also don't need the delalloc
5738          * mutex since we won't race with anybody.  We need this mostly to make
5739          * lockdep shut its filthy mouth.
5740          */
5741         if (btrfs_is_free_space_inode(inode)) {
5742                 flush = BTRFS_RESERVE_NO_FLUSH;
5743                 delalloc_lock = false;
5744         }
5745
5746         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5747             btrfs_transaction_in_commit(root->fs_info))
5748                 schedule_timeout(1);
5749
5750         if (delalloc_lock)
5751                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5752
5753         num_bytes = ALIGN(num_bytes, root->sectorsize);
5754
5755         spin_lock(&BTRFS_I(inode)->lock);
5756         nr_extents = (unsigned)div64_u64(num_bytes +
5757                                          BTRFS_MAX_EXTENT_SIZE - 1,
5758                                          BTRFS_MAX_EXTENT_SIZE);
5759         BTRFS_I(inode)->outstanding_extents += nr_extents;
5760         nr_extents = 0;
5761
5762         if (BTRFS_I(inode)->outstanding_extents >
5763             BTRFS_I(inode)->reserved_extents)
5764                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5765                         BTRFS_I(inode)->reserved_extents;
5766
5767         /*
5768          * Add an item to reserve for updating the inode when we complete the
5769          * delalloc io.
5770          */
5771         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5772                       &BTRFS_I(inode)->runtime_flags)) {
5773                 nr_extents++;
5774                 extra_reserve = 1;
5775         }
5776
5777         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5778         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5779         csum_bytes = BTRFS_I(inode)->csum_bytes;
5780         spin_unlock(&BTRFS_I(inode)->lock);
5781
5782         if (root->fs_info->quota_enabled) {
5783                 ret = btrfs_qgroup_reserve_meta(root,
5784                                 nr_extents * root->nodesize);
5785                 if (ret)
5786                         goto out_fail;
5787         }
5788
5789         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5790         if (unlikely(ret)) {
5791                 btrfs_qgroup_free_meta(root, nr_extents * root->nodesize);
5792                 goto out_fail;
5793         }
5794
5795         spin_lock(&BTRFS_I(inode)->lock);
5796         if (extra_reserve) {
5797                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5798                         &BTRFS_I(inode)->runtime_flags);
5799                 nr_extents--;
5800         }
5801         BTRFS_I(inode)->reserved_extents += nr_extents;
5802         spin_unlock(&BTRFS_I(inode)->lock);
5803
5804         if (delalloc_lock)
5805                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5806
5807         if (to_reserve)
5808                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5809                                               btrfs_ino(inode), to_reserve, 1);
5810         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5811
5812         return 0;
5813
5814 out_fail:
5815         spin_lock(&BTRFS_I(inode)->lock);
5816         dropped = drop_outstanding_extent(inode, num_bytes);
5817         /*
5818          * If the inodes csum_bytes is the same as the original
5819          * csum_bytes then we know we haven't raced with any free()ers
5820          * so we can just reduce our inodes csum bytes and carry on.
5821          */
5822         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5823                 calc_csum_metadata_size(inode, num_bytes, 0);
5824         } else {
5825                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5826                 u64 bytes;
5827
5828                 /*
5829                  * This is tricky, but first we need to figure out how much we
5830                  * freed from any free-ers that occurred during this
5831                  * reservation, so we reset ->csum_bytes to the csum_bytes
5832                  * before we dropped our lock, and then call the free for the
5833                  * number of bytes that were freed while we were trying our
5834                  * reservation.
5835                  */
5836                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5837                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5838                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5839
5840
5841                 /*
5842                  * Now we need to see how much we would have freed had we not
5843                  * been making this reservation and our ->csum_bytes were not
5844                  * artificially inflated.
5845                  */
5846                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5847                 bytes = csum_bytes - orig_csum_bytes;
5848                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5849
5850                 /*
5851                  * Now reset ->csum_bytes to what it should be.  If bytes is
5852                  * more than to_free then we would have freed more space had we
5853                  * not had an artificially high ->csum_bytes, so we need to free
5854                  * the remainder.  If bytes is the same or less then we don't
5855                  * need to do anything, the other free-ers did the correct
5856                  * thing.
5857                  */
5858                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5859                 if (bytes > to_free)
5860                         to_free = bytes - to_free;
5861                 else
5862                         to_free = 0;
5863         }
5864         spin_unlock(&BTRFS_I(inode)->lock);
5865         if (dropped)
5866                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5867
5868         if (to_free) {
5869                 btrfs_block_rsv_release(root, block_rsv, to_free);
5870                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5871                                               btrfs_ino(inode), to_free, 0);
5872         }
5873         if (delalloc_lock)
5874                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5875         return ret;
5876 }
5877
5878 /**
5879  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5880  * @inode: the inode to release the reservation for
5881  * @num_bytes: the number of bytes we're releasing
5882  *
5883  * This will release the metadata reservation for an inode.  This can be called
5884  * once we complete IO for a given set of bytes to release their metadata
5885  * reservations.
5886  */
5887 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5888 {
5889         struct btrfs_root *root = BTRFS_I(inode)->root;
5890         u64 to_free = 0;
5891         unsigned dropped;
5892
5893         num_bytes = ALIGN(num_bytes, root->sectorsize);
5894         spin_lock(&BTRFS_I(inode)->lock);
5895         dropped = drop_outstanding_extent(inode, num_bytes);
5896
5897         if (num_bytes)
5898                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5899         spin_unlock(&BTRFS_I(inode)->lock);
5900         if (dropped > 0)
5901                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5902
5903         if (btrfs_test_is_dummy_root(root))
5904                 return;
5905
5906         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5907                                       btrfs_ino(inode), to_free, 0);
5908
5909         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5910                                 to_free);
5911 }
5912
5913 /**
5914  * btrfs_delalloc_reserve_space - reserve data and metadata space for
5915  * delalloc
5916  * @inode: inode we're writing to
5917  * @start: start range we are writing to
5918  * @len: how long the range we are writing to
5919  *
5920  * TODO: This function will finally replace old btrfs_delalloc_reserve_space()
5921  *
5922  * This will do the following things
5923  *
5924  * o reserve space in data space info for num bytes
5925  *   and reserve precious corresponding qgroup space
5926  *   (Done in check_data_free_space)
5927  *
5928  * o reserve space for metadata space, based on the number of outstanding
5929  *   extents and how much csums will be needed
5930  *   also reserve metadata space in a per root over-reserve method.
5931  * o add to the inodes->delalloc_bytes
5932  * o add it to the fs_info's delalloc inodes list.
5933  *   (Above 3 all done in delalloc_reserve_metadata)
5934  *
5935  * Return 0 for success
5936  * Return <0 for error(-ENOSPC or -EQUOT)
5937  */
5938 int btrfs_delalloc_reserve_space(struct inode *inode, u64 start, u64 len)
5939 {
5940         int ret;
5941
5942         ret = btrfs_check_data_free_space(inode, start, len);
5943         if (ret < 0)
5944                 return ret;
5945         ret = btrfs_delalloc_reserve_metadata(inode, len);
5946         if (ret < 0)
5947                 btrfs_free_reserved_data_space(inode, start, len);
5948         return ret;
5949 }
5950
5951 /**
5952  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5953  * @inode: inode we're releasing space for
5954  * @start: start position of the space already reserved
5955  * @len: the len of the space already reserved
5956  *
5957  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5958  * called in the case that we don't need the metadata AND data reservations
5959  * anymore.  So if there is an error or we insert an inline extent.
5960  *
5961  * This function will release the metadata space that was not used and will
5962  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5963  * list if there are no delalloc bytes left.
5964  * Also it will handle the qgroup reserved space.
5965  */
5966 void btrfs_delalloc_release_space(struct inode *inode, u64 start, u64 len)
5967 {
5968         btrfs_delalloc_release_metadata(inode, len);
5969         btrfs_free_reserved_data_space(inode, start, len);
5970 }
5971
5972 static int update_block_group(struct btrfs_trans_handle *trans,
5973                               struct btrfs_root *root, u64 bytenr,
5974                               u64 num_bytes, int alloc)
5975 {
5976         struct btrfs_block_group_cache *cache = NULL;
5977         struct btrfs_fs_info *info = root->fs_info;
5978         u64 total = num_bytes;
5979         u64 old_val;
5980         u64 byte_in_group;
5981         int factor;
5982
5983         /* block accounting for super block */
5984         spin_lock(&info->delalloc_root_lock);
5985         old_val = btrfs_super_bytes_used(info->super_copy);
5986         if (alloc)
5987                 old_val += num_bytes;
5988         else
5989                 old_val -= num_bytes;
5990         btrfs_set_super_bytes_used(info->super_copy, old_val);
5991         spin_unlock(&info->delalloc_root_lock);
5992
5993         while (total) {
5994                 cache = btrfs_lookup_block_group(info, bytenr);
5995                 if (!cache)
5996                         return -ENOENT;
5997                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5998                                     BTRFS_BLOCK_GROUP_RAID1 |
5999                                     BTRFS_BLOCK_GROUP_RAID10))
6000                         factor = 2;
6001                 else
6002                         factor = 1;
6003                 /*
6004                  * If this block group has free space cache written out, we
6005                  * need to make sure to load it if we are removing space.  This
6006                  * is because we need the unpinning stage to actually add the
6007                  * space back to the block group, otherwise we will leak space.
6008                  */
6009                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6010                         cache_block_group(cache, 1);
6011
6012                 byte_in_group = bytenr - cache->key.objectid;
6013                 WARN_ON(byte_in_group > cache->key.offset);
6014
6015                 spin_lock(&cache->space_info->lock);
6016                 spin_lock(&cache->lock);
6017
6018                 if (btrfs_test_opt(root, SPACE_CACHE) &&
6019                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6020                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6021
6022                 old_val = btrfs_block_group_used(&cache->item);
6023                 num_bytes = min(total, cache->key.offset - byte_in_group);
6024                 if (alloc) {
6025                         old_val += num_bytes;
6026                         btrfs_set_block_group_used(&cache->item, old_val);
6027                         cache->reserved -= num_bytes;
6028                         cache->space_info->bytes_reserved -= num_bytes;
6029                         cache->space_info->bytes_used += num_bytes;
6030                         cache->space_info->disk_used += num_bytes * factor;
6031                         spin_unlock(&cache->lock);
6032                         spin_unlock(&cache->space_info->lock);
6033                 } else {
6034                         old_val -= num_bytes;
6035                         btrfs_set_block_group_used(&cache->item, old_val);
6036                         cache->pinned += num_bytes;
6037                         cache->space_info->bytes_pinned += num_bytes;
6038                         cache->space_info->bytes_used -= num_bytes;
6039                         cache->space_info->disk_used -= num_bytes * factor;
6040                         spin_unlock(&cache->lock);
6041                         spin_unlock(&cache->space_info->lock);
6042
6043                         set_extent_dirty(info->pinned_extents,
6044                                          bytenr, bytenr + num_bytes - 1,
6045                                          GFP_NOFS | __GFP_NOFAIL);
6046                 }
6047
6048                 spin_lock(&trans->transaction->dirty_bgs_lock);
6049                 if (list_empty(&cache->dirty_list)) {
6050                         list_add_tail(&cache->dirty_list,
6051                                       &trans->transaction->dirty_bgs);
6052                                 trans->transaction->num_dirty_bgs++;
6053                         btrfs_get_block_group(cache);
6054                 }
6055                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6056
6057                 /*
6058                  * No longer have used bytes in this block group, queue it for
6059                  * deletion. We do this after adding the block group to the
6060                  * dirty list to avoid races between cleaner kthread and space
6061                  * cache writeout.
6062                  */
6063                 if (!alloc && old_val == 0) {
6064                         spin_lock(&info->unused_bgs_lock);
6065                         if (list_empty(&cache->bg_list)) {
6066                                 btrfs_get_block_group(cache);
6067                                 list_add_tail(&cache->bg_list,
6068                                               &info->unused_bgs);
6069                         }
6070                         spin_unlock(&info->unused_bgs_lock);
6071                 }
6072
6073                 btrfs_put_block_group(cache);
6074                 total -= num_bytes;
6075                 bytenr += num_bytes;
6076         }
6077         return 0;
6078 }
6079
6080 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
6081 {
6082         struct btrfs_block_group_cache *cache;
6083         u64 bytenr;
6084
6085         spin_lock(&root->fs_info->block_group_cache_lock);
6086         bytenr = root->fs_info->first_logical_byte;
6087         spin_unlock(&root->fs_info->block_group_cache_lock);
6088
6089         if (bytenr < (u64)-1)
6090                 return bytenr;
6091
6092         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
6093         if (!cache)
6094                 return 0;
6095
6096         bytenr = cache->key.objectid;
6097         btrfs_put_block_group(cache);
6098
6099         return bytenr;
6100 }
6101
6102 static int pin_down_extent(struct btrfs_root *root,
6103                            struct btrfs_block_group_cache *cache,
6104                            u64 bytenr, u64 num_bytes, int reserved)
6105 {
6106         spin_lock(&cache->space_info->lock);
6107         spin_lock(&cache->lock);
6108         cache->pinned += num_bytes;
6109         cache->space_info->bytes_pinned += num_bytes;
6110         if (reserved) {
6111                 cache->reserved -= num_bytes;
6112                 cache->space_info->bytes_reserved -= num_bytes;
6113         }
6114         spin_unlock(&cache->lock);
6115         spin_unlock(&cache->space_info->lock);
6116
6117         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
6118                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6119         if (reserved)
6120                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
6121         return 0;
6122 }
6123
6124 /*
6125  * this function must be called within transaction
6126  */
6127 int btrfs_pin_extent(struct btrfs_root *root,
6128                      u64 bytenr, u64 num_bytes, int reserved)
6129 {
6130         struct btrfs_block_group_cache *cache;
6131
6132         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6133         BUG_ON(!cache); /* Logic error */
6134
6135         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
6136
6137         btrfs_put_block_group(cache);
6138         return 0;
6139 }
6140
6141 /*
6142  * this function must be called within transaction
6143  */
6144 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
6145                                     u64 bytenr, u64 num_bytes)
6146 {
6147         struct btrfs_block_group_cache *cache;
6148         int ret;
6149
6150         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6151         if (!cache)
6152                 return -EINVAL;
6153
6154         /*
6155          * pull in the free space cache (if any) so that our pin
6156          * removes the free space from the cache.  We have load_only set
6157          * to one because the slow code to read in the free extents does check
6158          * the pinned extents.
6159          */
6160         cache_block_group(cache, 1);
6161
6162         pin_down_extent(root, cache, bytenr, num_bytes, 0);
6163
6164         /* remove us from the free space cache (if we're there at all) */
6165         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6166         btrfs_put_block_group(cache);
6167         return ret;
6168 }
6169
6170 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
6171 {
6172         int ret;
6173         struct btrfs_block_group_cache *block_group;
6174         struct btrfs_caching_control *caching_ctl;
6175
6176         block_group = btrfs_lookup_block_group(root->fs_info, start);
6177         if (!block_group)
6178                 return -EINVAL;
6179
6180         cache_block_group(block_group, 0);
6181         caching_ctl = get_caching_control(block_group);
6182
6183         if (!caching_ctl) {
6184                 /* Logic error */
6185                 BUG_ON(!block_group_cache_done(block_group));
6186                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6187         } else {
6188                 mutex_lock(&caching_ctl->mutex);
6189
6190                 if (start >= caching_ctl->progress) {
6191                         ret = add_excluded_extent(root, start, num_bytes);
6192                 } else if (start + num_bytes <= caching_ctl->progress) {
6193                         ret = btrfs_remove_free_space(block_group,
6194                                                       start, num_bytes);
6195                 } else {
6196                         num_bytes = caching_ctl->progress - start;
6197                         ret = btrfs_remove_free_space(block_group,
6198                                                       start, num_bytes);
6199                         if (ret)
6200                                 goto out_lock;
6201
6202                         num_bytes = (start + num_bytes) -
6203                                 caching_ctl->progress;
6204                         start = caching_ctl->progress;
6205                         ret = add_excluded_extent(root, start, num_bytes);
6206                 }
6207 out_lock:
6208                 mutex_unlock(&caching_ctl->mutex);
6209                 put_caching_control(caching_ctl);
6210         }
6211         btrfs_put_block_group(block_group);
6212         return ret;
6213 }
6214
6215 int btrfs_exclude_logged_extents(struct btrfs_root *log,
6216                                  struct extent_buffer *eb)
6217 {
6218         struct btrfs_file_extent_item *item;
6219         struct btrfs_key key;
6220         int found_type;
6221         int i;
6222
6223         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
6224                 return 0;
6225
6226         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6227                 btrfs_item_key_to_cpu(eb, &key, i);
6228                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6229                         continue;
6230                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6231                 found_type = btrfs_file_extent_type(eb, item);
6232                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6233                         continue;
6234                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6235                         continue;
6236                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6237                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6238                 __exclude_logged_extent(log, key.objectid, key.offset);
6239         }
6240
6241         return 0;
6242 }
6243
6244 static void
6245 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6246 {
6247         atomic_inc(&bg->reservations);
6248 }
6249
6250 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6251                                         const u64 start)
6252 {
6253         struct btrfs_block_group_cache *bg;
6254
6255         bg = btrfs_lookup_block_group(fs_info, start);
6256         ASSERT(bg);
6257         if (atomic_dec_and_test(&bg->reservations))
6258                 wake_up_atomic_t(&bg->reservations);
6259         btrfs_put_block_group(bg);
6260 }
6261
6262 static int btrfs_wait_bg_reservations_atomic_t(atomic_t *a)
6263 {
6264         schedule();
6265         return 0;
6266 }
6267
6268 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6269 {
6270         struct btrfs_space_info *space_info = bg->space_info;
6271
6272         ASSERT(bg->ro);
6273
6274         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6275                 return;
6276
6277         /*
6278          * Our block group is read only but before we set it to read only,
6279          * some task might have had allocated an extent from it already, but it
6280          * has not yet created a respective ordered extent (and added it to a
6281          * root's list of ordered extents).
6282          * Therefore wait for any task currently allocating extents, since the
6283          * block group's reservations counter is incremented while a read lock
6284          * on the groups' semaphore is held and decremented after releasing
6285          * the read access on that semaphore and creating the ordered extent.
6286          */
6287         down_write(&space_info->groups_sem);
6288         up_write(&space_info->groups_sem);
6289
6290         wait_on_atomic_t(&bg->reservations,
6291                          btrfs_wait_bg_reservations_atomic_t,
6292                          TASK_UNINTERRUPTIBLE);
6293 }
6294
6295 /**
6296  * btrfs_update_reserved_bytes - update the block_group and space info counters
6297  * @cache:      The cache we are manipulating
6298  * @num_bytes:  The number of bytes in question
6299  * @reserve:    One of the reservation enums
6300  * @delalloc:   The blocks are allocated for the delalloc write
6301  *
6302  * This is called by the allocator when it reserves space, or by somebody who is
6303  * freeing space that was never actually used on disk.  For example if you
6304  * reserve some space for a new leaf in transaction A and before transaction A
6305  * commits you free that leaf, you call this with reserve set to 0 in order to
6306  * clear the reservation.
6307  *
6308  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
6309  * ENOSPC accounting.  For data we handle the reservation through clearing the
6310  * delalloc bits in the io_tree.  We have to do this since we could end up
6311  * allocating less disk space for the amount of data we have reserved in the
6312  * case of compression.
6313  *
6314  * If this is a reservation and the block group has become read only we cannot
6315  * make the reservation and return -EAGAIN, otherwise this function always
6316  * succeeds.
6317  */
6318 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
6319                                        u64 num_bytes, int reserve, int delalloc)
6320 {
6321         struct btrfs_space_info *space_info = cache->space_info;
6322         int ret = 0;
6323
6324         spin_lock(&space_info->lock);
6325         spin_lock(&cache->lock);
6326         if (reserve != RESERVE_FREE) {
6327                 if (cache->ro) {
6328                         ret = -EAGAIN;
6329                 } else {
6330                         cache->reserved += num_bytes;
6331                         space_info->bytes_reserved += num_bytes;
6332                         if (reserve == RESERVE_ALLOC) {
6333                                 trace_btrfs_space_reservation(cache->fs_info,
6334                                                 "space_info", space_info->flags,
6335                                                 num_bytes, 0);
6336                                 space_info->bytes_may_use -= num_bytes;
6337                         }
6338
6339                         if (delalloc)
6340                                 cache->delalloc_bytes += num_bytes;
6341                 }
6342         } else {
6343                 if (cache->ro)
6344                         space_info->bytes_readonly += num_bytes;
6345                 cache->reserved -= num_bytes;
6346                 space_info->bytes_reserved -= num_bytes;
6347
6348                 if (delalloc)
6349                         cache->delalloc_bytes -= num_bytes;
6350         }
6351         spin_unlock(&cache->lock);
6352         spin_unlock(&space_info->lock);
6353         return ret;
6354 }
6355
6356 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6357                                 struct btrfs_root *root)
6358 {
6359         struct btrfs_fs_info *fs_info = root->fs_info;
6360         struct btrfs_caching_control *next;
6361         struct btrfs_caching_control *caching_ctl;
6362         struct btrfs_block_group_cache *cache;
6363
6364         down_write(&fs_info->commit_root_sem);
6365
6366         list_for_each_entry_safe(caching_ctl, next,
6367                                  &fs_info->caching_block_groups, list) {
6368                 cache = caching_ctl->block_group;
6369                 if (block_group_cache_done(cache)) {
6370                         cache->last_byte_to_unpin = (u64)-1;
6371                         list_del_init(&caching_ctl->list);
6372                         put_caching_control(caching_ctl);
6373                 } else {
6374                         cache->last_byte_to_unpin = caching_ctl->progress;
6375                 }
6376         }
6377
6378         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6379                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6380         else
6381                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6382
6383         up_write(&fs_info->commit_root_sem);
6384
6385         update_global_block_rsv(fs_info);
6386 }
6387
6388 /*
6389  * Returns the free cluster for the given space info and sets empty_cluster to
6390  * what it should be based on the mount options.
6391  */
6392 static struct btrfs_free_cluster *
6393 fetch_cluster_info(struct btrfs_root *root, struct btrfs_space_info *space_info,
6394                    u64 *empty_cluster)
6395 {
6396         struct btrfs_free_cluster *ret = NULL;
6397         bool ssd = btrfs_test_opt(root, SSD);
6398
6399         *empty_cluster = 0;
6400         if (btrfs_mixed_space_info(space_info))
6401                 return ret;
6402
6403         if (ssd)
6404                 *empty_cluster = SZ_2M;
6405         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6406                 ret = &root->fs_info->meta_alloc_cluster;
6407                 if (!ssd)
6408                         *empty_cluster = SZ_64K;
6409         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) && ssd) {
6410                 ret = &root->fs_info->data_alloc_cluster;
6411         }
6412
6413         return ret;
6414 }
6415
6416 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
6417                               const bool return_free_space)
6418 {
6419         struct btrfs_fs_info *fs_info = root->fs_info;
6420         struct btrfs_block_group_cache *cache = NULL;
6421         struct btrfs_space_info *space_info;
6422         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6423         struct btrfs_free_cluster *cluster = NULL;
6424         u64 len;
6425         u64 total_unpinned = 0;
6426         u64 empty_cluster = 0;
6427         bool readonly;
6428
6429         while (start <= end) {
6430                 readonly = false;
6431                 if (!cache ||
6432                     start >= cache->key.objectid + cache->key.offset) {
6433                         if (cache)
6434                                 btrfs_put_block_group(cache);
6435                         total_unpinned = 0;
6436                         cache = btrfs_lookup_block_group(fs_info, start);
6437                         BUG_ON(!cache); /* Logic error */
6438
6439                         cluster = fetch_cluster_info(root,
6440                                                      cache->space_info,
6441                                                      &empty_cluster);
6442                         empty_cluster <<= 1;
6443                 }
6444
6445                 len = cache->key.objectid + cache->key.offset - start;
6446                 len = min(len, end + 1 - start);
6447
6448                 if (start < cache->last_byte_to_unpin) {
6449                         len = min(len, cache->last_byte_to_unpin - start);
6450                         if (return_free_space)
6451                                 btrfs_add_free_space(cache, start, len);
6452                 }
6453
6454                 start += len;
6455                 total_unpinned += len;
6456                 space_info = cache->space_info;
6457
6458                 /*
6459                  * If this space cluster has been marked as fragmented and we've
6460                  * unpinned enough in this block group to potentially allow a
6461                  * cluster to be created inside of it go ahead and clear the
6462                  * fragmented check.
6463                  */
6464                 if (cluster && cluster->fragmented &&
6465                     total_unpinned > empty_cluster) {
6466                         spin_lock(&cluster->lock);
6467                         cluster->fragmented = 0;
6468                         spin_unlock(&cluster->lock);
6469                 }
6470
6471                 spin_lock(&space_info->lock);
6472                 spin_lock(&cache->lock);
6473                 cache->pinned -= len;
6474                 space_info->bytes_pinned -= len;
6475                 space_info->max_extent_size = 0;
6476                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6477                 if (cache->ro) {
6478                         space_info->bytes_readonly += len;
6479                         readonly = true;
6480                 }
6481                 spin_unlock(&cache->lock);
6482                 if (!readonly && global_rsv->space_info == space_info) {
6483                         spin_lock(&global_rsv->lock);
6484                         if (!global_rsv->full) {
6485                                 len = min(len, global_rsv->size -
6486                                           global_rsv->reserved);
6487                                 global_rsv->reserved += len;
6488                                 space_info->bytes_may_use += len;
6489                                 if (global_rsv->reserved >= global_rsv->size)
6490                                         global_rsv->full = 1;
6491                         }
6492                         spin_unlock(&global_rsv->lock);
6493                 }
6494                 spin_unlock(&space_info->lock);
6495         }
6496
6497         if (cache)
6498                 btrfs_put_block_group(cache);
6499         return 0;
6500 }
6501
6502 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6503                                struct btrfs_root *root)
6504 {
6505         struct btrfs_fs_info *fs_info = root->fs_info;
6506         struct btrfs_block_group_cache *block_group, *tmp;
6507         struct list_head *deleted_bgs;
6508         struct extent_io_tree *unpin;
6509         u64 start;
6510         u64 end;
6511         int ret;
6512
6513         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6514                 unpin = &fs_info->freed_extents[1];
6515         else
6516                 unpin = &fs_info->freed_extents[0];
6517
6518         while (!trans->aborted) {
6519                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6520                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6521                                             EXTENT_DIRTY, NULL);
6522                 if (ret) {
6523                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6524                         break;
6525                 }
6526
6527                 if (btrfs_test_opt(root, DISCARD))
6528                         ret = btrfs_discard_extent(root, start,
6529                                                    end + 1 - start, NULL);
6530
6531                 clear_extent_dirty(unpin, start, end);
6532                 unpin_extent_range(root, start, end, true);
6533                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6534                 cond_resched();
6535         }
6536
6537         /*
6538          * Transaction is finished.  We don't need the lock anymore.  We
6539          * do need to clean up the block groups in case of a transaction
6540          * abort.
6541          */
6542         deleted_bgs = &trans->transaction->deleted_bgs;
6543         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6544                 u64 trimmed = 0;
6545
6546                 ret = -EROFS;
6547                 if (!trans->aborted)
6548                         ret = btrfs_discard_extent(root,
6549                                                    block_group->key.objectid,
6550                                                    block_group->key.offset,
6551                                                    &trimmed);
6552
6553                 list_del_init(&block_group->bg_list);
6554                 btrfs_put_block_group_trimming(block_group);
6555                 btrfs_put_block_group(block_group);
6556
6557                 if (ret) {
6558                         const char *errstr = btrfs_decode_error(ret);
6559                         btrfs_warn(fs_info,
6560                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6561                                    ret, errstr);
6562                 }
6563         }
6564
6565         return 0;
6566 }
6567
6568 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6569                              u64 owner, u64 root_objectid)
6570 {
6571         struct btrfs_space_info *space_info;
6572         u64 flags;
6573
6574         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6575                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6576                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6577                 else
6578                         flags = BTRFS_BLOCK_GROUP_METADATA;
6579         } else {
6580                 flags = BTRFS_BLOCK_GROUP_DATA;
6581         }
6582
6583         space_info = __find_space_info(fs_info, flags);
6584         BUG_ON(!space_info); /* Logic bug */
6585         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6586 }
6587
6588
6589 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6590                                 struct btrfs_root *root,
6591                                 struct btrfs_delayed_ref_node *node, u64 parent,
6592                                 u64 root_objectid, u64 owner_objectid,
6593                                 u64 owner_offset, int refs_to_drop,
6594                                 struct btrfs_delayed_extent_op *extent_op)
6595 {
6596         struct btrfs_key key;
6597         struct btrfs_path *path;
6598         struct btrfs_fs_info *info = root->fs_info;
6599         struct btrfs_root *extent_root = info->extent_root;
6600         struct extent_buffer *leaf;
6601         struct btrfs_extent_item *ei;
6602         struct btrfs_extent_inline_ref *iref;
6603         int ret;
6604         int is_data;
6605         int extent_slot = 0;
6606         int found_extent = 0;
6607         int num_to_del = 1;
6608         u32 item_size;
6609         u64 refs;
6610         u64 bytenr = node->bytenr;
6611         u64 num_bytes = node->num_bytes;
6612         int last_ref = 0;
6613         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6614                                                  SKINNY_METADATA);
6615
6616         path = btrfs_alloc_path();
6617         if (!path)
6618                 return -ENOMEM;
6619
6620         path->reada = READA_FORWARD;
6621         path->leave_spinning = 1;
6622
6623         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6624         BUG_ON(!is_data && refs_to_drop != 1);
6625
6626         if (is_data)
6627                 skinny_metadata = 0;
6628
6629         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6630                                     bytenr, num_bytes, parent,
6631                                     root_objectid, owner_objectid,
6632                                     owner_offset);
6633         if (ret == 0) {
6634                 extent_slot = path->slots[0];
6635                 while (extent_slot >= 0) {
6636                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6637                                               extent_slot);
6638                         if (key.objectid != bytenr)
6639                                 break;
6640                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6641                             key.offset == num_bytes) {
6642                                 found_extent = 1;
6643                                 break;
6644                         }
6645                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6646                             key.offset == owner_objectid) {
6647                                 found_extent = 1;
6648                                 break;
6649                         }
6650                         if (path->slots[0] - extent_slot > 5)
6651                                 break;
6652                         extent_slot--;
6653                 }
6654 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6655                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6656                 if (found_extent && item_size < sizeof(*ei))
6657                         found_extent = 0;
6658 #endif
6659                 if (!found_extent) {
6660                         BUG_ON(iref);
6661                         ret = remove_extent_backref(trans, extent_root, path,
6662                                                     NULL, refs_to_drop,
6663                                                     is_data, &last_ref);
6664                         if (ret) {
6665                                 btrfs_abort_transaction(trans, extent_root, ret);
6666                                 goto out;
6667                         }
6668                         btrfs_release_path(path);
6669                         path->leave_spinning = 1;
6670
6671                         key.objectid = bytenr;
6672                         key.type = BTRFS_EXTENT_ITEM_KEY;
6673                         key.offset = num_bytes;
6674
6675                         if (!is_data && skinny_metadata) {
6676                                 key.type = BTRFS_METADATA_ITEM_KEY;
6677                                 key.offset = owner_objectid;
6678                         }
6679
6680                         ret = btrfs_search_slot(trans, extent_root,
6681                                                 &key, path, -1, 1);
6682                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6683                                 /*
6684                                  * Couldn't find our skinny metadata item,
6685                                  * see if we have ye olde extent item.
6686                                  */
6687                                 path->slots[0]--;
6688                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6689                                                       path->slots[0]);
6690                                 if (key.objectid == bytenr &&
6691                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6692                                     key.offset == num_bytes)
6693                                         ret = 0;
6694                         }
6695
6696                         if (ret > 0 && skinny_metadata) {
6697                                 skinny_metadata = false;
6698                                 key.objectid = bytenr;
6699                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6700                                 key.offset = num_bytes;
6701                                 btrfs_release_path(path);
6702                                 ret = btrfs_search_slot(trans, extent_root,
6703                                                         &key, path, -1, 1);
6704                         }
6705
6706                         if (ret) {
6707                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6708                                         ret, bytenr);
6709                                 if (ret > 0)
6710                                         btrfs_print_leaf(extent_root,
6711                                                          path->nodes[0]);
6712                         }
6713                         if (ret < 0) {
6714                                 btrfs_abort_transaction(trans, extent_root, ret);
6715                                 goto out;
6716                         }
6717                         extent_slot = path->slots[0];
6718                 }
6719         } else if (WARN_ON(ret == -ENOENT)) {
6720                 btrfs_print_leaf(extent_root, path->nodes[0]);
6721                 btrfs_err(info,
6722                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6723                         bytenr, parent, root_objectid, owner_objectid,
6724                         owner_offset);
6725                 btrfs_abort_transaction(trans, extent_root, ret);
6726                 goto out;
6727         } else {
6728                 btrfs_abort_transaction(trans, extent_root, ret);
6729                 goto out;
6730         }
6731
6732         leaf = path->nodes[0];
6733         item_size = btrfs_item_size_nr(leaf, extent_slot);
6734 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6735         if (item_size < sizeof(*ei)) {
6736                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6737                 ret = convert_extent_item_v0(trans, extent_root, path,
6738                                              owner_objectid, 0);
6739                 if (ret < 0) {
6740                         btrfs_abort_transaction(trans, extent_root, ret);
6741                         goto out;
6742                 }
6743
6744                 btrfs_release_path(path);
6745                 path->leave_spinning = 1;
6746
6747                 key.objectid = bytenr;
6748                 key.type = BTRFS_EXTENT_ITEM_KEY;
6749                 key.offset = num_bytes;
6750
6751                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6752                                         -1, 1);
6753                 if (ret) {
6754                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6755                                 ret, bytenr);
6756                         btrfs_print_leaf(extent_root, path->nodes[0]);
6757                 }
6758                 if (ret < 0) {
6759                         btrfs_abort_transaction(trans, extent_root, ret);
6760                         goto out;
6761                 }
6762
6763                 extent_slot = path->slots[0];
6764                 leaf = path->nodes[0];
6765                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6766         }
6767 #endif
6768         BUG_ON(item_size < sizeof(*ei));
6769         ei = btrfs_item_ptr(leaf, extent_slot,
6770                             struct btrfs_extent_item);
6771         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6772             key.type == BTRFS_EXTENT_ITEM_KEY) {
6773                 struct btrfs_tree_block_info *bi;
6774                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6775                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6776                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6777         }
6778
6779         refs = btrfs_extent_refs(leaf, ei);
6780         if (refs < refs_to_drop) {
6781                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6782                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6783                 ret = -EINVAL;
6784                 btrfs_abort_transaction(trans, extent_root, ret);
6785                 goto out;
6786         }
6787         refs -= refs_to_drop;
6788
6789         if (refs > 0) {
6790                 if (extent_op)
6791                         __run_delayed_extent_op(extent_op, leaf, ei);
6792                 /*
6793                  * In the case of inline back ref, reference count will
6794                  * be updated by remove_extent_backref
6795                  */
6796                 if (iref) {
6797                         BUG_ON(!found_extent);
6798                 } else {
6799                         btrfs_set_extent_refs(leaf, ei, refs);
6800                         btrfs_mark_buffer_dirty(leaf);
6801                 }
6802                 if (found_extent) {
6803                         ret = remove_extent_backref(trans, extent_root, path,
6804                                                     iref, refs_to_drop,
6805                                                     is_data, &last_ref);
6806                         if (ret) {
6807                                 btrfs_abort_transaction(trans, extent_root, ret);
6808                                 goto out;
6809                         }
6810                 }
6811                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6812                                  root_objectid);
6813         } else {
6814                 if (found_extent) {
6815                         BUG_ON(is_data && refs_to_drop !=
6816                                extent_data_ref_count(path, iref));
6817                         if (iref) {
6818                                 BUG_ON(path->slots[0] != extent_slot);
6819                         } else {
6820                                 BUG_ON(path->slots[0] != extent_slot + 1);
6821                                 path->slots[0] = extent_slot;
6822                                 num_to_del = 2;
6823                         }
6824                 }
6825
6826                 last_ref = 1;
6827                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6828                                       num_to_del);
6829                 if (ret) {
6830                         btrfs_abort_transaction(trans, extent_root, ret);
6831                         goto out;
6832                 }
6833                 btrfs_release_path(path);
6834
6835                 if (is_data) {
6836                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6837                         if (ret) {
6838                                 btrfs_abort_transaction(trans, extent_root, ret);
6839                                 goto out;
6840                         }
6841                 }
6842
6843                 ret = add_to_free_space_tree(trans, root->fs_info, bytenr,
6844                                              num_bytes);
6845                 if (ret) {
6846                         btrfs_abort_transaction(trans, extent_root, ret);
6847                         goto out;
6848                 }
6849
6850                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6851                 if (ret) {
6852                         btrfs_abort_transaction(trans, extent_root, ret);
6853                         goto out;
6854                 }
6855         }
6856         btrfs_release_path(path);
6857
6858 out:
6859         btrfs_free_path(path);
6860         return ret;
6861 }
6862
6863 /*
6864  * when we free an block, it is possible (and likely) that we free the last
6865  * delayed ref for that extent as well.  This searches the delayed ref tree for
6866  * a given extent, and if there are no other delayed refs to be processed, it
6867  * removes it from the tree.
6868  */
6869 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6870                                       struct btrfs_root *root, u64 bytenr)
6871 {
6872         struct btrfs_delayed_ref_head *head;
6873         struct btrfs_delayed_ref_root *delayed_refs;
6874         int ret = 0;
6875
6876         delayed_refs = &trans->transaction->delayed_refs;
6877         spin_lock(&delayed_refs->lock);
6878         head = btrfs_find_delayed_ref_head(trans, bytenr);
6879         if (!head)
6880                 goto out_delayed_unlock;
6881
6882         spin_lock(&head->lock);
6883         if (!list_empty(&head->ref_list))
6884                 goto out;
6885
6886         if (head->extent_op) {
6887                 if (!head->must_insert_reserved)
6888                         goto out;
6889                 btrfs_free_delayed_extent_op(head->extent_op);
6890                 head->extent_op = NULL;
6891         }
6892
6893         /*
6894          * waiting for the lock here would deadlock.  If someone else has it
6895          * locked they are already in the process of dropping it anyway
6896          */
6897         if (!mutex_trylock(&head->mutex))
6898                 goto out;
6899
6900         /*
6901          * at this point we have a head with no other entries.  Go
6902          * ahead and process it.
6903          */
6904         head->node.in_tree = 0;
6905         rb_erase(&head->href_node, &delayed_refs->href_root);
6906
6907         atomic_dec(&delayed_refs->num_entries);
6908
6909         /*
6910          * we don't take a ref on the node because we're removing it from the
6911          * tree, so we just steal the ref the tree was holding.
6912          */
6913         delayed_refs->num_heads--;
6914         if (head->processing == 0)
6915                 delayed_refs->num_heads_ready--;
6916         head->processing = 0;
6917         spin_unlock(&head->lock);
6918         spin_unlock(&delayed_refs->lock);
6919
6920         BUG_ON(head->extent_op);
6921         if (head->must_insert_reserved)
6922                 ret = 1;
6923
6924         mutex_unlock(&head->mutex);
6925         btrfs_put_delayed_ref(&head->node);
6926         return ret;
6927 out:
6928         spin_unlock(&head->lock);
6929
6930 out_delayed_unlock:
6931         spin_unlock(&delayed_refs->lock);
6932         return 0;
6933 }
6934
6935 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6936                            struct btrfs_root *root,
6937                            struct extent_buffer *buf,
6938                            u64 parent, int last_ref)
6939 {
6940         int pin = 1;
6941         int ret;
6942
6943         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6944                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6945                                         buf->start, buf->len,
6946                                         parent, root->root_key.objectid,
6947                                         btrfs_header_level(buf),
6948                                         BTRFS_DROP_DELAYED_REF, NULL);
6949                 BUG_ON(ret); /* -ENOMEM */
6950         }
6951
6952         if (!last_ref)
6953                 return;
6954
6955         if (btrfs_header_generation(buf) == trans->transid) {
6956                 struct btrfs_block_group_cache *cache;
6957
6958                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6959                         ret = check_ref_cleanup(trans, root, buf->start);
6960                         if (!ret)
6961                                 goto out;
6962                 }
6963
6964                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6965
6966                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6967                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6968                         btrfs_put_block_group(cache);
6969                         goto out;
6970                 }
6971
6972                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6973
6974                 btrfs_add_free_space(cache, buf->start, buf->len);
6975                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6976                 btrfs_put_block_group(cache);
6977                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6978                 pin = 0;
6979         }
6980 out:
6981         if (pin)
6982                 add_pinned_bytes(root->fs_info, buf->len,
6983                                  btrfs_header_level(buf),
6984                                  root->root_key.objectid);
6985
6986         /*
6987          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6988          * anymore.
6989          */
6990         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6991 }
6992
6993 /* Can return -ENOMEM */
6994 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6995                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6996                       u64 owner, u64 offset)
6997 {
6998         int ret;
6999         struct btrfs_fs_info *fs_info = root->fs_info;
7000
7001         if (btrfs_test_is_dummy_root(root))
7002                 return 0;
7003
7004         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
7005
7006         /*
7007          * tree log blocks never actually go into the extent allocation
7008          * tree, just update pinning info and exit early.
7009          */
7010         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7011                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7012                 /* unlocks the pinned mutex */
7013                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
7014                 ret = 0;
7015         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7016                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7017                                         num_bytes,
7018                                         parent, root_objectid, (int)owner,
7019                                         BTRFS_DROP_DELAYED_REF, NULL);
7020         } else {
7021                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7022                                                 num_bytes,
7023                                                 parent, root_objectid, owner,
7024                                                 offset, 0,
7025                                                 BTRFS_DROP_DELAYED_REF, NULL);
7026         }
7027         return ret;
7028 }
7029
7030 /*
7031  * when we wait for progress in the block group caching, its because
7032  * our allocation attempt failed at least once.  So, we must sleep
7033  * and let some progress happen before we try again.
7034  *
7035  * This function will sleep at least once waiting for new free space to
7036  * show up, and then it will check the block group free space numbers
7037  * for our min num_bytes.  Another option is to have it go ahead
7038  * and look in the rbtree for a free extent of a given size, but this
7039  * is a good start.
7040  *
7041  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7042  * any of the information in this block group.
7043  */
7044 static noinline void
7045 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7046                                 u64 num_bytes)
7047 {
7048         struct btrfs_caching_control *caching_ctl;
7049
7050         caching_ctl = get_caching_control(cache);
7051         if (!caching_ctl)
7052                 return;
7053
7054         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7055                    (cache->free_space_ctl->free_space >= num_bytes));
7056
7057         put_caching_control(caching_ctl);
7058 }
7059
7060 static noinline int
7061 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7062 {
7063         struct btrfs_caching_control *caching_ctl;
7064         int ret = 0;
7065
7066         caching_ctl = get_caching_control(cache);
7067         if (!caching_ctl)
7068                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7069
7070         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7071         if (cache->cached == BTRFS_CACHE_ERROR)
7072                 ret = -EIO;
7073         put_caching_control(caching_ctl);
7074         return ret;
7075 }
7076
7077 int __get_raid_index(u64 flags)
7078 {
7079         if (flags & BTRFS_BLOCK_GROUP_RAID10)
7080                 return BTRFS_RAID_RAID10;
7081         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
7082                 return BTRFS_RAID_RAID1;
7083         else if (flags & BTRFS_BLOCK_GROUP_DUP)
7084                 return BTRFS_RAID_DUP;
7085         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
7086                 return BTRFS_RAID_RAID0;
7087         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
7088                 return BTRFS_RAID_RAID5;
7089         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
7090                 return BTRFS_RAID_RAID6;
7091
7092         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
7093 }
7094
7095 int get_block_group_index(struct btrfs_block_group_cache *cache)
7096 {
7097         return __get_raid_index(cache->flags);
7098 }
7099
7100 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7101         [BTRFS_RAID_RAID10]     = "raid10",
7102         [BTRFS_RAID_RAID1]      = "raid1",
7103         [BTRFS_RAID_DUP]        = "dup",
7104         [BTRFS_RAID_RAID0]      = "raid0",
7105         [BTRFS_RAID_SINGLE]     = "single",
7106         [BTRFS_RAID_RAID5]      = "raid5",
7107         [BTRFS_RAID_RAID6]      = "raid6",
7108 };
7109
7110 static const char *get_raid_name(enum btrfs_raid_types type)
7111 {
7112         if (type >= BTRFS_NR_RAID_TYPES)
7113                 return NULL;
7114
7115         return btrfs_raid_type_names[type];
7116 }
7117
7118 enum btrfs_loop_type {
7119         LOOP_CACHING_NOWAIT = 0,
7120         LOOP_CACHING_WAIT = 1,
7121         LOOP_ALLOC_CHUNK = 2,
7122         LOOP_NO_EMPTY_SIZE = 3,
7123 };
7124
7125 static inline void
7126 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7127                        int delalloc)
7128 {
7129         if (delalloc)
7130                 down_read(&cache->data_rwsem);
7131 }
7132
7133 static inline void
7134 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7135                        int delalloc)
7136 {
7137         btrfs_get_block_group(cache);
7138         if (delalloc)
7139                 down_read(&cache->data_rwsem);
7140 }
7141
7142 static struct btrfs_block_group_cache *
7143 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7144                    struct btrfs_free_cluster *cluster,
7145                    int delalloc)
7146 {
7147         struct btrfs_block_group_cache *used_bg = NULL;
7148
7149         spin_lock(&cluster->refill_lock);
7150         while (1) {
7151                 used_bg = cluster->block_group;
7152                 if (!used_bg)
7153                         return NULL;
7154
7155                 if (used_bg == block_group)
7156                         return used_bg;
7157
7158                 btrfs_get_block_group(used_bg);
7159
7160                 if (!delalloc)
7161                         return used_bg;
7162
7163                 if (down_read_trylock(&used_bg->data_rwsem))
7164                         return used_bg;
7165
7166                 spin_unlock(&cluster->refill_lock);
7167
7168                 down_read(&used_bg->data_rwsem);
7169
7170                 spin_lock(&cluster->refill_lock);
7171                 if (used_bg == cluster->block_group)
7172                         return used_bg;
7173
7174                 up_read(&used_bg->data_rwsem);
7175                 btrfs_put_block_group(used_bg);
7176         }
7177 }
7178
7179 static inline void
7180 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7181                          int delalloc)
7182 {
7183         if (delalloc)
7184                 up_read(&cache->data_rwsem);
7185         btrfs_put_block_group(cache);
7186 }
7187
7188 /*
7189  * walks the btree of allocated extents and find a hole of a given size.
7190  * The key ins is changed to record the hole:
7191  * ins->objectid == start position
7192  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7193  * ins->offset == the size of the hole.
7194  * Any available blocks before search_start are skipped.
7195  *
7196  * If there is no suitable free space, we will record the max size of
7197  * the free space extent currently.
7198  */
7199 static noinline int find_free_extent(struct btrfs_root *orig_root,
7200                                      u64 num_bytes, u64 empty_size,
7201                                      u64 hint_byte, struct btrfs_key *ins,
7202                                      u64 flags, int delalloc)
7203 {
7204         int ret = 0;
7205         struct btrfs_root *root = orig_root->fs_info->extent_root;
7206         struct btrfs_free_cluster *last_ptr = NULL;
7207         struct btrfs_block_group_cache *block_group = NULL;
7208         u64 search_start = 0;
7209         u64 max_extent_size = 0;
7210         u64 empty_cluster = 0;
7211         struct btrfs_space_info *space_info;
7212         int loop = 0;
7213         int index = __get_raid_index(flags);
7214         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
7215                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
7216         bool failed_cluster_refill = false;
7217         bool failed_alloc = false;
7218         bool use_cluster = true;
7219         bool have_caching_bg = false;
7220         bool orig_have_caching_bg = false;
7221         bool full_search = false;
7222
7223         WARN_ON(num_bytes < root->sectorsize);
7224         ins->type = BTRFS_EXTENT_ITEM_KEY;
7225         ins->objectid = 0;
7226         ins->offset = 0;
7227
7228         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
7229
7230         space_info = __find_space_info(root->fs_info, flags);
7231         if (!space_info) {
7232                 btrfs_err(root->fs_info, "No space info for %llu", flags);
7233                 return -ENOSPC;
7234         }
7235
7236         /*
7237          * If our free space is heavily fragmented we may not be able to make
7238          * big contiguous allocations, so instead of doing the expensive search
7239          * for free space, simply return ENOSPC with our max_extent_size so we
7240          * can go ahead and search for a more manageable chunk.
7241          *
7242          * If our max_extent_size is large enough for our allocation simply
7243          * disable clustering since we will likely not be able to find enough
7244          * space to create a cluster and induce latency trying.
7245          */
7246         if (unlikely(space_info->max_extent_size)) {
7247                 spin_lock(&space_info->lock);
7248                 if (space_info->max_extent_size &&
7249                     num_bytes > space_info->max_extent_size) {
7250                         ins->offset = space_info->max_extent_size;
7251                         spin_unlock(&space_info->lock);
7252                         return -ENOSPC;
7253                 } else if (space_info->max_extent_size) {
7254                         use_cluster = false;
7255                 }
7256                 spin_unlock(&space_info->lock);
7257         }
7258
7259         last_ptr = fetch_cluster_info(orig_root, space_info, &empty_cluster);
7260         if (last_ptr) {
7261                 spin_lock(&last_ptr->lock);
7262                 if (last_ptr->block_group)
7263                         hint_byte = last_ptr->window_start;
7264                 if (last_ptr->fragmented) {
7265                         /*
7266                          * We still set window_start so we can keep track of the
7267                          * last place we found an allocation to try and save
7268                          * some time.
7269                          */
7270                         hint_byte = last_ptr->window_start;
7271                         use_cluster = false;
7272                 }
7273                 spin_unlock(&last_ptr->lock);
7274         }
7275
7276         search_start = max(search_start, first_logical_byte(root, 0));
7277         search_start = max(search_start, hint_byte);
7278         if (search_start == hint_byte) {
7279                 block_group = btrfs_lookup_block_group(root->fs_info,
7280                                                        search_start);
7281                 /*
7282                  * we don't want to use the block group if it doesn't match our
7283                  * allocation bits, or if its not cached.
7284                  *
7285                  * However if we are re-searching with an ideal block group
7286                  * picked out then we don't care that the block group is cached.
7287                  */
7288                 if (block_group && block_group_bits(block_group, flags) &&
7289                     block_group->cached != BTRFS_CACHE_NO) {
7290                         down_read(&space_info->groups_sem);
7291                         if (list_empty(&block_group->list) ||
7292                             block_group->ro) {
7293                                 /*
7294                                  * someone is removing this block group,
7295                                  * we can't jump into the have_block_group
7296                                  * target because our list pointers are not
7297                                  * valid
7298                                  */
7299                                 btrfs_put_block_group(block_group);
7300                                 up_read(&space_info->groups_sem);
7301                         } else {
7302                                 index = get_block_group_index(block_group);
7303                                 btrfs_lock_block_group(block_group, delalloc);
7304                                 goto have_block_group;
7305                         }
7306                 } else if (block_group) {
7307                         btrfs_put_block_group(block_group);
7308                 }
7309         }
7310 search:
7311         have_caching_bg = false;
7312         if (index == 0 || index == __get_raid_index(flags))
7313                 full_search = true;
7314         down_read(&space_info->groups_sem);
7315         list_for_each_entry(block_group, &space_info->block_groups[index],
7316                             list) {
7317                 u64 offset;
7318                 int cached;
7319
7320                 btrfs_grab_block_group(block_group, delalloc);
7321                 search_start = block_group->key.objectid;
7322
7323                 /*
7324                  * this can happen if we end up cycling through all the
7325                  * raid types, but we want to make sure we only allocate
7326                  * for the proper type.
7327                  */
7328                 if (!block_group_bits(block_group, flags)) {
7329                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7330                                 BTRFS_BLOCK_GROUP_RAID1 |
7331                                 BTRFS_BLOCK_GROUP_RAID5 |
7332                                 BTRFS_BLOCK_GROUP_RAID6 |
7333                                 BTRFS_BLOCK_GROUP_RAID10;
7334
7335                         /*
7336                          * if they asked for extra copies and this block group
7337                          * doesn't provide them, bail.  This does allow us to
7338                          * fill raid0 from raid1.
7339                          */
7340                         if ((flags & extra) && !(block_group->flags & extra))
7341                                 goto loop;
7342                 }
7343
7344 have_block_group:
7345                 cached = block_group_cache_done(block_group);
7346                 if (unlikely(!cached)) {
7347                         have_caching_bg = true;
7348                         ret = cache_block_group(block_group, 0);
7349                         BUG_ON(ret < 0);
7350                         ret = 0;
7351                 }
7352
7353                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7354                         goto loop;
7355                 if (unlikely(block_group->ro))
7356                         goto loop;
7357
7358                 /*
7359                  * Ok we want to try and use the cluster allocator, so
7360                  * lets look there
7361                  */
7362                 if (last_ptr && use_cluster) {
7363                         struct btrfs_block_group_cache *used_block_group;
7364                         unsigned long aligned_cluster;
7365                         /*
7366                          * the refill lock keeps out other
7367                          * people trying to start a new cluster
7368                          */
7369                         used_block_group = btrfs_lock_cluster(block_group,
7370                                                               last_ptr,
7371                                                               delalloc);
7372                         if (!used_block_group)
7373                                 goto refill_cluster;
7374
7375                         if (used_block_group != block_group &&
7376                             (used_block_group->ro ||
7377                              !block_group_bits(used_block_group, flags)))
7378                                 goto release_cluster;
7379
7380                         offset = btrfs_alloc_from_cluster(used_block_group,
7381                                                 last_ptr,
7382                                                 num_bytes,
7383                                                 used_block_group->key.objectid,
7384                                                 &max_extent_size);
7385                         if (offset) {
7386                                 /* we have a block, we're done */
7387                                 spin_unlock(&last_ptr->refill_lock);
7388                                 trace_btrfs_reserve_extent_cluster(root,
7389                                                 used_block_group,
7390                                                 search_start, num_bytes);
7391                                 if (used_block_group != block_group) {
7392                                         btrfs_release_block_group(block_group,
7393                                                                   delalloc);
7394                                         block_group = used_block_group;
7395                                 }
7396                                 goto checks;
7397                         }
7398
7399                         WARN_ON(last_ptr->block_group != used_block_group);
7400 release_cluster:
7401                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7402                          * set up a new clusters, so lets just skip it
7403                          * and let the allocator find whatever block
7404                          * it can find.  If we reach this point, we
7405                          * will have tried the cluster allocator
7406                          * plenty of times and not have found
7407                          * anything, so we are likely way too
7408                          * fragmented for the clustering stuff to find
7409                          * anything.
7410                          *
7411                          * However, if the cluster is taken from the
7412                          * current block group, release the cluster
7413                          * first, so that we stand a better chance of
7414                          * succeeding in the unclustered
7415                          * allocation.  */
7416                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7417                             used_block_group != block_group) {
7418                                 spin_unlock(&last_ptr->refill_lock);
7419                                 btrfs_release_block_group(used_block_group,
7420                                                           delalloc);
7421                                 goto unclustered_alloc;
7422                         }
7423
7424                         /*
7425                          * this cluster didn't work out, free it and
7426                          * start over
7427                          */
7428                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7429
7430                         if (used_block_group != block_group)
7431                                 btrfs_release_block_group(used_block_group,
7432                                                           delalloc);
7433 refill_cluster:
7434                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7435                                 spin_unlock(&last_ptr->refill_lock);
7436                                 goto unclustered_alloc;
7437                         }
7438
7439                         aligned_cluster = max_t(unsigned long,
7440                                                 empty_cluster + empty_size,
7441                                               block_group->full_stripe_len);
7442
7443                         /* allocate a cluster in this block group */
7444                         ret = btrfs_find_space_cluster(root, block_group,
7445                                                        last_ptr, search_start,
7446                                                        num_bytes,
7447                                                        aligned_cluster);
7448                         if (ret == 0) {
7449                                 /*
7450                                  * now pull our allocation out of this
7451                                  * cluster
7452                                  */
7453                                 offset = btrfs_alloc_from_cluster(block_group,
7454                                                         last_ptr,
7455                                                         num_bytes,
7456                                                         search_start,
7457                                                         &max_extent_size);
7458                                 if (offset) {
7459                                         /* we found one, proceed */
7460                                         spin_unlock(&last_ptr->refill_lock);
7461                                         trace_btrfs_reserve_extent_cluster(root,
7462                                                 block_group, search_start,
7463                                                 num_bytes);
7464                                         goto checks;
7465                                 }
7466                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7467                                    && !failed_cluster_refill) {
7468                                 spin_unlock(&last_ptr->refill_lock);
7469
7470                                 failed_cluster_refill = true;
7471                                 wait_block_group_cache_progress(block_group,
7472                                        num_bytes + empty_cluster + empty_size);
7473                                 goto have_block_group;
7474                         }
7475
7476                         /*
7477                          * at this point we either didn't find a cluster
7478                          * or we weren't able to allocate a block from our
7479                          * cluster.  Free the cluster we've been trying
7480                          * to use, and go to the next block group
7481                          */
7482                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7483                         spin_unlock(&last_ptr->refill_lock);
7484                         goto loop;
7485                 }
7486
7487 unclustered_alloc:
7488                 /*
7489                  * We are doing an unclustered alloc, set the fragmented flag so
7490                  * we don't bother trying to setup a cluster again until we get
7491                  * more space.
7492                  */
7493                 if (unlikely(last_ptr)) {
7494                         spin_lock(&last_ptr->lock);
7495                         last_ptr->fragmented = 1;
7496                         spin_unlock(&last_ptr->lock);
7497                 }
7498                 spin_lock(&block_group->free_space_ctl->tree_lock);
7499                 if (cached &&
7500                     block_group->free_space_ctl->free_space <
7501                     num_bytes + empty_cluster + empty_size) {
7502                         if (block_group->free_space_ctl->free_space >
7503                             max_extent_size)
7504                                 max_extent_size =
7505                                         block_group->free_space_ctl->free_space;
7506                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7507                         goto loop;
7508                 }
7509                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7510
7511                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7512                                                     num_bytes, empty_size,
7513                                                     &max_extent_size);
7514                 /*
7515                  * If we didn't find a chunk, and we haven't failed on this
7516                  * block group before, and this block group is in the middle of
7517                  * caching and we are ok with waiting, then go ahead and wait
7518                  * for progress to be made, and set failed_alloc to true.
7519                  *
7520                  * If failed_alloc is true then we've already waited on this
7521                  * block group once and should move on to the next block group.
7522                  */
7523                 if (!offset && !failed_alloc && !cached &&
7524                     loop > LOOP_CACHING_NOWAIT) {
7525                         wait_block_group_cache_progress(block_group,
7526                                                 num_bytes + empty_size);
7527                         failed_alloc = true;
7528                         goto have_block_group;
7529                 } else if (!offset) {
7530                         goto loop;
7531                 }
7532 checks:
7533                 search_start = ALIGN(offset, root->stripesize);
7534
7535                 /* move on to the next group */
7536                 if (search_start + num_bytes >
7537                     block_group->key.objectid + block_group->key.offset) {
7538                         btrfs_add_free_space(block_group, offset, num_bytes);
7539                         goto loop;
7540                 }
7541
7542                 if (offset < search_start)
7543                         btrfs_add_free_space(block_group, offset,
7544                                              search_start - offset);
7545                 BUG_ON(offset > search_start);
7546
7547                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7548                                                   alloc_type, delalloc);
7549                 if (ret == -EAGAIN) {
7550                         btrfs_add_free_space(block_group, offset, num_bytes);
7551                         goto loop;
7552                 }
7553                 btrfs_inc_block_group_reservations(block_group);
7554
7555                 /* we are all good, lets return */
7556                 ins->objectid = search_start;
7557                 ins->offset = num_bytes;
7558
7559                 trace_btrfs_reserve_extent(orig_root, block_group,
7560                                            search_start, num_bytes);
7561                 btrfs_release_block_group(block_group, delalloc);
7562                 break;
7563 loop:
7564                 failed_cluster_refill = false;
7565                 failed_alloc = false;
7566                 BUG_ON(index != get_block_group_index(block_group));
7567                 btrfs_release_block_group(block_group, delalloc);
7568         }
7569         up_read(&space_info->groups_sem);
7570
7571         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7572                 && !orig_have_caching_bg)
7573                 orig_have_caching_bg = true;
7574
7575         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7576                 goto search;
7577
7578         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7579                 goto search;
7580
7581         /*
7582          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7583          *                      caching kthreads as we move along
7584          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7585          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7586          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7587          *                      again
7588          */
7589         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7590                 index = 0;
7591                 if (loop == LOOP_CACHING_NOWAIT) {
7592                         /*
7593                          * We want to skip the LOOP_CACHING_WAIT step if we
7594                          * don't have any uncached bgs and we've already done a
7595                          * full search through.
7596                          */
7597                         if (orig_have_caching_bg || !full_search)
7598                                 loop = LOOP_CACHING_WAIT;
7599                         else
7600                                 loop = LOOP_ALLOC_CHUNK;
7601                 } else {
7602                         loop++;
7603                 }
7604
7605                 if (loop == LOOP_ALLOC_CHUNK) {
7606                         struct btrfs_trans_handle *trans;
7607                         int exist = 0;
7608
7609                         trans = current->journal_info;
7610                         if (trans)
7611                                 exist = 1;
7612                         else
7613                                 trans = btrfs_join_transaction(root);
7614
7615                         if (IS_ERR(trans)) {
7616                                 ret = PTR_ERR(trans);
7617                                 goto out;
7618                         }
7619
7620                         ret = do_chunk_alloc(trans, root, flags,
7621                                              CHUNK_ALLOC_FORCE);
7622
7623                         /*
7624                          * If we can't allocate a new chunk we've already looped
7625                          * through at least once, move on to the NO_EMPTY_SIZE
7626                          * case.
7627                          */
7628                         if (ret == -ENOSPC)
7629                                 loop = LOOP_NO_EMPTY_SIZE;
7630
7631                         /*
7632                          * Do not bail out on ENOSPC since we
7633                          * can do more things.
7634                          */
7635                         if (ret < 0 && ret != -ENOSPC)
7636                                 btrfs_abort_transaction(trans,
7637                                                         root, ret);
7638                         else
7639                                 ret = 0;
7640                         if (!exist)
7641                                 btrfs_end_transaction(trans, root);
7642                         if (ret)
7643                                 goto out;
7644                 }
7645
7646                 if (loop == LOOP_NO_EMPTY_SIZE) {
7647                         /*
7648                          * Don't loop again if we already have no empty_size and
7649                          * no empty_cluster.
7650                          */
7651                         if (empty_size == 0 &&
7652                             empty_cluster == 0) {
7653                                 ret = -ENOSPC;
7654                                 goto out;
7655                         }
7656                         empty_size = 0;
7657                         empty_cluster = 0;
7658                 }
7659
7660                 goto search;
7661         } else if (!ins->objectid) {
7662                 ret = -ENOSPC;
7663         } else if (ins->objectid) {
7664                 if (!use_cluster && last_ptr) {
7665                         spin_lock(&last_ptr->lock);
7666                         last_ptr->window_start = ins->objectid;
7667                         spin_unlock(&last_ptr->lock);
7668                 }
7669                 ret = 0;
7670         }
7671 out:
7672         if (ret == -ENOSPC) {
7673                 spin_lock(&space_info->lock);
7674                 space_info->max_extent_size = max_extent_size;
7675                 spin_unlock(&space_info->lock);
7676                 ins->offset = max_extent_size;
7677         }
7678         return ret;
7679 }
7680
7681 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7682                             int dump_block_groups)
7683 {
7684         struct btrfs_block_group_cache *cache;
7685         int index = 0;
7686
7687         spin_lock(&info->lock);
7688         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7689                info->flags,
7690                info->total_bytes - info->bytes_used - info->bytes_pinned -
7691                info->bytes_reserved - info->bytes_readonly,
7692                (info->full) ? "" : "not ");
7693         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7694                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7695                info->total_bytes, info->bytes_used, info->bytes_pinned,
7696                info->bytes_reserved, info->bytes_may_use,
7697                info->bytes_readonly);
7698         spin_unlock(&info->lock);
7699
7700         if (!dump_block_groups)
7701                 return;
7702
7703         down_read(&info->groups_sem);
7704 again:
7705         list_for_each_entry(cache, &info->block_groups[index], list) {
7706                 spin_lock(&cache->lock);
7707                 printk(KERN_INFO "BTRFS: "
7708                            "block group %llu has %llu bytes, "
7709                            "%llu used %llu pinned %llu reserved %s\n",
7710                        cache->key.objectid, cache->key.offset,
7711                        btrfs_block_group_used(&cache->item), cache->pinned,
7712                        cache->reserved, cache->ro ? "[readonly]" : "");
7713                 btrfs_dump_free_space(cache, bytes);
7714                 spin_unlock(&cache->lock);
7715         }
7716         if (++index < BTRFS_NR_RAID_TYPES)
7717                 goto again;
7718         up_read(&info->groups_sem);
7719 }
7720
7721 int btrfs_reserve_extent(struct btrfs_root *root,
7722                          u64 num_bytes, u64 min_alloc_size,
7723                          u64 empty_size, u64 hint_byte,
7724                          struct btrfs_key *ins, int is_data, int delalloc)
7725 {
7726         bool final_tried = num_bytes == min_alloc_size;
7727         u64 flags;
7728         int ret;
7729
7730         flags = btrfs_get_alloc_profile(root, is_data);
7731 again:
7732         WARN_ON(num_bytes < root->sectorsize);
7733         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7734                                flags, delalloc);
7735         if (!ret && !is_data) {
7736                 btrfs_dec_block_group_reservations(root->fs_info,
7737                                                    ins->objectid);
7738         } else if (ret == -ENOSPC) {
7739                 if (!final_tried && ins->offset) {
7740                         num_bytes = min(num_bytes >> 1, ins->offset);
7741                         num_bytes = round_down(num_bytes, root->sectorsize);
7742                         num_bytes = max(num_bytes, min_alloc_size);
7743                         if (num_bytes == min_alloc_size)
7744                                 final_tried = true;
7745                         goto again;
7746                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7747                         struct btrfs_space_info *sinfo;
7748
7749                         sinfo = __find_space_info(root->fs_info, flags);
7750                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7751                                 flags, num_bytes);
7752                         if (sinfo)
7753                                 dump_space_info(sinfo, num_bytes, 1);
7754                 }
7755         }
7756
7757         return ret;
7758 }
7759
7760 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7761                                         u64 start, u64 len,
7762                                         int pin, int delalloc)
7763 {
7764         struct btrfs_block_group_cache *cache;
7765         int ret = 0;
7766
7767         cache = btrfs_lookup_block_group(root->fs_info, start);
7768         if (!cache) {
7769                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7770                         start);
7771                 return -ENOSPC;
7772         }
7773
7774         if (pin)
7775                 pin_down_extent(root, cache, start, len, 1);
7776         else {
7777                 if (btrfs_test_opt(root, DISCARD))
7778                         ret = btrfs_discard_extent(root, start, len, NULL);
7779                 btrfs_add_free_space(cache, start, len);
7780                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7781         }
7782
7783         btrfs_put_block_group(cache);
7784
7785         trace_btrfs_reserved_extent_free(root, start, len);
7786
7787         return ret;
7788 }
7789
7790 int btrfs_free_reserved_extent(struct btrfs_root *root,
7791                                u64 start, u64 len, int delalloc)
7792 {
7793         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7794 }
7795
7796 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7797                                        u64 start, u64 len)
7798 {
7799         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7800 }
7801
7802 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7803                                       struct btrfs_root *root,
7804                                       u64 parent, u64 root_objectid,
7805                                       u64 flags, u64 owner, u64 offset,
7806                                       struct btrfs_key *ins, int ref_mod)
7807 {
7808         int ret;
7809         struct btrfs_fs_info *fs_info = root->fs_info;
7810         struct btrfs_extent_item *extent_item;
7811         struct btrfs_extent_inline_ref *iref;
7812         struct btrfs_path *path;
7813         struct extent_buffer *leaf;
7814         int type;
7815         u32 size;
7816
7817         if (parent > 0)
7818                 type = BTRFS_SHARED_DATA_REF_KEY;
7819         else
7820                 type = BTRFS_EXTENT_DATA_REF_KEY;
7821
7822         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7823
7824         path = btrfs_alloc_path();
7825         if (!path)
7826                 return -ENOMEM;
7827
7828         path->leave_spinning = 1;
7829         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7830                                       ins, size);
7831         if (ret) {
7832                 btrfs_free_path(path);
7833                 return ret;
7834         }
7835
7836         leaf = path->nodes[0];
7837         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7838                                      struct btrfs_extent_item);
7839         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7840         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7841         btrfs_set_extent_flags(leaf, extent_item,
7842                                flags | BTRFS_EXTENT_FLAG_DATA);
7843
7844         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7845         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7846         if (parent > 0) {
7847                 struct btrfs_shared_data_ref *ref;
7848                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7849                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7850                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7851         } else {
7852                 struct btrfs_extent_data_ref *ref;
7853                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7854                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7855                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7856                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7857                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7858         }
7859
7860         btrfs_mark_buffer_dirty(path->nodes[0]);
7861         btrfs_free_path(path);
7862
7863         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
7864                                           ins->offset);
7865         if (ret)
7866                 return ret;
7867
7868         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7869         if (ret) { /* -ENOENT, logic error */
7870                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7871                         ins->objectid, ins->offset);
7872                 BUG();
7873         }
7874         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7875         return ret;
7876 }
7877
7878 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7879                                      struct btrfs_root *root,
7880                                      u64 parent, u64 root_objectid,
7881                                      u64 flags, struct btrfs_disk_key *key,
7882                                      int level, struct btrfs_key *ins)
7883 {
7884         int ret;
7885         struct btrfs_fs_info *fs_info = root->fs_info;
7886         struct btrfs_extent_item *extent_item;
7887         struct btrfs_tree_block_info *block_info;
7888         struct btrfs_extent_inline_ref *iref;
7889         struct btrfs_path *path;
7890         struct extent_buffer *leaf;
7891         u32 size = sizeof(*extent_item) + sizeof(*iref);
7892         u64 num_bytes = ins->offset;
7893         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7894                                                  SKINNY_METADATA);
7895
7896         if (!skinny_metadata)
7897                 size += sizeof(*block_info);
7898
7899         path = btrfs_alloc_path();
7900         if (!path) {
7901                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7902                                                    root->nodesize);
7903                 return -ENOMEM;
7904         }
7905
7906         path->leave_spinning = 1;
7907         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7908                                       ins, size);
7909         if (ret) {
7910                 btrfs_free_path(path);
7911                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7912                                                    root->nodesize);
7913                 return ret;
7914         }
7915
7916         leaf = path->nodes[0];
7917         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7918                                      struct btrfs_extent_item);
7919         btrfs_set_extent_refs(leaf, extent_item, 1);
7920         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7921         btrfs_set_extent_flags(leaf, extent_item,
7922                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7923
7924         if (skinny_metadata) {
7925                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7926                 num_bytes = root->nodesize;
7927         } else {
7928                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7929                 btrfs_set_tree_block_key(leaf, block_info, key);
7930                 btrfs_set_tree_block_level(leaf, block_info, level);
7931                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7932         }
7933
7934         if (parent > 0) {
7935                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7936                 btrfs_set_extent_inline_ref_type(leaf, iref,
7937                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7938                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7939         } else {
7940                 btrfs_set_extent_inline_ref_type(leaf, iref,
7941                                                  BTRFS_TREE_BLOCK_REF_KEY);
7942                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7943         }
7944
7945         btrfs_mark_buffer_dirty(leaf);
7946         btrfs_free_path(path);
7947
7948         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
7949                                           num_bytes);
7950         if (ret)
7951                 return ret;
7952
7953         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7954                                  1);
7955         if (ret) { /* -ENOENT, logic error */
7956                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7957                         ins->objectid, ins->offset);
7958                 BUG();
7959         }
7960
7961         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7962         return ret;
7963 }
7964
7965 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7966                                      struct btrfs_root *root,
7967                                      u64 root_objectid, u64 owner,
7968                                      u64 offset, u64 ram_bytes,
7969                                      struct btrfs_key *ins)
7970 {
7971         int ret;
7972
7973         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7974
7975         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7976                                          ins->offset, 0,
7977                                          root_objectid, owner, offset,
7978                                          ram_bytes, BTRFS_ADD_DELAYED_EXTENT,
7979                                          NULL);
7980         return ret;
7981 }
7982
7983 /*
7984  * this is used by the tree logging recovery code.  It records that
7985  * an extent has been allocated and makes sure to clear the free
7986  * space cache bits as well
7987  */
7988 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7989                                    struct btrfs_root *root,
7990                                    u64 root_objectid, u64 owner, u64 offset,
7991                                    struct btrfs_key *ins)
7992 {
7993         int ret;
7994         struct btrfs_block_group_cache *block_group;
7995
7996         /*
7997          * Mixed block groups will exclude before processing the log so we only
7998          * need to do the exclude dance if this fs isn't mixed.
7999          */
8000         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
8001                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
8002                 if (ret)
8003                         return ret;
8004         }
8005
8006         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
8007         if (!block_group)
8008                 return -EINVAL;
8009
8010         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
8011                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
8012         BUG_ON(ret); /* logic error */
8013         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
8014                                          0, owner, offset, ins, 1);
8015         btrfs_put_block_group(block_group);
8016         return ret;
8017 }
8018
8019 static struct extent_buffer *
8020 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8021                       u64 bytenr, int level)
8022 {
8023         struct extent_buffer *buf;
8024
8025         buf = btrfs_find_create_tree_block(root, bytenr);
8026         if (IS_ERR(buf))
8027                 return buf;
8028
8029         btrfs_set_header_generation(buf, trans->transid);
8030         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8031         btrfs_tree_lock(buf);
8032         clean_tree_block(trans, root->fs_info, buf);
8033         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8034
8035         btrfs_set_lock_blocking(buf);
8036         set_extent_buffer_uptodate(buf);
8037
8038         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8039                 buf->log_index = root->log_transid % 2;
8040                 /*
8041                  * we allow two log transactions at a time, use different
8042                  * EXENT bit to differentiate dirty pages.
8043                  */
8044                 if (buf->log_index == 0)
8045                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8046                                         buf->start + buf->len - 1, GFP_NOFS);
8047                 else
8048                         set_extent_new(&root->dirty_log_pages, buf->start,
8049                                         buf->start + buf->len - 1);
8050         } else {
8051                 buf->log_index = -1;
8052                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8053                          buf->start + buf->len - 1, GFP_NOFS);
8054         }
8055         trans->dirty = true;
8056         /* this returns a buffer locked for blocking */
8057         return buf;
8058 }
8059
8060 static struct btrfs_block_rsv *
8061 use_block_rsv(struct btrfs_trans_handle *trans,
8062               struct btrfs_root *root, u32 blocksize)
8063 {
8064         struct btrfs_block_rsv *block_rsv;
8065         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
8066         int ret;
8067         bool global_updated = false;
8068
8069         block_rsv = get_block_rsv(trans, root);
8070
8071         if (unlikely(block_rsv->size == 0))
8072                 goto try_reserve;
8073 again:
8074         ret = block_rsv_use_bytes(block_rsv, blocksize);
8075         if (!ret)
8076                 return block_rsv;
8077
8078         if (block_rsv->failfast)
8079                 return ERR_PTR(ret);
8080
8081         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8082                 global_updated = true;
8083                 update_global_block_rsv(root->fs_info);
8084                 goto again;
8085         }
8086
8087         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
8088                 static DEFINE_RATELIMIT_STATE(_rs,
8089                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8090                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8091                 if (__ratelimit(&_rs))
8092                         WARN(1, KERN_DEBUG
8093                                 "BTRFS: block rsv returned %d\n", ret);
8094         }
8095 try_reserve:
8096         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8097                                      BTRFS_RESERVE_NO_FLUSH);
8098         if (!ret)
8099                 return block_rsv;
8100         /*
8101          * If we couldn't reserve metadata bytes try and use some from
8102          * the global reserve if its space type is the same as the global
8103          * reservation.
8104          */
8105         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8106             block_rsv->space_info == global_rsv->space_info) {
8107                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8108                 if (!ret)
8109                         return global_rsv;
8110         }
8111         return ERR_PTR(ret);
8112 }
8113
8114 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8115                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8116 {
8117         block_rsv_add_bytes(block_rsv, blocksize, 0);
8118         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8119 }
8120
8121 /*
8122  * finds a free extent and does all the dirty work required for allocation
8123  * returns the tree buffer or an ERR_PTR on error.
8124  */
8125 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8126                                         struct btrfs_root *root,
8127                                         u64 parent, u64 root_objectid,
8128                                         struct btrfs_disk_key *key, int level,
8129                                         u64 hint, u64 empty_size)
8130 {
8131         struct btrfs_key ins;
8132         struct btrfs_block_rsv *block_rsv;
8133         struct extent_buffer *buf;
8134         struct btrfs_delayed_extent_op *extent_op;
8135         u64 flags = 0;
8136         int ret;
8137         u32 blocksize = root->nodesize;
8138         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8139                                                  SKINNY_METADATA);
8140
8141         if (btrfs_test_is_dummy_root(root)) {
8142                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8143                                             level);
8144                 if (!IS_ERR(buf))
8145                         root->alloc_bytenr += blocksize;
8146                 return buf;
8147         }
8148
8149         block_rsv = use_block_rsv(trans, root, blocksize);
8150         if (IS_ERR(block_rsv))
8151                 return ERR_CAST(block_rsv);
8152
8153         ret = btrfs_reserve_extent(root, blocksize, blocksize,
8154                                    empty_size, hint, &ins, 0, 0);
8155         if (ret)
8156                 goto out_unuse;
8157
8158         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8159         if (IS_ERR(buf)) {
8160                 ret = PTR_ERR(buf);
8161                 goto out_free_reserved;
8162         }
8163
8164         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8165                 if (parent == 0)
8166                         parent = ins.objectid;
8167                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8168         } else
8169                 BUG_ON(parent > 0);
8170
8171         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8172                 extent_op = btrfs_alloc_delayed_extent_op();
8173                 if (!extent_op) {
8174                         ret = -ENOMEM;
8175                         goto out_free_buf;
8176                 }
8177                 if (key)
8178                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8179                 else
8180                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8181                 extent_op->flags_to_set = flags;
8182                 extent_op->update_key = skinny_metadata ? false : true;
8183                 extent_op->update_flags = true;
8184                 extent_op->is_data = false;
8185                 extent_op->level = level;
8186
8187                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
8188                                                  ins.objectid, ins.offset,
8189                                                  parent, root_objectid, level,
8190                                                  BTRFS_ADD_DELAYED_EXTENT,
8191                                                  extent_op);
8192                 if (ret)
8193                         goto out_free_delayed;
8194         }
8195         return buf;
8196
8197 out_free_delayed:
8198         btrfs_free_delayed_extent_op(extent_op);
8199 out_free_buf:
8200         free_extent_buffer(buf);
8201 out_free_reserved:
8202         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
8203 out_unuse:
8204         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
8205         return ERR_PTR(ret);
8206 }
8207
8208 struct walk_control {
8209         u64 refs[BTRFS_MAX_LEVEL];
8210         u64 flags[BTRFS_MAX_LEVEL];
8211         struct btrfs_key update_progress;
8212         int stage;
8213         int level;
8214         int shared_level;
8215         int update_ref;
8216         int keep_locks;
8217         int reada_slot;
8218         int reada_count;
8219         int for_reloc;
8220 };
8221
8222 #define DROP_REFERENCE  1
8223 #define UPDATE_BACKREF  2
8224
8225 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8226                                      struct btrfs_root *root,
8227                                      struct walk_control *wc,
8228                                      struct btrfs_path *path)
8229 {
8230         u64 bytenr;
8231         u64 generation;
8232         u64 refs;
8233         u64 flags;
8234         u32 nritems;
8235         u32 blocksize;
8236         struct btrfs_key key;
8237         struct extent_buffer *eb;
8238         int ret;
8239         int slot;
8240         int nread = 0;
8241
8242         if (path->slots[wc->level] < wc->reada_slot) {
8243                 wc->reada_count = wc->reada_count * 2 / 3;
8244                 wc->reada_count = max(wc->reada_count, 2);
8245         } else {
8246                 wc->reada_count = wc->reada_count * 3 / 2;
8247                 wc->reada_count = min_t(int, wc->reada_count,
8248                                         BTRFS_NODEPTRS_PER_BLOCK(root));
8249         }
8250
8251         eb = path->nodes[wc->level];
8252         nritems = btrfs_header_nritems(eb);
8253         blocksize = root->nodesize;
8254
8255         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8256                 if (nread >= wc->reada_count)
8257                         break;
8258
8259                 cond_resched();
8260                 bytenr = btrfs_node_blockptr(eb, slot);
8261                 generation = btrfs_node_ptr_generation(eb, slot);
8262
8263                 if (slot == path->slots[wc->level])
8264                         goto reada;
8265
8266                 if (wc->stage == UPDATE_BACKREF &&
8267                     generation <= root->root_key.offset)
8268                         continue;
8269
8270                 /* We don't lock the tree block, it's OK to be racy here */
8271                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
8272                                                wc->level - 1, 1, &refs,
8273                                                &flags);
8274                 /* We don't care about errors in readahead. */
8275                 if (ret < 0)
8276                         continue;
8277                 BUG_ON(refs == 0);
8278
8279                 if (wc->stage == DROP_REFERENCE) {
8280                         if (refs == 1)
8281                                 goto reada;
8282
8283                         if (wc->level == 1 &&
8284                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8285                                 continue;
8286                         if (!wc->update_ref ||
8287                             generation <= root->root_key.offset)
8288                                 continue;
8289                         btrfs_node_key_to_cpu(eb, &key, slot);
8290                         ret = btrfs_comp_cpu_keys(&key,
8291                                                   &wc->update_progress);
8292                         if (ret < 0)
8293                                 continue;
8294                 } else {
8295                         if (wc->level == 1 &&
8296                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8297                                 continue;
8298                 }
8299 reada:
8300                 readahead_tree_block(root, bytenr);
8301                 nread++;
8302         }
8303         wc->reada_slot = slot;
8304 }
8305
8306 /*
8307  * These may not be seen by the usual inc/dec ref code so we have to
8308  * add them here.
8309  */
8310 static int record_one_subtree_extent(struct btrfs_trans_handle *trans,
8311                                      struct btrfs_root *root, u64 bytenr,
8312                                      u64 num_bytes)
8313 {
8314         struct btrfs_qgroup_extent_record *qrecord;
8315         struct btrfs_delayed_ref_root *delayed_refs;
8316
8317         qrecord = kmalloc(sizeof(*qrecord), GFP_NOFS);
8318         if (!qrecord)
8319                 return -ENOMEM;
8320
8321         qrecord->bytenr = bytenr;
8322         qrecord->num_bytes = num_bytes;
8323         qrecord->old_roots = NULL;
8324
8325         delayed_refs = &trans->transaction->delayed_refs;
8326         spin_lock(&delayed_refs->lock);
8327         if (btrfs_qgroup_insert_dirty_extent(delayed_refs, qrecord))
8328                 kfree(qrecord);
8329         spin_unlock(&delayed_refs->lock);
8330
8331         return 0;
8332 }
8333
8334 static int account_leaf_items(struct btrfs_trans_handle *trans,
8335                               struct btrfs_root *root,
8336                               struct extent_buffer *eb)
8337 {
8338         int nr = btrfs_header_nritems(eb);
8339         int i, extent_type, ret;
8340         struct btrfs_key key;
8341         struct btrfs_file_extent_item *fi;
8342         u64 bytenr, num_bytes;
8343
8344         /* We can be called directly from walk_up_proc() */
8345         if (!root->fs_info->quota_enabled)
8346                 return 0;
8347
8348         for (i = 0; i < nr; i++) {
8349                 btrfs_item_key_to_cpu(eb, &key, i);
8350
8351                 if (key.type != BTRFS_EXTENT_DATA_KEY)
8352                         continue;
8353
8354                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
8355                 /* filter out non qgroup-accountable extents  */
8356                 extent_type = btrfs_file_extent_type(eb, fi);
8357
8358                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
8359                         continue;
8360
8361                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
8362                 if (!bytenr)
8363                         continue;
8364
8365                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
8366
8367                 ret = record_one_subtree_extent(trans, root, bytenr, num_bytes);
8368                 if (ret)
8369                         return ret;
8370         }
8371         return 0;
8372 }
8373
8374 /*
8375  * Walk up the tree from the bottom, freeing leaves and any interior
8376  * nodes which have had all slots visited. If a node (leaf or
8377  * interior) is freed, the node above it will have it's slot
8378  * incremented. The root node will never be freed.
8379  *
8380  * At the end of this function, we should have a path which has all
8381  * slots incremented to the next position for a search. If we need to
8382  * read a new node it will be NULL and the node above it will have the
8383  * correct slot selected for a later read.
8384  *
8385  * If we increment the root nodes slot counter past the number of
8386  * elements, 1 is returned to signal completion of the search.
8387  */
8388 static int adjust_slots_upwards(struct btrfs_root *root,
8389                                 struct btrfs_path *path, int root_level)
8390 {
8391         int level = 0;
8392         int nr, slot;
8393         struct extent_buffer *eb;
8394
8395         if (root_level == 0)
8396                 return 1;
8397
8398         while (level <= root_level) {
8399                 eb = path->nodes[level];
8400                 nr = btrfs_header_nritems(eb);
8401                 path->slots[level]++;
8402                 slot = path->slots[level];
8403                 if (slot >= nr || level == 0) {
8404                         /*
8405                          * Don't free the root -  we will detect this
8406                          * condition after our loop and return a
8407                          * positive value for caller to stop walking the tree.
8408                          */
8409                         if (level != root_level) {
8410                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8411                                 path->locks[level] = 0;
8412
8413                                 free_extent_buffer(eb);
8414                                 path->nodes[level] = NULL;
8415                                 path->slots[level] = 0;
8416                         }
8417                 } else {
8418                         /*
8419                          * We have a valid slot to walk back down
8420                          * from. Stop here so caller can process these
8421                          * new nodes.
8422                          */
8423                         break;
8424                 }
8425
8426                 level++;
8427         }
8428
8429         eb = path->nodes[root_level];
8430         if (path->slots[root_level] >= btrfs_header_nritems(eb))
8431                 return 1;
8432
8433         return 0;
8434 }
8435
8436 /*
8437  * root_eb is the subtree root and is locked before this function is called.
8438  */
8439 static int account_shared_subtree(struct btrfs_trans_handle *trans,
8440                                   struct btrfs_root *root,
8441                                   struct extent_buffer *root_eb,
8442                                   u64 root_gen,
8443                                   int root_level)
8444 {
8445         int ret = 0;
8446         int level;
8447         struct extent_buffer *eb = root_eb;
8448         struct btrfs_path *path = NULL;
8449
8450         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
8451         BUG_ON(root_eb == NULL);
8452
8453         if (!root->fs_info->quota_enabled)
8454                 return 0;
8455
8456         if (!extent_buffer_uptodate(root_eb)) {
8457                 ret = btrfs_read_buffer(root_eb, root_gen);
8458                 if (ret)
8459                         goto out;
8460         }
8461
8462         if (root_level == 0) {
8463                 ret = account_leaf_items(trans, root, root_eb);
8464                 goto out;
8465         }
8466
8467         path = btrfs_alloc_path();
8468         if (!path)
8469                 return -ENOMEM;
8470
8471         /*
8472          * Walk down the tree.  Missing extent blocks are filled in as
8473          * we go. Metadata is accounted every time we read a new
8474          * extent block.
8475          *
8476          * When we reach a leaf, we account for file extent items in it,
8477          * walk back up the tree (adjusting slot pointers as we go)
8478          * and restart the search process.
8479          */
8480         extent_buffer_get(root_eb); /* For path */
8481         path->nodes[root_level] = root_eb;
8482         path->slots[root_level] = 0;
8483         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
8484 walk_down:
8485         level = root_level;
8486         while (level >= 0) {
8487                 if (path->nodes[level] == NULL) {
8488                         int parent_slot;
8489                         u64 child_gen;
8490                         u64 child_bytenr;
8491
8492                         /* We need to get child blockptr/gen from
8493                          * parent before we can read it. */
8494                         eb = path->nodes[level + 1];
8495                         parent_slot = path->slots[level + 1];
8496                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8497                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8498
8499                         eb = read_tree_block(root, child_bytenr, child_gen);
8500                         if (IS_ERR(eb)) {
8501                                 ret = PTR_ERR(eb);
8502                                 goto out;
8503                         } else if (!extent_buffer_uptodate(eb)) {
8504                                 free_extent_buffer(eb);
8505                                 ret = -EIO;
8506                                 goto out;
8507                         }
8508
8509                         path->nodes[level] = eb;
8510                         path->slots[level] = 0;
8511
8512                         btrfs_tree_read_lock(eb);
8513                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8514                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8515
8516                         ret = record_one_subtree_extent(trans, root, child_bytenr,
8517                                                         root->nodesize);
8518                         if (ret)
8519                                 goto out;
8520                 }
8521
8522                 if (level == 0) {
8523                         ret = account_leaf_items(trans, root, path->nodes[level]);
8524                         if (ret)
8525                                 goto out;
8526
8527                         /* Nonzero return here means we completed our search */
8528                         ret = adjust_slots_upwards(root, path, root_level);
8529                         if (ret)
8530                                 break;
8531
8532                         /* Restart search with new slots */
8533                         goto walk_down;
8534                 }
8535
8536                 level--;
8537         }
8538
8539         ret = 0;
8540 out:
8541         btrfs_free_path(path);
8542
8543         return ret;
8544 }
8545
8546 /*
8547  * helper to process tree block while walking down the tree.
8548  *
8549  * when wc->stage == UPDATE_BACKREF, this function updates
8550  * back refs for pointers in the block.
8551  *
8552  * NOTE: return value 1 means we should stop walking down.
8553  */
8554 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8555                                    struct btrfs_root *root,
8556                                    struct btrfs_path *path,
8557                                    struct walk_control *wc, int lookup_info)
8558 {
8559         int level = wc->level;
8560         struct extent_buffer *eb = path->nodes[level];
8561         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8562         int ret;
8563
8564         if (wc->stage == UPDATE_BACKREF &&
8565             btrfs_header_owner(eb) != root->root_key.objectid)
8566                 return 1;
8567
8568         /*
8569          * when reference count of tree block is 1, it won't increase
8570          * again. once full backref flag is set, we never clear it.
8571          */
8572         if (lookup_info &&
8573             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8574              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8575                 BUG_ON(!path->locks[level]);
8576                 ret = btrfs_lookup_extent_info(trans, root,
8577                                                eb->start, level, 1,
8578                                                &wc->refs[level],
8579                                                &wc->flags[level]);
8580                 BUG_ON(ret == -ENOMEM);
8581                 if (ret)
8582                         return ret;
8583                 BUG_ON(wc->refs[level] == 0);
8584         }
8585
8586         if (wc->stage == DROP_REFERENCE) {
8587                 if (wc->refs[level] > 1)
8588                         return 1;
8589
8590                 if (path->locks[level] && !wc->keep_locks) {
8591                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8592                         path->locks[level] = 0;
8593                 }
8594                 return 0;
8595         }
8596
8597         /* wc->stage == UPDATE_BACKREF */
8598         if (!(wc->flags[level] & flag)) {
8599                 BUG_ON(!path->locks[level]);
8600                 ret = btrfs_inc_ref(trans, root, eb, 1);
8601                 BUG_ON(ret); /* -ENOMEM */
8602                 ret = btrfs_dec_ref(trans, root, eb, 0);
8603                 BUG_ON(ret); /* -ENOMEM */
8604                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8605                                                   eb->len, flag,
8606                                                   btrfs_header_level(eb), 0);
8607                 BUG_ON(ret); /* -ENOMEM */
8608                 wc->flags[level] |= flag;
8609         }
8610
8611         /*
8612          * the block is shared by multiple trees, so it's not good to
8613          * keep the tree lock
8614          */
8615         if (path->locks[level] && level > 0) {
8616                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8617                 path->locks[level] = 0;
8618         }
8619         return 0;
8620 }
8621
8622 /*
8623  * helper to process tree block pointer.
8624  *
8625  * when wc->stage == DROP_REFERENCE, this function checks
8626  * reference count of the block pointed to. if the block
8627  * is shared and we need update back refs for the subtree
8628  * rooted at the block, this function changes wc->stage to
8629  * UPDATE_BACKREF. if the block is shared and there is no
8630  * need to update back, this function drops the reference
8631  * to the block.
8632  *
8633  * NOTE: return value 1 means we should stop walking down.
8634  */
8635 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8636                                  struct btrfs_root *root,
8637                                  struct btrfs_path *path,
8638                                  struct walk_control *wc, int *lookup_info)
8639 {
8640         u64 bytenr;
8641         u64 generation;
8642         u64 parent;
8643         u32 blocksize;
8644         struct btrfs_key key;
8645         struct extent_buffer *next;
8646         int level = wc->level;
8647         int reada = 0;
8648         int ret = 0;
8649         bool need_account = false;
8650
8651         generation = btrfs_node_ptr_generation(path->nodes[level],
8652                                                path->slots[level]);
8653         /*
8654          * if the lower level block was created before the snapshot
8655          * was created, we know there is no need to update back refs
8656          * for the subtree
8657          */
8658         if (wc->stage == UPDATE_BACKREF &&
8659             generation <= root->root_key.offset) {
8660                 *lookup_info = 1;
8661                 return 1;
8662         }
8663
8664         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8665         blocksize = root->nodesize;
8666
8667         next = btrfs_find_tree_block(root->fs_info, bytenr);
8668         if (!next) {
8669                 next = btrfs_find_create_tree_block(root, bytenr);
8670                 if (IS_ERR(next))
8671                         return PTR_ERR(next);
8672
8673                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8674                                                level - 1);
8675                 reada = 1;
8676         }
8677         btrfs_tree_lock(next);
8678         btrfs_set_lock_blocking(next);
8679
8680         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8681                                        &wc->refs[level - 1],
8682                                        &wc->flags[level - 1]);
8683         if (ret < 0) {
8684                 btrfs_tree_unlock(next);
8685                 return ret;
8686         }
8687
8688         if (unlikely(wc->refs[level - 1] == 0)) {
8689                 btrfs_err(root->fs_info, "Missing references.");
8690                 BUG();
8691         }
8692         *lookup_info = 0;
8693
8694         if (wc->stage == DROP_REFERENCE) {
8695                 if (wc->refs[level - 1] > 1) {
8696                         need_account = true;
8697                         if (level == 1 &&
8698                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8699                                 goto skip;
8700
8701                         if (!wc->update_ref ||
8702                             generation <= root->root_key.offset)
8703                                 goto skip;
8704
8705                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8706                                               path->slots[level]);
8707                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8708                         if (ret < 0)
8709                                 goto skip;
8710
8711                         wc->stage = UPDATE_BACKREF;
8712                         wc->shared_level = level - 1;
8713                 }
8714         } else {
8715                 if (level == 1 &&
8716                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8717                         goto skip;
8718         }
8719
8720         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8721                 btrfs_tree_unlock(next);
8722                 free_extent_buffer(next);
8723                 next = NULL;
8724                 *lookup_info = 1;
8725         }
8726
8727         if (!next) {
8728                 if (reada && level == 1)
8729                         reada_walk_down(trans, root, wc, path);
8730                 next = read_tree_block(root, bytenr, generation);
8731                 if (IS_ERR(next)) {
8732                         return PTR_ERR(next);
8733                 } else if (!extent_buffer_uptodate(next)) {
8734                         free_extent_buffer(next);
8735                         return -EIO;
8736                 }
8737                 btrfs_tree_lock(next);
8738                 btrfs_set_lock_blocking(next);
8739         }
8740
8741         level--;
8742         BUG_ON(level != btrfs_header_level(next));
8743         path->nodes[level] = next;
8744         path->slots[level] = 0;
8745         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8746         wc->level = level;
8747         if (wc->level == 1)
8748                 wc->reada_slot = 0;
8749         return 0;
8750 skip:
8751         wc->refs[level - 1] = 0;
8752         wc->flags[level - 1] = 0;
8753         if (wc->stage == DROP_REFERENCE) {
8754                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8755                         parent = path->nodes[level]->start;
8756                 } else {
8757                         BUG_ON(root->root_key.objectid !=
8758                                btrfs_header_owner(path->nodes[level]));
8759                         parent = 0;
8760                 }
8761
8762                 if (need_account) {
8763                         ret = account_shared_subtree(trans, root, next,
8764                                                      generation, level - 1);
8765                         if (ret) {
8766                                 btrfs_err_rl(root->fs_info,
8767                                         "Error "
8768                                         "%d accounting shared subtree. Quota "
8769                                         "is out of sync, rescan required.",
8770                                         ret);
8771                         }
8772                 }
8773                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8774                                 root->root_key.objectid, level - 1, 0);
8775                 BUG_ON(ret); /* -ENOMEM */
8776         }
8777         btrfs_tree_unlock(next);
8778         free_extent_buffer(next);
8779         *lookup_info = 1;
8780         return 1;
8781 }
8782
8783 /*
8784  * helper to process tree block while walking up the tree.
8785  *
8786  * when wc->stage == DROP_REFERENCE, this function drops
8787  * reference count on the block.
8788  *
8789  * when wc->stage == UPDATE_BACKREF, this function changes
8790  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8791  * to UPDATE_BACKREF previously while processing the block.
8792  *
8793  * NOTE: return value 1 means we should stop walking up.
8794  */
8795 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8796                                  struct btrfs_root *root,
8797                                  struct btrfs_path *path,
8798                                  struct walk_control *wc)
8799 {
8800         int ret;
8801         int level = wc->level;
8802         struct extent_buffer *eb = path->nodes[level];
8803         u64 parent = 0;
8804
8805         if (wc->stage == UPDATE_BACKREF) {
8806                 BUG_ON(wc->shared_level < level);
8807                 if (level < wc->shared_level)
8808                         goto out;
8809
8810                 ret = find_next_key(path, level + 1, &wc->update_progress);
8811                 if (ret > 0)
8812                         wc->update_ref = 0;
8813
8814                 wc->stage = DROP_REFERENCE;
8815                 wc->shared_level = -1;
8816                 path->slots[level] = 0;
8817
8818                 /*
8819                  * check reference count again if the block isn't locked.
8820                  * we should start walking down the tree again if reference
8821                  * count is one.
8822                  */
8823                 if (!path->locks[level]) {
8824                         BUG_ON(level == 0);
8825                         btrfs_tree_lock(eb);
8826                         btrfs_set_lock_blocking(eb);
8827                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8828
8829                         ret = btrfs_lookup_extent_info(trans, root,
8830                                                        eb->start, level, 1,
8831                                                        &wc->refs[level],
8832                                                        &wc->flags[level]);
8833                         if (ret < 0) {
8834                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8835                                 path->locks[level] = 0;
8836                                 return ret;
8837                         }
8838                         BUG_ON(wc->refs[level] == 0);
8839                         if (wc->refs[level] == 1) {
8840                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8841                                 path->locks[level] = 0;
8842                                 return 1;
8843                         }
8844                 }
8845         }
8846
8847         /* wc->stage == DROP_REFERENCE */
8848         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8849
8850         if (wc->refs[level] == 1) {
8851                 if (level == 0) {
8852                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8853                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8854                         else
8855                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8856                         BUG_ON(ret); /* -ENOMEM */
8857                         ret = account_leaf_items(trans, root, eb);
8858                         if (ret) {
8859                                 btrfs_err_rl(root->fs_info,
8860                                         "error "
8861                                         "%d accounting leaf items. Quota "
8862                                         "is out of sync, rescan required.",
8863                                         ret);
8864                         }
8865                 }
8866                 /* make block locked assertion in clean_tree_block happy */
8867                 if (!path->locks[level] &&
8868                     btrfs_header_generation(eb) == trans->transid) {
8869                         btrfs_tree_lock(eb);
8870                         btrfs_set_lock_blocking(eb);
8871                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8872                 }
8873                 clean_tree_block(trans, root->fs_info, eb);
8874         }
8875
8876         if (eb == root->node) {
8877                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8878                         parent = eb->start;
8879                 else
8880                         BUG_ON(root->root_key.objectid !=
8881                                btrfs_header_owner(eb));
8882         } else {
8883                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8884                         parent = path->nodes[level + 1]->start;
8885                 else
8886                         BUG_ON(root->root_key.objectid !=
8887                                btrfs_header_owner(path->nodes[level + 1]));
8888         }
8889
8890         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8891 out:
8892         wc->refs[level] = 0;
8893         wc->flags[level] = 0;
8894         return 0;
8895 }
8896
8897 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8898                                    struct btrfs_root *root,
8899                                    struct btrfs_path *path,
8900                                    struct walk_control *wc)
8901 {
8902         int level = wc->level;
8903         int lookup_info = 1;
8904         int ret;
8905
8906         while (level >= 0) {
8907                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8908                 if (ret > 0)
8909                         break;
8910
8911                 if (level == 0)
8912                         break;
8913
8914                 if (path->slots[level] >=
8915                     btrfs_header_nritems(path->nodes[level]))
8916                         break;
8917
8918                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8919                 if (ret > 0) {
8920                         path->slots[level]++;
8921                         continue;
8922                 } else if (ret < 0)
8923                         return ret;
8924                 level = wc->level;
8925         }
8926         return 0;
8927 }
8928
8929 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8930                                  struct btrfs_root *root,
8931                                  struct btrfs_path *path,
8932                                  struct walk_control *wc, int max_level)
8933 {
8934         int level = wc->level;
8935         int ret;
8936
8937         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8938         while (level < max_level && path->nodes[level]) {
8939                 wc->level = level;
8940                 if (path->slots[level] + 1 <
8941                     btrfs_header_nritems(path->nodes[level])) {
8942                         path->slots[level]++;
8943                         return 0;
8944                 } else {
8945                         ret = walk_up_proc(trans, root, path, wc);
8946                         if (ret > 0)
8947                                 return 0;
8948
8949                         if (path->locks[level]) {
8950                                 btrfs_tree_unlock_rw(path->nodes[level],
8951                                                      path->locks[level]);
8952                                 path->locks[level] = 0;
8953                         }
8954                         free_extent_buffer(path->nodes[level]);
8955                         path->nodes[level] = NULL;
8956                         level++;
8957                 }
8958         }
8959         return 1;
8960 }
8961
8962 /*
8963  * drop a subvolume tree.
8964  *
8965  * this function traverses the tree freeing any blocks that only
8966  * referenced by the tree.
8967  *
8968  * when a shared tree block is found. this function decreases its
8969  * reference count by one. if update_ref is true, this function
8970  * also make sure backrefs for the shared block and all lower level
8971  * blocks are properly updated.
8972  *
8973  * If called with for_reloc == 0, may exit early with -EAGAIN
8974  */
8975 int btrfs_drop_snapshot(struct btrfs_root *root,
8976                          struct btrfs_block_rsv *block_rsv, int update_ref,
8977                          int for_reloc)
8978 {
8979         struct btrfs_path *path;
8980         struct btrfs_trans_handle *trans;
8981         struct btrfs_root *tree_root = root->fs_info->tree_root;
8982         struct btrfs_root_item *root_item = &root->root_item;
8983         struct walk_control *wc;
8984         struct btrfs_key key;
8985         int err = 0;
8986         int ret;
8987         int level;
8988         bool root_dropped = false;
8989
8990         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8991
8992         path = btrfs_alloc_path();
8993         if (!path) {
8994                 err = -ENOMEM;
8995                 goto out;
8996         }
8997
8998         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8999         if (!wc) {
9000                 btrfs_free_path(path);
9001                 err = -ENOMEM;
9002                 goto out;
9003         }
9004
9005         trans = btrfs_start_transaction(tree_root, 0);
9006         if (IS_ERR(trans)) {
9007                 err = PTR_ERR(trans);
9008                 goto out_free;
9009         }
9010
9011         if (block_rsv)
9012                 trans->block_rsv = block_rsv;
9013
9014         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9015                 level = btrfs_header_level(root->node);
9016                 path->nodes[level] = btrfs_lock_root_node(root);
9017                 btrfs_set_lock_blocking(path->nodes[level]);
9018                 path->slots[level] = 0;
9019                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9020                 memset(&wc->update_progress, 0,
9021                        sizeof(wc->update_progress));
9022         } else {
9023                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9024                 memcpy(&wc->update_progress, &key,
9025                        sizeof(wc->update_progress));
9026
9027                 level = root_item->drop_level;
9028                 BUG_ON(level == 0);
9029                 path->lowest_level = level;
9030                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9031                 path->lowest_level = 0;
9032                 if (ret < 0) {
9033                         err = ret;
9034                         goto out_end_trans;
9035                 }
9036                 WARN_ON(ret > 0);
9037
9038                 /*
9039                  * unlock our path, this is safe because only this
9040                  * function is allowed to delete this snapshot
9041                  */
9042                 btrfs_unlock_up_safe(path, 0);
9043
9044                 level = btrfs_header_level(root->node);
9045                 while (1) {
9046                         btrfs_tree_lock(path->nodes[level]);
9047                         btrfs_set_lock_blocking(path->nodes[level]);
9048                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9049
9050                         ret = btrfs_lookup_extent_info(trans, root,
9051                                                 path->nodes[level]->start,
9052                                                 level, 1, &wc->refs[level],
9053                                                 &wc->flags[level]);
9054                         if (ret < 0) {
9055                                 err = ret;
9056                                 goto out_end_trans;
9057                         }
9058                         BUG_ON(wc->refs[level] == 0);
9059
9060                         if (level == root_item->drop_level)
9061                                 break;
9062
9063                         btrfs_tree_unlock(path->nodes[level]);
9064                         path->locks[level] = 0;
9065                         WARN_ON(wc->refs[level] != 1);
9066                         level--;
9067                 }
9068         }
9069
9070         wc->level = level;
9071         wc->shared_level = -1;
9072         wc->stage = DROP_REFERENCE;
9073         wc->update_ref = update_ref;
9074         wc->keep_locks = 0;
9075         wc->for_reloc = for_reloc;
9076         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9077
9078         while (1) {
9079
9080                 ret = walk_down_tree(trans, root, path, wc);
9081                 if (ret < 0) {
9082                         err = ret;
9083                         break;
9084                 }
9085
9086                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9087                 if (ret < 0) {
9088                         err = ret;
9089                         break;
9090                 }
9091
9092                 if (ret > 0) {
9093                         BUG_ON(wc->stage != DROP_REFERENCE);
9094                         break;
9095                 }
9096
9097                 if (wc->stage == DROP_REFERENCE) {
9098                         level = wc->level;
9099                         btrfs_node_key(path->nodes[level],
9100                                        &root_item->drop_progress,
9101                                        path->slots[level]);
9102                         root_item->drop_level = level;
9103                 }
9104
9105                 BUG_ON(wc->level == 0);
9106                 if (btrfs_should_end_transaction(trans, tree_root) ||
9107                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
9108                         ret = btrfs_update_root(trans, tree_root,
9109                                                 &root->root_key,
9110                                                 root_item);
9111                         if (ret) {
9112                                 btrfs_abort_transaction(trans, tree_root, ret);
9113                                 err = ret;
9114                                 goto out_end_trans;
9115                         }
9116
9117                         btrfs_end_transaction_throttle(trans, tree_root);
9118                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
9119                                 pr_debug("BTRFS: drop snapshot early exit\n");
9120                                 err = -EAGAIN;
9121                                 goto out_free;
9122                         }
9123
9124                         trans = btrfs_start_transaction(tree_root, 0);
9125                         if (IS_ERR(trans)) {
9126                                 err = PTR_ERR(trans);
9127                                 goto out_free;
9128                         }
9129                         if (block_rsv)
9130                                 trans->block_rsv = block_rsv;
9131                 }
9132         }
9133         btrfs_release_path(path);
9134         if (err)
9135                 goto out_end_trans;
9136
9137         ret = btrfs_del_root(trans, tree_root, &root->root_key);
9138         if (ret) {
9139                 btrfs_abort_transaction(trans, tree_root, ret);
9140                 goto out_end_trans;
9141         }
9142
9143         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9144                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9145                                       NULL, NULL);
9146                 if (ret < 0) {
9147                         btrfs_abort_transaction(trans, tree_root, ret);
9148                         err = ret;
9149                         goto out_end_trans;
9150                 } else if (ret > 0) {
9151                         /* if we fail to delete the orphan item this time
9152                          * around, it'll get picked up the next time.
9153                          *
9154                          * The most common failure here is just -ENOENT.
9155                          */
9156                         btrfs_del_orphan_item(trans, tree_root,
9157                                               root->root_key.objectid);
9158                 }
9159         }
9160
9161         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9162                 btrfs_add_dropped_root(trans, root);
9163         } else {
9164                 free_extent_buffer(root->node);
9165                 free_extent_buffer(root->commit_root);
9166                 btrfs_put_fs_root(root);
9167         }
9168         root_dropped = true;
9169 out_end_trans:
9170         btrfs_end_transaction_throttle(trans, tree_root);
9171 out_free:
9172         kfree(wc);
9173         btrfs_free_path(path);
9174 out:
9175         /*
9176          * So if we need to stop dropping the snapshot for whatever reason we
9177          * need to make sure to add it back to the dead root list so that we
9178          * keep trying to do the work later.  This also cleans up roots if we
9179          * don't have it in the radix (like when we recover after a power fail
9180          * or unmount) so we don't leak memory.
9181          */
9182         if (!for_reloc && root_dropped == false)
9183                 btrfs_add_dead_root(root);
9184         if (err && err != -EAGAIN)
9185                 btrfs_handle_fs_error(root->fs_info, err, NULL);
9186         return err;
9187 }
9188
9189 /*
9190  * drop subtree rooted at tree block 'node'.
9191  *
9192  * NOTE: this function will unlock and release tree block 'node'
9193  * only used by relocation code
9194  */
9195 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9196                         struct btrfs_root *root,
9197                         struct extent_buffer *node,
9198                         struct extent_buffer *parent)
9199 {
9200         struct btrfs_path *path;
9201         struct walk_control *wc;
9202         int level;
9203         int parent_level;
9204         int ret = 0;
9205         int wret;
9206
9207         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9208
9209         path = btrfs_alloc_path();
9210         if (!path)
9211                 return -ENOMEM;
9212
9213         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9214         if (!wc) {
9215                 btrfs_free_path(path);
9216                 return -ENOMEM;
9217         }
9218
9219         btrfs_assert_tree_locked(parent);
9220         parent_level = btrfs_header_level(parent);
9221         extent_buffer_get(parent);
9222         path->nodes[parent_level] = parent;
9223         path->slots[parent_level] = btrfs_header_nritems(parent);
9224
9225         btrfs_assert_tree_locked(node);
9226         level = btrfs_header_level(node);
9227         path->nodes[level] = node;
9228         path->slots[level] = 0;
9229         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9230
9231         wc->refs[parent_level] = 1;
9232         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9233         wc->level = level;
9234         wc->shared_level = -1;
9235         wc->stage = DROP_REFERENCE;
9236         wc->update_ref = 0;
9237         wc->keep_locks = 1;
9238         wc->for_reloc = 1;
9239         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9240
9241         while (1) {
9242                 wret = walk_down_tree(trans, root, path, wc);
9243                 if (wret < 0) {
9244                         ret = wret;
9245                         break;
9246                 }
9247
9248                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9249                 if (wret < 0)
9250                         ret = wret;
9251                 if (wret != 0)
9252                         break;
9253         }
9254
9255         kfree(wc);
9256         btrfs_free_path(path);
9257         return ret;
9258 }
9259
9260 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
9261 {
9262         u64 num_devices;
9263         u64 stripped;
9264
9265         /*
9266          * if restripe for this chunk_type is on pick target profile and
9267          * return, otherwise do the usual balance
9268          */
9269         stripped = get_restripe_target(root->fs_info, flags);
9270         if (stripped)
9271                 return extended_to_chunk(stripped);
9272
9273         num_devices = root->fs_info->fs_devices->rw_devices;
9274
9275         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9276                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9277                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9278
9279         if (num_devices == 1) {
9280                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9281                 stripped = flags & ~stripped;
9282
9283                 /* turn raid0 into single device chunks */
9284                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9285                         return stripped;
9286
9287                 /* turn mirroring into duplication */
9288                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9289                              BTRFS_BLOCK_GROUP_RAID10))
9290                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9291         } else {
9292                 /* they already had raid on here, just return */
9293                 if (flags & stripped)
9294                         return flags;
9295
9296                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9297                 stripped = flags & ~stripped;
9298
9299                 /* switch duplicated blocks with raid1 */
9300                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9301                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9302
9303                 /* this is drive concat, leave it alone */
9304         }
9305
9306         return flags;
9307 }
9308
9309 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9310 {
9311         struct btrfs_space_info *sinfo = cache->space_info;
9312         u64 num_bytes;
9313         u64 min_allocable_bytes;
9314         int ret = -ENOSPC;
9315
9316         /*
9317          * We need some metadata space and system metadata space for
9318          * allocating chunks in some corner cases until we force to set
9319          * it to be readonly.
9320          */
9321         if ((sinfo->flags &
9322              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9323             !force)
9324                 min_allocable_bytes = SZ_1M;
9325         else
9326                 min_allocable_bytes = 0;
9327
9328         spin_lock(&sinfo->lock);
9329         spin_lock(&cache->lock);
9330
9331         if (cache->ro) {
9332                 cache->ro++;
9333                 ret = 0;
9334                 goto out;
9335         }
9336
9337         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9338                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9339
9340         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
9341             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
9342             min_allocable_bytes <= sinfo->total_bytes) {
9343                 sinfo->bytes_readonly += num_bytes;
9344                 cache->ro++;
9345                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9346                 ret = 0;
9347         }
9348 out:
9349         spin_unlock(&cache->lock);
9350         spin_unlock(&sinfo->lock);
9351         return ret;
9352 }
9353
9354 int btrfs_inc_block_group_ro(struct btrfs_root *root,
9355                              struct btrfs_block_group_cache *cache)
9356
9357 {
9358         struct btrfs_trans_handle *trans;
9359         u64 alloc_flags;
9360         int ret;
9361
9362 again:
9363         trans = btrfs_join_transaction(root);
9364         if (IS_ERR(trans))
9365                 return PTR_ERR(trans);
9366
9367         /*
9368          * we're not allowed to set block groups readonly after the dirty
9369          * block groups cache has started writing.  If it already started,
9370          * back off and let this transaction commit
9371          */
9372         mutex_lock(&root->fs_info->ro_block_group_mutex);
9373         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9374                 u64 transid = trans->transid;
9375
9376                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
9377                 btrfs_end_transaction(trans, root);
9378
9379                 ret = btrfs_wait_for_commit(root, transid);
9380                 if (ret)
9381                         return ret;
9382                 goto again;
9383         }
9384
9385         /*
9386          * if we are changing raid levels, try to allocate a corresponding
9387          * block group with the new raid level.
9388          */
9389         alloc_flags = update_block_group_flags(root, cache->flags);
9390         if (alloc_flags != cache->flags) {
9391                 ret = do_chunk_alloc(trans, root, alloc_flags,
9392                                      CHUNK_ALLOC_FORCE);
9393                 /*
9394                  * ENOSPC is allowed here, we may have enough space
9395                  * already allocated at the new raid level to
9396                  * carry on
9397                  */
9398                 if (ret == -ENOSPC)
9399                         ret = 0;
9400                 if (ret < 0)
9401                         goto out;
9402         }
9403
9404         ret = inc_block_group_ro(cache, 0);
9405         if (!ret)
9406                 goto out;
9407         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
9408         ret = do_chunk_alloc(trans, root, alloc_flags,
9409                              CHUNK_ALLOC_FORCE);
9410         if (ret < 0)
9411                 goto out;
9412         ret = inc_block_group_ro(cache, 0);
9413 out:
9414         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9415                 alloc_flags = update_block_group_flags(root, cache->flags);
9416                 lock_chunks(root->fs_info->chunk_root);
9417                 check_system_chunk(trans, root, alloc_flags);
9418                 unlock_chunks(root->fs_info->chunk_root);
9419         }
9420         mutex_unlock(&root->fs_info->ro_block_group_mutex);
9421
9422         btrfs_end_transaction(trans, root);
9423         return ret;
9424 }
9425
9426 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9427                             struct btrfs_root *root, u64 type)
9428 {
9429         u64 alloc_flags = get_alloc_profile(root, type);
9430         return do_chunk_alloc(trans, root, alloc_flags,
9431                               CHUNK_ALLOC_FORCE);
9432 }
9433
9434 /*
9435  * helper to account the unused space of all the readonly block group in the
9436  * space_info. takes mirrors into account.
9437  */
9438 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9439 {
9440         struct btrfs_block_group_cache *block_group;
9441         u64 free_bytes = 0;
9442         int factor;
9443
9444         /* It's df, we don't care if it's racy */
9445         if (list_empty(&sinfo->ro_bgs))
9446                 return 0;
9447
9448         spin_lock(&sinfo->lock);
9449         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9450                 spin_lock(&block_group->lock);
9451
9452                 if (!block_group->ro) {
9453                         spin_unlock(&block_group->lock);
9454                         continue;
9455                 }
9456
9457                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9458                                           BTRFS_BLOCK_GROUP_RAID10 |
9459                                           BTRFS_BLOCK_GROUP_DUP))
9460                         factor = 2;
9461                 else
9462                         factor = 1;
9463
9464                 free_bytes += (block_group->key.offset -
9465                                btrfs_block_group_used(&block_group->item)) *
9466                                factor;
9467
9468                 spin_unlock(&block_group->lock);
9469         }
9470         spin_unlock(&sinfo->lock);
9471
9472         return free_bytes;
9473 }
9474
9475 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9476                               struct btrfs_block_group_cache *cache)
9477 {
9478         struct btrfs_space_info *sinfo = cache->space_info;
9479         u64 num_bytes;
9480
9481         BUG_ON(!cache->ro);
9482
9483         spin_lock(&sinfo->lock);
9484         spin_lock(&cache->lock);
9485         if (!--cache->ro) {
9486                 num_bytes = cache->key.offset - cache->reserved -
9487                             cache->pinned - cache->bytes_super -
9488                             btrfs_block_group_used(&cache->item);
9489                 sinfo->bytes_readonly -= num_bytes;
9490                 list_del_init(&cache->ro_list);
9491         }
9492         spin_unlock(&cache->lock);
9493         spin_unlock(&sinfo->lock);
9494 }
9495
9496 /*
9497  * checks to see if its even possible to relocate this block group.
9498  *
9499  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9500  * ok to go ahead and try.
9501  */
9502 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9503 {
9504         struct btrfs_block_group_cache *block_group;
9505         struct btrfs_space_info *space_info;
9506         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9507         struct btrfs_device *device;
9508         struct btrfs_trans_handle *trans;
9509         u64 min_free;
9510         u64 dev_min = 1;
9511         u64 dev_nr = 0;
9512         u64 target;
9513         int debug;
9514         int index;
9515         int full = 0;
9516         int ret = 0;
9517
9518         debug = btrfs_test_opt(root, ENOSPC_DEBUG);
9519
9520         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9521
9522         /* odd, couldn't find the block group, leave it alone */
9523         if (!block_group) {
9524                 if (debug)
9525                         btrfs_warn(root->fs_info,
9526                                    "can't find block group for bytenr %llu",
9527                                    bytenr);
9528                 return -1;
9529         }
9530
9531         min_free = btrfs_block_group_used(&block_group->item);
9532
9533         /* no bytes used, we're good */
9534         if (!min_free)
9535                 goto out;
9536
9537         space_info = block_group->space_info;
9538         spin_lock(&space_info->lock);
9539
9540         full = space_info->full;
9541
9542         /*
9543          * if this is the last block group we have in this space, we can't
9544          * relocate it unless we're able to allocate a new chunk below.
9545          *
9546          * Otherwise, we need to make sure we have room in the space to handle
9547          * all of the extents from this block group.  If we can, we're good
9548          */
9549         if ((space_info->total_bytes != block_group->key.offset) &&
9550             (space_info->bytes_used + space_info->bytes_reserved +
9551              space_info->bytes_pinned + space_info->bytes_readonly +
9552              min_free < space_info->total_bytes)) {
9553                 spin_unlock(&space_info->lock);
9554                 goto out;
9555         }
9556         spin_unlock(&space_info->lock);
9557
9558         /*
9559          * ok we don't have enough space, but maybe we have free space on our
9560          * devices to allocate new chunks for relocation, so loop through our
9561          * alloc devices and guess if we have enough space.  if this block
9562          * group is going to be restriped, run checks against the target
9563          * profile instead of the current one.
9564          */
9565         ret = -1;
9566
9567         /*
9568          * index:
9569          *      0: raid10
9570          *      1: raid1
9571          *      2: dup
9572          *      3: raid0
9573          *      4: single
9574          */
9575         target = get_restripe_target(root->fs_info, block_group->flags);
9576         if (target) {
9577                 index = __get_raid_index(extended_to_chunk(target));
9578         } else {
9579                 /*
9580                  * this is just a balance, so if we were marked as full
9581                  * we know there is no space for a new chunk
9582                  */
9583                 if (full) {
9584                         if (debug)
9585                                 btrfs_warn(root->fs_info,
9586                                         "no space to alloc new chunk for block group %llu",
9587                                         block_group->key.objectid);
9588                         goto out;
9589                 }
9590
9591                 index = get_block_group_index(block_group);
9592         }
9593
9594         if (index == BTRFS_RAID_RAID10) {
9595                 dev_min = 4;
9596                 /* Divide by 2 */
9597                 min_free >>= 1;
9598         } else if (index == BTRFS_RAID_RAID1) {
9599                 dev_min = 2;
9600         } else if (index == BTRFS_RAID_DUP) {
9601                 /* Multiply by 2 */
9602                 min_free <<= 1;
9603         } else if (index == BTRFS_RAID_RAID0) {
9604                 dev_min = fs_devices->rw_devices;
9605                 min_free = div64_u64(min_free, dev_min);
9606         }
9607
9608         /* We need to do this so that we can look at pending chunks */
9609         trans = btrfs_join_transaction(root);
9610         if (IS_ERR(trans)) {
9611                 ret = PTR_ERR(trans);
9612                 goto out;
9613         }
9614
9615         mutex_lock(&root->fs_info->chunk_mutex);
9616         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9617                 u64 dev_offset;
9618
9619                 /*
9620                  * check to make sure we can actually find a chunk with enough
9621                  * space to fit our block group in.
9622                  */
9623                 if (device->total_bytes > device->bytes_used + min_free &&
9624                     !device->is_tgtdev_for_dev_replace) {
9625                         ret = find_free_dev_extent(trans, device, min_free,
9626                                                    &dev_offset, NULL);
9627                         if (!ret)
9628                                 dev_nr++;
9629
9630                         if (dev_nr >= dev_min)
9631                                 break;
9632
9633                         ret = -1;
9634                 }
9635         }
9636         if (debug && ret == -1)
9637                 btrfs_warn(root->fs_info,
9638                         "no space to allocate a new chunk for block group %llu",
9639                         block_group->key.objectid);
9640         mutex_unlock(&root->fs_info->chunk_mutex);
9641         btrfs_end_transaction(trans, root);
9642 out:
9643         btrfs_put_block_group(block_group);
9644         return ret;
9645 }
9646
9647 static int find_first_block_group(struct btrfs_root *root,
9648                 struct btrfs_path *path, struct btrfs_key *key)
9649 {
9650         int ret = 0;
9651         struct btrfs_key found_key;
9652         struct extent_buffer *leaf;
9653         int slot;
9654
9655         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9656         if (ret < 0)
9657                 goto out;
9658
9659         while (1) {
9660                 slot = path->slots[0];
9661                 leaf = path->nodes[0];
9662                 if (slot >= btrfs_header_nritems(leaf)) {
9663                         ret = btrfs_next_leaf(root, path);
9664                         if (ret == 0)
9665                                 continue;
9666                         if (ret < 0)
9667                                 goto out;
9668                         break;
9669                 }
9670                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9671
9672                 if (found_key.objectid >= key->objectid &&
9673                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9674                         ret = 0;
9675                         goto out;
9676                 }
9677                 path->slots[0]++;
9678         }
9679 out:
9680         return ret;
9681 }
9682
9683 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9684 {
9685         struct btrfs_block_group_cache *block_group;
9686         u64 last = 0;
9687
9688         while (1) {
9689                 struct inode *inode;
9690
9691                 block_group = btrfs_lookup_first_block_group(info, last);
9692                 while (block_group) {
9693                         spin_lock(&block_group->lock);
9694                         if (block_group->iref)
9695                                 break;
9696                         spin_unlock(&block_group->lock);
9697                         block_group = next_block_group(info->tree_root,
9698                                                        block_group);
9699                 }
9700                 if (!block_group) {
9701                         if (last == 0)
9702                                 break;
9703                         last = 0;
9704                         continue;
9705                 }
9706
9707                 inode = block_group->inode;
9708                 block_group->iref = 0;
9709                 block_group->inode = NULL;
9710                 spin_unlock(&block_group->lock);
9711                 iput(inode);
9712                 last = block_group->key.objectid + block_group->key.offset;
9713                 btrfs_put_block_group(block_group);
9714         }
9715 }
9716
9717 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9718 {
9719         struct btrfs_block_group_cache *block_group;
9720         struct btrfs_space_info *space_info;
9721         struct btrfs_caching_control *caching_ctl;
9722         struct rb_node *n;
9723
9724         down_write(&info->commit_root_sem);
9725         while (!list_empty(&info->caching_block_groups)) {
9726                 caching_ctl = list_entry(info->caching_block_groups.next,
9727                                          struct btrfs_caching_control, list);
9728                 list_del(&caching_ctl->list);
9729                 put_caching_control(caching_ctl);
9730         }
9731         up_write(&info->commit_root_sem);
9732
9733         spin_lock(&info->unused_bgs_lock);
9734         while (!list_empty(&info->unused_bgs)) {
9735                 block_group = list_first_entry(&info->unused_bgs,
9736                                                struct btrfs_block_group_cache,
9737                                                bg_list);
9738                 list_del_init(&block_group->bg_list);
9739                 btrfs_put_block_group(block_group);
9740         }
9741         spin_unlock(&info->unused_bgs_lock);
9742
9743         spin_lock(&info->block_group_cache_lock);
9744         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9745                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9746                                        cache_node);
9747                 rb_erase(&block_group->cache_node,
9748                          &info->block_group_cache_tree);
9749                 RB_CLEAR_NODE(&block_group->cache_node);
9750                 spin_unlock(&info->block_group_cache_lock);
9751
9752                 down_write(&block_group->space_info->groups_sem);
9753                 list_del(&block_group->list);
9754                 up_write(&block_group->space_info->groups_sem);
9755
9756                 if (block_group->cached == BTRFS_CACHE_STARTED)
9757                         wait_block_group_cache_done(block_group);
9758
9759                 /*
9760                  * We haven't cached this block group, which means we could
9761                  * possibly have excluded extents on this block group.
9762                  */
9763                 if (block_group->cached == BTRFS_CACHE_NO ||
9764                     block_group->cached == BTRFS_CACHE_ERROR)
9765                         free_excluded_extents(info->extent_root, block_group);
9766
9767                 btrfs_remove_free_space_cache(block_group);
9768                 btrfs_put_block_group(block_group);
9769
9770                 spin_lock(&info->block_group_cache_lock);
9771         }
9772         spin_unlock(&info->block_group_cache_lock);
9773
9774         /* now that all the block groups are freed, go through and
9775          * free all the space_info structs.  This is only called during
9776          * the final stages of unmount, and so we know nobody is
9777          * using them.  We call synchronize_rcu() once before we start,
9778          * just to be on the safe side.
9779          */
9780         synchronize_rcu();
9781
9782         release_global_block_rsv(info);
9783
9784         while (!list_empty(&info->space_info)) {
9785                 int i;
9786
9787                 space_info = list_entry(info->space_info.next,
9788                                         struct btrfs_space_info,
9789                                         list);
9790                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9791                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9792                             space_info->bytes_reserved > 0 ||
9793                             space_info->bytes_may_use > 0)) {
9794                                 dump_space_info(space_info, 0, 0);
9795                         }
9796                 }
9797                 list_del(&space_info->list);
9798                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9799                         struct kobject *kobj;
9800                         kobj = space_info->block_group_kobjs[i];
9801                         space_info->block_group_kobjs[i] = NULL;
9802                         if (kobj) {
9803                                 kobject_del(kobj);
9804                                 kobject_put(kobj);
9805                         }
9806                 }
9807                 kobject_del(&space_info->kobj);
9808                 kobject_put(&space_info->kobj);
9809         }
9810         return 0;
9811 }
9812
9813 static void __link_block_group(struct btrfs_space_info *space_info,
9814                                struct btrfs_block_group_cache *cache)
9815 {
9816         int index = get_block_group_index(cache);
9817         bool first = false;
9818
9819         down_write(&space_info->groups_sem);
9820         if (list_empty(&space_info->block_groups[index]))
9821                 first = true;
9822         list_add_tail(&cache->list, &space_info->block_groups[index]);
9823         up_write(&space_info->groups_sem);
9824
9825         if (first) {
9826                 struct raid_kobject *rkobj;
9827                 int ret;
9828
9829                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9830                 if (!rkobj)
9831                         goto out_err;
9832                 rkobj->raid_type = index;
9833                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9834                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9835                                   "%s", get_raid_name(index));
9836                 if (ret) {
9837                         kobject_put(&rkobj->kobj);
9838                         goto out_err;
9839                 }
9840                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9841         }
9842
9843         return;
9844 out_err:
9845         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9846 }
9847
9848 static struct btrfs_block_group_cache *
9849 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9850 {
9851         struct btrfs_block_group_cache *cache;
9852
9853         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9854         if (!cache)
9855                 return NULL;
9856
9857         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9858                                         GFP_NOFS);
9859         if (!cache->free_space_ctl) {
9860                 kfree(cache);
9861                 return NULL;
9862         }
9863
9864         cache->key.objectid = start;
9865         cache->key.offset = size;
9866         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9867
9868         cache->sectorsize = root->sectorsize;
9869         cache->fs_info = root->fs_info;
9870         cache->full_stripe_len = btrfs_full_stripe_len(root,
9871                                                &root->fs_info->mapping_tree,
9872                                                start);
9873         set_free_space_tree_thresholds(cache);
9874
9875         atomic_set(&cache->count, 1);
9876         spin_lock_init(&cache->lock);
9877         init_rwsem(&cache->data_rwsem);
9878         INIT_LIST_HEAD(&cache->list);
9879         INIT_LIST_HEAD(&cache->cluster_list);
9880         INIT_LIST_HEAD(&cache->bg_list);
9881         INIT_LIST_HEAD(&cache->ro_list);
9882         INIT_LIST_HEAD(&cache->dirty_list);
9883         INIT_LIST_HEAD(&cache->io_list);
9884         btrfs_init_free_space_ctl(cache);
9885         atomic_set(&cache->trimming, 0);
9886         mutex_init(&cache->free_space_lock);
9887
9888         return cache;
9889 }
9890
9891 int btrfs_read_block_groups(struct btrfs_root *root)
9892 {
9893         struct btrfs_path *path;
9894         int ret;
9895         struct btrfs_block_group_cache *cache;
9896         struct btrfs_fs_info *info = root->fs_info;
9897         struct btrfs_space_info *space_info;
9898         struct btrfs_key key;
9899         struct btrfs_key found_key;
9900         struct extent_buffer *leaf;
9901         int need_clear = 0;
9902         u64 cache_gen;
9903
9904         root = info->extent_root;
9905         key.objectid = 0;
9906         key.offset = 0;
9907         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9908         path = btrfs_alloc_path();
9909         if (!path)
9910                 return -ENOMEM;
9911         path->reada = READA_FORWARD;
9912
9913         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9914         if (btrfs_test_opt(root, SPACE_CACHE) &&
9915             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9916                 need_clear = 1;
9917         if (btrfs_test_opt(root, CLEAR_CACHE))
9918                 need_clear = 1;
9919
9920         while (1) {
9921                 ret = find_first_block_group(root, path, &key);
9922                 if (ret > 0)
9923                         break;
9924                 if (ret != 0)
9925                         goto error;
9926
9927                 leaf = path->nodes[0];
9928                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9929
9930                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9931                                                        found_key.offset);
9932                 if (!cache) {
9933                         ret = -ENOMEM;
9934                         goto error;
9935                 }
9936
9937                 if (need_clear) {
9938                         /*
9939                          * When we mount with old space cache, we need to
9940                          * set BTRFS_DC_CLEAR and set dirty flag.
9941                          *
9942                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9943                          *    truncate the old free space cache inode and
9944                          *    setup a new one.
9945                          * b) Setting 'dirty flag' makes sure that we flush
9946                          *    the new space cache info onto disk.
9947                          */
9948                         if (btrfs_test_opt(root, SPACE_CACHE))
9949                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9950                 }
9951
9952                 read_extent_buffer(leaf, &cache->item,
9953                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9954                                    sizeof(cache->item));
9955                 cache->flags = btrfs_block_group_flags(&cache->item);
9956
9957                 key.objectid = found_key.objectid + found_key.offset;
9958                 btrfs_release_path(path);
9959
9960                 /*
9961                  * We need to exclude the super stripes now so that the space
9962                  * info has super bytes accounted for, otherwise we'll think
9963                  * we have more space than we actually do.
9964                  */
9965                 ret = exclude_super_stripes(root, cache);
9966                 if (ret) {
9967                         /*
9968                          * We may have excluded something, so call this just in
9969                          * case.
9970                          */
9971                         free_excluded_extents(root, cache);
9972                         btrfs_put_block_group(cache);
9973                         goto error;
9974                 }
9975
9976                 /*
9977                  * check for two cases, either we are full, and therefore
9978                  * don't need to bother with the caching work since we won't
9979                  * find any space, or we are empty, and we can just add all
9980                  * the space in and be done with it.  This saves us _alot_ of
9981                  * time, particularly in the full case.
9982                  */
9983                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9984                         cache->last_byte_to_unpin = (u64)-1;
9985                         cache->cached = BTRFS_CACHE_FINISHED;
9986                         free_excluded_extents(root, cache);
9987                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9988                         cache->last_byte_to_unpin = (u64)-1;
9989                         cache->cached = BTRFS_CACHE_FINISHED;
9990                         add_new_free_space(cache, root->fs_info,
9991                                            found_key.objectid,
9992                                            found_key.objectid +
9993                                            found_key.offset);
9994                         free_excluded_extents(root, cache);
9995                 }
9996
9997                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9998                 if (ret) {
9999                         btrfs_remove_free_space_cache(cache);
10000                         btrfs_put_block_group(cache);
10001                         goto error;
10002                 }
10003
10004                 ret = update_space_info(info, cache->flags, found_key.offset,
10005                                         btrfs_block_group_used(&cache->item),
10006                                         cache->bytes_super, &space_info);
10007                 if (ret) {
10008                         btrfs_remove_free_space_cache(cache);
10009                         spin_lock(&info->block_group_cache_lock);
10010                         rb_erase(&cache->cache_node,
10011                                  &info->block_group_cache_tree);
10012                         RB_CLEAR_NODE(&cache->cache_node);
10013                         spin_unlock(&info->block_group_cache_lock);
10014                         btrfs_put_block_group(cache);
10015                         goto error;
10016                 }
10017
10018                 cache->space_info = space_info;
10019
10020                 __link_block_group(space_info, cache);
10021
10022                 set_avail_alloc_bits(root->fs_info, cache->flags);
10023                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
10024                         inc_block_group_ro(cache, 1);
10025                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10026                         spin_lock(&info->unused_bgs_lock);
10027                         /* Should always be true but just in case. */
10028                         if (list_empty(&cache->bg_list)) {
10029                                 btrfs_get_block_group(cache);
10030                                 list_add_tail(&cache->bg_list,
10031                                               &info->unused_bgs);
10032                         }
10033                         spin_unlock(&info->unused_bgs_lock);
10034                 }
10035         }
10036
10037         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
10038                 if (!(get_alloc_profile(root, space_info->flags) &
10039                       (BTRFS_BLOCK_GROUP_RAID10 |
10040                        BTRFS_BLOCK_GROUP_RAID1 |
10041                        BTRFS_BLOCK_GROUP_RAID5 |
10042                        BTRFS_BLOCK_GROUP_RAID6 |
10043                        BTRFS_BLOCK_GROUP_DUP)))
10044                         continue;
10045                 /*
10046                  * avoid allocating from un-mirrored block group if there are
10047                  * mirrored block groups.
10048                  */
10049                 list_for_each_entry(cache,
10050                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10051                                 list)
10052                         inc_block_group_ro(cache, 1);
10053                 list_for_each_entry(cache,
10054                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10055                                 list)
10056                         inc_block_group_ro(cache, 1);
10057         }
10058
10059         init_global_block_rsv(info);
10060         ret = 0;
10061 error:
10062         btrfs_free_path(path);
10063         return ret;
10064 }
10065
10066 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
10067                                        struct btrfs_root *root)
10068 {
10069         struct btrfs_block_group_cache *block_group, *tmp;
10070         struct btrfs_root *extent_root = root->fs_info->extent_root;
10071         struct btrfs_block_group_item item;
10072         struct btrfs_key key;
10073         int ret = 0;
10074         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10075
10076         trans->can_flush_pending_bgs = false;
10077         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10078                 if (ret)
10079                         goto next;
10080
10081                 spin_lock(&block_group->lock);
10082                 memcpy(&item, &block_group->item, sizeof(item));
10083                 memcpy(&key, &block_group->key, sizeof(key));
10084                 spin_unlock(&block_group->lock);
10085
10086                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10087                                         sizeof(item));
10088                 if (ret)
10089                         btrfs_abort_transaction(trans, extent_root, ret);
10090                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
10091                                                key.objectid, key.offset);
10092                 if (ret)
10093                         btrfs_abort_transaction(trans, extent_root, ret);
10094                 add_block_group_free_space(trans, root->fs_info, block_group);
10095                 /* already aborted the transaction if it failed. */
10096 next:
10097                 list_del_init(&block_group->bg_list);
10098         }
10099         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10100 }
10101
10102 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10103                            struct btrfs_root *root, u64 bytes_used,
10104                            u64 type, u64 chunk_objectid, u64 chunk_offset,
10105                            u64 size)
10106 {
10107         int ret;
10108         struct btrfs_root *extent_root;
10109         struct btrfs_block_group_cache *cache;
10110         extent_root = root->fs_info->extent_root;
10111
10112         btrfs_set_log_full_commit(root->fs_info, trans);
10113
10114         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
10115         if (!cache)
10116                 return -ENOMEM;
10117
10118         btrfs_set_block_group_used(&cache->item, bytes_used);
10119         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
10120         btrfs_set_block_group_flags(&cache->item, type);
10121
10122         cache->flags = type;
10123         cache->last_byte_to_unpin = (u64)-1;
10124         cache->cached = BTRFS_CACHE_FINISHED;
10125         cache->needs_free_space = 1;
10126         ret = exclude_super_stripes(root, cache);
10127         if (ret) {
10128                 /*
10129                  * We may have excluded something, so call this just in
10130                  * case.
10131                  */
10132                 free_excluded_extents(root, cache);
10133                 btrfs_put_block_group(cache);
10134                 return ret;
10135         }
10136
10137         add_new_free_space(cache, root->fs_info, chunk_offset,
10138                            chunk_offset + size);
10139
10140         free_excluded_extents(root, cache);
10141
10142 #ifdef CONFIG_BTRFS_DEBUG
10143         if (btrfs_should_fragment_free_space(root, cache)) {
10144                 u64 new_bytes_used = size - bytes_used;
10145
10146                 bytes_used += new_bytes_used >> 1;
10147                 fragment_free_space(root, cache);
10148         }
10149 #endif
10150         /*
10151          * Call to ensure the corresponding space_info object is created and
10152          * assigned to our block group, but don't update its counters just yet.
10153          * We want our bg to be added to the rbtree with its ->space_info set.
10154          */
10155         ret = update_space_info(root->fs_info, cache->flags, 0, 0, 0,
10156                                 &cache->space_info);
10157         if (ret) {
10158                 btrfs_remove_free_space_cache(cache);
10159                 btrfs_put_block_group(cache);
10160                 return ret;
10161         }
10162
10163         ret = btrfs_add_block_group_cache(root->fs_info, cache);
10164         if (ret) {
10165                 btrfs_remove_free_space_cache(cache);
10166                 btrfs_put_block_group(cache);
10167                 return ret;
10168         }
10169
10170         /*
10171          * Now that our block group has its ->space_info set and is inserted in
10172          * the rbtree, update the space info's counters.
10173          */
10174         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
10175                                 cache->bytes_super, &cache->space_info);
10176         if (ret) {
10177                 btrfs_remove_free_space_cache(cache);
10178                 spin_lock(&root->fs_info->block_group_cache_lock);
10179                 rb_erase(&cache->cache_node,
10180                          &root->fs_info->block_group_cache_tree);
10181                 RB_CLEAR_NODE(&cache->cache_node);
10182                 spin_unlock(&root->fs_info->block_group_cache_lock);
10183                 btrfs_put_block_group(cache);
10184                 return ret;
10185         }
10186         update_global_block_rsv(root->fs_info);
10187
10188         __link_block_group(cache->space_info, cache);
10189
10190         list_add_tail(&cache->bg_list, &trans->new_bgs);
10191
10192         set_avail_alloc_bits(extent_root->fs_info, type);
10193         return 0;
10194 }
10195
10196 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10197 {
10198         u64 extra_flags = chunk_to_extended(flags) &
10199                                 BTRFS_EXTENDED_PROFILE_MASK;
10200
10201         write_seqlock(&fs_info->profiles_lock);
10202         if (flags & BTRFS_BLOCK_GROUP_DATA)
10203                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10204         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10205                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10206         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10207                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10208         write_sequnlock(&fs_info->profiles_lock);
10209 }
10210
10211 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10212                              struct btrfs_root *root, u64 group_start,
10213                              struct extent_map *em)
10214 {
10215         struct btrfs_path *path;
10216         struct btrfs_block_group_cache *block_group;
10217         struct btrfs_free_cluster *cluster;
10218         struct btrfs_root *tree_root = root->fs_info->tree_root;
10219         struct btrfs_key key;
10220         struct inode *inode;
10221         struct kobject *kobj = NULL;
10222         int ret;
10223         int index;
10224         int factor;
10225         struct btrfs_caching_control *caching_ctl = NULL;
10226         bool remove_em;
10227
10228         root = root->fs_info->extent_root;
10229
10230         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
10231         BUG_ON(!block_group);
10232         BUG_ON(!block_group->ro);
10233
10234         /*
10235          * Free the reserved super bytes from this block group before
10236          * remove it.
10237          */
10238         free_excluded_extents(root, block_group);
10239
10240         memcpy(&key, &block_group->key, sizeof(key));
10241         index = get_block_group_index(block_group);
10242         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10243                                   BTRFS_BLOCK_GROUP_RAID1 |
10244                                   BTRFS_BLOCK_GROUP_RAID10))
10245                 factor = 2;
10246         else
10247                 factor = 1;
10248
10249         /* make sure this block group isn't part of an allocation cluster */
10250         cluster = &root->fs_info->data_alloc_cluster;
10251         spin_lock(&cluster->refill_lock);
10252         btrfs_return_cluster_to_free_space(block_group, cluster);
10253         spin_unlock(&cluster->refill_lock);
10254
10255         /*
10256          * make sure this block group isn't part of a metadata
10257          * allocation cluster
10258          */
10259         cluster = &root->fs_info->meta_alloc_cluster;
10260         spin_lock(&cluster->refill_lock);
10261         btrfs_return_cluster_to_free_space(block_group, cluster);
10262         spin_unlock(&cluster->refill_lock);
10263
10264         path = btrfs_alloc_path();
10265         if (!path) {
10266                 ret = -ENOMEM;
10267                 goto out;
10268         }
10269
10270         /*
10271          * get the inode first so any iput calls done for the io_list
10272          * aren't the final iput (no unlinks allowed now)
10273          */
10274         inode = lookup_free_space_inode(tree_root, block_group, path);
10275
10276         mutex_lock(&trans->transaction->cache_write_mutex);
10277         /*
10278          * make sure our free spache cache IO is done before remove the
10279          * free space inode
10280          */
10281         spin_lock(&trans->transaction->dirty_bgs_lock);
10282         if (!list_empty(&block_group->io_list)) {
10283                 list_del_init(&block_group->io_list);
10284
10285                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10286
10287                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10288                 btrfs_wait_cache_io(root, trans, block_group,
10289                                     &block_group->io_ctl, path,
10290                                     block_group->key.objectid);
10291                 btrfs_put_block_group(block_group);
10292                 spin_lock(&trans->transaction->dirty_bgs_lock);
10293         }
10294
10295         if (!list_empty(&block_group->dirty_list)) {
10296                 list_del_init(&block_group->dirty_list);
10297                 btrfs_put_block_group(block_group);
10298         }
10299         spin_unlock(&trans->transaction->dirty_bgs_lock);
10300         mutex_unlock(&trans->transaction->cache_write_mutex);
10301
10302         if (!IS_ERR(inode)) {
10303                 ret = btrfs_orphan_add(trans, inode);
10304                 if (ret) {
10305                         btrfs_add_delayed_iput(inode);
10306                         goto out;
10307                 }
10308                 clear_nlink(inode);
10309                 /* One for the block groups ref */
10310                 spin_lock(&block_group->lock);
10311                 if (block_group->iref) {
10312                         block_group->iref = 0;
10313                         block_group->inode = NULL;
10314                         spin_unlock(&block_group->lock);
10315                         iput(inode);
10316                 } else {
10317                         spin_unlock(&block_group->lock);
10318                 }
10319                 /* One for our lookup ref */
10320                 btrfs_add_delayed_iput(inode);
10321         }
10322
10323         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10324         key.offset = block_group->key.objectid;
10325         key.type = 0;
10326
10327         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10328         if (ret < 0)
10329                 goto out;
10330         if (ret > 0)
10331                 btrfs_release_path(path);
10332         if (ret == 0) {
10333                 ret = btrfs_del_item(trans, tree_root, path);
10334                 if (ret)
10335                         goto out;
10336                 btrfs_release_path(path);
10337         }
10338
10339         spin_lock(&root->fs_info->block_group_cache_lock);
10340         rb_erase(&block_group->cache_node,
10341                  &root->fs_info->block_group_cache_tree);
10342         RB_CLEAR_NODE(&block_group->cache_node);
10343
10344         if (root->fs_info->first_logical_byte == block_group->key.objectid)
10345                 root->fs_info->first_logical_byte = (u64)-1;
10346         spin_unlock(&root->fs_info->block_group_cache_lock);
10347
10348         down_write(&block_group->space_info->groups_sem);
10349         /*
10350          * we must use list_del_init so people can check to see if they
10351          * are still on the list after taking the semaphore
10352          */
10353         list_del_init(&block_group->list);
10354         if (list_empty(&block_group->space_info->block_groups[index])) {
10355                 kobj = block_group->space_info->block_group_kobjs[index];
10356                 block_group->space_info->block_group_kobjs[index] = NULL;
10357                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
10358         }
10359         up_write(&block_group->space_info->groups_sem);
10360         if (kobj) {
10361                 kobject_del(kobj);
10362                 kobject_put(kobj);
10363         }
10364
10365         if (block_group->has_caching_ctl)
10366                 caching_ctl = get_caching_control(block_group);
10367         if (block_group->cached == BTRFS_CACHE_STARTED)
10368                 wait_block_group_cache_done(block_group);
10369         if (block_group->has_caching_ctl) {
10370                 down_write(&root->fs_info->commit_root_sem);
10371                 if (!caching_ctl) {
10372                         struct btrfs_caching_control *ctl;
10373
10374                         list_for_each_entry(ctl,
10375                                     &root->fs_info->caching_block_groups, list)
10376                                 if (ctl->block_group == block_group) {
10377                                         caching_ctl = ctl;
10378                                         atomic_inc(&caching_ctl->count);
10379                                         break;
10380                                 }
10381                 }
10382                 if (caching_ctl)
10383                         list_del_init(&caching_ctl->list);
10384                 up_write(&root->fs_info->commit_root_sem);
10385                 if (caching_ctl) {
10386                         /* Once for the caching bgs list and once for us. */
10387                         put_caching_control(caching_ctl);
10388                         put_caching_control(caching_ctl);
10389                 }
10390         }
10391
10392         spin_lock(&trans->transaction->dirty_bgs_lock);
10393         if (!list_empty(&block_group->dirty_list)) {
10394                 WARN_ON(1);
10395         }
10396         if (!list_empty(&block_group->io_list)) {
10397                 WARN_ON(1);
10398         }
10399         spin_unlock(&trans->transaction->dirty_bgs_lock);
10400         btrfs_remove_free_space_cache(block_group);
10401
10402         spin_lock(&block_group->space_info->lock);
10403         list_del_init(&block_group->ro_list);
10404
10405         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
10406                 WARN_ON(block_group->space_info->total_bytes
10407                         < block_group->key.offset);
10408                 WARN_ON(block_group->space_info->bytes_readonly
10409                         < block_group->key.offset);
10410                 WARN_ON(block_group->space_info->disk_total
10411                         < block_group->key.offset * factor);
10412         }
10413         block_group->space_info->total_bytes -= block_group->key.offset;
10414         block_group->space_info->bytes_readonly -= block_group->key.offset;
10415         block_group->space_info->disk_total -= block_group->key.offset * factor;
10416
10417         spin_unlock(&block_group->space_info->lock);
10418
10419         memcpy(&key, &block_group->key, sizeof(key));
10420
10421         lock_chunks(root);
10422         if (!list_empty(&em->list)) {
10423                 /* We're in the transaction->pending_chunks list. */
10424                 free_extent_map(em);
10425         }
10426         spin_lock(&block_group->lock);
10427         block_group->removed = 1;
10428         /*
10429          * At this point trimming can't start on this block group, because we
10430          * removed the block group from the tree fs_info->block_group_cache_tree
10431          * so no one can't find it anymore and even if someone already got this
10432          * block group before we removed it from the rbtree, they have already
10433          * incremented block_group->trimming - if they didn't, they won't find
10434          * any free space entries because we already removed them all when we
10435          * called btrfs_remove_free_space_cache().
10436          *
10437          * And we must not remove the extent map from the fs_info->mapping_tree
10438          * to prevent the same logical address range and physical device space
10439          * ranges from being reused for a new block group. This is because our
10440          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10441          * completely transactionless, so while it is trimming a range the
10442          * currently running transaction might finish and a new one start,
10443          * allowing for new block groups to be created that can reuse the same
10444          * physical device locations unless we take this special care.
10445          *
10446          * There may also be an implicit trim operation if the file system
10447          * is mounted with -odiscard. The same protections must remain
10448          * in place until the extents have been discarded completely when
10449          * the transaction commit has completed.
10450          */
10451         remove_em = (atomic_read(&block_group->trimming) == 0);
10452         /*
10453          * Make sure a trimmer task always sees the em in the pinned_chunks list
10454          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10455          * before checking block_group->removed).
10456          */
10457         if (!remove_em) {
10458                 /*
10459                  * Our em might be in trans->transaction->pending_chunks which
10460                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10461                  * and so is the fs_info->pinned_chunks list.
10462                  *
10463                  * So at this point we must be holding the chunk_mutex to avoid
10464                  * any races with chunk allocation (more specifically at
10465                  * volumes.c:contains_pending_extent()), to ensure it always
10466                  * sees the em, either in the pending_chunks list or in the
10467                  * pinned_chunks list.
10468                  */
10469                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10470         }
10471         spin_unlock(&block_group->lock);
10472
10473         if (remove_em) {
10474                 struct extent_map_tree *em_tree;
10475
10476                 em_tree = &root->fs_info->mapping_tree.map_tree;
10477                 write_lock(&em_tree->lock);
10478                 /*
10479                  * The em might be in the pending_chunks list, so make sure the
10480                  * chunk mutex is locked, since remove_extent_mapping() will
10481                  * delete us from that list.
10482                  */
10483                 remove_extent_mapping(em_tree, em);
10484                 write_unlock(&em_tree->lock);
10485                 /* once for the tree */
10486                 free_extent_map(em);
10487         }
10488
10489         unlock_chunks(root);
10490
10491         ret = remove_block_group_free_space(trans, root->fs_info, block_group);
10492         if (ret)
10493                 goto out;
10494
10495         btrfs_put_block_group(block_group);
10496         btrfs_put_block_group(block_group);
10497
10498         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10499         if (ret > 0)
10500                 ret = -EIO;
10501         if (ret < 0)
10502                 goto out;
10503
10504         ret = btrfs_del_item(trans, root, path);
10505 out:
10506         btrfs_free_path(path);
10507         return ret;
10508 }
10509
10510 struct btrfs_trans_handle *
10511 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10512                                      const u64 chunk_offset)
10513 {
10514         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10515         struct extent_map *em;
10516         struct map_lookup *map;
10517         unsigned int num_items;
10518
10519         read_lock(&em_tree->lock);
10520         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10521         read_unlock(&em_tree->lock);
10522         ASSERT(em && em->start == chunk_offset);
10523
10524         /*
10525          * We need to reserve 3 + N units from the metadata space info in order
10526          * to remove a block group (done at btrfs_remove_chunk() and at
10527          * btrfs_remove_block_group()), which are used for:
10528          *
10529          * 1 unit for adding the free space inode's orphan (located in the tree
10530          * of tree roots).
10531          * 1 unit for deleting the block group item (located in the extent
10532          * tree).
10533          * 1 unit for deleting the free space item (located in tree of tree
10534          * roots).
10535          * N units for deleting N device extent items corresponding to each
10536          * stripe (located in the device tree).
10537          *
10538          * In order to remove a block group we also need to reserve units in the
10539          * system space info in order to update the chunk tree (update one or
10540          * more device items and remove one chunk item), but this is done at
10541          * btrfs_remove_chunk() through a call to check_system_chunk().
10542          */
10543         map = em->map_lookup;
10544         num_items = 3 + map->num_stripes;
10545         free_extent_map(em);
10546
10547         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10548                                                            num_items, 1);
10549 }
10550
10551 /*
10552  * Process the unused_bgs list and remove any that don't have any allocated
10553  * space inside of them.
10554  */
10555 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10556 {
10557         struct btrfs_block_group_cache *block_group;
10558         struct btrfs_space_info *space_info;
10559         struct btrfs_root *root = fs_info->extent_root;
10560         struct btrfs_trans_handle *trans;
10561         int ret = 0;
10562
10563         if (!fs_info->open)
10564                 return;
10565
10566         spin_lock(&fs_info->unused_bgs_lock);
10567         while (!list_empty(&fs_info->unused_bgs)) {
10568                 u64 start, end;
10569                 int trimming;
10570
10571                 block_group = list_first_entry(&fs_info->unused_bgs,
10572                                                struct btrfs_block_group_cache,
10573                                                bg_list);
10574                 list_del_init(&block_group->bg_list);
10575
10576                 space_info = block_group->space_info;
10577
10578                 if (ret || btrfs_mixed_space_info(space_info)) {
10579                         btrfs_put_block_group(block_group);
10580                         continue;
10581                 }
10582                 spin_unlock(&fs_info->unused_bgs_lock);
10583
10584                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10585
10586                 /* Don't want to race with allocators so take the groups_sem */
10587                 down_write(&space_info->groups_sem);
10588                 spin_lock(&block_group->lock);
10589                 if (block_group->reserved ||
10590                     btrfs_block_group_used(&block_group->item) ||
10591                     block_group->ro ||
10592                     list_is_singular(&block_group->list)) {
10593                         /*
10594                          * We want to bail if we made new allocations or have
10595                          * outstanding allocations in this block group.  We do
10596                          * the ro check in case balance is currently acting on
10597                          * this block group.
10598                          */
10599                         spin_unlock(&block_group->lock);
10600                         up_write(&space_info->groups_sem);
10601                         goto next;
10602                 }
10603                 spin_unlock(&block_group->lock);
10604
10605                 /* We don't want to force the issue, only flip if it's ok. */
10606                 ret = inc_block_group_ro(block_group, 0);
10607                 up_write(&space_info->groups_sem);
10608                 if (ret < 0) {
10609                         ret = 0;
10610                         goto next;
10611                 }
10612
10613                 /*
10614                  * Want to do this before we do anything else so we can recover
10615                  * properly if we fail to join the transaction.
10616                  */
10617                 trans = btrfs_start_trans_remove_block_group(fs_info,
10618                                                      block_group->key.objectid);
10619                 if (IS_ERR(trans)) {
10620                         btrfs_dec_block_group_ro(root, block_group);
10621                         ret = PTR_ERR(trans);
10622                         goto next;
10623                 }
10624
10625                 /*
10626                  * We could have pending pinned extents for this block group,
10627                  * just delete them, we don't care about them anymore.
10628                  */
10629                 start = block_group->key.objectid;
10630                 end = start + block_group->key.offset - 1;
10631                 /*
10632                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10633                  * btrfs_finish_extent_commit(). If we are at transaction N,
10634                  * another task might be running finish_extent_commit() for the
10635                  * previous transaction N - 1, and have seen a range belonging
10636                  * to the block group in freed_extents[] before we were able to
10637                  * clear the whole block group range from freed_extents[]. This
10638                  * means that task can lookup for the block group after we
10639                  * unpinned it from freed_extents[] and removed it, leading to
10640                  * a BUG_ON() at btrfs_unpin_extent_range().
10641                  */
10642                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10643                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10644                                   EXTENT_DIRTY);
10645                 if (ret) {
10646                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10647                         btrfs_dec_block_group_ro(root, block_group);
10648                         goto end_trans;
10649                 }
10650                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10651                                   EXTENT_DIRTY);
10652                 if (ret) {
10653                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10654                         btrfs_dec_block_group_ro(root, block_group);
10655                         goto end_trans;
10656                 }
10657                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10658
10659                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10660                 spin_lock(&space_info->lock);
10661                 spin_lock(&block_group->lock);
10662
10663                 space_info->bytes_pinned -= block_group->pinned;
10664                 space_info->bytes_readonly += block_group->pinned;
10665                 percpu_counter_add(&space_info->total_bytes_pinned,
10666                                    -block_group->pinned);
10667                 block_group->pinned = 0;
10668
10669                 spin_unlock(&block_group->lock);
10670                 spin_unlock(&space_info->lock);
10671
10672                 /* DISCARD can flip during remount */
10673                 trimming = btrfs_test_opt(root, DISCARD);
10674
10675                 /* Implicit trim during transaction commit. */
10676                 if (trimming)
10677                         btrfs_get_block_group_trimming(block_group);
10678
10679                 /*
10680                  * Btrfs_remove_chunk will abort the transaction if things go
10681                  * horribly wrong.
10682                  */
10683                 ret = btrfs_remove_chunk(trans, root,
10684                                          block_group->key.objectid);
10685
10686                 if (ret) {
10687                         if (trimming)
10688                                 btrfs_put_block_group_trimming(block_group);
10689                         goto end_trans;
10690                 }
10691
10692                 /*
10693                  * If we're not mounted with -odiscard, we can just forget
10694                  * about this block group. Otherwise we'll need to wait
10695                  * until transaction commit to do the actual discard.
10696                  */
10697                 if (trimming) {
10698                         spin_lock(&fs_info->unused_bgs_lock);
10699                         /*
10700                          * A concurrent scrub might have added us to the list
10701                          * fs_info->unused_bgs, so use a list_move operation
10702                          * to add the block group to the deleted_bgs list.
10703                          */
10704                         list_move(&block_group->bg_list,
10705                                   &trans->transaction->deleted_bgs);
10706                         spin_unlock(&fs_info->unused_bgs_lock);
10707                         btrfs_get_block_group(block_group);
10708                 }
10709 end_trans:
10710                 btrfs_end_transaction(trans, root);
10711 next:
10712                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10713                 btrfs_put_block_group(block_group);
10714                 spin_lock(&fs_info->unused_bgs_lock);
10715         }
10716         spin_unlock(&fs_info->unused_bgs_lock);
10717 }
10718
10719 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10720 {
10721         struct btrfs_space_info *space_info;
10722         struct btrfs_super_block *disk_super;
10723         u64 features;
10724         u64 flags;
10725         int mixed = 0;
10726         int ret;
10727
10728         disk_super = fs_info->super_copy;
10729         if (!btrfs_super_root(disk_super))
10730                 return -EINVAL;
10731
10732         features = btrfs_super_incompat_flags(disk_super);
10733         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10734                 mixed = 1;
10735
10736         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10737         ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10738         if (ret)
10739                 goto out;
10740
10741         if (mixed) {
10742                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10743                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10744         } else {
10745                 flags = BTRFS_BLOCK_GROUP_METADATA;
10746                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10747                 if (ret)
10748                         goto out;
10749
10750                 flags = BTRFS_BLOCK_GROUP_DATA;
10751                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10752         }
10753 out:
10754         return ret;
10755 }
10756
10757 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10758 {
10759         return unpin_extent_range(root, start, end, false);
10760 }
10761
10762 /*
10763  * It used to be that old block groups would be left around forever.
10764  * Iterating over them would be enough to trim unused space.  Since we
10765  * now automatically remove them, we also need to iterate over unallocated
10766  * space.
10767  *
10768  * We don't want a transaction for this since the discard may take a
10769  * substantial amount of time.  We don't require that a transaction be
10770  * running, but we do need to take a running transaction into account
10771  * to ensure that we're not discarding chunks that were released in
10772  * the current transaction.
10773  *
10774  * Holding the chunks lock will prevent other threads from allocating
10775  * or releasing chunks, but it won't prevent a running transaction
10776  * from committing and releasing the memory that the pending chunks
10777  * list head uses.  For that, we need to take a reference to the
10778  * transaction.
10779  */
10780 static int btrfs_trim_free_extents(struct btrfs_device *device,
10781                                    u64 minlen, u64 *trimmed)
10782 {
10783         u64 start = 0, len = 0;
10784         int ret;
10785
10786         *trimmed = 0;
10787
10788         /* Not writeable = nothing to do. */
10789         if (!device->writeable)
10790                 return 0;
10791
10792         /* No free space = nothing to do. */
10793         if (device->total_bytes <= device->bytes_used)
10794                 return 0;
10795
10796         ret = 0;
10797
10798         while (1) {
10799                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
10800                 struct btrfs_transaction *trans;
10801                 u64 bytes;
10802
10803                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10804                 if (ret)
10805                         return ret;
10806
10807                 down_read(&fs_info->commit_root_sem);
10808
10809                 spin_lock(&fs_info->trans_lock);
10810                 trans = fs_info->running_transaction;
10811                 if (trans)
10812                         atomic_inc(&trans->use_count);
10813                 spin_unlock(&fs_info->trans_lock);
10814
10815                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10816                                                  &start, &len);
10817                 if (trans)
10818                         btrfs_put_transaction(trans);
10819
10820                 if (ret) {
10821                         up_read(&fs_info->commit_root_sem);
10822                         mutex_unlock(&fs_info->chunk_mutex);
10823                         if (ret == -ENOSPC)
10824                                 ret = 0;
10825                         break;
10826                 }
10827
10828                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10829                 up_read(&fs_info->commit_root_sem);
10830                 mutex_unlock(&fs_info->chunk_mutex);
10831
10832                 if (ret)
10833                         break;
10834
10835                 start += len;
10836                 *trimmed += bytes;
10837
10838                 if (fatal_signal_pending(current)) {
10839                         ret = -ERESTARTSYS;
10840                         break;
10841                 }
10842
10843                 cond_resched();
10844         }
10845
10846         return ret;
10847 }
10848
10849 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10850 {
10851         struct btrfs_fs_info *fs_info = root->fs_info;
10852         struct btrfs_block_group_cache *cache = NULL;
10853         struct btrfs_device *device;
10854         struct list_head *devices;
10855         u64 group_trimmed;
10856         u64 start;
10857         u64 end;
10858         u64 trimmed = 0;
10859         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10860         int ret = 0;
10861
10862         /*
10863          * try to trim all FS space, our block group may start from non-zero.
10864          */
10865         if (range->len == total_bytes)
10866                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10867         else
10868                 cache = btrfs_lookup_block_group(fs_info, range->start);
10869
10870         while (cache) {
10871                 if (cache->key.objectid >= (range->start + range->len)) {
10872                         btrfs_put_block_group(cache);
10873                         break;
10874                 }
10875
10876                 start = max(range->start, cache->key.objectid);
10877                 end = min(range->start + range->len,
10878                                 cache->key.objectid + cache->key.offset);
10879
10880                 if (end - start >= range->minlen) {
10881                         if (!block_group_cache_done(cache)) {
10882                                 ret = cache_block_group(cache, 0);
10883                                 if (ret) {
10884                                         btrfs_put_block_group(cache);
10885                                         break;
10886                                 }
10887                                 ret = wait_block_group_cache_done(cache);
10888                                 if (ret) {
10889                                         btrfs_put_block_group(cache);
10890                                         break;
10891                                 }
10892                         }
10893                         ret = btrfs_trim_block_group(cache,
10894                                                      &group_trimmed,
10895                                                      start,
10896                                                      end,
10897                                                      range->minlen);
10898
10899                         trimmed += group_trimmed;
10900                         if (ret) {
10901                                 btrfs_put_block_group(cache);
10902                                 break;
10903                         }
10904                 }
10905
10906                 cache = next_block_group(fs_info->tree_root, cache);
10907         }
10908
10909         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
10910         devices = &root->fs_info->fs_devices->alloc_list;
10911         list_for_each_entry(device, devices, dev_alloc_list) {
10912                 ret = btrfs_trim_free_extents(device, range->minlen,
10913                                               &group_trimmed);
10914                 if (ret)
10915                         break;
10916
10917                 trimmed += group_trimmed;
10918         }
10919         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
10920
10921         range->len = trimmed;
10922         return ret;
10923 }
10924
10925 /*
10926  * btrfs_{start,end}_write_no_snapshoting() are similar to
10927  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10928  * data into the page cache through nocow before the subvolume is snapshoted,
10929  * but flush the data into disk after the snapshot creation, or to prevent
10930  * operations while snapshoting is ongoing and that cause the snapshot to be
10931  * inconsistent (writes followed by expanding truncates for example).
10932  */
10933 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10934 {
10935         percpu_counter_dec(&root->subv_writers->counter);
10936         /*
10937          * Make sure counter is updated before we wake up waiters.
10938          */
10939         smp_mb();
10940         if (waitqueue_active(&root->subv_writers->wait))
10941                 wake_up(&root->subv_writers->wait);
10942 }
10943
10944 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10945 {
10946         if (atomic_read(&root->will_be_snapshoted))
10947                 return 0;
10948
10949         percpu_counter_inc(&root->subv_writers->counter);
10950         /*
10951          * Make sure counter is updated before we check for snapshot creation.
10952          */
10953         smp_mb();
10954         if (atomic_read(&root->will_be_snapshoted)) {
10955                 btrfs_end_write_no_snapshoting(root);
10956                 return 0;
10957         }
10958         return 1;
10959 }
10960
10961 static int wait_snapshoting_atomic_t(atomic_t *a)
10962 {
10963         schedule();
10964         return 0;
10965 }
10966
10967 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
10968 {
10969         while (true) {
10970                 int ret;
10971
10972                 ret = btrfs_start_write_no_snapshoting(root);
10973                 if (ret)
10974                         break;
10975                 wait_on_atomic_t(&root->will_be_snapshoted,
10976                                  wait_snapshoting_atomic_t,
10977                                  TASK_UNINTERRUPTIBLE);
10978         }
10979 }