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 static int update_block_group(struct btrfs_trans_handle *trans,
  64                               struct btrfs_root *root, u64 bytenr,
  65                               u64 num_bytes, int alloc);
  66 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  67                                 struct btrfs_root *root,
  68                                 struct btrfs_delayed_ref_node *node, u64 parent,
  69                                 u64 root_objectid, u64 owner_objectid,
  70                                 u64 owner_offset, int refs_to_drop,
  71                                 struct btrfs_delayed_extent_op *extra_op);
  72 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  73                                     struct extent_buffer *leaf,
  74                                     struct btrfs_extent_item *ei);
  75 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  76                                       struct btrfs_root *root,
  77                                       u64 parent, u64 root_objectid,
  78                                       u64 flags, u64 owner, u64 offset,
  79                                       struct btrfs_key *ins, int ref_mod);
  80 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  81                                      struct btrfs_root *root,
  82                                      u64 parent, u64 root_objectid,
  83                                      u64 flags, struct btrfs_disk_key *key,
  84                                      int level, struct btrfs_key *ins);
  85 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  86                           struct btrfs_root *extent_root, u64 flags,
  87                           int force);
  88 static int find_next_key(struct btrfs_path *path, int level,
  89                          struct btrfs_key *key);
  90 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
  91                             int dump_block_groups);
  92 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
  93                                     u64 ram_bytes, u64 num_bytes, int delalloc);
  94 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
  95                                      u64 num_bytes, int delalloc);
  96 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  97                                u64 num_bytes);
  98 int btrfs_pin_extent(struct btrfs_root *root,
  99                      u64 bytenr, u64 num_bytes, int reserved);
 100 static int __reserve_metadata_bytes(struct btrfs_root *root,
 101                                     struct btrfs_space_info *space_info,
 102                                     u64 orig_bytes,
 103                                     enum btrfs_reserve_flush_enum flush);
 104 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
 105                                      struct btrfs_space_info *space_info,
 106                                      u64 num_bytes);
 107 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 108                                      struct btrfs_space_info *space_info,
 109                                      u64 num_bytes);
 110
 111 static noinline int
 112 block_group_cache_done(struct btrfs_block_group_cache *cache)
 113 {
 114         smp_mb();
 115         return cache->cached == BTRFS_CACHE_FINISHED ||
 116                 cache->cached == BTRFS_CACHE_ERROR;
 117 }
 118
 119 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 120 {
 121         return (cache->flags & bits) == bits;
 122 }
 123
 124 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 125 {
 126         atomic_inc(&cache->count);
 127 }
 128
 129 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 130 {
 131         if (atomic_dec_and_test(&cache->count)) {
 132                 WARN_ON(cache->pinned > 0);
 133                 WARN_ON(cache->reserved > 0);
 134                 kfree(cache->free_space_ctl);
 135                 kfree(cache);
 136         }
 137 }
 138
 139 /*
 140  * this adds the block group to the fs_info rb tree for the block group
 141  * cache
 142  */
 143 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 144                                 struct btrfs_block_group_cache *block_group)
 145 {
 146         struct rb_node **p;
 147         struct rb_node *parent = NULL;
 148         struct btrfs_block_group_cache *cache;
 149
 150         spin_lock(&info->block_group_cache_lock);
 151         p = &info->block_group_cache_tree.rb_node;
 152
 153         while (*p) {
 154                 parent = *p;
 155                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 156                                  cache_node);
 157                 if (block_group->key.objectid < cache->key.objectid) {
 158                         p = &(*p)->rb_left;
 159                 } else if (block_group->key.objectid > cache->key.objectid) {
 160                         p = &(*p)->rb_right;
 161                 } else {
 162                         spin_unlock(&info->block_group_cache_lock);
 163                         return -EEXIST;
 164                 }
 165         }
 166
 167         rb_link_node(&block_group->cache_node, parent, p);
 168         rb_insert_color(&block_group->cache_node,
 169                         &info->block_group_cache_tree);
 170
 171         if (info->first_logical_byte > block_group->key.objectid)
 172                 info->first_logical_byte = block_group->key.objectid;
 173
 174         spin_unlock(&info->block_group_cache_lock);
 175
 176         return 0;
 177 }
 178
 179 /*
 180  * This will return the block group at or after bytenr if contains is 0, else
 181  * it will return the block group that contains the bytenr
 182  */
 183 static struct btrfs_block_group_cache *
 184 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 185                               int contains)
 186 {
 187         struct btrfs_block_group_cache *cache, *ret = NULL;
 188         struct rb_node *n;
 189         u64 end, start;
 190
 191         spin_lock(&info->block_group_cache_lock);
 192         n = info->block_group_cache_tree.rb_node;
 193
 194         while (n) {
 195                 cache = rb_entry(n, struct btrfs_block_group_cache,
 196                                  cache_node);
 197                 end = cache->key.objectid + cache->key.offset - 1;
 198                 start = cache->key.objectid;
 199
 200                 if (bytenr < start) {
 201                         if (!contains && (!ret || start < ret->key.objectid))
 202                                 ret = cache;
 203                         n = n->rb_left;
 204                 } else if (bytenr > start) {
 205                         if (contains && bytenr <= end) {
 206                                 ret = cache;
 207                                 break;
 208                         }
 209                         n = n->rb_right;
 210                 } else {
 211                         ret = cache;
 212                         break;
 213                 }
 214         }
 215         if (ret) {
 216                 btrfs_get_block_group(ret);
 217                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 218                         info->first_logical_byte = ret->key.objectid;
 219         }
 220         spin_unlock(&info->block_group_cache_lock);
 221
 222         return ret;
 223 }
 224
 225 static int add_excluded_extent(struct btrfs_root *root,
 226                                u64 start, u64 num_bytes)
 227 {
 228         u64 end = start + num_bytes - 1;
 229         set_extent_bits(&root->fs_info->freed_extents[0],
 230                         start, end, EXTENT_UPTODATE);
 231         set_extent_bits(&root->fs_info->freed_extents[1],
 232                         start, end, EXTENT_UPTODATE);
 233         return 0;
 234 }
 235
 236 static void free_excluded_extents(struct btrfs_root *root,
 237                                   struct btrfs_block_group_cache *cache)
 238 {
 239         u64 start, end;
 240
 241         start = cache->key.objectid;
 242         end = start + cache->key.offset - 1;
 243
 244         clear_extent_bits(&root->fs_info->freed_extents[0],
 245                           start, end, EXTENT_UPTODATE);
 246         clear_extent_bits(&root->fs_info->freed_extents[1],
 247                           start, end, EXTENT_UPTODATE);
 248 }
 249
 250 static int exclude_super_stripes(struct btrfs_root *root,
 251                                  struct btrfs_block_group_cache *cache)
 252 {
 253         u64 bytenr;
 254         u64 *logical;
 255         int stripe_len;
 256         int i, nr, ret;
 257
 258         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 259                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 260                 cache->bytes_super += stripe_len;
 261                 ret = add_excluded_extent(root, cache->key.objectid,
 262                                           stripe_len);
 263                 if (ret)
 264                         return ret;
 265         }
 266
 267         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 268                 bytenr = btrfs_sb_offset(i);
 269                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 270                                        cache->key.objectid, bytenr,
 271                                        0, &logical, &nr, &stripe_len);
 272                 if (ret)
 273                         return ret;
 274
 275                 while (nr--) {
 276                         u64 start, len;
 277
 278                         if (logical[nr] > cache->key.objectid +
 279                             cache->key.offset)
 280                                 continue;
 281
 282                         if (logical[nr] + stripe_len <= cache->key.objectid)
 283                                 continue;
 284
 285                         start = logical[nr];
 286                         if (start < cache->key.objectid) {
 287                                 start = cache->key.objectid;
 288                                 len = (logical[nr] + stripe_len) - start;
 289                         } else {
 290                                 len = min_t(u64, stripe_len,
 291                                             cache->key.objectid +
 292                                             cache->key.offset - start);
 293                         }
 294
 295                         cache->bytes_super += len;
 296                         ret = add_excluded_extent(root, start, len);
 297                         if (ret) {
 298                                 kfree(logical);
 299                                 return ret;
 300                         }
 301                 }
 302
 303                 kfree(logical);
 304         }
 305         return 0;
 306 }
 307
 308 static struct btrfs_caching_control *
 309 get_caching_control(struct btrfs_block_group_cache *cache)
 310 {
 311         struct btrfs_caching_control *ctl;
 312
 313         spin_lock(&cache->lock);
 314         if (!cache->caching_ctl) {
 315                 spin_unlock(&cache->lock);
 316                 return NULL;
 317         }
 318
 319         ctl = cache->caching_ctl;
 320         atomic_inc(&ctl->count);
 321         spin_unlock(&cache->lock);
 322         return ctl;
 323 }
 324
 325 static void put_caching_control(struct btrfs_caching_control *ctl)
 326 {
 327         if (atomic_dec_and_test(&ctl->count))
 328                 kfree(ctl);
 329 }
 330
 331 #ifdef CONFIG_BTRFS_DEBUG
 332 static void fragment_free_space(struct btrfs_root *root,
 333                                 struct btrfs_block_group_cache *block_group)
 334 {
 335         u64 start = block_group->key.objectid;
 336         u64 len = block_group->key.offset;
 337         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 338                 root->nodesize : root->sectorsize;
 339         u64 step = chunk << 1;
 340
 341         while (len > chunk) {
 342                 btrfs_remove_free_space(block_group, start, chunk);
 343                 start += step;
 344                 if (len < step)
 345                         len = 0;
 346                 else
 347                         len -= step;
 348         }
 349 }
 350 #endif
 351
 352 /*
 353  * this is only called by cache_block_group, since we could have freed extents
 354  * we need to check the pinned_extents for any extents that can't be used yet
 355  * since their free space will be released as soon as the transaction commits.
 356  */
 357 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 358                        struct btrfs_fs_info *info, u64 start, u64 end)
 359 {
 360         u64 extent_start, extent_end, size, total_added = 0;
 361         int ret;
 362
 363         while (start < end) {
 364                 ret = find_first_extent_bit(info->pinned_extents, start,
 365                                             &extent_start, &extent_end,
 366                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 367                                             NULL);
 368                 if (ret)
 369                         break;
 370
 371                 if (extent_start <= start) {
 372                         start = extent_end + 1;
 373                 } else if (extent_start > start && extent_start < end) {
 374                         size = extent_start - start;
 375                         total_added += size;
 376                         ret = btrfs_add_free_space(block_group, start,
 377                                                    size);
 378                         BUG_ON(ret); /* -ENOMEM or logic error */
 379                         start = extent_end + 1;
 380                 } else {
 381                         break;
 382                 }
 383         }
 384
 385         if (start < end) {
 386                 size = end - start;
 387                 total_added += size;
 388                 ret = btrfs_add_free_space(block_group, start, size);
 389                 BUG_ON(ret); /* -ENOMEM or logic error */
 390         }
 391
 392         return total_added;
 393 }
 394
 395 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 396 {
 397         struct btrfs_block_group_cache *block_group;
 398         struct btrfs_fs_info *fs_info;
 399         struct btrfs_root *extent_root;
 400         struct btrfs_path *path;
 401         struct extent_buffer *leaf;
 402         struct btrfs_key key;
 403         u64 total_found = 0;
 404         u64 last = 0;
 405         u32 nritems;
 406         int ret;
 407         bool wakeup = true;
 408
 409         block_group = caching_ctl->block_group;
 410         fs_info = block_group->fs_info;
 411         extent_root = fs_info->extent_root;
 412
 413         path = btrfs_alloc_path();
 414         if (!path)
 415                 return -ENOMEM;
 416
 417         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 418
 419 #ifdef CONFIG_BTRFS_DEBUG
 420         /*
 421          * If we're fragmenting we don't want to make anybody think we can
 422          * allocate from this block group until we've had a chance to fragment
 423          * the free space.
 424          */
 425         if (btrfs_should_fragment_free_space(extent_root, block_group))
 426                 wakeup = false;
 427 #endif
 428         /*
 429          * We don't want to deadlock with somebody trying to allocate a new
 430          * extent for the extent root while also trying to search the extent
 431          * root to add free space.  So we skip locking and search the commit
 432          * root, since its read-only
 433          */
 434         path->skip_locking = 1;
 435         path->search_commit_root = 1;
 436         path->reada = READA_FORWARD;
 437
 438         key.objectid = last;
 439         key.offset = 0;
 440         key.type = BTRFS_EXTENT_ITEM_KEY;
 441
 442 next:
 443         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 444         if (ret < 0)
 445                 goto out;
 446
 447         leaf = path->nodes[0];
 448         nritems = btrfs_header_nritems(leaf);
 449
 450         while (1) {
 451                 if (btrfs_fs_closing(fs_info) > 1) {
 452                         last = (u64)-1;
 453                         break;
 454                 }
 455
 456                 if (path->slots[0] < nritems) {
 457                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 458                 } else {
 459                         ret = find_next_key(path, 0, &key);
 460                         if (ret)
 461                                 break;
 462
 463                         if (need_resched() ||
 464                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 465                                 if (wakeup)
 466                                         caching_ctl->progress = last;
 467                                 btrfs_release_path(path);
 468                                 up_read(&fs_info->commit_root_sem);
 469                                 mutex_unlock(&caching_ctl->mutex);
 470                                 cond_resched();
 471                                 mutex_lock(&caching_ctl->mutex);
 472                                 down_read(&fs_info->commit_root_sem);
 473                                 goto next;
 474                         }
 475
 476                         ret = btrfs_next_leaf(extent_root, path);
 477                         if (ret < 0)
 478                                 goto out;
 479                         if (ret)
 480                                 break;
 481                         leaf = path->nodes[0];
 482                         nritems = btrfs_header_nritems(leaf);
 483                         continue;
 484                 }
 485
 486                 if (key.objectid < last) {
 487                         key.objectid = last;
 488                         key.offset = 0;
 489                         key.type = BTRFS_EXTENT_ITEM_KEY;
 490
 491                         if (wakeup)
 492                                 caching_ctl->progress = last;
 493                         btrfs_release_path(path);
 494                         goto next;
 495                 }
 496
 497                 if (key.objectid < block_group->key.objectid) {
 498                         path->slots[0]++;
 499                         continue;
 500                 }
 501
 502                 if (key.objectid >= block_group->key.objectid +
 503                     block_group->key.offset)
 504                         break;
 505
 506                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 507                     key.type == BTRFS_METADATA_ITEM_KEY) {
 508                         total_found += add_new_free_space(block_group,
 509                                                           fs_info, last,
 510                                                           key.objectid);
 511                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 512                                 last = key.objectid +
 513                                         fs_info->tree_root->nodesize;
 514                         else
 515                                 last = key.objectid + key.offset;
 516
 517                         if (total_found > CACHING_CTL_WAKE_UP) {
 518                                 total_found = 0;
 519                                 if (wakeup)
 520                                         wake_up(&caching_ctl->wait);
 521                         }
 522                 }
 523                 path->slots[0]++;
 524         }
 525         ret = 0;
 526
 527         total_found += add_new_free_space(block_group, fs_info, last,
 528                                           block_group->key.objectid +
 529                                           block_group->key.offset);
 530         caching_ctl->progress = (u64)-1;
 531
 532 out:
 533         btrfs_free_path(path);
 534         return ret;
 535 }
 536
 537 static noinline void caching_thread(struct btrfs_work *work)
 538 {
 539         struct btrfs_block_group_cache *block_group;
 540         struct btrfs_fs_info *fs_info;
 541         struct btrfs_caching_control *caching_ctl;
 542         struct btrfs_root *extent_root;
 543         int ret;
 544
 545         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 546         block_group = caching_ctl->block_group;
 547         fs_info = block_group->fs_info;
 548         extent_root = fs_info->extent_root;
 549
 550         mutex_lock(&caching_ctl->mutex);
 551         down_read(&fs_info->commit_root_sem);
 552
 553         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 554                 ret = load_free_space_tree(caching_ctl);
 555         else
 556                 ret = load_extent_tree_free(caching_ctl);
 557
 558         spin_lock(&block_group->lock);
 559         block_group->caching_ctl = NULL;
 560         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 561         spin_unlock(&block_group->lock);
 562
 563 #ifdef CONFIG_BTRFS_DEBUG
 564         if (btrfs_should_fragment_free_space(extent_root, block_group)) {
 565                 u64 bytes_used;
 566
 567                 spin_lock(&block_group->space_info->lock);
 568                 spin_lock(&block_group->lock);
 569                 bytes_used = block_group->key.offset -
 570                         btrfs_block_group_used(&block_group->item);
 571                 block_group->space_info->bytes_used += bytes_used >> 1;
 572                 spin_unlock(&block_group->lock);
 573                 spin_unlock(&block_group->space_info->lock);
 574                 fragment_free_space(extent_root, block_group);
 575         }
 576 #endif
 577
 578         caching_ctl->progress = (u64)-1;
 579
 580         up_read(&fs_info->commit_root_sem);
 581         free_excluded_extents(fs_info->extent_root, block_group);
 582         mutex_unlock(&caching_ctl->mutex);
 583
 584         wake_up(&caching_ctl->wait);
 585
 586         put_caching_control(caching_ctl);
 587         btrfs_put_block_group(block_group);
 588 }
 589
 590 static int cache_block_group(struct btrfs_block_group_cache *cache,
 591                              int load_cache_only)
 592 {
 593         DEFINE_WAIT(wait);
 594         struct btrfs_fs_info *fs_info = cache->fs_info;
 595         struct btrfs_caching_control *caching_ctl;
 596         int ret = 0;
 597
 598         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 599         if (!caching_ctl)
 600                 return -ENOMEM;
 601
 602         INIT_LIST_HEAD(&caching_ctl->list);
 603         mutex_init(&caching_ctl->mutex);
 604         init_waitqueue_head(&caching_ctl->wait);
 605         caching_ctl->block_group = cache;
 606         caching_ctl->progress = cache->key.objectid;
 607         atomic_set(&caching_ctl->count, 1);
 608         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 609                         caching_thread, NULL, NULL);
 610
 611         spin_lock(&cache->lock);
 612         /*
 613          * This should be a rare occasion, but this could happen I think in the
 614          * case where one thread starts to load the space cache info, and then
 615          * some other thread starts a transaction commit which tries to do an
 616          * allocation while the other thread is still loading the space cache
 617          * info.  The previous loop should have kept us from choosing this block
 618          * group, but if we've moved to the state where we will wait on caching
 619          * block groups we need to first check if we're doing a fast load here,
 620          * so we can wait for it to finish, otherwise we could end up allocating
 621          * from a block group who's cache gets evicted for one reason or
 622          * another.
 623          */
 624         while (cache->cached == BTRFS_CACHE_FAST) {
 625                 struct btrfs_caching_control *ctl;
 626
 627                 ctl = cache->caching_ctl;
 628                 atomic_inc(&ctl->count);
 629                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 630                 spin_unlock(&cache->lock);
 631
 632                 schedule();
 633
 634                 finish_wait(&ctl->wait, &wait);
 635                 put_caching_control(ctl);
 636                 spin_lock(&cache->lock);
 637         }
 638
 639         if (cache->cached != BTRFS_CACHE_NO) {
 640                 spin_unlock(&cache->lock);
 641                 kfree(caching_ctl);
 642                 return 0;
 643         }
 644         WARN_ON(cache->caching_ctl);
 645         cache->caching_ctl = caching_ctl;
 646         cache->cached = BTRFS_CACHE_FAST;
 647         spin_unlock(&cache->lock);
 648
 649         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 650                 mutex_lock(&caching_ctl->mutex);
 651                 ret = load_free_space_cache(fs_info, cache);
 652
 653                 spin_lock(&cache->lock);
 654                 if (ret == 1) {
 655                         cache->caching_ctl = NULL;
 656                         cache->cached = BTRFS_CACHE_FINISHED;
 657                         cache->last_byte_to_unpin = (u64)-1;
 658                         caching_ctl->progress = (u64)-1;
 659                 } else {
 660                         if (load_cache_only) {
 661                                 cache->caching_ctl = NULL;
 662                                 cache->cached = BTRFS_CACHE_NO;
 663                         } else {
 664                                 cache->cached = BTRFS_CACHE_STARTED;
 665                                 cache->has_caching_ctl = 1;
 666                         }
 667                 }
 668                 spin_unlock(&cache->lock);
 669 #ifdef CONFIG_BTRFS_DEBUG
 670                 if (ret == 1 &&
 671                     btrfs_should_fragment_free_space(fs_info->extent_root,
 672                                                      cache)) {
 673                         u64 bytes_used;
 674
 675                         spin_lock(&cache->space_info->lock);
 676                         spin_lock(&cache->lock);
 677                         bytes_used = cache->key.offset -
 678                                 btrfs_block_group_used(&cache->item);
 679                         cache->space_info->bytes_used += bytes_used >> 1;
 680                         spin_unlock(&cache->lock);
 681                         spin_unlock(&cache->space_info->lock);
 682                         fragment_free_space(fs_info->extent_root, cache);
 683                 }
 684 #endif
 685                 mutex_unlock(&caching_ctl->mutex);
 686
 687                 wake_up(&caching_ctl->wait);
 688                 if (ret == 1) {
 689                         put_caching_control(caching_ctl);
 690                         free_excluded_extents(fs_info->extent_root, cache);
 691                         return 0;
 692                 }
 693         } else {
 694                 /*
 695                  * We're either using the free space tree or no caching at all.
 696                  * Set cached to the appropriate value and wakeup any waiters.
 697                  */
 698                 spin_lock(&cache->lock);
 699                 if (load_cache_only) {
 700                         cache->caching_ctl = NULL;
 701                         cache->cached = BTRFS_CACHE_NO;
 702                 } else {
 703                         cache->cached = BTRFS_CACHE_STARTED;
 704                         cache->has_caching_ctl = 1;
 705                 }
 706                 spin_unlock(&cache->lock);
 707                 wake_up(&caching_ctl->wait);
 708         }
 709
 710         if (load_cache_only) {
 711                 put_caching_control(caching_ctl);
 712                 return 0;
 713         }
 714
 715         down_write(&fs_info->commit_root_sem);
 716         atomic_inc(&caching_ctl->count);
 717         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 718         up_write(&fs_info->commit_root_sem);
 719
 720         btrfs_get_block_group(cache);
 721
 722         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 723
 724         return ret;
 725 }
 726
 727 /*
 728  * return the block group that starts at or after bytenr
 729  */
 730 static struct btrfs_block_group_cache *
 731 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 732 {
 733         struct btrfs_block_group_cache *cache;
 734
 735         cache = block_group_cache_tree_search(info, bytenr, 0);
 736
 737         return cache;
 738 }
 739
 740 /*
 741  * return the block group that contains the given bytenr
 742  */
 743 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 744                                                  struct btrfs_fs_info *info,
 745                                                  u64 bytenr)
 746 {
 747         struct btrfs_block_group_cache *cache;
 748
 749         cache = block_group_cache_tree_search(info, bytenr, 1);
 750
 751         return cache;
 752 }
 753
 754 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 755                                                   u64 flags)
 756 {
 757         struct list_head *head = &info->space_info;
 758         struct btrfs_space_info *found;
 759
 760         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 761
 762         rcu_read_lock();
 763         list_for_each_entry_rcu(found, head, list) {
 764                 if (found->flags & flags) {
 765                         rcu_read_unlock();
 766                         return found;
 767                 }
 768         }
 769         rcu_read_unlock();
 770         return NULL;
 771 }
 772
 773 /*
 774  * after adding space to the filesystem, we need to clear the full flags
 775  * on all the space infos.
 776  */
 777 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 778 {
 779         struct list_head *head = &info->space_info;
 780         struct btrfs_space_info *found;
 781
 782         rcu_read_lock();
 783         list_for_each_entry_rcu(found, head, list)
 784                 found->full = 0;
 785         rcu_read_unlock();
 786 }
 787
 788 /* simple helper to search for an existing data extent at a given offset */
 789 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 790 {
 791         int ret;
 792         struct btrfs_key key;
 793         struct btrfs_path *path;
 794
 795         path = btrfs_alloc_path();
 796         if (!path)
 797                 return -ENOMEM;
 798
 799         key.objectid = start;
 800         key.offset = len;
 801         key.type = BTRFS_EXTENT_ITEM_KEY;
 802         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 803                                 0, 0);
 804         btrfs_free_path(path);
 805         return ret;
 806 }
 807
 808 /*
 809  * helper function to lookup reference count and flags of a tree block.
 810  *
 811  * the head node for delayed ref is used to store the sum of all the
 812  * reference count modifications queued up in the rbtree. the head
 813  * node may also store the extent flags to set. This way you can check
 814  * to see what the reference count and extent flags would be if all of
 815  * the delayed refs are not processed.
 816  */
 817 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 818                              struct btrfs_root *root, u64 bytenr,
 819                              u64 offset, int metadata, u64 *refs, u64 *flags)
 820 {
 821         struct btrfs_delayed_ref_head *head;
 822         struct btrfs_delayed_ref_root *delayed_refs;
 823         struct btrfs_path *path;
 824         struct btrfs_extent_item *ei;
 825         struct extent_buffer *leaf;
 826         struct btrfs_key key;
 827         u32 item_size;
 828         u64 num_refs;
 829         u64 extent_flags;
 830         int ret;
 831
 832         /*
 833          * If we don't have skinny metadata, don't bother doing anything
 834          * different
 835          */
 836         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 837                 offset = root->nodesize;
 838                 metadata = 0;
 839         }
 840
 841         path = btrfs_alloc_path();
 842         if (!path)
 843                 return -ENOMEM;
 844
 845         if (!trans) {
 846                 path->skip_locking = 1;
 847                 path->search_commit_root = 1;
 848         }
 849
 850 search_again:
 851         key.objectid = bytenr;
 852         key.offset = offset;
 853         if (metadata)
 854                 key.type = BTRFS_METADATA_ITEM_KEY;
 855         else
 856                 key.type = BTRFS_EXTENT_ITEM_KEY;
 857
 858         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 859                                 &key, path, 0, 0);
 860         if (ret < 0)
 861                 goto out_free;
 862
 863         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 864                 if (path->slots[0]) {
 865                         path->slots[0]--;
 866                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 867                                               path->slots[0]);
 868                         if (key.objectid == bytenr &&
 869                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 870                             key.offset == root->nodesize)
 871                                 ret = 0;
 872                 }
 873         }
 874
 875         if (ret == 0) {
 876                 leaf = path->nodes[0];
 877                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 878                 if (item_size >= sizeof(*ei)) {
 879                         ei = btrfs_item_ptr(leaf, path->slots[0],
 880                                             struct btrfs_extent_item);
 881                         num_refs = btrfs_extent_refs(leaf, ei);
 882                         extent_flags = btrfs_extent_flags(leaf, ei);
 883                 } else {
 884 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 885                         struct btrfs_extent_item_v0 *ei0;
 886                         BUG_ON(item_size != sizeof(*ei0));
 887                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 888                                              struct btrfs_extent_item_v0);
 889                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 890                         /* FIXME: this isn't correct for data */
 891                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 892 #else
 893                         BUG();
 894 #endif
 895                 }
 896                 BUG_ON(num_refs == 0);
 897         } else {
 898                 num_refs = 0;
 899                 extent_flags = 0;
 900                 ret = 0;
 901         }
 902
 903         if (!trans)
 904                 goto out;
 905
 906         delayed_refs = &trans->transaction->delayed_refs;
 907         spin_lock(&delayed_refs->lock);
 908         head = btrfs_find_delayed_ref_head(trans, bytenr);
 909         if (head) {
 910                 if (!mutex_trylock(&head->mutex)) {
 911                         atomic_inc(&head->node.refs);
 912                         spin_unlock(&delayed_refs->lock);
 913
 914                         btrfs_release_path(path);
 915
 916                         /*
 917                          * Mutex was contended, block until it's released and try
 918                          * again
 919                          */
 920                         mutex_lock(&head->mutex);
 921                         mutex_unlock(&head->mutex);
 922                         btrfs_put_delayed_ref(&head->node);
 923                         goto search_again;
 924                 }
 925                 spin_lock(&head->lock);
 926                 if (head->extent_op && head->extent_op->update_flags)
 927                         extent_flags |= head->extent_op->flags_to_set;
 928                 else
 929                         BUG_ON(num_refs == 0);
 930
 931                 num_refs += head->node.ref_mod;
 932                 spin_unlock(&head->lock);
 933                 mutex_unlock(&head->mutex);
 934         }
 935         spin_unlock(&delayed_refs->lock);
 936 out:
 937         WARN_ON(num_refs == 0);
 938         if (refs)
 939                 *refs = num_refs;
 940         if (flags)
 941                 *flags = extent_flags;
 942 out_free:
 943         btrfs_free_path(path);
 944         return ret;
 945 }
 946
 947 /*
 948  * Back reference rules.  Back refs have three main goals:
 949  *
 950  * 1) differentiate between all holders of references to an extent so that
 951  *    when a reference is dropped we can make sure it was a valid reference
 952  *    before freeing the extent.
 953  *
 954  * 2) Provide enough information to quickly find the holders of an extent
 955  *    if we notice a given block is corrupted or bad.
 956  *
 957  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 958  *    maintenance.  This is actually the same as #2, but with a slightly
 959  *    different use case.
 960  *
 961  * There are two kinds of back refs. The implicit back refs is optimized
 962  * for pointers in non-shared tree blocks. For a given pointer in a block,
 963  * back refs of this kind provide information about the block's owner tree
 964  * and the pointer's key. These information allow us to find the block by
 965  * b-tree searching. The full back refs is for pointers in tree blocks not
 966  * referenced by their owner trees. The location of tree block is recorded
 967  * in the back refs. Actually the full back refs is generic, and can be
 968  * used in all cases the implicit back refs is used. The major shortcoming
 969  * of the full back refs is its overhead. Every time a tree block gets
 970  * COWed, we have to update back refs entry for all pointers in it.
 971  *
 972  * For a newly allocated tree block, we use implicit back refs for
 973  * pointers in it. This means most tree related operations only involve
 974  * implicit back refs. For a tree block created in old transaction, the
 975  * only way to drop a reference to it is COW it. So we can detect the
 976  * event that tree block loses its owner tree's reference and do the
 977  * back refs conversion.
 978  *
 979  * When a tree block is COWed through a tree, there are four cases:
 980  *
 981  * The reference count of the block is one and the tree is the block's
 982  * owner tree. Nothing to do in this case.
 983  *
 984  * The reference count of the block is one and the tree is not the
 985  * block's owner tree. In this case, full back refs is used for pointers
 986  * in the block. Remove these full back refs, add implicit back refs for
 987  * every pointers in the new block.
 988  *
 989  * The reference count of the block is greater than one and the tree is
 990  * the block's owner tree. In this case, implicit back refs is used for
 991  * pointers in the block. Add full back refs for every pointers in the
 992  * block, increase lower level extents' reference counts. The original
 993  * implicit back refs are entailed to the new block.
 994  *
 995  * The reference count of the block is greater than one and the tree is
 996  * not the block's owner tree. Add implicit back refs for every pointer in
 997  * the new block, increase lower level extents' reference count.
 998  *
 999  * Back Reference Key composing:
1000  *
1001  * The key objectid corresponds to the first byte in the extent,
1002  * The key type is used to differentiate between types of back refs.
1003  * There are different meanings of the key offset for different types
1004  * of back refs.
1005  *
1006  * File extents can be referenced by:
1007  *
1008  * - multiple snapshots, subvolumes, or different generations in one subvol
1009  * - different files inside a single subvolume
1010  * - different offsets inside a file (bookend extents in file.c)
1011  *
1012  * The extent ref structure for the implicit back refs has fields for:
1013  *
1014  * - Objectid of the subvolume root
1015  * - objectid of the file holding the reference
1016  * - original offset in the file
1017  * - how many bookend extents
1018  *
1019  * The key offset for the implicit back refs is hash of the first
1020  * three fields.
1021  *
1022  * The extent ref structure for the full back refs has field for:
1023  *
1024  * - number of pointers in the tree leaf
1025  *
1026  * The key offset for the implicit back refs is the first byte of
1027  * the tree leaf
1028  *
1029  * When a file extent is allocated, The implicit back refs is used.
1030  * the fields are filled in:
1031  *
1032  *     (root_key.objectid, inode objectid, offset in file, 1)
1033  *
1034  * When a file extent is removed file truncation, we find the
1035  * corresponding implicit back refs and check the following fields:
1036  *
1037  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1038  *
1039  * Btree extents can be referenced by:
1040  *
1041  * - Different subvolumes
1042  *
1043  * Both the implicit back refs and the full back refs for tree blocks
1044  * only consist of key. The key offset for the implicit back refs is
1045  * objectid of block's owner tree. The key offset for the full back refs
1046  * is the first byte of parent block.
1047  *
1048  * When implicit back refs is used, information about the lowest key and
1049  * level of the tree block are required. These information are stored in
1050  * tree block info structure.
1051  */
1052
1053 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1054 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1055                                   struct btrfs_root *root,
1056                                   struct btrfs_path *path,
1057                                   u64 owner, u32 extra_size)
1058 {
1059         struct btrfs_extent_item *item;
1060         struct btrfs_extent_item_v0 *ei0;
1061         struct btrfs_extent_ref_v0 *ref0;
1062         struct btrfs_tree_block_info *bi;
1063         struct extent_buffer *leaf;
1064         struct btrfs_key key;
1065         struct btrfs_key found_key;
1066         u32 new_size = sizeof(*item);
1067         u64 refs;
1068         int ret;
1069
1070         leaf = path->nodes[0];
1071         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1072
1073         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1074         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1075                              struct btrfs_extent_item_v0);
1076         refs = btrfs_extent_refs_v0(leaf, ei0);
1077
1078         if (owner == (u64)-1) {
1079                 while (1) {
1080                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1081                                 ret = btrfs_next_leaf(root, path);
1082                                 if (ret < 0)
1083                                         return ret;
1084                                 BUG_ON(ret > 0); /* Corruption */
1085                                 leaf = path->nodes[0];
1086                         }
1087                         btrfs_item_key_to_cpu(leaf, &found_key,
1088                                               path->slots[0]);
1089                         BUG_ON(key.objectid != found_key.objectid);
1090                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1091                                 path->slots[0]++;
1092                                 continue;
1093                         }
1094                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1095                                               struct btrfs_extent_ref_v0);
1096                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1097                         break;
1098                 }
1099         }
1100         btrfs_release_path(path);
1101
1102         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1103                 new_size += sizeof(*bi);
1104
1105         new_size -= sizeof(*ei0);
1106         ret = btrfs_search_slot(trans, root, &key, path,
1107                                 new_size + extra_size, 1);
1108         if (ret < 0)
1109                 return ret;
1110         BUG_ON(ret); /* Corruption */
1111
1112         btrfs_extend_item(root, path, new_size);
1113
1114         leaf = path->nodes[0];
1115         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1116         btrfs_set_extent_refs(leaf, item, refs);
1117         /* FIXME: get real generation */
1118         btrfs_set_extent_generation(leaf, item, 0);
1119         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1120                 btrfs_set_extent_flags(leaf, item,
1121                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1122                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1123                 bi = (struct btrfs_tree_block_info *)(item + 1);
1124                 /* FIXME: get first key of the block */
1125                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1126                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1127         } else {
1128                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1129         }
1130         btrfs_mark_buffer_dirty(leaf);
1131         return 0;
1132 }
1133 #endif
1134
1135 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1136 {
1137         u32 high_crc = ~(u32)0;
1138         u32 low_crc = ~(u32)0;
1139         __le64 lenum;
1140
1141         lenum = cpu_to_le64(root_objectid);
1142         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1143         lenum = cpu_to_le64(owner);
1144         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1145         lenum = cpu_to_le64(offset);
1146         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1147
1148         return ((u64)high_crc << 31) ^ (u64)low_crc;
1149 }
1150
1151 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1152                                      struct btrfs_extent_data_ref *ref)
1153 {
1154         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1155                                     btrfs_extent_data_ref_objectid(leaf, ref),
1156                                     btrfs_extent_data_ref_offset(leaf, ref));
1157 }
1158
1159 static int match_extent_data_ref(struct extent_buffer *leaf,
1160                                  struct btrfs_extent_data_ref *ref,
1161                                  u64 root_objectid, u64 owner, u64 offset)
1162 {
1163         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1164             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1165             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1166                 return 0;
1167         return 1;
1168 }
1169
1170 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1171                                            struct btrfs_root *root,
1172                                            struct btrfs_path *path,
1173                                            u64 bytenr, u64 parent,
1174                                            u64 root_objectid,
1175                                            u64 owner, u64 offset)
1176 {
1177         struct btrfs_key key;
1178         struct btrfs_extent_data_ref *ref;
1179         struct extent_buffer *leaf;
1180         u32 nritems;
1181         int ret;
1182         int recow;
1183         int err = -ENOENT;
1184
1185         key.objectid = bytenr;
1186         if (parent) {
1187                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1188                 key.offset = parent;
1189         } else {
1190                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1191                 key.offset = hash_extent_data_ref(root_objectid,
1192                                                   owner, offset);
1193         }
1194 again:
1195         recow = 0;
1196         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1197         if (ret < 0) {
1198                 err = ret;
1199                 goto fail;
1200         }
1201
1202         if (parent) {
1203                 if (!ret)
1204                         return 0;
1205 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1206                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1207                 btrfs_release_path(path);
1208                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1209                 if (ret < 0) {
1210                         err = ret;
1211                         goto fail;
1212                 }
1213                 if (!ret)
1214                         return 0;
1215 #endif
1216                 goto fail;
1217         }
1218
1219         leaf = path->nodes[0];
1220         nritems = btrfs_header_nritems(leaf);
1221         while (1) {
1222                 if (path->slots[0] >= nritems) {
1223                         ret = btrfs_next_leaf(root, path);
1224                         if (ret < 0)
1225                                 err = ret;
1226                         if (ret)
1227                                 goto fail;
1228
1229                         leaf = path->nodes[0];
1230                         nritems = btrfs_header_nritems(leaf);
1231                         recow = 1;
1232                 }
1233
1234                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1235                 if (key.objectid != bytenr ||
1236                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1237                         goto fail;
1238
1239                 ref = btrfs_item_ptr(leaf, path->slots[0],
1240                                      struct btrfs_extent_data_ref);
1241
1242                 if (match_extent_data_ref(leaf, ref, root_objectid,
1243                                           owner, offset)) {
1244                         if (recow) {
1245                                 btrfs_release_path(path);
1246                                 goto again;
1247                         }
1248                         err = 0;
1249                         break;
1250                 }
1251                 path->slots[0]++;
1252         }
1253 fail:
1254         return err;
1255 }
1256
1257 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1258                                            struct btrfs_root *root,
1259                                            struct btrfs_path *path,
1260                                            u64 bytenr, u64 parent,
1261                                            u64 root_objectid, u64 owner,
1262                                            u64 offset, int refs_to_add)
1263 {
1264         struct btrfs_key key;
1265         struct extent_buffer *leaf;
1266         u32 size;
1267         u32 num_refs;
1268         int ret;
1269
1270         key.objectid = bytenr;
1271         if (parent) {
1272                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1273                 key.offset = parent;
1274                 size = sizeof(struct btrfs_shared_data_ref);
1275         } else {
1276                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1277                 key.offset = hash_extent_data_ref(root_objectid,
1278                                                   owner, offset);
1279                 size = sizeof(struct btrfs_extent_data_ref);
1280         }
1281
1282         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1283         if (ret && ret != -EEXIST)
1284                 goto fail;
1285
1286         leaf = path->nodes[0];
1287         if (parent) {
1288                 struct btrfs_shared_data_ref *ref;
1289                 ref = btrfs_item_ptr(leaf, path->slots[0],
1290                                      struct btrfs_shared_data_ref);
1291                 if (ret == 0) {
1292                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1293                 } else {
1294                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1295                         num_refs += refs_to_add;
1296                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1297                 }
1298         } else {
1299                 struct btrfs_extent_data_ref *ref;
1300                 while (ret == -EEXIST) {
1301                         ref = btrfs_item_ptr(leaf, path->slots[0],
1302                                              struct btrfs_extent_data_ref);
1303                         if (match_extent_data_ref(leaf, ref, root_objectid,
1304                                                   owner, offset))
1305                                 break;
1306                         btrfs_release_path(path);
1307                         key.offset++;
1308                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1309                                                       size);
1310                         if (ret && ret != -EEXIST)
1311                                 goto fail;
1312
1313                         leaf = path->nodes[0];
1314                 }
1315                 ref = btrfs_item_ptr(leaf, path->slots[0],
1316                                      struct btrfs_extent_data_ref);
1317                 if (ret == 0) {
1318                         btrfs_set_extent_data_ref_root(leaf, ref,
1319                                                        root_objectid);
1320                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1321                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1322                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1323                 } else {
1324                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1325                         num_refs += refs_to_add;
1326                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1327                 }
1328         }
1329         btrfs_mark_buffer_dirty(leaf);
1330         ret = 0;
1331 fail:
1332         btrfs_release_path(path);
1333         return ret;
1334 }
1335
1336 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1337                                            struct btrfs_root *root,
1338                                            struct btrfs_path *path,
1339                                            int refs_to_drop, int *last_ref)
1340 {
1341         struct btrfs_key key;
1342         struct btrfs_extent_data_ref *ref1 = NULL;
1343         struct btrfs_shared_data_ref *ref2 = NULL;
1344         struct extent_buffer *leaf;
1345         u32 num_refs = 0;
1346         int ret = 0;
1347
1348         leaf = path->nodes[0];
1349         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1350
1351         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1352                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1353                                       struct btrfs_extent_data_ref);
1354                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1355         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1356                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1357                                       struct btrfs_shared_data_ref);
1358                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1359 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1360         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1361                 struct btrfs_extent_ref_v0 *ref0;
1362                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1363                                       struct btrfs_extent_ref_v0);
1364                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1365 #endif
1366         } else {
1367                 BUG();
1368         }
1369
1370         BUG_ON(num_refs < refs_to_drop);
1371         num_refs -= refs_to_drop;
1372
1373         if (num_refs == 0) {
1374                 ret = btrfs_del_item(trans, root, path);
1375                 *last_ref = 1;
1376         } else {
1377                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1378                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1379                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1380                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1381 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1382                 else {
1383                         struct btrfs_extent_ref_v0 *ref0;
1384                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1385                                         struct btrfs_extent_ref_v0);
1386                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1387                 }
1388 #endif
1389                 btrfs_mark_buffer_dirty(leaf);
1390         }
1391         return ret;
1392 }
1393
1394 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1395                                           struct btrfs_extent_inline_ref *iref)
1396 {
1397         struct btrfs_key key;
1398         struct extent_buffer *leaf;
1399         struct btrfs_extent_data_ref *ref1;
1400         struct btrfs_shared_data_ref *ref2;
1401         u32 num_refs = 0;
1402
1403         leaf = path->nodes[0];
1404         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1405         if (iref) {
1406                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1407                     BTRFS_EXTENT_DATA_REF_KEY) {
1408                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1409                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1410                 } else {
1411                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1412                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1413                 }
1414         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1415                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1416                                       struct btrfs_extent_data_ref);
1417                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1418         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1419                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1420                                       struct btrfs_shared_data_ref);
1421                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1422 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1423         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1424                 struct btrfs_extent_ref_v0 *ref0;
1425                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1426                                       struct btrfs_extent_ref_v0);
1427                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1428 #endif
1429         } else {
1430                 WARN_ON(1);
1431         }
1432         return num_refs;
1433 }
1434
1435 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1436                                           struct btrfs_root *root,
1437                                           struct btrfs_path *path,
1438                                           u64 bytenr, u64 parent,
1439                                           u64 root_objectid)
1440 {
1441         struct btrfs_key key;
1442         int ret;
1443
1444         key.objectid = bytenr;
1445         if (parent) {
1446                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1447                 key.offset = parent;
1448         } else {
1449                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1450                 key.offset = root_objectid;
1451         }
1452
1453         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1454         if (ret > 0)
1455                 ret = -ENOENT;
1456 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1457         if (ret == -ENOENT && parent) {
1458                 btrfs_release_path(path);
1459                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1460                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1461                 if (ret > 0)
1462                         ret = -ENOENT;
1463         }
1464 #endif
1465         return ret;
1466 }
1467
1468 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1469                                           struct btrfs_root *root,
1470                                           struct btrfs_path *path,
1471                                           u64 bytenr, u64 parent,
1472                                           u64 root_objectid)
1473 {
1474         struct btrfs_key key;
1475         int ret;
1476
1477         key.objectid = bytenr;
1478         if (parent) {
1479                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1480                 key.offset = parent;
1481         } else {
1482                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1483                 key.offset = root_objectid;
1484         }
1485
1486         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1487         btrfs_release_path(path);
1488         return ret;
1489 }
1490
1491 static inline int extent_ref_type(u64 parent, u64 owner)
1492 {
1493         int type;
1494         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1495                 if (parent > 0)
1496                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1497                 else
1498                         type = BTRFS_TREE_BLOCK_REF_KEY;
1499         } else {
1500                 if (parent > 0)
1501                         type = BTRFS_SHARED_DATA_REF_KEY;
1502                 else
1503                         type = BTRFS_EXTENT_DATA_REF_KEY;
1504         }
1505         return type;
1506 }
1507
1508 static int find_next_key(struct btrfs_path *path, int level,
1509                          struct btrfs_key *key)
1510
1511 {
1512         for (; level < BTRFS_MAX_LEVEL; level++) {
1513                 if (!path->nodes[level])
1514                         break;
1515                 if (path->slots[level] + 1 >=
1516                     btrfs_header_nritems(path->nodes[level]))
1517                         continue;
1518                 if (level == 0)
1519                         btrfs_item_key_to_cpu(path->nodes[level], key,
1520                                               path->slots[level] + 1);
1521                 else
1522                         btrfs_node_key_to_cpu(path->nodes[level], key,
1523                                               path->slots[level] + 1);
1524                 return 0;
1525         }
1526         return 1;
1527 }
1528
1529 /*
1530  * look for inline back ref. if back ref is found, *ref_ret is set
1531  * to the address of inline back ref, and 0 is returned.
1532  *
1533  * if back ref isn't found, *ref_ret is set to the address where it
1534  * should be inserted, and -ENOENT is returned.
1535  *
1536  * if insert is true and there are too many inline back refs, the path
1537  * points to the extent item, and -EAGAIN is returned.
1538  *
1539  * NOTE: inline back refs are ordered in the same way that back ref
1540  *       items in the tree are ordered.
1541  */
1542 static noinline_for_stack
1543 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1544                                  struct btrfs_root *root,
1545                                  struct btrfs_path *path,
1546                                  struct btrfs_extent_inline_ref **ref_ret,
1547                                  u64 bytenr, u64 num_bytes,
1548                                  u64 parent, u64 root_objectid,
1549                                  u64 owner, u64 offset, int insert)
1550 {
1551         struct btrfs_key key;
1552         struct extent_buffer *leaf;
1553         struct btrfs_extent_item *ei;
1554         struct btrfs_extent_inline_ref *iref;
1555         u64 flags;
1556         u64 item_size;
1557         unsigned long ptr;
1558         unsigned long end;
1559         int extra_size;
1560         int type;
1561         int want;
1562         int ret;
1563         int err = 0;
1564         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1565                                                  SKINNY_METADATA);
1566
1567         key.objectid = bytenr;
1568         key.type = BTRFS_EXTENT_ITEM_KEY;
1569         key.offset = num_bytes;
1570
1571         want = extent_ref_type(parent, owner);
1572         if (insert) {
1573                 extra_size = btrfs_extent_inline_ref_size(want);
1574                 path->keep_locks = 1;
1575         } else
1576                 extra_size = -1;
1577
1578         /*
1579          * Owner is our parent level, so we can just add one to get the level
1580          * for the block we are interested in.
1581          */
1582         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1583                 key.type = BTRFS_METADATA_ITEM_KEY;
1584                 key.offset = owner;
1585         }
1586
1587 again:
1588         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1589         if (ret < 0) {
1590                 err = ret;
1591                 goto out;
1592         }
1593
1594         /*
1595          * We may be a newly converted file system which still has the old fat
1596          * extent entries for metadata, so try and see if we have one of those.
1597          */
1598         if (ret > 0 && skinny_metadata) {
1599                 skinny_metadata = false;
1600                 if (path->slots[0]) {
1601                         path->slots[0]--;
1602                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1603                                               path->slots[0]);
1604                         if (key.objectid == bytenr &&
1605                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1606                             key.offset == num_bytes)
1607                                 ret = 0;
1608                 }
1609                 if (ret) {
1610                         key.objectid = bytenr;
1611                         key.type = BTRFS_EXTENT_ITEM_KEY;
1612                         key.offset = num_bytes;
1613                         btrfs_release_path(path);
1614                         goto again;
1615                 }
1616         }
1617
1618         if (ret && !insert) {
1619                 err = -ENOENT;
1620                 goto out;
1621         } else if (WARN_ON(ret)) {
1622                 err = -EIO;
1623                 goto out;
1624         }
1625
1626         leaf = path->nodes[0];
1627         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1628 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1629         if (item_size < sizeof(*ei)) {
1630                 if (!insert) {
1631                         err = -ENOENT;
1632                         goto out;
1633                 }
1634                 ret = convert_extent_item_v0(trans, root, path, owner,
1635                                              extra_size);
1636                 if (ret < 0) {
1637                         err = ret;
1638                         goto out;
1639                 }
1640                 leaf = path->nodes[0];
1641                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1642         }
1643 #endif
1644         BUG_ON(item_size < sizeof(*ei));
1645
1646         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1647         flags = btrfs_extent_flags(leaf, ei);
1648
1649         ptr = (unsigned long)(ei + 1);
1650         end = (unsigned long)ei + item_size;
1651
1652         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1653                 ptr += sizeof(struct btrfs_tree_block_info);
1654                 BUG_ON(ptr > end);
1655         }
1656
1657         err = -ENOENT;
1658         while (1) {
1659                 if (ptr >= end) {
1660                         WARN_ON(ptr > end);
1661                         break;
1662                 }
1663                 iref = (struct btrfs_extent_inline_ref *)ptr;
1664                 type = btrfs_extent_inline_ref_type(leaf, iref);
1665                 if (want < type)
1666                         break;
1667                 if (want > type) {
1668                         ptr += btrfs_extent_inline_ref_size(type);
1669                         continue;
1670                 }
1671
1672                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1673                         struct btrfs_extent_data_ref *dref;
1674                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1675                         if (match_extent_data_ref(leaf, dref, root_objectid,
1676                                                   owner, offset)) {
1677                                 err = 0;
1678                                 break;
1679                         }
1680                         if (hash_extent_data_ref_item(leaf, dref) <
1681                             hash_extent_data_ref(root_objectid, owner, offset))
1682                                 break;
1683                 } else {
1684                         u64 ref_offset;
1685                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1686                         if (parent > 0) {
1687                                 if (parent == ref_offset) {
1688                                         err = 0;
1689                                         break;
1690                                 }
1691                                 if (ref_offset < parent)
1692                                         break;
1693                         } else {
1694                                 if (root_objectid == ref_offset) {
1695                                         err = 0;
1696                                         break;
1697                                 }
1698                                 if (ref_offset < root_objectid)
1699                                         break;
1700                         }
1701                 }
1702                 ptr += btrfs_extent_inline_ref_size(type);
1703         }
1704         if (err == -ENOENT && insert) {
1705                 if (item_size + extra_size >=
1706                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1707                         err = -EAGAIN;
1708                         goto out;
1709                 }
1710                 /*
1711                  * To add new inline back ref, we have to make sure
1712                  * there is no corresponding back ref item.
1713                  * For simplicity, we just do not add new inline back
1714                  * ref if there is any kind of item for this block
1715                  */
1716                 if (find_next_key(path, 0, &key) == 0 &&
1717                     key.objectid == bytenr &&
1718                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1719                         err = -EAGAIN;
1720                         goto out;
1721                 }
1722         }
1723         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1724 out:
1725         if (insert) {
1726                 path->keep_locks = 0;
1727                 btrfs_unlock_up_safe(path, 1);
1728         }
1729         return err;
1730 }
1731
1732 /*
1733  * helper to add new inline back ref
1734  */
1735 static noinline_for_stack
1736 void setup_inline_extent_backref(struct btrfs_root *root,
1737                                  struct btrfs_path *path,
1738                                  struct btrfs_extent_inline_ref *iref,
1739                                  u64 parent, u64 root_objectid,
1740                                  u64 owner, u64 offset, int refs_to_add,
1741                                  struct btrfs_delayed_extent_op *extent_op)
1742 {
1743         struct extent_buffer *leaf;
1744         struct btrfs_extent_item *ei;
1745         unsigned long ptr;
1746         unsigned long end;
1747         unsigned long item_offset;
1748         u64 refs;
1749         int size;
1750         int type;
1751
1752         leaf = path->nodes[0];
1753         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1754         item_offset = (unsigned long)iref - (unsigned long)ei;
1755
1756         type = extent_ref_type(parent, owner);
1757         size = btrfs_extent_inline_ref_size(type);
1758
1759         btrfs_extend_item(root, path, size);
1760
1761         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1762         refs = btrfs_extent_refs(leaf, ei);
1763         refs += refs_to_add;
1764         btrfs_set_extent_refs(leaf, ei, refs);
1765         if (extent_op)
1766                 __run_delayed_extent_op(extent_op, leaf, ei);
1767
1768         ptr = (unsigned long)ei + item_offset;
1769         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1770         if (ptr < end - size)
1771                 memmove_extent_buffer(leaf, ptr + size, ptr,
1772                                       end - size - ptr);
1773
1774         iref = (struct btrfs_extent_inline_ref *)ptr;
1775         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1776         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1777                 struct btrfs_extent_data_ref *dref;
1778                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1779                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1780                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1781                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1782                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1783         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1784                 struct btrfs_shared_data_ref *sref;
1785                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1786                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1787                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1788         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1789                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1790         } else {
1791                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1792         }
1793         btrfs_mark_buffer_dirty(leaf);
1794 }
1795
1796 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1797                                  struct btrfs_root *root,
1798                                  struct btrfs_path *path,
1799                                  struct btrfs_extent_inline_ref **ref_ret,
1800                                  u64 bytenr, u64 num_bytes, u64 parent,
1801                                  u64 root_objectid, u64 owner, u64 offset)
1802 {
1803         int ret;
1804
1805         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1806                                            bytenr, num_bytes, parent,
1807                                            root_objectid, owner, offset, 0);
1808         if (ret != -ENOENT)
1809                 return ret;
1810
1811         btrfs_release_path(path);
1812         *ref_ret = NULL;
1813
1814         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1815                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1816                                             root_objectid);
1817         } else {
1818                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1819                                              root_objectid, owner, offset);
1820         }
1821         return ret;
1822 }
1823
1824 /*
1825  * helper to update/remove inline back ref
1826  */
1827 static noinline_for_stack
1828 void update_inline_extent_backref(struct btrfs_root *root,
1829                                   struct btrfs_path *path,
1830                                   struct btrfs_extent_inline_ref *iref,
1831                                   int refs_to_mod,
1832                                   struct btrfs_delayed_extent_op *extent_op,
1833                                   int *last_ref)
1834 {
1835         struct extent_buffer *leaf;
1836         struct btrfs_extent_item *ei;
1837         struct btrfs_extent_data_ref *dref = NULL;
1838         struct btrfs_shared_data_ref *sref = NULL;
1839         unsigned long ptr;
1840         unsigned long end;
1841         u32 item_size;
1842         int size;
1843         int type;
1844         u64 refs;
1845
1846         leaf = path->nodes[0];
1847         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1848         refs = btrfs_extent_refs(leaf, ei);
1849         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1850         refs += refs_to_mod;
1851         btrfs_set_extent_refs(leaf, ei, refs);
1852         if (extent_op)
1853                 __run_delayed_extent_op(extent_op, leaf, ei);
1854
1855         type = btrfs_extent_inline_ref_type(leaf, iref);
1856
1857         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1858                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1859                 refs = btrfs_extent_data_ref_count(leaf, dref);
1860         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1861                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1862                 refs = btrfs_shared_data_ref_count(leaf, sref);
1863         } else {
1864                 refs = 1;
1865                 BUG_ON(refs_to_mod != -1);
1866         }
1867
1868         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1869         refs += refs_to_mod;
1870
1871         if (refs > 0) {
1872                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1873                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1874                 else
1875                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1876         } else {
1877                 *last_ref = 1;
1878                 size =  btrfs_extent_inline_ref_size(type);
1879                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1880                 ptr = (unsigned long)iref;
1881                 end = (unsigned long)ei + item_size;
1882                 if (ptr + size < end)
1883                         memmove_extent_buffer(leaf, ptr, ptr + size,
1884                                               end - ptr - size);
1885                 item_size -= size;
1886                 btrfs_truncate_item(root, path, item_size, 1);
1887         }
1888         btrfs_mark_buffer_dirty(leaf);
1889 }
1890
1891 static noinline_for_stack
1892 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1893                                  struct btrfs_root *root,
1894                                  struct btrfs_path *path,
1895                                  u64 bytenr, u64 num_bytes, u64 parent,
1896                                  u64 root_objectid, u64 owner,
1897                                  u64 offset, int refs_to_add,
1898                                  struct btrfs_delayed_extent_op *extent_op)
1899 {
1900         struct btrfs_extent_inline_ref *iref;
1901         int ret;
1902
1903         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1904                                            bytenr, num_bytes, parent,
1905                                            root_objectid, owner, offset, 1);
1906         if (ret == 0) {
1907                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1908                 update_inline_extent_backref(root, path, iref,
1909                                              refs_to_add, extent_op, NULL);
1910         } else if (ret == -ENOENT) {
1911                 setup_inline_extent_backref(root, path, iref, parent,
1912                                             root_objectid, owner, offset,
1913                                             refs_to_add, extent_op);
1914                 ret = 0;
1915         }
1916         return ret;
1917 }
1918
1919 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1920                                  struct btrfs_root *root,
1921                                  struct btrfs_path *path,
1922                                  u64 bytenr, u64 parent, u64 root_objectid,
1923                                  u64 owner, u64 offset, int refs_to_add)
1924 {
1925         int ret;
1926         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1927                 BUG_ON(refs_to_add != 1);
1928                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1929                                             parent, root_objectid);
1930         } else {
1931                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1932                                              parent, root_objectid,
1933                                              owner, offset, refs_to_add);
1934         }
1935         return ret;
1936 }
1937
1938 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1939                                  struct btrfs_root *root,
1940                                  struct btrfs_path *path,
1941                                  struct btrfs_extent_inline_ref *iref,
1942                                  int refs_to_drop, int is_data, int *last_ref)
1943 {
1944         int ret = 0;
1945
1946         BUG_ON(!is_data && refs_to_drop != 1);
1947         if (iref) {
1948                 update_inline_extent_backref(root, path, iref,
1949                                              -refs_to_drop, NULL, last_ref);
1950         } else if (is_data) {
1951                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1952                                              last_ref);
1953         } else {
1954                 *last_ref = 1;
1955                 ret = btrfs_del_item(trans, root, path);
1956         }
1957         return ret;
1958 }
1959
1960 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1961 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1962                                u64 *discarded_bytes)
1963 {
1964         int j, ret = 0;
1965         u64 bytes_left, end;
1966         u64 aligned_start = ALIGN(start, 1 << 9);
1967
1968         if (WARN_ON(start != aligned_start)) {
1969                 len -= aligned_start - start;
1970                 len = round_down(len, 1 << 9);
1971                 start = aligned_start;
1972         }
1973
1974         *discarded_bytes = 0;
1975
1976         if (!len)
1977                 return 0;
1978
1979         end = start + len;
1980         bytes_left = len;
1981
1982         /* Skip any superblocks on this device. */
1983         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1984                 u64 sb_start = btrfs_sb_offset(j);
1985                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1986                 u64 size = sb_start - start;
1987
1988                 if (!in_range(sb_start, start, bytes_left) &&
1989                     !in_range(sb_end, start, bytes_left) &&
1990                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1991                         continue;
1992
1993                 /*
1994                  * Superblock spans beginning of range.  Adjust start and
1995                  * try again.
1996                  */
1997                 if (sb_start <= start) {
1998                         start += sb_end - start;
1999                         if (start > end) {
2000                                 bytes_left = 0;
2001                                 break;
2002                         }
2003                         bytes_left = end - start;
2004                         continue;
2005                 }
2006
2007                 if (size) {
2008                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2009                                                    GFP_NOFS, 0);
2010                         if (!ret)
2011                                 *discarded_bytes += size;
2012                         else if (ret != -EOPNOTSUPP)
2013                                 return ret;
2014                 }
2015
2016                 start = sb_end;
2017                 if (start > end) {
2018                         bytes_left = 0;
2019                         break;
2020                 }
2021                 bytes_left = end - start;
2022         }
2023
2024         if (bytes_left) {
2025                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2026                                            GFP_NOFS, 0);
2027                 if (!ret)
2028                         *discarded_bytes += bytes_left;
2029         }
2030         return ret;
2031 }
2032
2033 int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
2034                          u64 num_bytes, u64 *actual_bytes)
2035 {
2036         int ret;
2037         u64 discarded_bytes = 0;
2038         struct btrfs_bio *bbio = NULL;
2039
2040
2041         /*
2042          * Avoid races with device replace and make sure our bbio has devices
2043          * associated to its stripes that don't go away while we are discarding.
2044          */
2045         btrfs_bio_counter_inc_blocked(root->fs_info);
2046         /* Tell the block device(s) that the sectors can be discarded */
2047         ret = btrfs_map_block(root->fs_info, REQ_OP_DISCARD,
2048                               bytenr, &num_bytes, &bbio, 0);
2049         /* Error condition is -ENOMEM */
2050         if (!ret) {
2051                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2052                 int i;
2053
2054
2055                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2056                         u64 bytes;
2057                         if (!stripe->dev->can_discard)
2058                                 continue;
2059
2060                         ret = btrfs_issue_discard(stripe->dev->bdev,
2061                                                   stripe->physical,
2062                                                   stripe->length,
2063                                                   &bytes);
2064                         if (!ret)
2065                                 discarded_bytes += bytes;
2066                         else if (ret != -EOPNOTSUPP)
2067                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2068
2069                         /*
2070                          * Just in case we get back EOPNOTSUPP for some reason,
2071                          * just ignore the return value so we don't screw up
2072                          * people calling discard_extent.
2073                          */
2074                         ret = 0;
2075                 }
2076                 btrfs_put_bbio(bbio);
2077         }
2078         btrfs_bio_counter_dec(root->fs_info);
2079
2080         if (actual_bytes)
2081                 *actual_bytes = discarded_bytes;
2082
2083
2084         if (ret == -EOPNOTSUPP)
2085                 ret = 0;
2086         return ret;
2087 }
2088
2089 /* Can return -ENOMEM */
2090 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2091                          struct btrfs_root *root,
2092                          u64 bytenr, u64 num_bytes, u64 parent,
2093                          u64 root_objectid, u64 owner, u64 offset)
2094 {
2095         int ret;
2096         struct btrfs_fs_info *fs_info = root->fs_info;
2097
2098         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2099                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2100
2101         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2102                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2103                                         num_bytes,
2104                                         parent, root_objectid, (int)owner,
2105                                         BTRFS_ADD_DELAYED_REF, NULL);
2106         } else {
2107                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2108                                         num_bytes, parent, root_objectid,
2109                                         owner, offset, 0,
2110                                         BTRFS_ADD_DELAYED_REF, NULL);
2111         }
2112         return ret;
2113 }
2114
2115 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2116                                   struct btrfs_root *root,
2117                                   struct btrfs_delayed_ref_node *node,
2118                                   u64 parent, u64 root_objectid,
2119                                   u64 owner, u64 offset, int refs_to_add,
2120                                   struct btrfs_delayed_extent_op *extent_op)
2121 {
2122         struct btrfs_fs_info *fs_info = root->fs_info;
2123         struct btrfs_path *path;
2124         struct extent_buffer *leaf;
2125         struct btrfs_extent_item *item;
2126         struct btrfs_key key;
2127         u64 bytenr = node->bytenr;
2128         u64 num_bytes = node->num_bytes;
2129         u64 refs;
2130         int ret;
2131
2132         path = btrfs_alloc_path();
2133         if (!path)
2134                 return -ENOMEM;
2135
2136         path->reada = READA_FORWARD;
2137         path->leave_spinning = 1;
2138         /* this will setup the path even if it fails to insert the back ref */
2139         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2140                                            bytenr, num_bytes, parent,
2141                                            root_objectid, owner, offset,
2142                                            refs_to_add, extent_op);
2143         if ((ret < 0 && ret != -EAGAIN) || !ret)
2144                 goto out;
2145
2146         /*
2147          * Ok we had -EAGAIN which means we didn't have space to insert and
2148          * inline extent ref, so just update the reference count and add a
2149          * normal backref.
2150          */
2151         leaf = path->nodes[0];
2152         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2153         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2154         refs = btrfs_extent_refs(leaf, item);
2155         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2156         if (extent_op)
2157                 __run_delayed_extent_op(extent_op, leaf, item);
2158
2159         btrfs_mark_buffer_dirty(leaf);
2160         btrfs_release_path(path);
2161
2162         path->reada = READA_FORWARD;
2163         path->leave_spinning = 1;
2164         /* now insert the actual backref */
2165         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2166                                     path, bytenr, parent, root_objectid,
2167                                     owner, offset, refs_to_add);
2168         if (ret)
2169                 btrfs_abort_transaction(trans, ret);
2170 out:
2171         btrfs_free_path(path);
2172         return ret;
2173 }
2174
2175 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2176                                 struct btrfs_root *root,
2177                                 struct btrfs_delayed_ref_node *node,
2178                                 struct btrfs_delayed_extent_op *extent_op,
2179                                 int insert_reserved)
2180 {
2181         int ret = 0;
2182         struct btrfs_delayed_data_ref *ref;
2183         struct btrfs_key ins;
2184         u64 parent = 0;
2185         u64 ref_root = 0;
2186         u64 flags = 0;
2187
2188         ins.objectid = node->bytenr;
2189         ins.offset = node->num_bytes;
2190         ins.type = BTRFS_EXTENT_ITEM_KEY;
2191
2192         ref = btrfs_delayed_node_to_data_ref(node);
2193         trace_run_delayed_data_ref(root->fs_info, node, ref, node->action);
2194
2195         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2196                 parent = ref->parent;
2197         ref_root = ref->root;
2198
2199         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2200                 if (extent_op)
2201                         flags |= extent_op->flags_to_set;
2202                 ret = alloc_reserved_file_extent(trans, root,
2203                                                  parent, ref_root, flags,
2204                                                  ref->objectid, ref->offset,
2205                                                  &ins, node->ref_mod);
2206         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2207                 ret = __btrfs_inc_extent_ref(trans, root, node, parent,
2208                                              ref_root, ref->objectid,
2209                                              ref->offset, node->ref_mod,
2210                                              extent_op);
2211         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2212                 ret = __btrfs_free_extent(trans, root, node, parent,
2213                                           ref_root, ref->objectid,
2214                                           ref->offset, node->ref_mod,
2215                                           extent_op);
2216         } else {
2217                 BUG();
2218         }
2219         return ret;
2220 }
2221
2222 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2223                                     struct extent_buffer *leaf,
2224                                     struct btrfs_extent_item *ei)
2225 {
2226         u64 flags = btrfs_extent_flags(leaf, ei);
2227         if (extent_op->update_flags) {
2228                 flags |= extent_op->flags_to_set;
2229                 btrfs_set_extent_flags(leaf, ei, flags);
2230         }
2231
2232         if (extent_op->update_key) {
2233                 struct btrfs_tree_block_info *bi;
2234                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2235                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2236                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2237         }
2238 }
2239
2240 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2241                                  struct btrfs_root *root,
2242                                  struct btrfs_delayed_ref_node *node,
2243                                  struct btrfs_delayed_extent_op *extent_op)
2244 {
2245         struct btrfs_key key;
2246         struct btrfs_path *path;
2247         struct btrfs_extent_item *ei;
2248         struct extent_buffer *leaf;
2249         u32 item_size;
2250         int ret;
2251         int err = 0;
2252         int metadata = !extent_op->is_data;
2253
2254         if (trans->aborted)
2255                 return 0;
2256
2257         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2258                 metadata = 0;
2259
2260         path = btrfs_alloc_path();
2261         if (!path)
2262                 return -ENOMEM;
2263
2264         key.objectid = node->bytenr;
2265
2266         if (metadata) {
2267                 key.type = BTRFS_METADATA_ITEM_KEY;
2268                 key.offset = extent_op->level;
2269         } else {
2270                 key.type = BTRFS_EXTENT_ITEM_KEY;
2271                 key.offset = node->num_bytes;
2272         }
2273
2274 again:
2275         path->reada = READA_FORWARD;
2276         path->leave_spinning = 1;
2277         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2278                                 path, 0, 1);
2279         if (ret < 0) {
2280                 err = ret;
2281                 goto out;
2282         }
2283         if (ret > 0) {
2284                 if (metadata) {
2285                         if (path->slots[0] > 0) {
2286                                 path->slots[0]--;
2287                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2288                                                       path->slots[0]);
2289                                 if (key.objectid == node->bytenr &&
2290                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2291                                     key.offset == node->num_bytes)
2292                                         ret = 0;
2293                         }
2294                         if (ret > 0) {
2295                                 btrfs_release_path(path);
2296                                 metadata = 0;
2297
2298                                 key.objectid = node->bytenr;
2299                                 key.offset = node->num_bytes;
2300                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2301                                 goto again;
2302                         }
2303                 } else {
2304                         err = -EIO;
2305                         goto out;
2306                 }
2307         }
2308
2309         leaf = path->nodes[0];
2310         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2311 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2312         if (item_size < sizeof(*ei)) {
2313                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2314                                              path, (u64)-1, 0);
2315                 if (ret < 0) {
2316                         err = ret;
2317                         goto out;
2318                 }
2319                 leaf = path->nodes[0];
2320                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2321         }
2322 #endif
2323         BUG_ON(item_size < sizeof(*ei));
2324         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2325         __run_delayed_extent_op(extent_op, leaf, ei);
2326
2327         btrfs_mark_buffer_dirty(leaf);
2328 out:
2329         btrfs_free_path(path);
2330         return err;
2331 }
2332
2333 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2334                                 struct btrfs_root *root,
2335                                 struct btrfs_delayed_ref_node *node,
2336                                 struct btrfs_delayed_extent_op *extent_op,
2337                                 int insert_reserved)
2338 {
2339         int ret = 0;
2340         struct btrfs_delayed_tree_ref *ref;
2341         struct btrfs_key ins;
2342         u64 parent = 0;
2343         u64 ref_root = 0;
2344         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2345                                                  SKINNY_METADATA);
2346
2347         ref = btrfs_delayed_node_to_tree_ref(node);
2348         trace_run_delayed_tree_ref(root->fs_info, node, ref, node->action);
2349
2350         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2351                 parent = ref->parent;
2352         ref_root = ref->root;
2353
2354         ins.objectid = node->bytenr;
2355         if (skinny_metadata) {
2356                 ins.offset = ref->level;
2357                 ins.type = BTRFS_METADATA_ITEM_KEY;
2358         } else {
2359                 ins.offset = node->num_bytes;
2360                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2361         }
2362
2363         BUG_ON(node->ref_mod != 1);
2364         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2365                 BUG_ON(!extent_op || !extent_op->update_flags);
2366                 ret = alloc_reserved_tree_block(trans, root,
2367                                                 parent, ref_root,
2368                                                 extent_op->flags_to_set,
2369                                                 &extent_op->key,
2370                                                 ref->level, &ins);
2371         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2372                 ret = __btrfs_inc_extent_ref(trans, root, node,
2373                                              parent, ref_root,
2374                                              ref->level, 0, 1,
2375                                              extent_op);
2376         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2377                 ret = __btrfs_free_extent(trans, root, node,
2378                                           parent, ref_root,
2379                                           ref->level, 0, 1, extent_op);
2380         } else {
2381                 BUG();
2382         }
2383         return ret;
2384 }
2385
2386 /* helper function to actually process a single delayed ref entry */
2387 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2388                                struct btrfs_root *root,
2389                                struct btrfs_delayed_ref_node *node,
2390                                struct btrfs_delayed_extent_op *extent_op,
2391                                int insert_reserved)
2392 {
2393         int ret = 0;
2394
2395         if (trans->aborted) {
2396                 if (insert_reserved)
2397                         btrfs_pin_extent(root, node->bytenr,
2398                                          node->num_bytes, 1);
2399                 return 0;
2400         }
2401
2402         if (btrfs_delayed_ref_is_head(node)) {
2403                 struct btrfs_delayed_ref_head *head;
2404                 /*
2405                  * we've hit the end of the chain and we were supposed
2406                  * to insert this extent into the tree.  But, it got
2407                  * deleted before we ever needed to insert it, so all
2408                  * we have to do is clean up the accounting
2409                  */
2410                 BUG_ON(extent_op);
2411                 head = btrfs_delayed_node_to_head(node);
2412                 trace_run_delayed_ref_head(root->fs_info, node, head,
2413                                            node->action);
2414
2415                 if (insert_reserved) {
2416                         btrfs_pin_extent(root, node->bytenr,
2417                                          node->num_bytes, 1);
2418                         if (head->is_data) {
2419                                 ret = btrfs_del_csums(trans, root,
2420                                                       node->bytenr,
2421                                                       node->num_bytes);
2422                         }
2423                 }
2424
2425                 /* Also free its reserved qgroup space */
2426                 btrfs_qgroup_free_delayed_ref(root->fs_info,
2427                                               head->qgroup_ref_root,
2428                                               head->qgroup_reserved);
2429                 return ret;
2430         }
2431
2432         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2433             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2434                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2435                                            insert_reserved);
2436         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2437                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2438                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2439                                            insert_reserved);
2440         else
2441                 BUG();
2442         return ret;
2443 }
2444
2445 static inline struct btrfs_delayed_ref_node *
2446 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2447 {
2448         struct btrfs_delayed_ref_node *ref;
2449
2450         if (list_empty(&head->ref_list))
2451                 return NULL;
2452
2453         /*
2454          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2455          * This is to prevent a ref count from going down to zero, which deletes
2456          * the extent item from the extent tree, when there still are references
2457          * to add, which would fail because they would not find the extent item.
2458          */
2459         list_for_each_entry(ref, &head->ref_list, list) {
2460                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2461                         return ref;
2462         }
2463
2464         return list_entry(head->ref_list.next, struct btrfs_delayed_ref_node,
2465                           list);
2466 }
2467
2468 /*
2469  * Returns 0 on success or if called with an already aborted transaction.
2470  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2471  */
2472 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2473                                              struct btrfs_root *root,
2474                                              unsigned long nr)
2475 {
2476         struct btrfs_delayed_ref_root *delayed_refs;
2477         struct btrfs_delayed_ref_node *ref;
2478         struct btrfs_delayed_ref_head *locked_ref = NULL;
2479         struct btrfs_delayed_extent_op *extent_op;
2480         struct btrfs_fs_info *fs_info = root->fs_info;
2481         ktime_t start = ktime_get();
2482         int ret;
2483         unsigned long count = 0;
2484         unsigned long actual_count = 0;
2485         int must_insert_reserved = 0;
2486
2487         delayed_refs = &trans->transaction->delayed_refs;
2488         while (1) {
2489                 if (!locked_ref) {
2490                         if (count >= nr)
2491                                 break;
2492
2493                         spin_lock(&delayed_refs->lock);
2494                         locked_ref = btrfs_select_ref_head(trans);
2495                         if (!locked_ref) {
2496                                 spin_unlock(&delayed_refs->lock);
2497                                 break;
2498                         }
2499
2500                         /* grab the lock that says we are going to process
2501                          * all the refs for this head */
2502                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2503                         spin_unlock(&delayed_refs->lock);
2504                         /*
2505                          * we may have dropped the spin lock to get the head
2506                          * mutex lock, and that might have given someone else
2507                          * time to free the head.  If that's true, it has been
2508                          * removed from our list and we can move on.
2509                          */
2510                         if (ret == -EAGAIN) {
2511                                 locked_ref = NULL;
2512                                 count++;
2513                                 continue;
2514                         }
2515                 }
2516
2517                 /*
2518                  * We need to try and merge add/drops of the same ref since we
2519                  * can run into issues with relocate dropping the implicit ref
2520                  * and then it being added back again before the drop can
2521                  * finish.  If we merged anything we need to re-loop so we can
2522                  * get a good ref.
2523                  * Or we can get node references of the same type that weren't
2524                  * merged when created due to bumps in the tree mod seq, and
2525                  * we need to merge them to prevent adding an inline extent
2526                  * backref before dropping it (triggering a BUG_ON at
2527                  * insert_inline_extent_backref()).
2528                  */
2529                 spin_lock(&locked_ref->lock);
2530                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2531                                          locked_ref);
2532
2533                 /*
2534                  * locked_ref is the head node, so we have to go one
2535                  * node back for any delayed ref updates
2536                  */
2537                 ref = select_delayed_ref(locked_ref);
2538
2539                 if (ref && ref->seq &&
2540                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2541                         spin_unlock(&locked_ref->lock);
2542                         btrfs_delayed_ref_unlock(locked_ref);
2543                         spin_lock(&delayed_refs->lock);
2544                         locked_ref->processing = 0;
2545                         delayed_refs->num_heads_ready++;
2546                         spin_unlock(&delayed_refs->lock);
2547                         locked_ref = NULL;
2548                         cond_resched();
2549                         count++;
2550                         continue;
2551                 }
2552
2553                 /*
2554                  * record the must insert reserved flag before we
2555                  * drop the spin lock.
2556                  */
2557                 must_insert_reserved = locked_ref->must_insert_reserved;
2558                 locked_ref->must_insert_reserved = 0;
2559
2560                 extent_op = locked_ref->extent_op;
2561                 locked_ref->extent_op = NULL;
2562
2563                 if (!ref) {
2564
2565
2566                         /* All delayed refs have been processed, Go ahead
2567                          * and send the head node to run_one_delayed_ref,
2568                          * so that any accounting fixes can happen
2569                          */
2570                         ref = &locked_ref->node;
2571
2572                         if (extent_op && must_insert_reserved) {
2573                                 btrfs_free_delayed_extent_op(extent_op);
2574                                 extent_op = NULL;
2575                         }
2576
2577                         if (extent_op) {
2578                                 spin_unlock(&locked_ref->lock);
2579                                 ret = run_delayed_extent_op(trans, root,
2580                                                             ref, extent_op);
2581                                 btrfs_free_delayed_extent_op(extent_op);
2582
2583                                 if (ret) {
2584                                         /*
2585                                          * Need to reset must_insert_reserved if
2586                                          * there was an error so the abort stuff
2587                                          * can cleanup the reserved space
2588                                          * properly.
2589                                          */
2590                                         if (must_insert_reserved)
2591                                                 locked_ref->must_insert_reserved = 1;
2592                                         locked_ref->processing = 0;
2593                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2594                                         btrfs_delayed_ref_unlock(locked_ref);
2595                                         return ret;
2596                                 }
2597                                 continue;
2598                         }
2599
2600                         /*
2601                          * Need to drop our head ref lock and re-acquire the
2602                          * delayed ref lock and then re-check to make sure
2603                          * nobody got added.
2604                          */
2605                         spin_unlock(&locked_ref->lock);
2606                         spin_lock(&delayed_refs->lock);
2607                         spin_lock(&locked_ref->lock);
2608                         if (!list_empty(&locked_ref->ref_list) ||
2609                             locked_ref->extent_op) {
2610                                 spin_unlock(&locked_ref->lock);
2611                                 spin_unlock(&delayed_refs->lock);
2612                                 continue;
2613                         }
2614                         ref->in_tree = 0;
2615                         delayed_refs->num_heads--;
2616                         rb_erase(&locked_ref->href_node,
2617                                  &delayed_refs->href_root);
2618                         spin_unlock(&delayed_refs->lock);
2619                 } else {
2620                         actual_count++;
2621                         ref->in_tree = 0;
2622                         list_del(&ref->list);
2623                 }
2624                 atomic_dec(&delayed_refs->num_entries);
2625
2626                 if (!btrfs_delayed_ref_is_head(ref)) {
2627                         /*
2628                          * when we play the delayed ref, also correct the
2629                          * ref_mod on head
2630                          */
2631                         switch (ref->action) {
2632                         case BTRFS_ADD_DELAYED_REF:
2633                         case BTRFS_ADD_DELAYED_EXTENT:
2634                                 locked_ref->node.ref_mod -= ref->ref_mod;
2635                                 break;
2636                         case BTRFS_DROP_DELAYED_REF:
2637                                 locked_ref->node.ref_mod += ref->ref_mod;
2638                                 break;
2639                         default:
2640                                 WARN_ON(1);
2641                         }
2642                 }
2643                 spin_unlock(&locked_ref->lock);
2644
2645                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2646                                           must_insert_reserved);
2647
2648                 btrfs_free_delayed_extent_op(extent_op);
2649                 if (ret) {
2650                         locked_ref->processing = 0;
2651                         btrfs_delayed_ref_unlock(locked_ref);
2652                         btrfs_put_delayed_ref(ref);
2653                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2654                         return ret;
2655                 }
2656
2657                 /*
2658                  * If this node is a head, that means all the refs in this head
2659                  * have been dealt with, and we will pick the next head to deal
2660                  * with, so we must unlock the head and drop it from the cluster
2661                  * list before we release it.
2662                  */
2663                 if (btrfs_delayed_ref_is_head(ref)) {
2664                         if (locked_ref->is_data &&
2665                             locked_ref->total_ref_mod < 0) {
2666                                 spin_lock(&delayed_refs->lock);
2667                                 delayed_refs->pending_csums -= ref->num_bytes;
2668                                 spin_unlock(&delayed_refs->lock);
2669                         }
2670                         btrfs_delayed_ref_unlock(locked_ref);
2671                         locked_ref = NULL;
2672                 }
2673                 btrfs_put_delayed_ref(ref);
2674                 count++;
2675                 cond_resched();
2676         }
2677
2678         /*
2679          * We don't want to include ref heads since we can have empty ref heads
2680          * and those will drastically skew our runtime down since we just do
2681          * accounting, no actual extent tree updates.
2682          */
2683         if (actual_count > 0) {
2684                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2685                 u64 avg;
2686
2687                 /*
2688                  * We weigh the current average higher than our current runtime
2689                  * to avoid large swings in the average.
2690                  */
2691                 spin_lock(&delayed_refs->lock);
2692                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2693                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2694                 spin_unlock(&delayed_refs->lock);
2695         }
2696         return 0;
2697 }
2698
2699 #ifdef SCRAMBLE_DELAYED_REFS
2700 /*
2701  * Normally delayed refs get processed in ascending bytenr order. This
2702  * correlates in most cases to the order added. To expose dependencies on this
2703  * order, we start to process the tree in the middle instead of the beginning
2704  */
2705 static u64 find_middle(struct rb_root *root)
2706 {
2707         struct rb_node *n = root->rb_node;
2708         struct btrfs_delayed_ref_node *entry;
2709         int alt = 1;
2710         u64 middle;
2711         u64 first = 0, last = 0;
2712
2713         n = rb_first(root);
2714         if (n) {
2715                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2716                 first = entry->bytenr;
2717         }
2718         n = rb_last(root);
2719         if (n) {
2720                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2721                 last = entry->bytenr;
2722         }
2723         n = root->rb_node;
2724
2725         while (n) {
2726                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2727                 WARN_ON(!entry->in_tree);
2728
2729                 middle = entry->bytenr;
2730
2731                 if (alt)
2732                         n = n->rb_left;
2733                 else
2734                         n = n->rb_right;
2735
2736                 alt = 1 - alt;
2737         }
2738         return middle;
2739 }
2740 #endif
2741
2742 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2743 {
2744         u64 num_bytes;
2745
2746         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2747                              sizeof(struct btrfs_extent_inline_ref));
2748         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2749                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2750
2751         /*
2752          * We don't ever fill up leaves all the way so multiply by 2 just to be
2753          * closer to what we're really going to want to use.
2754          */
2755         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2756 }
2757
2758 /*
2759  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2760  * would require to store the csums for that many bytes.
2761  */
2762 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2763 {
2764         u64 csum_size;
2765         u64 num_csums_per_leaf;
2766         u64 num_csums;
2767
2768         csum_size = BTRFS_MAX_ITEM_SIZE(root);
2769         num_csums_per_leaf = div64_u64(csum_size,
2770                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2771         num_csums = div64_u64(csum_bytes, root->sectorsize);
2772         num_csums += num_csums_per_leaf - 1;
2773         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2774         return num_csums;
2775 }
2776
2777 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2778                                        struct btrfs_root *root)
2779 {
2780         struct btrfs_block_rsv *global_rsv;
2781         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2782         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2783         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2784         u64 num_bytes, num_dirty_bgs_bytes;
2785         int ret = 0;
2786
2787         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2788         num_heads = heads_to_leaves(root, num_heads);
2789         if (num_heads > 1)
2790                 num_bytes += (num_heads - 1) * root->nodesize;
2791         num_bytes <<= 1;
2792         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2793         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2794                                                              num_dirty_bgs);
2795         global_rsv = &root->fs_info->global_block_rsv;
2796
2797         /*
2798          * If we can't allocate any more chunks lets make sure we have _lots_ of
2799          * wiggle room since running delayed refs can create more delayed refs.
2800          */
2801         if (global_rsv->space_info->full) {
2802                 num_dirty_bgs_bytes <<= 1;
2803                 num_bytes <<= 1;
2804         }
2805
2806         spin_lock(&global_rsv->lock);
2807         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2808                 ret = 1;
2809         spin_unlock(&global_rsv->lock);
2810         return ret;
2811 }
2812
2813 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2814                                        struct btrfs_root *root)
2815 {
2816         struct btrfs_fs_info *fs_info = root->fs_info;
2817         u64 num_entries =
2818                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2819         u64 avg_runtime;
2820         u64 val;
2821
2822         smp_mb();
2823         avg_runtime = fs_info->avg_delayed_ref_runtime;
2824         val = num_entries * avg_runtime;
2825         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2826                 return 1;
2827         if (val >= NSEC_PER_SEC / 2)
2828                 return 2;
2829
2830         return btrfs_check_space_for_delayed_refs(trans, root);
2831 }
2832
2833 struct async_delayed_refs {
2834         struct btrfs_root *root;
2835         u64 transid;
2836         int count;
2837         int error;
2838         int sync;
2839         struct completion wait;
2840         struct btrfs_work work;
2841 };
2842
2843 static void delayed_ref_async_start(struct btrfs_work *work)
2844 {
2845         struct async_delayed_refs *async;
2846         struct btrfs_trans_handle *trans;
2847         int ret;
2848
2849         async = container_of(work, struct async_delayed_refs, work);
2850
2851         /* if the commit is already started, we don't need to wait here */
2852         if (btrfs_transaction_blocked(async->root->fs_info))
2853                 goto done;
2854
2855         trans = btrfs_join_transaction(async->root);
2856         if (IS_ERR(trans)) {
2857                 async->error = PTR_ERR(trans);
2858                 goto done;
2859         }
2860
2861         /*
2862          * trans->sync means that when we call end_transaction, we won't
2863          * wait on delayed refs
2864          */
2865         trans->sync = true;
2866
2867         /* Don't bother flushing if we got into a different transaction */
2868         if (trans->transid > async->transid)
2869                 goto end;
2870
2871         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2872         if (ret)
2873                 async->error = ret;
2874 end:
2875         ret = btrfs_end_transaction(trans, async->root);
2876         if (ret && !async->error)
2877                 async->error = ret;
2878 done:
2879         if (async->sync)
2880                 complete(&async->wait);
2881         else
2882                 kfree(async);
2883 }
2884
2885 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2886                                  unsigned long count, u64 transid, int wait)
2887 {
2888         struct async_delayed_refs *async;
2889         int ret;
2890
2891         async = kmalloc(sizeof(*async), GFP_NOFS);
2892         if (!async)
2893                 return -ENOMEM;
2894
2895         async->root = root->fs_info->tree_root;
2896         async->count = count;
2897         async->error = 0;
2898         async->transid = transid;
2899         if (wait)
2900                 async->sync = 1;
2901         else
2902                 async->sync = 0;
2903         init_completion(&async->wait);
2904
2905         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2906                         delayed_ref_async_start, NULL, NULL);
2907
2908         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2909
2910         if (wait) {
2911                 wait_for_completion(&async->wait);
2912                 ret = async->error;
2913                 kfree(async);
2914                 return ret;
2915         }
2916         return 0;
2917 }
2918
2919 /*
2920  * this starts processing the delayed reference count updates and
2921  * extent insertions we have queued up so far.  count can be
2922  * 0, which means to process everything in the tree at the start
2923  * of the run (but not newly added entries), or it can be some target
2924  * number you'd like to process.
2925  *
2926  * Returns 0 on success or if called with an aborted transaction
2927  * Returns <0 on error and aborts the transaction
2928  */
2929 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2930                            struct btrfs_root *root, unsigned long count)
2931 {
2932         struct rb_node *node;
2933         struct btrfs_delayed_ref_root *delayed_refs;
2934         struct btrfs_delayed_ref_head *head;
2935         int ret;
2936         int run_all = count == (unsigned long)-1;
2937         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2938
2939         /* We'll clean this up in btrfs_cleanup_transaction */
2940         if (trans->aborted)
2941                 return 0;
2942
2943         if (root->fs_info->creating_free_space_tree)
2944                 return 0;
2945
2946         if (root == root->fs_info->extent_root)
2947                 root = root->fs_info->tree_root;
2948
2949         delayed_refs = &trans->transaction->delayed_refs;
2950         if (count == 0)
2951                 count = atomic_read(&delayed_refs->num_entries) * 2;
2952
2953 again:
2954 #ifdef SCRAMBLE_DELAYED_REFS
2955         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2956 #endif
2957         trans->can_flush_pending_bgs = false;
2958         ret = __btrfs_run_delayed_refs(trans, root, count);
2959         if (ret < 0) {
2960                 btrfs_abort_transaction(trans, ret);
2961                 return ret;
2962         }
2963
2964         if (run_all) {
2965                 if (!list_empty(&trans->new_bgs))
2966                         btrfs_create_pending_block_groups(trans, root);
2967
2968                 spin_lock(&delayed_refs->lock);
2969                 node = rb_first(&delayed_refs->href_root);
2970                 if (!node) {
2971                         spin_unlock(&delayed_refs->lock);
2972                         goto out;
2973                 }
2974                 count = (unsigned long)-1;
2975
2976                 while (node) {
2977                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2978                                         href_node);
2979                         if (btrfs_delayed_ref_is_head(&head->node)) {
2980                                 struct btrfs_delayed_ref_node *ref;
2981
2982                                 ref = &head->node;
2983                                 atomic_inc(&ref->refs);
2984
2985                                 spin_unlock(&delayed_refs->lock);
2986                                 /*
2987                                  * Mutex was contended, block until it's
2988                                  * released and try again
2989                                  */
2990                                 mutex_lock(&head->mutex);
2991                                 mutex_unlock(&head->mutex);
2992
2993                                 btrfs_put_delayed_ref(ref);
2994                                 cond_resched();
2995                                 goto again;
2996                         } else {
2997                                 WARN_ON(1);
2998                         }
2999                         node = rb_next(node);
3000                 }
3001                 spin_unlock(&delayed_refs->lock);
3002                 cond_resched();
3003                 goto again;
3004         }
3005 out:
3006         assert_qgroups_uptodate(trans);
3007         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3008         return 0;
3009 }
3010
3011 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3012                                 struct btrfs_root *root,
3013                                 u64 bytenr, u64 num_bytes, u64 flags,
3014                                 int level, int is_data)
3015 {
3016         struct btrfs_delayed_extent_op *extent_op;
3017         int ret;
3018
3019         extent_op = btrfs_alloc_delayed_extent_op();
3020         if (!extent_op)
3021                 return -ENOMEM;
3022
3023         extent_op->flags_to_set = flags;
3024         extent_op->update_flags = true;
3025         extent_op->update_key = false;
3026         extent_op->is_data = is_data ? true : false;
3027         extent_op->level = level;
3028
3029         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
3030                                           num_bytes, extent_op);
3031         if (ret)
3032                 btrfs_free_delayed_extent_op(extent_op);
3033         return ret;
3034 }
3035
3036 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
3037                                       struct btrfs_root *root,
3038                                       struct btrfs_path *path,
3039                                       u64 objectid, u64 offset, u64 bytenr)
3040 {
3041         struct btrfs_delayed_ref_head *head;
3042         struct btrfs_delayed_ref_node *ref;
3043         struct btrfs_delayed_data_ref *data_ref;
3044         struct btrfs_delayed_ref_root *delayed_refs;
3045         int ret = 0;
3046
3047         delayed_refs = &trans->transaction->delayed_refs;
3048         spin_lock(&delayed_refs->lock);
3049         head = btrfs_find_delayed_ref_head(trans, bytenr);
3050         if (!head) {
3051                 spin_unlock(&delayed_refs->lock);
3052                 return 0;
3053         }
3054
3055         if (!mutex_trylock(&head->mutex)) {
3056                 atomic_inc(&head->node.refs);
3057                 spin_unlock(&delayed_refs->lock);
3058
3059                 btrfs_release_path(path);
3060
3061                 /*
3062                  * Mutex was contended, block until it's released and let
3063                  * caller try again
3064                  */
3065                 mutex_lock(&head->mutex);
3066                 mutex_unlock(&head->mutex);
3067                 btrfs_put_delayed_ref(&head->node);
3068                 return -EAGAIN;
3069         }
3070         spin_unlock(&delayed_refs->lock);
3071
3072         spin_lock(&head->lock);
3073         list_for_each_entry(ref, &head->ref_list, list) {
3074                 /* If it's a shared ref we know a cross reference exists */
3075                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3076                         ret = 1;
3077                         break;
3078                 }
3079
3080                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3081
3082                 /*
3083                  * If our ref doesn't match the one we're currently looking at
3084                  * then we have a cross reference.
3085                  */
3086                 if (data_ref->root != root->root_key.objectid ||
3087                     data_ref->objectid != objectid ||
3088                     data_ref->offset != offset) {
3089                         ret = 1;
3090                         break;
3091                 }
3092         }
3093         spin_unlock(&head->lock);
3094         mutex_unlock(&head->mutex);
3095         return ret;
3096 }
3097
3098 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
3099                                         struct btrfs_root *root,
3100                                         struct btrfs_path *path,
3101                                         u64 objectid, u64 offset, u64 bytenr)
3102 {
3103         struct btrfs_root *extent_root = root->fs_info->extent_root;
3104         struct extent_buffer *leaf;
3105         struct btrfs_extent_data_ref *ref;
3106         struct btrfs_extent_inline_ref *iref;
3107         struct btrfs_extent_item *ei;
3108         struct btrfs_key key;
3109         u32 item_size;
3110         int ret;
3111
3112         key.objectid = bytenr;
3113         key.offset = (u64)-1;
3114         key.type = BTRFS_EXTENT_ITEM_KEY;
3115
3116         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3117         if (ret < 0)
3118                 goto out;
3119         BUG_ON(ret == 0); /* Corruption */
3120
3121         ret = -ENOENT;
3122         if (path->slots[0] == 0)
3123                 goto out;
3124
3125         path->slots[0]--;
3126         leaf = path->nodes[0];
3127         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3128
3129         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3130                 goto out;
3131
3132         ret = 1;
3133         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3134 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3135         if (item_size < sizeof(*ei)) {
3136                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3137                 goto out;
3138         }
3139 #endif
3140         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3141
3142         if (item_size != sizeof(*ei) +
3143             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3144                 goto out;
3145
3146         if (btrfs_extent_generation(leaf, ei) <=
3147             btrfs_root_last_snapshot(&root->root_item))
3148                 goto out;
3149
3150         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3151         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3152             BTRFS_EXTENT_DATA_REF_KEY)
3153                 goto out;
3154
3155         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3156         if (btrfs_extent_refs(leaf, ei) !=
3157             btrfs_extent_data_ref_count(leaf, ref) ||
3158             btrfs_extent_data_ref_root(leaf, ref) !=
3159             root->root_key.objectid ||
3160             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3161             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3162                 goto out;
3163
3164         ret = 0;
3165 out:
3166         return ret;
3167 }
3168
3169 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3170                           struct btrfs_root *root,
3171                           u64 objectid, u64 offset, u64 bytenr)
3172 {
3173         struct btrfs_path *path;
3174         int ret;
3175         int ret2;
3176
3177         path = btrfs_alloc_path();
3178         if (!path)
3179                 return -ENOENT;
3180
3181         do {
3182                 ret = check_committed_ref(trans, root, path, objectid,
3183                                           offset, bytenr);
3184                 if (ret && ret != -ENOENT)
3185                         goto out;
3186
3187                 ret2 = check_delayed_ref(trans, root, path, objectid,
3188                                          offset, bytenr);
3189         } while (ret2 == -EAGAIN);
3190
3191         if (ret2 && ret2 != -ENOENT) {
3192                 ret = ret2;
3193                 goto out;
3194         }
3195
3196         if (ret != -ENOENT || ret2 != -ENOENT)
3197                 ret = 0;
3198 out:
3199         btrfs_free_path(path);
3200         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3201                 WARN_ON(ret > 0);
3202         return ret;
3203 }
3204
3205 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3206                            struct btrfs_root *root,
3207                            struct extent_buffer *buf,
3208                            int full_backref, int inc)
3209 {
3210         u64 bytenr;
3211         u64 num_bytes;
3212         u64 parent;
3213         u64 ref_root;
3214         u32 nritems;
3215         struct btrfs_key key;
3216         struct btrfs_file_extent_item *fi;
3217         int i;
3218         int level;
3219         int ret = 0;
3220         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3221                             u64, u64, u64, u64, u64, u64);
3222
3223
3224         if (btrfs_is_testing(root->fs_info))
3225                 return 0;
3226
3227         ref_root = btrfs_header_owner(buf);
3228         nritems = btrfs_header_nritems(buf);
3229         level = btrfs_header_level(buf);
3230
3231         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3232                 return 0;
3233
3234         if (inc)
3235                 process_func = btrfs_inc_extent_ref;
3236         else
3237                 process_func = btrfs_free_extent;
3238
3239         if (full_backref)
3240                 parent = buf->start;
3241         else
3242                 parent = 0;
3243
3244         for (i = 0; i < nritems; i++) {
3245                 if (level == 0) {
3246                         btrfs_item_key_to_cpu(buf, &key, i);
3247                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3248                                 continue;
3249                         fi = btrfs_item_ptr(buf, i,
3250                                             struct btrfs_file_extent_item);
3251                         if (btrfs_file_extent_type(buf, fi) ==
3252                             BTRFS_FILE_EXTENT_INLINE)
3253                                 continue;
3254                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3255                         if (bytenr == 0)
3256                                 continue;
3257
3258                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3259                         key.offset -= btrfs_file_extent_offset(buf, fi);
3260                         ret = process_func(trans, root, bytenr, num_bytes,
3261                                            parent, ref_root, key.objectid,
3262                                            key.offset);
3263                         if (ret)
3264                                 goto fail;
3265                 } else {
3266                         bytenr = btrfs_node_blockptr(buf, i);
3267                         num_bytes = root->nodesize;
3268                         ret = process_func(trans, root, bytenr, num_bytes,
3269                                            parent, ref_root, level - 1, 0);
3270                         if (ret)
3271                                 goto fail;
3272                 }
3273         }
3274         return 0;
3275 fail:
3276         return ret;
3277 }
3278
3279 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3280                   struct extent_buffer *buf, int full_backref)
3281 {
3282         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3283 }
3284
3285 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3286                   struct extent_buffer *buf, int full_backref)
3287 {
3288         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3289 }
3290
3291 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3292                                  struct btrfs_root *root,
3293                                  struct btrfs_path *path,
3294                                  struct btrfs_block_group_cache *cache)
3295 {
3296         int ret;
3297         struct btrfs_root *extent_root = root->fs_info->extent_root;
3298         unsigned long bi;
3299         struct extent_buffer *leaf;
3300
3301         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3302         if (ret) {
3303                 if (ret > 0)
3304                         ret = -ENOENT;
3305                 goto fail;
3306         }
3307
3308         leaf = path->nodes[0];
3309         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3310         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3311         btrfs_mark_buffer_dirty(leaf);
3312 fail:
3313         btrfs_release_path(path);
3314         return ret;
3315
3316 }
3317
3318 static struct btrfs_block_group_cache *
3319 next_block_group(struct btrfs_root *root,
3320                  struct btrfs_block_group_cache *cache)
3321 {
3322         struct rb_node *node;
3323
3324         spin_lock(&root->fs_info->block_group_cache_lock);
3325
3326         /* If our block group was removed, we need a full search. */
3327         if (RB_EMPTY_NODE(&cache->cache_node)) {
3328                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3329
3330                 spin_unlock(&root->fs_info->block_group_cache_lock);
3331                 btrfs_put_block_group(cache);
3332                 cache = btrfs_lookup_first_block_group(root->fs_info,
3333                                                        next_bytenr);
3334                 return cache;
3335         }
3336         node = rb_next(&cache->cache_node);
3337         btrfs_put_block_group(cache);
3338         if (node) {
3339                 cache = rb_entry(node, struct btrfs_block_group_cache,
3340                                  cache_node);
3341                 btrfs_get_block_group(cache);
3342         } else
3343                 cache = NULL;
3344         spin_unlock(&root->fs_info->block_group_cache_lock);
3345         return cache;
3346 }
3347
3348 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3349                             struct btrfs_trans_handle *trans,
3350                             struct btrfs_path *path)
3351 {
3352         struct btrfs_root *root = block_group->fs_info->tree_root;
3353         struct inode *inode = NULL;
3354         u64 alloc_hint = 0;
3355         int dcs = BTRFS_DC_ERROR;
3356         u64 num_pages = 0;
3357         int retries = 0;
3358         int ret = 0;
3359
3360         /*
3361          * If this block group is smaller than 100 megs don't bother caching the
3362          * block group.
3363          */
3364         if (block_group->key.offset < (100 * SZ_1M)) {
3365                 spin_lock(&block_group->lock);
3366                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3367                 spin_unlock(&block_group->lock);
3368                 return 0;
3369         }
3370
3371         if (trans->aborted)
3372                 return 0;
3373 again:
3374         inode = lookup_free_space_inode(root, block_group, path);
3375         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3376                 ret = PTR_ERR(inode);
3377                 btrfs_release_path(path);
3378                 goto out;
3379         }
3380
3381         if (IS_ERR(inode)) {
3382                 BUG_ON(retries);
3383                 retries++;
3384
3385                 if (block_group->ro)
3386                         goto out_free;
3387
3388                 ret = create_free_space_inode(root, trans, block_group, path);
3389                 if (ret)
3390                         goto out_free;
3391                 goto again;
3392         }
3393
3394         /* We've already setup this transaction, go ahead and exit */
3395         if (block_group->cache_generation == trans->transid &&
3396             i_size_read(inode)) {
3397                 dcs = BTRFS_DC_SETUP;
3398                 goto out_put;
3399         }
3400
3401         /*
3402          * We want to set the generation to 0, that way if anything goes wrong
3403          * from here on out we know not to trust this cache when we load up next
3404          * time.
3405          */
3406         BTRFS_I(inode)->generation = 0;
3407         ret = btrfs_update_inode(trans, root, inode);
3408         if (ret) {
3409                 /*
3410                  * So theoretically we could recover from this, simply set the
3411                  * super cache generation to 0 so we know to invalidate the
3412                  * cache, but then we'd have to keep track of the block groups
3413                  * that fail this way so we know we _have_ to reset this cache
3414                  * before the next commit or risk reading stale cache.  So to
3415                  * limit our exposure to horrible edge cases lets just abort the
3416                  * transaction, this only happens in really bad situations
3417                  * anyway.
3418                  */
3419                 btrfs_abort_transaction(trans, ret);
3420                 goto out_put;
3421         }
3422         WARN_ON(ret);
3423
3424         if (i_size_read(inode) > 0) {
3425                 ret = btrfs_check_trunc_cache_free_space(root,
3426                                         &root->fs_info->global_block_rsv);
3427                 if (ret)
3428                         goto out_put;
3429
3430                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3431                 if (ret)
3432                         goto out_put;
3433         }
3434
3435         spin_lock(&block_group->lock);
3436         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3437             !btrfs_test_opt(root->fs_info, SPACE_CACHE)) {
3438                 /*
3439                  * don't bother trying to write stuff out _if_
3440                  * a) we're not cached,
3441                  * b) we're with nospace_cache mount option.
3442                  */
3443                 dcs = BTRFS_DC_WRITTEN;
3444                 spin_unlock(&block_group->lock);
3445                 goto out_put;
3446         }
3447         spin_unlock(&block_group->lock);
3448
3449         /*
3450          * We hit an ENOSPC when setting up the cache in this transaction, just
3451          * skip doing the setup, we've already cleared the cache so we're safe.
3452          */
3453         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3454                 ret = -ENOSPC;
3455                 goto out_put;
3456         }
3457
3458         /*
3459          * Try to preallocate enough space based on how big the block group is.
3460          * Keep in mind this has to include any pinned space which could end up
3461          * taking up quite a bit since it's not folded into the other space
3462          * cache.
3463          */
3464         num_pages = div_u64(block_group->key.offset, SZ_256M);
3465         if (!num_pages)
3466                 num_pages = 1;
3467
3468         num_pages *= 16;
3469         num_pages *= PAGE_SIZE;
3470
3471         ret = btrfs_check_data_free_space(inode, 0, num_pages);
3472         if (ret)
3473                 goto out_put;
3474
3475         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3476                                               num_pages, num_pages,
3477                                               &alloc_hint);
3478         /*
3479          * Our cache requires contiguous chunks so that we don't modify a bunch
3480          * of metadata or split extents when writing the cache out, which means
3481          * we can enospc if we are heavily fragmented in addition to just normal
3482          * out of space conditions.  So if we hit this just skip setting up any
3483          * other block groups for this transaction, maybe we'll unpin enough
3484          * space the next time around.
3485          */
3486         if (!ret)
3487                 dcs = BTRFS_DC_SETUP;
3488         else if (ret == -ENOSPC)
3489                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3490
3491 out_put:
3492         iput(inode);
3493 out_free:
3494         btrfs_release_path(path);
3495 out:
3496         spin_lock(&block_group->lock);
3497         if (!ret && dcs == BTRFS_DC_SETUP)
3498                 block_group->cache_generation = trans->transid;
3499         block_group->disk_cache_state = dcs;
3500         spin_unlock(&block_group->lock);
3501
3502         return ret;
3503 }
3504
3505 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3506                             struct btrfs_root *root)
3507 {
3508         struct btrfs_block_group_cache *cache, *tmp;
3509         struct btrfs_transaction *cur_trans = trans->transaction;
3510         struct btrfs_path *path;
3511
3512         if (list_empty(&cur_trans->dirty_bgs) ||
3513             !btrfs_test_opt(root->fs_info, SPACE_CACHE))
3514                 return 0;
3515
3516         path = btrfs_alloc_path();
3517         if (!path)
3518                 return -ENOMEM;
3519
3520         /* Could add new block groups, use _safe just in case */
3521         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3522                                  dirty_list) {
3523                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3524                         cache_save_setup(cache, trans, path);
3525         }
3526
3527         btrfs_free_path(path);
3528         return 0;
3529 }
3530
3531 /*
3532  * transaction commit does final block group cache writeback during a
3533  * critical section where nothing is allowed to change the FS.  This is
3534  * required in order for the cache to actually match the block group,
3535  * but can introduce a lot of latency into the commit.
3536  *
3537  * So, btrfs_start_dirty_block_groups is here to kick off block group
3538  * cache IO.  There's a chance we'll have to redo some of it if the
3539  * block group changes again during the commit, but it greatly reduces
3540  * the commit latency by getting rid of the easy block groups while
3541  * we're still allowing others to join the commit.
3542  */
3543 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3544                                    struct btrfs_root *root)
3545 {
3546         struct btrfs_block_group_cache *cache;
3547         struct btrfs_transaction *cur_trans = trans->transaction;
3548         int ret = 0;
3549         int should_put;
3550         struct btrfs_path *path = NULL;
3551         LIST_HEAD(dirty);
3552         struct list_head *io = &cur_trans->io_bgs;
3553         int num_started = 0;
3554         int loops = 0;
3555
3556         spin_lock(&cur_trans->dirty_bgs_lock);
3557         if (list_empty(&cur_trans->dirty_bgs)) {
3558                 spin_unlock(&cur_trans->dirty_bgs_lock);
3559                 return 0;
3560         }
3561         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3562         spin_unlock(&cur_trans->dirty_bgs_lock);
3563
3564 again:
3565         /*
3566          * make sure all the block groups on our dirty list actually
3567          * exist
3568          */
3569         btrfs_create_pending_block_groups(trans, root);
3570
3571         if (!path) {
3572                 path = btrfs_alloc_path();
3573                 if (!path)
3574                         return -ENOMEM;
3575         }
3576
3577         /*
3578          * cache_write_mutex is here only to save us from balance or automatic
3579          * removal of empty block groups deleting this block group while we are
3580          * writing out the cache
3581          */
3582         mutex_lock(&trans->transaction->cache_write_mutex);
3583         while (!list_empty(&dirty)) {
3584                 cache = list_first_entry(&dirty,
3585                                          struct btrfs_block_group_cache,
3586                                          dirty_list);
3587                 /*
3588                  * this can happen if something re-dirties a block
3589                  * group that is already under IO.  Just wait for it to
3590                  * finish and then do it all again
3591                  */
3592                 if (!list_empty(&cache->io_list)) {
3593                         list_del_init(&cache->io_list);
3594                         btrfs_wait_cache_io(root, trans, cache,
3595                                             &cache->io_ctl, path,
3596                                             cache->key.objectid);
3597                         btrfs_put_block_group(cache);
3598                 }
3599
3600
3601                 /*
3602                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3603                  * if it should update the cache_state.  Don't delete
3604                  * until after we wait.
3605                  *
3606                  * Since we're not running in the commit critical section
3607                  * we need the dirty_bgs_lock to protect from update_block_group
3608                  */
3609                 spin_lock(&cur_trans->dirty_bgs_lock);
3610                 list_del_init(&cache->dirty_list);
3611                 spin_unlock(&cur_trans->dirty_bgs_lock);
3612
3613                 should_put = 1;
3614
3615                 cache_save_setup(cache, trans, path);
3616
3617                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3618                         cache->io_ctl.inode = NULL;
3619                         ret = btrfs_write_out_cache(root, trans, cache, path);
3620                         if (ret == 0 && cache->io_ctl.inode) {
3621                                 num_started++;
3622                                 should_put = 0;
3623
3624                                 /*
3625                                  * the cache_write_mutex is protecting
3626                                  * the io_list
3627                                  */
3628                                 list_add_tail(&cache->io_list, io);
3629                         } else {
3630                                 /*
3631                                  * if we failed to write the cache, the
3632                                  * generation will be bad and life goes on
3633                                  */
3634                                 ret = 0;
3635                         }
3636                 }
3637                 if (!ret) {
3638                         ret = write_one_cache_group(trans, root, path, cache);
3639                         /*
3640                          * Our block group might still be attached to the list
3641                          * of new block groups in the transaction handle of some
3642                          * other task (struct btrfs_trans_handle->new_bgs). This
3643                          * means its block group item isn't yet in the extent
3644                          * tree. If this happens ignore the error, as we will
3645                          * try again later in the critical section of the
3646                          * transaction commit.
3647                          */
3648                         if (ret == -ENOENT) {
3649                                 ret = 0;
3650                                 spin_lock(&cur_trans->dirty_bgs_lock);
3651                                 if (list_empty(&cache->dirty_list)) {
3652                                         list_add_tail(&cache->dirty_list,
3653                                                       &cur_trans->dirty_bgs);
3654                                         btrfs_get_block_group(cache);
3655                                 }
3656                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3657                         } else if (ret) {
3658                                 btrfs_abort_transaction(trans, ret);
3659                         }
3660                 }
3661
3662                 /* if its not on the io list, we need to put the block group */
3663                 if (should_put)
3664                         btrfs_put_block_group(cache);
3665
3666                 if (ret)
3667                         break;
3668
3669                 /*
3670                  * Avoid blocking other tasks for too long. It might even save
3671                  * us from writing caches for block groups that are going to be
3672                  * removed.
3673                  */
3674                 mutex_unlock(&trans->transaction->cache_write_mutex);
3675                 mutex_lock(&trans->transaction->cache_write_mutex);
3676         }
3677         mutex_unlock(&trans->transaction->cache_write_mutex);
3678
3679         /*
3680          * go through delayed refs for all the stuff we've just kicked off
3681          * and then loop back (just once)
3682          */
3683         ret = btrfs_run_delayed_refs(trans, root, 0);
3684         if (!ret && loops == 0) {
3685                 loops++;
3686                 spin_lock(&cur_trans->dirty_bgs_lock);
3687                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3688                 /*
3689                  * dirty_bgs_lock protects us from concurrent block group
3690                  * deletes too (not just cache_write_mutex).
3691                  */
3692                 if (!list_empty(&dirty)) {
3693                         spin_unlock(&cur_trans->dirty_bgs_lock);
3694                         goto again;
3695                 }
3696                 spin_unlock(&cur_trans->dirty_bgs_lock);
3697         }
3698
3699         btrfs_free_path(path);
3700         return ret;
3701 }
3702
3703 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3704                                    struct btrfs_root *root)
3705 {
3706         struct btrfs_block_group_cache *cache;
3707         struct btrfs_transaction *cur_trans = trans->transaction;
3708         int ret = 0;
3709         int should_put;
3710         struct btrfs_path *path;
3711         struct list_head *io = &cur_trans->io_bgs;
3712         int num_started = 0;
3713
3714         path = btrfs_alloc_path();
3715         if (!path)
3716                 return -ENOMEM;
3717
3718         /*
3719          * Even though we are in the critical section of the transaction commit,
3720          * we can still have concurrent tasks adding elements to this
3721          * transaction's list of dirty block groups. These tasks correspond to
3722          * endio free space workers started when writeback finishes for a
3723          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3724          * allocate new block groups as a result of COWing nodes of the root
3725          * tree when updating the free space inode. The writeback for the space
3726          * caches is triggered by an earlier call to
3727          * btrfs_start_dirty_block_groups() and iterations of the following
3728          * loop.
3729          * Also we want to do the cache_save_setup first and then run the
3730          * delayed refs to make sure we have the best chance at doing this all
3731          * in one shot.
3732          */
3733         spin_lock(&cur_trans->dirty_bgs_lock);
3734         while (!list_empty(&cur_trans->dirty_bgs)) {
3735                 cache = list_first_entry(&cur_trans->dirty_bgs,
3736                                          struct btrfs_block_group_cache,
3737                                          dirty_list);
3738
3739                 /*
3740                  * this can happen if cache_save_setup re-dirties a block
3741                  * group that is already under IO.  Just wait for it to
3742                  * finish and then do it all again
3743                  */
3744                 if (!list_empty(&cache->io_list)) {
3745                         spin_unlock(&cur_trans->dirty_bgs_lock);
3746                         list_del_init(&cache->io_list);
3747                         btrfs_wait_cache_io(root, trans, cache,
3748                                             &cache->io_ctl, path,
3749                                             cache->key.objectid);
3750                         btrfs_put_block_group(cache);
3751                         spin_lock(&cur_trans->dirty_bgs_lock);
3752                 }
3753
3754                 /*
3755                  * don't remove from the dirty list until after we've waited
3756                  * on any pending IO
3757                  */
3758                 list_del_init(&cache->dirty_list);
3759                 spin_unlock(&cur_trans->dirty_bgs_lock);
3760                 should_put = 1;
3761
3762                 cache_save_setup(cache, trans, path);
3763
3764                 if (!ret)
3765                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3766
3767                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3768                         cache->io_ctl.inode = NULL;
3769                         ret = btrfs_write_out_cache(root, trans, cache, path);
3770                         if (ret == 0 && cache->io_ctl.inode) {
3771                                 num_started++;
3772                                 should_put = 0;
3773                                 list_add_tail(&cache->io_list, io);
3774                         } else {
3775                                 /*
3776                                  * if we failed to write the cache, the
3777                                  * generation will be bad and life goes on
3778                                  */
3779                                 ret = 0;
3780                         }
3781                 }
3782                 if (!ret) {
3783                         ret = write_one_cache_group(trans, root, path, cache);
3784                         /*
3785                          * One of the free space endio workers might have
3786                          * created a new block group while updating a free space
3787                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3788                          * and hasn't released its transaction handle yet, in
3789                          * which case the new block group is still attached to
3790                          * its transaction handle and its creation has not
3791                          * finished yet (no block group item in the extent tree
3792                          * yet, etc). If this is the case, wait for all free
3793                          * space endio workers to finish and retry. This is a
3794                          * a very rare case so no need for a more efficient and
3795                          * complex approach.
3796                          */
3797                         if (ret == -ENOENT) {
3798                                 wait_event(cur_trans->writer_wait,
3799                                    atomic_read(&cur_trans->num_writers) == 1);
3800                                 ret = write_one_cache_group(trans, root, path,
3801                                                             cache);
3802                         }
3803                         if (ret)
3804                                 btrfs_abort_transaction(trans, ret);
3805                 }
3806
3807                 /* if its not on the io list, we need to put the block group */
3808                 if (should_put)
3809                         btrfs_put_block_group(cache);
3810                 spin_lock(&cur_trans->dirty_bgs_lock);
3811         }
3812         spin_unlock(&cur_trans->dirty_bgs_lock);
3813
3814         while (!list_empty(io)) {
3815                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3816                                          io_list);
3817                 list_del_init(&cache->io_list);
3818                 btrfs_wait_cache_io(root, trans, cache,
3819                                     &cache->io_ctl, path, cache->key.objectid);
3820                 btrfs_put_block_group(cache);
3821         }
3822
3823         btrfs_free_path(path);
3824         return ret;
3825 }
3826
3827 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3828 {
3829         struct btrfs_block_group_cache *block_group;
3830         int readonly = 0;
3831
3832         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3833         if (!block_group || block_group->ro)
3834                 readonly = 1;
3835         if (block_group)
3836                 btrfs_put_block_group(block_group);
3837         return readonly;
3838 }
3839
3840 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3841 {
3842         struct btrfs_block_group_cache *bg;
3843         bool ret = true;
3844
3845         bg = btrfs_lookup_block_group(fs_info, bytenr);
3846         if (!bg)
3847                 return false;
3848
3849         spin_lock(&bg->lock);
3850         if (bg->ro)
3851                 ret = false;
3852         else
3853                 atomic_inc(&bg->nocow_writers);
3854         spin_unlock(&bg->lock);
3855
3856         /* no put on block group, done by btrfs_dec_nocow_writers */
3857         if (!ret)
3858                 btrfs_put_block_group(bg);
3859
3860         return ret;
3861
3862 }
3863
3864 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3865 {
3866         struct btrfs_block_group_cache *bg;
3867
3868         bg = btrfs_lookup_block_group(fs_info, bytenr);
3869         ASSERT(bg);
3870         if (atomic_dec_and_test(&bg->nocow_writers))
3871                 wake_up_atomic_t(&bg->nocow_writers);
3872         /*
3873          * Once for our lookup and once for the lookup done by a previous call
3874          * to btrfs_inc_nocow_writers()
3875          */
3876         btrfs_put_block_group(bg);
3877         btrfs_put_block_group(bg);
3878 }
3879
3880 static int btrfs_wait_nocow_writers_atomic_t(atomic_t *a)
3881 {
3882         schedule();
3883         return 0;
3884 }
3885
3886 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3887 {
3888         wait_on_atomic_t(&bg->nocow_writers,
3889                          btrfs_wait_nocow_writers_atomic_t,
3890                          TASK_UNINTERRUPTIBLE);
3891 }
3892
3893 static const char *alloc_name(u64 flags)
3894 {
3895         switch (flags) {
3896         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3897                 return "mixed";
3898         case BTRFS_BLOCK_GROUP_METADATA:
3899                 return "metadata";
3900         case BTRFS_BLOCK_GROUP_DATA:
3901                 return "data";
3902         case BTRFS_BLOCK_GROUP_SYSTEM:
3903                 return "system";
3904         default:
3905                 WARN_ON(1);
3906                 return "invalid-combination";
3907         };
3908 }
3909
3910 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3911                              u64 total_bytes, u64 bytes_used,
3912                              u64 bytes_readonly,
3913                              struct btrfs_space_info **space_info)
3914 {
3915         struct btrfs_space_info *found;
3916         int i;
3917         int factor;
3918         int ret;
3919
3920         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3921                      BTRFS_BLOCK_GROUP_RAID10))
3922                 factor = 2;
3923         else
3924                 factor = 1;
3925
3926         found = __find_space_info(info, flags);
3927         if (found) {
3928                 spin_lock(&found->lock);
3929                 found->total_bytes += total_bytes;
3930                 found->disk_total += total_bytes * factor;
3931                 found->bytes_used += bytes_used;
3932                 found->disk_used += bytes_used * factor;
3933                 found->bytes_readonly += bytes_readonly;
3934                 if (total_bytes > 0)
3935                         found->full = 0;
3936                 space_info_add_new_bytes(info, found, total_bytes -
3937                                          bytes_used - bytes_readonly);
3938                 spin_unlock(&found->lock);
3939                 *space_info = found;
3940                 return 0;
3941         }
3942         found = kzalloc(sizeof(*found), GFP_NOFS);
3943         if (!found)
3944                 return -ENOMEM;
3945
3946         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3947         if (ret) {
3948                 kfree(found);
3949                 return ret;
3950         }
3951
3952         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3953                 INIT_LIST_HEAD(&found->block_groups[i]);
3954         init_rwsem(&found->groups_sem);
3955         spin_lock_init(&found->lock);
3956         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3957         found->total_bytes = total_bytes;
3958         found->disk_total = total_bytes * factor;
3959         found->bytes_used = bytes_used;
3960         found->disk_used = bytes_used * factor;
3961         found->bytes_pinned = 0;
3962         found->bytes_reserved = 0;
3963         found->bytes_readonly = bytes_readonly;
3964         found->bytes_may_use = 0;
3965         found->full = 0;
3966         found->max_extent_size = 0;
3967         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3968         found->chunk_alloc = 0;
3969         found->flush = 0;
3970         init_waitqueue_head(&found->wait);
3971         INIT_LIST_HEAD(&found->ro_bgs);
3972         INIT_LIST_HEAD(&found->tickets);
3973         INIT_LIST_HEAD(&found->priority_tickets);
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         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4275         ret = btrfs_qgroup_reserve_data(inode, start, len);
4276         if (ret)
4277                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4278         return ret;
4279 }
4280
4281 /*
4282  * Called if we need to clear a data reservation for this inode
4283  * Normally in a error case.
4284  *
4285  * This one will *NOT* use accurate qgroup reserved space API, just for case
4286  * which we can't sleep and is sure it won't affect qgroup reserved space.
4287  * Like clear_bit_hook().
4288  */
4289 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4290                                             u64 len)
4291 {
4292         struct btrfs_root *root = BTRFS_I(inode)->root;
4293         struct btrfs_space_info *data_sinfo;
4294
4295         /* Make sure the range is aligned to sectorsize */
4296         len = round_up(start + len, root->sectorsize) -
4297               round_down(start, root->sectorsize);
4298         start = round_down(start, root->sectorsize);
4299
4300         data_sinfo = root->fs_info->data_sinfo;
4301         spin_lock(&data_sinfo->lock);
4302         if (WARN_ON(data_sinfo->bytes_may_use < len))
4303                 data_sinfo->bytes_may_use = 0;
4304         else
4305                 data_sinfo->bytes_may_use -= len;
4306         trace_btrfs_space_reservation(root->fs_info, "space_info",
4307                                       data_sinfo->flags, len, 0);
4308         spin_unlock(&data_sinfo->lock);
4309 }
4310
4311 /*
4312  * Called if we need to clear a data reservation for this inode
4313  * Normally in a error case.
4314  *
4315  * This one will handle the per-inode data rsv map for accurate reserved
4316  * space framework.
4317  */
4318 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4319 {
4320         btrfs_free_reserved_data_space_noquota(inode, start, len);
4321         btrfs_qgroup_free_data(inode, start, len);
4322 }
4323
4324 static void force_metadata_allocation(struct btrfs_fs_info *info)
4325 {
4326         struct list_head *head = &info->space_info;
4327         struct btrfs_space_info *found;
4328
4329         rcu_read_lock();
4330         list_for_each_entry_rcu(found, head, list) {
4331                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4332                         found->force_alloc = CHUNK_ALLOC_FORCE;
4333         }
4334         rcu_read_unlock();
4335 }
4336
4337 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4338 {
4339         return (global->size << 1);
4340 }
4341
4342 static int should_alloc_chunk(struct btrfs_root *root,
4343                               struct btrfs_space_info *sinfo, int force)
4344 {
4345         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4346         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4347         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4348         u64 thresh;
4349
4350         if (force == CHUNK_ALLOC_FORCE)
4351                 return 1;
4352
4353         /*
4354          * We need to take into account the global rsv because for all intents
4355          * and purposes it's used space.  Don't worry about locking the
4356          * global_rsv, it doesn't change except when the transaction commits.
4357          */
4358         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4359                 num_allocated += calc_global_rsv_need_space(global_rsv);
4360
4361         /*
4362          * in limited mode, we want to have some free space up to
4363          * about 1% of the FS size.
4364          */
4365         if (force == CHUNK_ALLOC_LIMITED) {
4366                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4367                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4368
4369                 if (num_bytes - num_allocated < thresh)
4370                         return 1;
4371         }
4372
4373         if (num_allocated + SZ_2M < div_factor(num_bytes, 8))
4374                 return 0;
4375         return 1;
4376 }
4377
4378 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4379 {
4380         u64 num_dev;
4381
4382         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4383                     BTRFS_BLOCK_GROUP_RAID0 |
4384                     BTRFS_BLOCK_GROUP_RAID5 |
4385                     BTRFS_BLOCK_GROUP_RAID6))
4386                 num_dev = root->fs_info->fs_devices->rw_devices;
4387         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4388                 num_dev = 2;
4389         else
4390                 num_dev = 1;    /* DUP or single */
4391
4392         return num_dev;
4393 }
4394
4395 /*
4396  * If @is_allocation is true, reserve space in the system space info necessary
4397  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4398  * removing a chunk.
4399  */
4400 void check_system_chunk(struct btrfs_trans_handle *trans,
4401                         struct btrfs_root *root,
4402                         u64 type)
4403 {
4404         struct btrfs_space_info *info;
4405         u64 left;
4406         u64 thresh;
4407         int ret = 0;
4408         u64 num_devs;
4409
4410         /*
4411          * Needed because we can end up allocating a system chunk and for an
4412          * atomic and race free space reservation in the chunk block reserve.
4413          */
4414         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4415
4416         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4417         spin_lock(&info->lock);
4418         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4419                 info->bytes_reserved - info->bytes_readonly -
4420                 info->bytes_may_use;
4421         spin_unlock(&info->lock);
4422
4423         num_devs = get_profile_num_devs(root, type);
4424
4425         /* num_devs device items to update and 1 chunk item to add or remove */
4426         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4427                 btrfs_calc_trans_metadata_size(root, 1);
4428
4429         if (left < thresh && btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
4430                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4431                         left, thresh, type);
4432                 dump_space_info(info, 0, 0);
4433         }
4434
4435         if (left < thresh) {
4436                 u64 flags;
4437
4438                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4439                 /*
4440                  * Ignore failure to create system chunk. We might end up not
4441                  * needing it, as we might not need to COW all nodes/leafs from
4442                  * the paths we visit in the chunk tree (they were already COWed
4443                  * or created in the current transaction for example).
4444                  */
4445                 ret = btrfs_alloc_chunk(trans, root, flags);
4446         }
4447
4448         if (!ret) {
4449                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4450                                           &root->fs_info->chunk_block_rsv,
4451                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4452                 if (!ret)
4453                         trans->chunk_bytes_reserved += thresh;
4454         }
4455 }
4456
4457 /*
4458  * If force is CHUNK_ALLOC_FORCE:
4459  *    - return 1 if it successfully allocates a chunk,
4460  *    - return errors including -ENOSPC otherwise.
4461  * If force is NOT CHUNK_ALLOC_FORCE:
4462  *    - return 0 if it doesn't need to allocate a new chunk,
4463  *    - return 1 if it successfully allocates a chunk,
4464  *    - return errors including -ENOSPC otherwise.
4465  */
4466 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4467                           struct btrfs_root *extent_root, u64 flags, int force)
4468 {
4469         struct btrfs_space_info *space_info;
4470         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4471         int wait_for_alloc = 0;
4472         int ret = 0;
4473
4474         /* Don't re-enter if we're already allocating a chunk */
4475         if (trans->allocating_chunk)
4476                 return -ENOSPC;
4477
4478         space_info = __find_space_info(extent_root->fs_info, flags);
4479         if (!space_info) {
4480                 ret = update_space_info(extent_root->fs_info, flags,
4481                                         0, 0, 0, &space_info);
4482                 BUG_ON(ret); /* -ENOMEM */
4483         }
4484         BUG_ON(!space_info); /* Logic error */
4485
4486 again:
4487         spin_lock(&space_info->lock);
4488         if (force < space_info->force_alloc)
4489                 force = space_info->force_alloc;
4490         if (space_info->full) {
4491                 if (should_alloc_chunk(extent_root, space_info, force))
4492                         ret = -ENOSPC;
4493                 else
4494                         ret = 0;
4495                 spin_unlock(&space_info->lock);
4496                 return ret;
4497         }
4498
4499         if (!should_alloc_chunk(extent_root, space_info, force)) {
4500                 spin_unlock(&space_info->lock);
4501                 return 0;
4502         } else if (space_info->chunk_alloc) {
4503                 wait_for_alloc = 1;
4504         } else {
4505                 space_info->chunk_alloc = 1;
4506         }
4507
4508         spin_unlock(&space_info->lock);
4509
4510         mutex_lock(&fs_info->chunk_mutex);
4511
4512         /*
4513          * The chunk_mutex is held throughout the entirety of a chunk
4514          * allocation, so once we've acquired the chunk_mutex we know that the
4515          * other guy is done and we need to recheck and see if we should
4516          * allocate.
4517          */
4518         if (wait_for_alloc) {
4519                 mutex_unlock(&fs_info->chunk_mutex);
4520                 wait_for_alloc = 0;
4521                 goto again;
4522         }
4523
4524         trans->allocating_chunk = true;
4525
4526         /*
4527          * If we have mixed data/metadata chunks we want to make sure we keep
4528          * allocating mixed chunks instead of individual chunks.
4529          */
4530         if (btrfs_mixed_space_info(space_info))
4531                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4532
4533         /*
4534          * if we're doing a data chunk, go ahead and make sure that
4535          * we keep a reasonable number of metadata chunks allocated in the
4536          * FS as well.
4537          */
4538         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4539                 fs_info->data_chunk_allocations++;
4540                 if (!(fs_info->data_chunk_allocations %
4541                       fs_info->metadata_ratio))
4542                         force_metadata_allocation(fs_info);
4543         }
4544
4545         /*
4546          * Check if we have enough space in SYSTEM chunk because we may need
4547          * to update devices.
4548          */
4549         check_system_chunk(trans, extent_root, flags);
4550
4551         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4552         trans->allocating_chunk = false;
4553
4554         spin_lock(&space_info->lock);
4555         if (ret < 0 && ret != -ENOSPC)
4556                 goto out;
4557         if (ret)
4558                 space_info->full = 1;
4559         else
4560                 ret = 1;
4561
4562         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4563 out:
4564         space_info->chunk_alloc = 0;
4565         spin_unlock(&space_info->lock);
4566         mutex_unlock(&fs_info->chunk_mutex);
4567         /*
4568          * When we allocate a new chunk we reserve space in the chunk block
4569          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4570          * add new nodes/leafs to it if we end up needing to do it when
4571          * inserting the chunk item and updating device items as part of the
4572          * second phase of chunk allocation, performed by
4573          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4574          * large number of new block groups to create in our transaction
4575          * handle's new_bgs list to avoid exhausting the chunk block reserve
4576          * in extreme cases - like having a single transaction create many new
4577          * block groups when starting to write out the free space caches of all
4578          * the block groups that were made dirty during the lifetime of the
4579          * transaction.
4580          */
4581         if (trans->can_flush_pending_bgs &&
4582             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4583                 btrfs_create_pending_block_groups(trans, extent_root);
4584                 btrfs_trans_release_chunk_metadata(trans);
4585         }
4586         return ret;
4587 }
4588
4589 static int can_overcommit(struct btrfs_root *root,
4590                           struct btrfs_space_info *space_info, u64 bytes,
4591                           enum btrfs_reserve_flush_enum flush)
4592 {
4593         struct btrfs_block_rsv *global_rsv;
4594         u64 profile;
4595         u64 space_size;
4596         u64 avail;
4597         u64 used;
4598
4599         /* Don't overcommit when in mixed mode. */
4600         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4601                 return 0;
4602
4603         BUG_ON(root->fs_info == NULL);
4604         global_rsv = &root->fs_info->global_block_rsv;
4605         profile = btrfs_get_alloc_profile(root, 0);
4606         used = space_info->bytes_used + space_info->bytes_reserved +
4607                 space_info->bytes_pinned + space_info->bytes_readonly;
4608
4609         /*
4610          * We only want to allow over committing if we have lots of actual space
4611          * free, but if we don't have enough space to handle the global reserve
4612          * space then we could end up having a real enospc problem when trying
4613          * to allocate a chunk or some other such important allocation.
4614          */
4615         spin_lock(&global_rsv->lock);
4616         space_size = calc_global_rsv_need_space(global_rsv);
4617         spin_unlock(&global_rsv->lock);
4618         if (used + space_size >= space_info->total_bytes)
4619                 return 0;
4620
4621         used += space_info->bytes_may_use;
4622
4623         spin_lock(&root->fs_info->free_chunk_lock);
4624         avail = root->fs_info->free_chunk_space;
4625         spin_unlock(&root->fs_info->free_chunk_lock);
4626
4627         /*
4628          * If we have dup, raid1 or raid10 then only half of the free
4629          * space is actually useable.  For raid56, the space info used
4630          * doesn't include the parity drive, so we don't have to
4631          * change the math
4632          */
4633         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4634                        BTRFS_BLOCK_GROUP_RAID1 |
4635                        BTRFS_BLOCK_GROUP_RAID10))
4636                 avail >>= 1;
4637
4638         /*
4639          * If we aren't flushing all things, let us overcommit up to
4640          * 1/2th of the space. If we can flush, don't let us overcommit
4641          * too much, let it overcommit up to 1/8 of the space.
4642          */
4643         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4644                 avail >>= 3;
4645         else
4646                 avail >>= 1;
4647
4648         if (used + bytes < space_info->total_bytes + avail)
4649                 return 1;
4650         return 0;
4651 }
4652
4653 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4654                                          unsigned long nr_pages, int nr_items)
4655 {
4656         struct super_block *sb = root->fs_info->sb;
4657
4658         if (down_read_trylock(&sb->s_umount)) {
4659                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4660                 up_read(&sb->s_umount);
4661         } else {
4662                 /*
4663                  * We needn't worry the filesystem going from r/w to r/o though
4664                  * we don't acquire ->s_umount mutex, because the filesystem
4665                  * should guarantee the delalloc inodes list be empty after
4666                  * the filesystem is readonly(all dirty pages are written to
4667                  * the disk).
4668                  */
4669                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4670                 if (!current->journal_info)
4671                         btrfs_wait_ordered_roots(root->fs_info, nr_items,
4672                                                  0, (u64)-1);
4673         }
4674 }
4675
4676 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4677 {
4678         u64 bytes;
4679         int nr;
4680
4681         bytes = btrfs_calc_trans_metadata_size(root, 1);
4682         nr = (int)div64_u64(to_reclaim, bytes);
4683         if (!nr)
4684                 nr = 1;
4685         return nr;
4686 }
4687
4688 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4689
4690 /*
4691  * shrink metadata reservation for delalloc
4692  */
4693 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4694                             bool wait_ordered)
4695 {
4696         struct btrfs_block_rsv *block_rsv;
4697         struct btrfs_space_info *space_info;
4698         struct btrfs_trans_handle *trans;
4699         u64 delalloc_bytes;
4700         u64 max_reclaim;
4701         long time_left;
4702         unsigned long nr_pages;
4703         int loops;
4704         int items;
4705         enum btrfs_reserve_flush_enum flush;
4706
4707         /* Calc the number of the pages we need flush for space reservation */
4708         items = calc_reclaim_items_nr(root, to_reclaim);
4709         to_reclaim = (u64)items * EXTENT_SIZE_PER_ITEM;
4710
4711         trans = (struct btrfs_trans_handle *)current->journal_info;
4712         block_rsv = &root->fs_info->delalloc_block_rsv;
4713         space_info = block_rsv->space_info;
4714
4715         delalloc_bytes = percpu_counter_sum_positive(
4716                                                 &root->fs_info->delalloc_bytes);
4717         if (delalloc_bytes == 0) {
4718                 if (trans)
4719                         return;
4720                 if (wait_ordered)
4721                         btrfs_wait_ordered_roots(root->fs_info, items,
4722                                                  0, (u64)-1);
4723                 return;
4724         }
4725
4726         loops = 0;
4727         while (delalloc_bytes && loops < 3) {
4728                 max_reclaim = min(delalloc_bytes, to_reclaim);
4729                 nr_pages = max_reclaim >> PAGE_SHIFT;
4730                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4731                 /*
4732                  * We need to wait for the async pages to actually start before
4733                  * we do anything.
4734                  */
4735                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4736                 if (!max_reclaim)
4737                         goto skip_async;
4738
4739                 if (max_reclaim <= nr_pages)
4740                         max_reclaim = 0;
4741                 else
4742                         max_reclaim -= nr_pages;
4743
4744                 wait_event(root->fs_info->async_submit_wait,
4745                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4746                            (int)max_reclaim);
4747 skip_async:
4748                 if (!trans)
4749                         flush = BTRFS_RESERVE_FLUSH_ALL;
4750                 else
4751                         flush = BTRFS_RESERVE_NO_FLUSH;
4752                 spin_lock(&space_info->lock);
4753                 if (can_overcommit(root, space_info, orig, flush)) {
4754                         spin_unlock(&space_info->lock);
4755                         break;
4756                 }
4757                 if (list_empty(&space_info->tickets) &&
4758                     list_empty(&space_info->priority_tickets)) {
4759                         spin_unlock(&space_info->lock);
4760                         break;
4761                 }
4762                 spin_unlock(&space_info->lock);
4763
4764                 loops++;
4765                 if (wait_ordered && !trans) {
4766                         btrfs_wait_ordered_roots(root->fs_info, items,
4767                                                  0, (u64)-1);
4768                 } else {
4769                         time_left = schedule_timeout_killable(1);
4770                         if (time_left)
4771                                 break;
4772                 }
4773                 delalloc_bytes = percpu_counter_sum_positive(
4774                                                 &root->fs_info->delalloc_bytes);
4775         }
4776 }
4777
4778 /**
4779  * maybe_commit_transaction - possibly commit the transaction if its ok to
4780  * @root - the root we're allocating for
4781  * @bytes - the number of bytes we want to reserve
4782  * @force - force the commit
4783  *
4784  * This will check to make sure that committing the transaction will actually
4785  * get us somewhere and then commit the transaction if it does.  Otherwise it
4786  * will return -ENOSPC.
4787  */
4788 static int may_commit_transaction(struct btrfs_root *root,
4789                                   struct btrfs_space_info *space_info,
4790                                   u64 bytes, int force)
4791 {
4792         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4793         struct btrfs_trans_handle *trans;
4794
4795         trans = (struct btrfs_trans_handle *)current->journal_info;
4796         if (trans)
4797                 return -EAGAIN;
4798
4799         if (force)
4800                 goto commit;
4801
4802         /* See if there is enough pinned space to make this reservation */
4803         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4804                                    bytes) >= 0)
4805                 goto commit;
4806
4807         /*
4808          * See if there is some space in the delayed insertion reservation for
4809          * this reservation.
4810          */
4811         if (space_info != delayed_rsv->space_info)
4812                 return -ENOSPC;
4813
4814         spin_lock(&delayed_rsv->lock);
4815         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4816                                    bytes - delayed_rsv->size) >= 0) {
4817                 spin_unlock(&delayed_rsv->lock);
4818                 return -ENOSPC;
4819         }
4820         spin_unlock(&delayed_rsv->lock);
4821
4822 commit:
4823         trans = btrfs_join_transaction(root);
4824         if (IS_ERR(trans))
4825                 return -ENOSPC;
4826
4827         return btrfs_commit_transaction(trans, root);
4828 }
4829
4830 struct reserve_ticket {
4831         u64 bytes;
4832         int error;
4833         struct list_head list;
4834         wait_queue_head_t wait;
4835 };
4836
4837 static int flush_space(struct btrfs_root *root,
4838                        struct btrfs_space_info *space_info, u64 num_bytes,
4839                        u64 orig_bytes, int state)
4840 {
4841         struct btrfs_trans_handle *trans;
4842         int nr;
4843         int ret = 0;
4844
4845         switch (state) {
4846         case FLUSH_DELAYED_ITEMS_NR:
4847         case FLUSH_DELAYED_ITEMS:
4848                 if (state == FLUSH_DELAYED_ITEMS_NR)
4849                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4850                 else
4851                         nr = -1;
4852
4853                 trans = btrfs_join_transaction(root);
4854                 if (IS_ERR(trans)) {
4855                         ret = PTR_ERR(trans);
4856                         break;
4857                 }
4858                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4859                 btrfs_end_transaction(trans, root);
4860                 break;
4861         case FLUSH_DELALLOC:
4862         case FLUSH_DELALLOC_WAIT:
4863                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4864                                 state == FLUSH_DELALLOC_WAIT);
4865                 break;
4866         case ALLOC_CHUNK:
4867                 trans = btrfs_join_transaction(root);
4868                 if (IS_ERR(trans)) {
4869                         ret = PTR_ERR(trans);
4870                         break;
4871                 }
4872                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4873                                      btrfs_get_alloc_profile(root, 0),
4874                                      CHUNK_ALLOC_NO_FORCE);
4875                 btrfs_end_transaction(trans, root);
4876                 if (ret > 0 || ret == -ENOSPC)
4877                         ret = 0;
4878                 break;
4879         case COMMIT_TRANS:
4880                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4881                 break;
4882         default:
4883                 ret = -ENOSPC;
4884                 break;
4885         }
4886
4887         trace_btrfs_flush_space(root->fs_info, space_info->flags, num_bytes,
4888                                 orig_bytes, state, ret);
4889         return ret;
4890 }
4891
4892 static inline u64
4893 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4894                                  struct btrfs_space_info *space_info)
4895 {
4896         struct reserve_ticket *ticket;
4897         u64 used;
4898         u64 expected;
4899         u64 to_reclaim = 0;
4900
4901         list_for_each_entry(ticket, &space_info->tickets, list)
4902                 to_reclaim += ticket->bytes;
4903         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4904                 to_reclaim += ticket->bytes;
4905         if (to_reclaim)
4906                 return to_reclaim;
4907
4908         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4909         if (can_overcommit(root, space_info, to_reclaim,
4910                            BTRFS_RESERVE_FLUSH_ALL))
4911                 return 0;
4912
4913         used = space_info->bytes_used + space_info->bytes_reserved +
4914                space_info->bytes_pinned + space_info->bytes_readonly +
4915                space_info->bytes_may_use;
4916         if (can_overcommit(root, space_info, SZ_1M, BTRFS_RESERVE_FLUSH_ALL))
4917                 expected = div_factor_fine(space_info->total_bytes, 95);
4918         else
4919                 expected = div_factor_fine(space_info->total_bytes, 90);
4920
4921         if (used > expected)
4922                 to_reclaim = used - expected;
4923         else
4924                 to_reclaim = 0;
4925         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4926                                      space_info->bytes_reserved);
4927         return to_reclaim;
4928 }
4929
4930 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4931                                         struct btrfs_root *root, u64 used)
4932 {
4933         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4934
4935         /* If we're just plain full then async reclaim just slows us down. */
4936         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4937                 return 0;
4938
4939         if (!btrfs_calc_reclaim_metadata_size(root, space_info))
4940                 return 0;
4941
4942         return (used >= thresh && !btrfs_fs_closing(root->fs_info) &&
4943                 !test_bit(BTRFS_FS_STATE_REMOUNTING,
4944                           &root->fs_info->fs_state));
4945 }
4946
4947 static void wake_all_tickets(struct list_head *head)
4948 {
4949         struct reserve_ticket *ticket;
4950
4951         while (!list_empty(head)) {
4952                 ticket = list_first_entry(head, struct reserve_ticket, list);
4953                 list_del_init(&ticket->list);
4954                 ticket->error = -ENOSPC;
4955                 wake_up(&ticket->wait);
4956         }
4957 }
4958
4959 /*
4960  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4961  * will loop and continuously try to flush as long as we are making progress.
4962  * We count progress as clearing off tickets each time we have to loop.
4963  */
4964 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4965 {
4966         struct btrfs_fs_info *fs_info;
4967         struct btrfs_space_info *space_info;
4968         u64 to_reclaim;
4969         int flush_state;
4970         int commit_cycles = 0;
4971         u64 last_tickets_id;
4972
4973         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4974         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4975
4976         spin_lock(&space_info->lock);
4977         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4978                                                       space_info);
4979         if (!to_reclaim) {
4980                 space_info->flush = 0;
4981                 spin_unlock(&space_info->lock);
4982                 return;
4983         }
4984         last_tickets_id = space_info->tickets_id;
4985         spin_unlock(&space_info->lock);
4986
4987         flush_state = FLUSH_DELAYED_ITEMS_NR;
4988         do {
4989                 struct reserve_ticket *ticket;
4990                 int ret;
4991
4992                 ret = flush_space(fs_info->fs_root, space_info, to_reclaim,
4993                             to_reclaim, flush_state);
4994                 spin_lock(&space_info->lock);
4995                 if (list_empty(&space_info->tickets)) {
4996                         space_info->flush = 0;
4997                         spin_unlock(&space_info->lock);
4998                         return;
4999                 }
5000                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5001                                                               space_info);
5002                 ticket = list_first_entry(&space_info->tickets,
5003                                           struct reserve_ticket, list);
5004                 if (last_tickets_id == space_info->tickets_id) {
5005                         flush_state++;
5006                 } else {
5007                         last_tickets_id = space_info->tickets_id;
5008                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5009                         if (commit_cycles)
5010                                 commit_cycles--;
5011                 }
5012
5013                 if (flush_state > COMMIT_TRANS) {
5014                         commit_cycles++;
5015                         if (commit_cycles > 2) {
5016                                 wake_all_tickets(&space_info->tickets);
5017                                 space_info->flush = 0;
5018                         } else {
5019                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5020                         }
5021                 }
5022                 spin_unlock(&space_info->lock);
5023         } while (flush_state <= COMMIT_TRANS);
5024 }
5025
5026 void btrfs_init_async_reclaim_work(struct work_struct *work)
5027 {
5028         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5029 }
5030
5031 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5032                                             struct btrfs_space_info *space_info,
5033                                             struct reserve_ticket *ticket)
5034 {
5035         u64 to_reclaim;
5036         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5037
5038         spin_lock(&space_info->lock);
5039         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5040                                                       space_info);
5041         if (!to_reclaim) {
5042                 spin_unlock(&space_info->lock);
5043                 return;
5044         }
5045         spin_unlock(&space_info->lock);
5046
5047         do {
5048                 flush_space(fs_info->fs_root, space_info, to_reclaim,
5049                             to_reclaim, flush_state);
5050                 flush_state++;
5051                 spin_lock(&space_info->lock);
5052                 if (ticket->bytes == 0) {
5053                         spin_unlock(&space_info->lock);
5054                         return;
5055                 }
5056                 spin_unlock(&space_info->lock);
5057
5058                 /*
5059                  * Priority flushers can't wait on delalloc without
5060                  * deadlocking.
5061                  */
5062                 if (flush_state == FLUSH_DELALLOC ||
5063                     flush_state == FLUSH_DELALLOC_WAIT)
5064                         flush_state = ALLOC_CHUNK;
5065         } while (flush_state < COMMIT_TRANS);
5066 }
5067
5068 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5069                                struct btrfs_space_info *space_info,
5070                                struct reserve_ticket *ticket, u64 orig_bytes)
5071
5072 {
5073         DEFINE_WAIT(wait);
5074         int ret = 0;
5075
5076         spin_lock(&space_info->lock);
5077         while (ticket->bytes > 0 && ticket->error == 0) {
5078                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5079                 if (ret) {
5080                         ret = -EINTR;
5081                         break;
5082                 }
5083                 spin_unlock(&space_info->lock);
5084
5085                 schedule();
5086
5087                 finish_wait(&ticket->wait, &wait);
5088                 spin_lock(&space_info->lock);
5089         }
5090         if (!ret)
5091                 ret = ticket->error;
5092         if (!list_empty(&ticket->list))
5093                 list_del_init(&ticket->list);
5094         if (ticket->bytes && ticket->bytes < orig_bytes) {
5095                 u64 num_bytes = orig_bytes - ticket->bytes;
5096                 space_info->bytes_may_use -= num_bytes;
5097                 trace_btrfs_space_reservation(fs_info, "space_info",
5098                                               space_info->flags, num_bytes, 0);
5099         }
5100         spin_unlock(&space_info->lock);
5101
5102         return ret;
5103 }
5104
5105 /**
5106  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5107  * @root - the root we're allocating for
5108  * @space_info - the space info we want to allocate from
5109  * @orig_bytes - the number of bytes we want
5110  * @flush - whether or not we can flush to make our reservation
5111  *
5112  * This will reserve orig_bytes number of bytes from the space info associated
5113  * with the block_rsv.  If there is not enough space it will make an attempt to
5114  * flush out space to make room.  It will do this by flushing delalloc if
5115  * possible or committing the transaction.  If flush is 0 then no attempts to
5116  * regain reservations will be made and this will fail if there is not enough
5117  * space already.
5118  */
5119 static int __reserve_metadata_bytes(struct btrfs_root *root,
5120                                     struct btrfs_space_info *space_info,
5121                                     u64 orig_bytes,
5122                                     enum btrfs_reserve_flush_enum flush)
5123 {
5124         struct reserve_ticket ticket;
5125         u64 used;
5126         int ret = 0;
5127
5128         ASSERT(orig_bytes);
5129         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5130
5131         spin_lock(&space_info->lock);
5132         ret = -ENOSPC;
5133         used = space_info->bytes_used + space_info->bytes_reserved +
5134                 space_info->bytes_pinned + space_info->bytes_readonly +
5135                 space_info->bytes_may_use;
5136
5137         /*
5138          * If we have enough space then hooray, make our reservation and carry
5139          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5140          * If not things get more complicated.
5141          */
5142         if (used + orig_bytes <= space_info->total_bytes) {
5143                 space_info->bytes_may_use += orig_bytes;
5144                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5145                                               space_info->flags, orig_bytes,
5146                                               1);
5147                 ret = 0;
5148         } else if (can_overcommit(root, space_info, orig_bytes, flush)) {
5149                 space_info->bytes_may_use += orig_bytes;
5150                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5151                                               space_info->flags, orig_bytes,
5152                                               1);
5153                 ret = 0;
5154         }
5155
5156         /*
5157          * If we couldn't make a reservation then setup our reservation ticket
5158          * and kick the async worker if it's not already running.
5159          *
5160          * If we are a priority flusher then we just need to add our ticket to
5161          * the list and we will do our own flushing further down.
5162          */
5163         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5164                 ticket.bytes = orig_bytes;
5165                 ticket.error = 0;
5166                 init_waitqueue_head(&ticket.wait);
5167                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5168                         list_add_tail(&ticket.list, &space_info->tickets);
5169                         if (!space_info->flush) {
5170                                 space_info->flush = 1;
5171                                 trace_btrfs_trigger_flush(root->fs_info,
5172                                                           space_info->flags,
5173                                                           orig_bytes, flush,
5174                                                           "enospc");
5175                                 queue_work(system_unbound_wq,
5176                                            &root->fs_info->async_reclaim_work);
5177                         }
5178                 } else {
5179                         list_add_tail(&ticket.list,
5180                                       &space_info->priority_tickets);
5181                 }
5182         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5183                 used += orig_bytes;
5184                 /*
5185                  * We will do the space reservation dance during log replay,
5186                  * which means we won't have fs_info->fs_root set, so don't do
5187                  * the async reclaim as we will panic.
5188                  */
5189                 if (!root->fs_info->log_root_recovering &&
5190                     need_do_async_reclaim(space_info, root, used) &&
5191                     !work_busy(&root->fs_info->async_reclaim_work)) {
5192                         trace_btrfs_trigger_flush(root->fs_info,
5193                                                   space_info->flags,
5194                                                   orig_bytes, flush,
5195                                                   "preempt");
5196                         queue_work(system_unbound_wq,
5197                                    &root->fs_info->async_reclaim_work);
5198                 }
5199         }
5200         spin_unlock(&space_info->lock);
5201         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5202                 return ret;
5203
5204         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5205                 return wait_reserve_ticket(root->fs_info, space_info, &ticket,
5206                                            orig_bytes);
5207
5208         ret = 0;
5209         priority_reclaim_metadata_space(root->fs_info, space_info, &ticket);
5210         spin_lock(&space_info->lock);
5211         if (ticket.bytes) {
5212                 if (ticket.bytes < orig_bytes) {
5213                         u64 num_bytes = orig_bytes - ticket.bytes;
5214                         space_info->bytes_may_use -= num_bytes;
5215                         trace_btrfs_space_reservation(root->fs_info,
5216                                         "space_info", space_info->flags,
5217                                         num_bytes, 0);
5218
5219                 }
5220                 list_del_init(&ticket.list);
5221                 ret = -ENOSPC;
5222         }
5223         spin_unlock(&space_info->lock);
5224         ASSERT(list_empty(&ticket.list));
5225         return ret;
5226 }
5227
5228 /**
5229  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5230  * @root - the root we're allocating for
5231  * @block_rsv - the block_rsv we're allocating for
5232  * @orig_bytes - the number of bytes we want
5233  * @flush - whether or not we can flush to make our reservation
5234  *
5235  * This will reserve orgi_bytes number of bytes from the space info associated
5236  * with the block_rsv.  If there is not enough space it will make an attempt to
5237  * flush out space to make room.  It will do this by flushing delalloc if
5238  * possible or committing the transaction.  If flush is 0 then no attempts to
5239  * regain reservations will be made and this will fail if there is not enough
5240  * space already.
5241  */
5242 static int reserve_metadata_bytes(struct btrfs_root *root,
5243                                   struct btrfs_block_rsv *block_rsv,
5244                                   u64 orig_bytes,
5245                                   enum btrfs_reserve_flush_enum flush)
5246 {
5247         int ret;
5248
5249         ret = __reserve_metadata_bytes(root, block_rsv->space_info, orig_bytes,
5250                                        flush);
5251         if (ret == -ENOSPC &&
5252             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5253                 struct btrfs_block_rsv *global_rsv =
5254                         &root->fs_info->global_block_rsv;
5255
5256                 if (block_rsv != global_rsv &&
5257                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5258                         ret = 0;
5259         }
5260         if (ret == -ENOSPC)
5261                 trace_btrfs_space_reservation(root->fs_info,
5262                                               "space_info:enospc",
5263                                               block_rsv->space_info->flags,
5264                                               orig_bytes, 1);
5265         return ret;
5266 }
5267
5268 static struct btrfs_block_rsv *get_block_rsv(
5269                                         const struct btrfs_trans_handle *trans,
5270                                         const struct btrfs_root *root)
5271 {
5272         struct btrfs_block_rsv *block_rsv = NULL;
5273
5274         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5275             (root == root->fs_info->csum_root && trans->adding_csums) ||
5276              (root == root->fs_info->uuid_root))
5277                 block_rsv = trans->block_rsv;
5278
5279         if (!block_rsv)
5280                 block_rsv = root->block_rsv;
5281
5282         if (!block_rsv)
5283                 block_rsv = &root->fs_info->empty_block_rsv;
5284
5285         return block_rsv;
5286 }
5287
5288 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5289                                u64 num_bytes)
5290 {
5291         int ret = -ENOSPC;
5292         spin_lock(&block_rsv->lock);
5293         if (block_rsv->reserved >= num_bytes) {
5294                 block_rsv->reserved -= num_bytes;
5295                 if (block_rsv->reserved < block_rsv->size)
5296                         block_rsv->full = 0;
5297                 ret = 0;
5298         }
5299         spin_unlock(&block_rsv->lock);
5300         return ret;
5301 }
5302
5303 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5304                                 u64 num_bytes, int update_size)
5305 {
5306         spin_lock(&block_rsv->lock);
5307         block_rsv->reserved += num_bytes;
5308         if (update_size)
5309                 block_rsv->size += num_bytes;
5310         else if (block_rsv->reserved >= block_rsv->size)
5311                 block_rsv->full = 1;
5312         spin_unlock(&block_rsv->lock);
5313 }
5314
5315 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5316                              struct btrfs_block_rsv *dest, u64 num_bytes,
5317                              int min_factor)
5318 {
5319         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5320         u64 min_bytes;
5321
5322         if (global_rsv->space_info != dest->space_info)
5323                 return -ENOSPC;
5324
5325         spin_lock(&global_rsv->lock);
5326         min_bytes = div_factor(global_rsv->size, min_factor);
5327         if (global_rsv->reserved < min_bytes + num_bytes) {
5328                 spin_unlock(&global_rsv->lock);
5329                 return -ENOSPC;
5330         }
5331         global_rsv->reserved -= num_bytes;
5332         if (global_rsv->reserved < global_rsv->size)
5333                 global_rsv->full = 0;
5334         spin_unlock(&global_rsv->lock);
5335
5336         block_rsv_add_bytes(dest, num_bytes, 1);
5337         return 0;
5338 }
5339
5340 /*
5341  * This is for space we already have accounted in space_info->bytes_may_use, so
5342  * basically when we're returning space from block_rsv's.
5343  */
5344 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5345                                      struct btrfs_space_info *space_info,
5346                                      u64 num_bytes)
5347 {
5348         struct reserve_ticket *ticket;
5349         struct list_head *head;
5350         u64 used;
5351         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5352         bool check_overcommit = false;
5353
5354         spin_lock(&space_info->lock);
5355         head = &space_info->priority_tickets;
5356
5357         /*
5358          * If we are over our limit then we need to check and see if we can
5359          * overcommit, and if we can't then we just need to free up our space
5360          * and not satisfy any requests.
5361          */
5362         used = space_info->bytes_used + space_info->bytes_reserved +
5363                 space_info->bytes_pinned + space_info->bytes_readonly +
5364                 space_info->bytes_may_use;
5365         if (used - num_bytes >= space_info->total_bytes)
5366                 check_overcommit = true;
5367 again:
5368         while (!list_empty(head) && num_bytes) {
5369                 ticket = list_first_entry(head, struct reserve_ticket,
5370                                           list);
5371                 /*
5372                  * We use 0 bytes because this space is already reserved, so
5373                  * adding the ticket space would be a double count.
5374                  */
5375                 if (check_overcommit &&
5376                     !can_overcommit(fs_info->extent_root, space_info, 0,
5377                                     flush))
5378                         break;
5379                 if (num_bytes >= ticket->bytes) {
5380                         list_del_init(&ticket->list);
5381                         num_bytes -= ticket->bytes;
5382                         ticket->bytes = 0;
5383                         space_info->tickets_id++;
5384                         wake_up(&ticket->wait);
5385                 } else {
5386                         ticket->bytes -= num_bytes;
5387                         num_bytes = 0;
5388                 }
5389         }
5390
5391         if (num_bytes && head == &space_info->priority_tickets) {
5392                 head = &space_info->tickets;
5393                 flush = BTRFS_RESERVE_FLUSH_ALL;
5394                 goto again;
5395         }
5396         space_info->bytes_may_use -= num_bytes;
5397         trace_btrfs_space_reservation(fs_info, "space_info",
5398                                       space_info->flags, num_bytes, 0);
5399         spin_unlock(&space_info->lock);
5400 }
5401
5402 /*
5403  * This is for newly allocated space that isn't accounted in
5404  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5405  * we use this helper.
5406  */
5407 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5408                                      struct btrfs_space_info *space_info,
5409                                      u64 num_bytes)
5410 {
5411         struct reserve_ticket *ticket;
5412         struct list_head *head = &space_info->priority_tickets;
5413
5414 again:
5415         while (!list_empty(head) && num_bytes) {
5416                 ticket = list_first_entry(head, struct reserve_ticket,
5417                                           list);
5418                 if (num_bytes >= ticket->bytes) {
5419                         trace_btrfs_space_reservation(fs_info, "space_info",
5420                                                       space_info->flags,
5421                                                       ticket->bytes, 1);
5422                         list_del_init(&ticket->list);
5423                         num_bytes -= ticket->bytes;
5424                         space_info->bytes_may_use += ticket->bytes;
5425                         ticket->bytes = 0;
5426                         space_info->tickets_id++;
5427                         wake_up(&ticket->wait);
5428                 } else {
5429                         trace_btrfs_space_reservation(fs_info, "space_info",
5430                                                       space_info->flags,
5431                                                       num_bytes, 1);
5432                         space_info->bytes_may_use += num_bytes;
5433                         ticket->bytes -= num_bytes;
5434                         num_bytes = 0;
5435                 }
5436         }
5437
5438         if (num_bytes && head == &space_info->priority_tickets) {
5439                 head = &space_info->tickets;
5440                 goto again;
5441         }
5442 }
5443
5444 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5445                                     struct btrfs_block_rsv *block_rsv,
5446                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5447 {
5448         struct btrfs_space_info *space_info = block_rsv->space_info;
5449
5450         spin_lock(&block_rsv->lock);
5451         if (num_bytes == (u64)-1)
5452                 num_bytes = block_rsv->size;
5453         block_rsv->size -= num_bytes;
5454         if (block_rsv->reserved >= block_rsv->size) {
5455                 num_bytes = block_rsv->reserved - block_rsv->size;
5456                 block_rsv->reserved = block_rsv->size;
5457                 block_rsv->full = 1;
5458         } else {
5459                 num_bytes = 0;
5460         }
5461         spin_unlock(&block_rsv->lock);
5462
5463         if (num_bytes > 0) {
5464                 if (dest) {
5465                         spin_lock(&dest->lock);
5466                         if (!dest->full) {
5467                                 u64 bytes_to_add;
5468
5469                                 bytes_to_add = dest->size - dest->reserved;
5470                                 bytes_to_add = min(num_bytes, bytes_to_add);
5471                                 dest->reserved += bytes_to_add;
5472                                 if (dest->reserved >= dest->size)
5473                                         dest->full = 1;
5474                                 num_bytes -= bytes_to_add;
5475                         }
5476                         spin_unlock(&dest->lock);
5477                 }
5478                 if (num_bytes)
5479                         space_info_add_old_bytes(fs_info, space_info,
5480                                                  num_bytes);
5481         }
5482 }
5483
5484 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5485                             struct btrfs_block_rsv *dst, u64 num_bytes,
5486                             int update_size)
5487 {
5488         int ret;
5489
5490         ret = block_rsv_use_bytes(src, num_bytes);
5491         if (ret)
5492                 return ret;
5493
5494         block_rsv_add_bytes(dst, num_bytes, update_size);
5495         return 0;
5496 }
5497
5498 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5499 {
5500         memset(rsv, 0, sizeof(*rsv));
5501         spin_lock_init(&rsv->lock);
5502         rsv->type = type;
5503 }
5504
5505 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5506                                               unsigned short type)
5507 {
5508         struct btrfs_block_rsv *block_rsv;
5509         struct btrfs_fs_info *fs_info = root->fs_info;
5510
5511         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5512         if (!block_rsv)
5513                 return NULL;
5514
5515         btrfs_init_block_rsv(block_rsv, type);
5516         block_rsv->space_info = __find_space_info(fs_info,
5517                                                   BTRFS_BLOCK_GROUP_METADATA);
5518         return block_rsv;
5519 }
5520
5521 void btrfs_free_block_rsv(struct btrfs_root *root,
5522                           struct btrfs_block_rsv *rsv)
5523 {
5524         if (!rsv)
5525                 return;
5526         btrfs_block_rsv_release(root, rsv, (u64)-1);
5527         kfree(rsv);
5528 }
5529
5530 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5531 {
5532         kfree(rsv);
5533 }
5534
5535 int btrfs_block_rsv_add(struct btrfs_root *root,
5536                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5537                         enum btrfs_reserve_flush_enum flush)
5538 {
5539         int ret;
5540
5541         if (num_bytes == 0)
5542                 return 0;
5543
5544         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5545         if (!ret) {
5546                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5547                 return 0;
5548         }
5549
5550         return ret;
5551 }
5552
5553 int btrfs_block_rsv_check(struct btrfs_root *root,
5554                           struct btrfs_block_rsv *block_rsv, int min_factor)
5555 {
5556         u64 num_bytes = 0;
5557         int ret = -ENOSPC;
5558
5559         if (!block_rsv)
5560                 return 0;
5561
5562         spin_lock(&block_rsv->lock);
5563         num_bytes = div_factor(block_rsv->size, min_factor);
5564         if (block_rsv->reserved >= num_bytes)
5565                 ret = 0;
5566         spin_unlock(&block_rsv->lock);
5567
5568         return ret;
5569 }
5570
5571 int btrfs_block_rsv_refill(struct btrfs_root *root,
5572                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5573                            enum btrfs_reserve_flush_enum flush)
5574 {
5575         u64 num_bytes = 0;
5576         int ret = -ENOSPC;
5577
5578         if (!block_rsv)
5579                 return 0;
5580
5581         spin_lock(&block_rsv->lock);
5582         num_bytes = min_reserved;
5583         if (block_rsv->reserved >= num_bytes)
5584                 ret = 0;
5585         else
5586                 num_bytes -= block_rsv->reserved;
5587         spin_unlock(&block_rsv->lock);
5588
5589         if (!ret)
5590                 return 0;
5591
5592         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5593         if (!ret) {
5594                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5595                 return 0;
5596         }
5597
5598         return ret;
5599 }
5600
5601 void btrfs_block_rsv_release(struct btrfs_root *root,
5602                              struct btrfs_block_rsv *block_rsv,
5603                              u64 num_bytes)
5604 {
5605         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5606         if (global_rsv == block_rsv ||
5607             block_rsv->space_info != global_rsv->space_info)
5608                 global_rsv = NULL;
5609         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5610                                 num_bytes);
5611 }
5612
5613 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5614 {
5615         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5616         struct btrfs_space_info *sinfo = block_rsv->space_info;
5617         u64 num_bytes;
5618
5619         /*
5620          * The global block rsv is based on the size of the extent tree, the
5621          * checksum tree and the root tree.  If the fs is empty we want to set
5622          * it to a minimal amount for safety.
5623          */
5624         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5625                 btrfs_root_used(&fs_info->csum_root->root_item) +
5626                 btrfs_root_used(&fs_info->tree_root->root_item);
5627         num_bytes = max_t(u64, num_bytes, SZ_16M);
5628
5629         spin_lock(&sinfo->lock);
5630         spin_lock(&block_rsv->lock);
5631
5632         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5633
5634         if (block_rsv->reserved < block_rsv->size) {
5635                 num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5636                         sinfo->bytes_reserved + sinfo->bytes_readonly +
5637                         sinfo->bytes_may_use;
5638                 if (sinfo->total_bytes > num_bytes) {
5639                         num_bytes = sinfo->total_bytes - num_bytes;
5640                         num_bytes = min(num_bytes,
5641                                         block_rsv->size - block_rsv->reserved);
5642                         block_rsv->reserved += num_bytes;
5643                         sinfo->bytes_may_use += num_bytes;
5644                         trace_btrfs_space_reservation(fs_info, "space_info",
5645                                                       sinfo->flags, num_bytes,
5646                                                       1);
5647                 }
5648         } else if (block_rsv->reserved > block_rsv->size) {
5649                 num_bytes = block_rsv->reserved - block_rsv->size;
5650                 sinfo->bytes_may_use -= num_bytes;
5651                 trace_btrfs_space_reservation(fs_info, "space_info",
5652                                       sinfo->flags, num_bytes, 0);
5653                 block_rsv->reserved = block_rsv->size;
5654         }
5655
5656         if (block_rsv->reserved == block_rsv->size)
5657                 block_rsv->full = 1;
5658         else
5659                 block_rsv->full = 0;
5660
5661         spin_unlock(&block_rsv->lock);
5662         spin_unlock(&sinfo->lock);
5663 }
5664
5665 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5666 {
5667         struct btrfs_space_info *space_info;
5668
5669         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5670         fs_info->chunk_block_rsv.space_info = space_info;
5671
5672         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5673         fs_info->global_block_rsv.space_info = space_info;
5674         fs_info->delalloc_block_rsv.space_info = space_info;
5675         fs_info->trans_block_rsv.space_info = space_info;
5676         fs_info->empty_block_rsv.space_info = space_info;
5677         fs_info->delayed_block_rsv.space_info = space_info;
5678
5679         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5680         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5681         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5682         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5683         if (fs_info->quota_root)
5684                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5685         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5686
5687         update_global_block_rsv(fs_info);
5688 }
5689
5690 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5691 {
5692         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5693                                 (u64)-1);
5694         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5695         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5696         WARN_ON(fs_info->trans_block_rsv.size > 0);
5697         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5698         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5699         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5700         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5701         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5702 }
5703
5704 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5705                                   struct btrfs_root *root)
5706 {
5707         if (!trans->block_rsv)
5708                 return;
5709
5710         if (!trans->bytes_reserved)
5711                 return;
5712
5713         trace_btrfs_space_reservation(root->fs_info, "transaction",
5714                                       trans->transid, trans->bytes_reserved, 0);
5715         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5716         trans->bytes_reserved = 0;
5717 }
5718
5719 /*
5720  * To be called after all the new block groups attached to the transaction
5721  * handle have been created (btrfs_create_pending_block_groups()).
5722  */
5723 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5724 {
5725         struct btrfs_fs_info *fs_info = trans->fs_info;
5726
5727         if (!trans->chunk_bytes_reserved)
5728                 return;
5729
5730         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5731
5732         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5733                                 trans->chunk_bytes_reserved);
5734         trans->chunk_bytes_reserved = 0;
5735 }
5736
5737 /* Can only return 0 or -ENOSPC */
5738 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5739                                   struct inode *inode)
5740 {
5741         struct btrfs_root *root = BTRFS_I(inode)->root;
5742         /*
5743          * We always use trans->block_rsv here as we will have reserved space
5744          * for our orphan when starting the transaction, using get_block_rsv()
5745          * here will sometimes make us choose the wrong block rsv as we could be
5746          * doing a reloc inode for a non refcounted root.
5747          */
5748         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5749         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5750
5751         /*
5752          * We need to hold space in order to delete our orphan item once we've
5753          * added it, so this takes the reservation so we can release it later
5754          * when we are truly done with the orphan item.
5755          */
5756         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5757         trace_btrfs_space_reservation(root->fs_info, "orphan",
5758                                       btrfs_ino(inode), num_bytes, 1);
5759         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5760 }
5761
5762 void btrfs_orphan_release_metadata(struct inode *inode)
5763 {
5764         struct btrfs_root *root = BTRFS_I(inode)->root;
5765         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5766         trace_btrfs_space_reservation(root->fs_info, "orphan",
5767                                       btrfs_ino(inode), num_bytes, 0);
5768         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5769 }
5770
5771 /*
5772  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5773  * root: the root of the parent directory
5774  * rsv: block reservation
5775  * items: the number of items that we need do reservation
5776  * qgroup_reserved: used to return the reserved size in qgroup
5777  *
5778  * This function is used to reserve the space for snapshot/subvolume
5779  * creation and deletion. Those operations are different with the
5780  * common file/directory operations, they change two fs/file trees
5781  * and root tree, the number of items that the qgroup reserves is
5782  * different with the free space reservation. So we can not use
5783  * the space reservation mechanism in start_transaction().
5784  */
5785 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5786                                      struct btrfs_block_rsv *rsv,
5787                                      int items,
5788                                      u64 *qgroup_reserved,
5789                                      bool use_global_rsv)
5790 {
5791         u64 num_bytes;
5792         int ret;
5793         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5794
5795         if (root->fs_info->quota_enabled) {
5796                 /* One for parent inode, two for dir entries */
5797                 num_bytes = 3 * root->nodesize;
5798                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5799                 if (ret)
5800                         return ret;
5801         } else {
5802                 num_bytes = 0;
5803         }
5804
5805         *qgroup_reserved = num_bytes;
5806
5807         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5808         rsv->space_info = __find_space_info(root->fs_info,
5809                                             BTRFS_BLOCK_GROUP_METADATA);
5810         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5811                                   BTRFS_RESERVE_FLUSH_ALL);
5812
5813         if (ret == -ENOSPC && use_global_rsv)
5814                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5815
5816         if (ret && *qgroup_reserved)
5817                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5818
5819         return ret;
5820 }
5821
5822 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5823                                       struct btrfs_block_rsv *rsv,
5824                                       u64 qgroup_reserved)
5825 {
5826         btrfs_block_rsv_release(root, rsv, (u64)-1);
5827 }
5828
5829 /**
5830  * drop_outstanding_extent - drop an outstanding extent
5831  * @inode: the inode we're dropping the extent for
5832  * @num_bytes: the number of bytes we're releasing.
5833  *
5834  * This is called when we are freeing up an outstanding extent, either called
5835  * after an error or after an extent is written.  This will return the number of
5836  * reserved extents that need to be freed.  This must be called with
5837  * BTRFS_I(inode)->lock held.
5838  */
5839 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5840 {
5841         unsigned drop_inode_space = 0;
5842         unsigned dropped_extents = 0;
5843         unsigned num_extents = 0;
5844
5845         num_extents = (unsigned)div64_u64(num_bytes +
5846                                           BTRFS_MAX_EXTENT_SIZE - 1,
5847                                           BTRFS_MAX_EXTENT_SIZE);
5848         ASSERT(num_extents);
5849         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5850         BTRFS_I(inode)->outstanding_extents -= num_extents;
5851
5852         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5853             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5854                                &BTRFS_I(inode)->runtime_flags))
5855                 drop_inode_space = 1;
5856
5857         /*
5858          * If we have more or the same amount of outstanding extents than we have
5859          * reserved then we need to leave the reserved extents count alone.
5860          */
5861         if (BTRFS_I(inode)->outstanding_extents >=
5862             BTRFS_I(inode)->reserved_extents)
5863                 return drop_inode_space;
5864
5865         dropped_extents = BTRFS_I(inode)->reserved_extents -
5866                 BTRFS_I(inode)->outstanding_extents;
5867         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5868         return dropped_extents + drop_inode_space;
5869 }
5870
5871 /**
5872  * calc_csum_metadata_size - return the amount of metadata space that must be
5873  *      reserved/freed for the given bytes.
5874  * @inode: the inode we're manipulating
5875  * @num_bytes: the number of bytes in question
5876  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5877  *
5878  * This adjusts the number of csum_bytes in the inode and then returns the
5879  * correct amount of metadata that must either be reserved or freed.  We
5880  * calculate how many checksums we can fit into one leaf and then divide the
5881  * number of bytes that will need to be checksumed by this value to figure out
5882  * how many checksums will be required.  If we are adding bytes then the number
5883  * may go up and we will return the number of additional bytes that must be
5884  * reserved.  If it is going down we will return the number of bytes that must
5885  * be freed.
5886  *
5887  * This must be called with BTRFS_I(inode)->lock held.
5888  */
5889 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5890                                    int reserve)
5891 {
5892         struct btrfs_root *root = BTRFS_I(inode)->root;
5893         u64 old_csums, num_csums;
5894
5895         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5896             BTRFS_I(inode)->csum_bytes == 0)
5897                 return 0;
5898
5899         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5900         if (reserve)
5901                 BTRFS_I(inode)->csum_bytes += num_bytes;
5902         else
5903                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5904         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5905
5906         /* No change, no need to reserve more */
5907         if (old_csums == num_csums)
5908                 return 0;
5909
5910         if (reserve)
5911                 return btrfs_calc_trans_metadata_size(root,
5912                                                       num_csums - old_csums);
5913
5914         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5915 }
5916
5917 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5918 {
5919         struct btrfs_root *root = BTRFS_I(inode)->root;
5920         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5921         u64 to_reserve = 0;
5922         u64 csum_bytes;
5923         unsigned nr_extents = 0;
5924         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5925         int ret = 0;
5926         bool delalloc_lock = true;
5927         u64 to_free = 0;
5928         unsigned dropped;
5929         bool release_extra = false;
5930
5931         /* If we are a free space inode we need to not flush since we will be in
5932          * the middle of a transaction commit.  We also don't need the delalloc
5933          * mutex since we won't race with anybody.  We need this mostly to make
5934          * lockdep shut its filthy mouth.
5935          *
5936          * If we have a transaction open (can happen if we call truncate_block
5937          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5938          */
5939         if (btrfs_is_free_space_inode(inode)) {
5940                 flush = BTRFS_RESERVE_NO_FLUSH;
5941                 delalloc_lock = false;
5942         } else if (current->journal_info) {
5943                 flush = BTRFS_RESERVE_FLUSH_LIMIT;
5944         }
5945
5946         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5947             btrfs_transaction_in_commit(root->fs_info))
5948                 schedule_timeout(1);
5949
5950         if (delalloc_lock)
5951                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5952
5953         num_bytes = ALIGN(num_bytes, root->sectorsize);
5954
5955         spin_lock(&BTRFS_I(inode)->lock);
5956         nr_extents = (unsigned)div64_u64(num_bytes +
5957                                          BTRFS_MAX_EXTENT_SIZE - 1,
5958                                          BTRFS_MAX_EXTENT_SIZE);
5959         BTRFS_I(inode)->outstanding_extents += nr_extents;
5960
5961         nr_extents = 0;
5962         if (BTRFS_I(inode)->outstanding_extents >
5963             BTRFS_I(inode)->reserved_extents)
5964                 nr_extents += BTRFS_I(inode)->outstanding_extents -
5965                         BTRFS_I(inode)->reserved_extents;
5966
5967         /* We always want to reserve a slot for updating the inode. */
5968         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents + 1);
5969         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5970         csum_bytes = BTRFS_I(inode)->csum_bytes;
5971         spin_unlock(&BTRFS_I(inode)->lock);
5972
5973         if (root->fs_info->quota_enabled) {
5974                 ret = btrfs_qgroup_reserve_meta(root,
5975                                 nr_extents * root->nodesize);
5976                 if (ret)
5977                         goto out_fail;
5978         }
5979
5980         ret = btrfs_block_rsv_add(root, block_rsv, to_reserve, flush);
5981         if (unlikely(ret)) {
5982                 btrfs_qgroup_free_meta(root, nr_extents * root->nodesize);
5983                 goto out_fail;
5984         }
5985
5986         spin_lock(&BTRFS_I(inode)->lock);
5987         if (test_and_set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5988                              &BTRFS_I(inode)->runtime_flags)) {
5989                 to_reserve -= btrfs_calc_trans_metadata_size(root, 1);
5990                 release_extra = true;
5991         }
5992         BTRFS_I(inode)->reserved_extents += nr_extents;
5993         spin_unlock(&BTRFS_I(inode)->lock);
5994
5995         if (delalloc_lock)
5996                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5997
5998         if (to_reserve)
5999                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
6000                                               btrfs_ino(inode), to_reserve, 1);
6001         if (release_extra)
6002                 btrfs_block_rsv_release(root, block_rsv,
6003                                         btrfs_calc_trans_metadata_size(root,
6004                                                                        1));
6005         return 0;
6006
6007 out_fail:
6008         spin_lock(&BTRFS_I(inode)->lock);
6009         dropped = drop_outstanding_extent(inode, num_bytes);
6010         /*
6011          * If the inodes csum_bytes is the same as the original
6012          * csum_bytes then we know we haven't raced with any free()ers
6013          * so we can just reduce our inodes csum bytes and carry on.
6014          */
6015         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
6016                 calc_csum_metadata_size(inode, num_bytes, 0);
6017         } else {
6018                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
6019                 u64 bytes;
6020
6021                 /*
6022                  * This is tricky, but first we need to figure out how much we
6023                  * freed from any free-ers that occurred during this
6024                  * reservation, so we reset ->csum_bytes to the csum_bytes
6025                  * before we dropped our lock, and then call the free for the
6026                  * number of bytes that were freed while we were trying our
6027                  * reservation.
6028                  */
6029                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
6030                 BTRFS_I(inode)->csum_bytes = csum_bytes;
6031                 to_free = calc_csum_metadata_size(inode, bytes, 0);
6032
6033
6034                 /*
6035                  * Now we need to see how much we would have freed had we not
6036                  * been making this reservation and our ->csum_bytes were not
6037                  * artificially inflated.
6038                  */
6039                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
6040                 bytes = csum_bytes - orig_csum_bytes;
6041                 bytes = calc_csum_metadata_size(inode, bytes, 0);
6042
6043                 /*
6044                  * Now reset ->csum_bytes to what it should be.  If bytes is
6045                  * more than to_free then we would have freed more space had we
6046                  * not had an artificially high ->csum_bytes, so we need to free
6047                  * the remainder.  If bytes is the same or less then we don't
6048                  * need to do anything, the other free-ers did the correct
6049                  * thing.
6050                  */
6051                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
6052                 if (bytes > to_free)
6053                         to_free = bytes - to_free;
6054                 else
6055                         to_free = 0;
6056         }
6057         spin_unlock(&BTRFS_I(inode)->lock);
6058         if (dropped)
6059                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
6060
6061         if (to_free) {
6062                 btrfs_block_rsv_release(root, block_rsv, to_free);
6063                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
6064                                               btrfs_ino(inode), to_free, 0);
6065         }
6066         if (delalloc_lock)
6067                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6068         return ret;
6069 }
6070
6071 /**
6072  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6073  * @inode: the inode to release the reservation for
6074  * @num_bytes: the number of bytes we're releasing
6075  *
6076  * This will release the metadata reservation for an inode.  This can be called
6077  * once we complete IO for a given set of bytes to release their metadata
6078  * reservations.
6079  */
6080 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
6081 {
6082         struct btrfs_root *root = BTRFS_I(inode)->root;
6083         u64 to_free = 0;
6084         unsigned dropped;
6085
6086         num_bytes = ALIGN(num_bytes, root->sectorsize);
6087         spin_lock(&BTRFS_I(inode)->lock);
6088         dropped = drop_outstanding_extent(inode, num_bytes);
6089
6090         if (num_bytes)
6091                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
6092         spin_unlock(&BTRFS_I(inode)->lock);
6093         if (dropped > 0)
6094                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
6095
6096         if (btrfs_is_testing(root->fs_info))
6097                 return;
6098
6099         trace_btrfs_space_reservation(root->fs_info, "delalloc",
6100                                       btrfs_ino(inode), to_free, 0);
6101
6102         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
6103                                 to_free);
6104 }
6105
6106 /**
6107  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6108  * delalloc
6109  * @inode: inode we're writing to
6110  * @start: start range we are writing to
6111  * @len: how long the range we are writing to
6112  *
6113  * TODO: This function will finally replace old btrfs_delalloc_reserve_space()
6114  *
6115  * This will do the following things
6116  *
6117  * o reserve space in data space info for num bytes
6118  *   and reserve precious corresponding qgroup space
6119  *   (Done in check_data_free_space)
6120  *
6121  * o reserve space for metadata space, based on the number of outstanding
6122  *   extents and how much csums will be needed
6123  *   also reserve metadata space in a per root over-reserve method.
6124  * o add to the inodes->delalloc_bytes
6125  * o add it to the fs_info's delalloc inodes list.
6126  *   (Above 3 all done in delalloc_reserve_metadata)
6127  *
6128  * Return 0 for success
6129  * Return <0 for error(-ENOSPC or -EQUOT)
6130  */
6131 int btrfs_delalloc_reserve_space(struct inode *inode, u64 start, u64 len)
6132 {
6133         int ret;
6134
6135         ret = btrfs_check_data_free_space(inode, start, len);
6136         if (ret < 0)
6137                 return ret;
6138         ret = btrfs_delalloc_reserve_metadata(inode, len);
6139         if (ret < 0)
6140                 btrfs_free_reserved_data_space(inode, start, len);
6141         return ret;
6142 }
6143
6144 /**
6145  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6146  * @inode: inode we're releasing space for
6147  * @start: start position of the space already reserved
6148  * @len: the len of the space already reserved
6149  *
6150  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
6151  * called in the case that we don't need the metadata AND data reservations
6152  * anymore.  So if there is an error or we insert an inline extent.
6153  *
6154  * This function will release the metadata space that was not used and will
6155  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6156  * list if there are no delalloc bytes left.
6157  * Also it will handle the qgroup reserved space.
6158  */
6159 void btrfs_delalloc_release_space(struct inode *inode, u64 start, u64 len)
6160 {
6161         btrfs_delalloc_release_metadata(inode, len);
6162         btrfs_free_reserved_data_space(inode, start, len);
6163 }
6164
6165 static int update_block_group(struct btrfs_trans_handle *trans,
6166                               struct btrfs_root *root, u64 bytenr,
6167                               u64 num_bytes, int alloc)
6168 {
6169         struct btrfs_block_group_cache *cache = NULL;
6170         struct btrfs_fs_info *info = root->fs_info;
6171         u64 total = num_bytes;
6172         u64 old_val;
6173         u64 byte_in_group;
6174         int factor;
6175
6176         /* block accounting for super block */
6177         spin_lock(&info->delalloc_root_lock);
6178         old_val = btrfs_super_bytes_used(info->super_copy);
6179         if (alloc)
6180                 old_val += num_bytes;
6181         else
6182                 old_val -= num_bytes;
6183         btrfs_set_super_bytes_used(info->super_copy, old_val);
6184         spin_unlock(&info->delalloc_root_lock);
6185
6186         while (total) {
6187                 cache = btrfs_lookup_block_group(info, bytenr);
6188                 if (!cache)
6189                         return -ENOENT;
6190                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6191                                     BTRFS_BLOCK_GROUP_RAID1 |
6192                                     BTRFS_BLOCK_GROUP_RAID10))
6193                         factor = 2;
6194                 else
6195                         factor = 1;
6196                 /*
6197                  * If this block group has free space cache written out, we
6198                  * need to make sure to load it if we are removing space.  This
6199                  * is because we need the unpinning stage to actually add the
6200                  * space back to the block group, otherwise we will leak space.
6201                  */
6202                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6203                         cache_block_group(cache, 1);
6204
6205                 byte_in_group = bytenr - cache->key.objectid;
6206                 WARN_ON(byte_in_group > cache->key.offset);
6207
6208                 spin_lock(&cache->space_info->lock);
6209                 spin_lock(&cache->lock);
6210
6211                 if (btrfs_test_opt(root->fs_info, SPACE_CACHE) &&
6212                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6213                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6214
6215                 old_val = btrfs_block_group_used(&cache->item);
6216                 num_bytes = min(total, cache->key.offset - byte_in_group);
6217                 if (alloc) {
6218                         old_val += num_bytes;
6219                         btrfs_set_block_group_used(&cache->item, old_val);
6220                         cache->reserved -= num_bytes;
6221                         cache->space_info->bytes_reserved -= num_bytes;
6222                         cache->space_info->bytes_used += num_bytes;
6223                         cache->space_info->disk_used += num_bytes * factor;
6224                         spin_unlock(&cache->lock);
6225                         spin_unlock(&cache->space_info->lock);
6226                 } else {
6227                         old_val -= num_bytes;
6228                         btrfs_set_block_group_used(&cache->item, old_val);
6229                         cache->pinned += num_bytes;
6230                         cache->space_info->bytes_pinned += num_bytes;
6231                         cache->space_info->bytes_used -= num_bytes;
6232                         cache->space_info->disk_used -= num_bytes * factor;
6233                         spin_unlock(&cache->lock);
6234                         spin_unlock(&cache->space_info->lock);
6235
6236                         trace_btrfs_space_reservation(root->fs_info, "pinned",
6237                                                       cache->space_info->flags,
6238                                                       num_bytes, 1);
6239                         set_extent_dirty(info->pinned_extents,
6240                                          bytenr, bytenr + num_bytes - 1,
6241                                          GFP_NOFS | __GFP_NOFAIL);
6242                 }
6243
6244                 spin_lock(&trans->transaction->dirty_bgs_lock);
6245                 if (list_empty(&cache->dirty_list)) {
6246                         list_add_tail(&cache->dirty_list,
6247                                       &trans->transaction->dirty_bgs);
6248                                 trans->transaction->num_dirty_bgs++;
6249                         btrfs_get_block_group(cache);
6250                 }
6251                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6252
6253                 /*
6254                  * No longer have used bytes in this block group, queue it for
6255                  * deletion. We do this after adding the block group to the
6256                  * dirty list to avoid races between cleaner kthread and space
6257                  * cache writeout.
6258                  */
6259                 if (!alloc && old_val == 0) {
6260                         spin_lock(&info->unused_bgs_lock);
6261                         if (list_empty(&cache->bg_list)) {
6262                                 btrfs_get_block_group(cache);
6263                                 list_add_tail(&cache->bg_list,
6264                                               &info->unused_bgs);
6265                         }
6266                         spin_unlock(&info->unused_bgs_lock);
6267                 }
6268
6269                 btrfs_put_block_group(cache);
6270                 total -= num_bytes;
6271                 bytenr += num_bytes;
6272         }
6273         return 0;
6274 }
6275
6276 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
6277 {
6278         struct btrfs_block_group_cache *cache;
6279         u64 bytenr;
6280
6281         spin_lock(&root->fs_info->block_group_cache_lock);
6282         bytenr = root->fs_info->first_logical_byte;
6283         spin_unlock(&root->fs_info->block_group_cache_lock);
6284
6285         if (bytenr < (u64)-1)
6286                 return bytenr;
6287
6288         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
6289         if (!cache)
6290                 return 0;
6291
6292         bytenr = cache->key.objectid;
6293         btrfs_put_block_group(cache);
6294
6295         return bytenr;
6296 }
6297
6298 static int pin_down_extent(struct btrfs_root *root,
6299                            struct btrfs_block_group_cache *cache,
6300                            u64 bytenr, u64 num_bytes, int reserved)
6301 {
6302         spin_lock(&cache->space_info->lock);
6303         spin_lock(&cache->lock);
6304         cache->pinned += num_bytes;
6305         cache->space_info->bytes_pinned += num_bytes;
6306         if (reserved) {
6307                 cache->reserved -= num_bytes;
6308                 cache->space_info->bytes_reserved -= num_bytes;
6309         }
6310         spin_unlock(&cache->lock);
6311         spin_unlock(&cache->space_info->lock);
6312
6313         trace_btrfs_space_reservation(root->fs_info, "pinned",
6314                                       cache->space_info->flags, num_bytes, 1);
6315         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
6316                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6317         return 0;
6318 }
6319
6320 /*
6321  * this function must be called within transaction
6322  */
6323 int btrfs_pin_extent(struct btrfs_root *root,
6324                      u64 bytenr, u64 num_bytes, int reserved)
6325 {
6326         struct btrfs_block_group_cache *cache;
6327
6328         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6329         BUG_ON(!cache); /* Logic error */
6330
6331         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
6332
6333         btrfs_put_block_group(cache);
6334         return 0;
6335 }
6336
6337 /*
6338  * this function must be called within transaction
6339  */
6340 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
6341                                     u64 bytenr, u64 num_bytes)
6342 {
6343         struct btrfs_block_group_cache *cache;
6344         int ret;
6345
6346         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6347         if (!cache)
6348                 return -EINVAL;
6349
6350         /*
6351          * pull in the free space cache (if any) so that our pin
6352          * removes the free space from the cache.  We have load_only set
6353          * to one because the slow code to read in the free extents does check
6354          * the pinned extents.
6355          */
6356         cache_block_group(cache, 1);
6357
6358         pin_down_extent(root, cache, bytenr, num_bytes, 0);
6359
6360         /* remove us from the free space cache (if we're there at all) */
6361         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6362         btrfs_put_block_group(cache);
6363         return ret;
6364 }
6365
6366 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
6367 {
6368         int ret;
6369         struct btrfs_block_group_cache *block_group;
6370         struct btrfs_caching_control *caching_ctl;
6371
6372         block_group = btrfs_lookup_block_group(root->fs_info, start);
6373         if (!block_group)
6374                 return -EINVAL;
6375
6376         cache_block_group(block_group, 0);
6377         caching_ctl = get_caching_control(block_group);
6378
6379         if (!caching_ctl) {
6380                 /* Logic error */
6381                 BUG_ON(!block_group_cache_done(block_group));
6382                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6383         } else {
6384                 mutex_lock(&caching_ctl->mutex);
6385
6386                 if (start >= caching_ctl->progress) {
6387                         ret = add_excluded_extent(root, start, num_bytes);
6388                 } else if (start + num_bytes <= caching_ctl->progress) {
6389                         ret = btrfs_remove_free_space(block_group,
6390                                                       start, num_bytes);
6391                 } else {
6392                         num_bytes = caching_ctl->progress - start;
6393                         ret = btrfs_remove_free_space(block_group,
6394                                                       start, num_bytes);
6395                         if (ret)
6396                                 goto out_lock;
6397
6398                         num_bytes = (start + num_bytes) -
6399                                 caching_ctl->progress;
6400                         start = caching_ctl->progress;
6401                         ret = add_excluded_extent(root, start, num_bytes);
6402                 }
6403 out_lock:
6404                 mutex_unlock(&caching_ctl->mutex);
6405                 put_caching_control(caching_ctl);
6406         }
6407         btrfs_put_block_group(block_group);
6408         return ret;
6409 }
6410
6411 int btrfs_exclude_logged_extents(struct btrfs_root *log,
6412                                  struct extent_buffer *eb)
6413 {
6414         struct btrfs_file_extent_item *item;
6415         struct btrfs_key key;
6416         int found_type;
6417         int i;
6418
6419         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
6420                 return 0;
6421
6422         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6423                 btrfs_item_key_to_cpu(eb, &key, i);
6424                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6425                         continue;
6426                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6427                 found_type = btrfs_file_extent_type(eb, item);
6428                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6429                         continue;
6430                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6431                         continue;
6432                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6433                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6434                 __exclude_logged_extent(log, key.objectid, key.offset);
6435         }
6436
6437         return 0;
6438 }
6439
6440 static void
6441 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6442 {
6443         atomic_inc(&bg->reservations);
6444 }
6445
6446 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6447                                         const u64 start)
6448 {
6449         struct btrfs_block_group_cache *bg;
6450
6451         bg = btrfs_lookup_block_group(fs_info, start);
6452         ASSERT(bg);
6453         if (atomic_dec_and_test(&bg->reservations))
6454                 wake_up_atomic_t(&bg->reservations);
6455         btrfs_put_block_group(bg);
6456 }
6457
6458 static int btrfs_wait_bg_reservations_atomic_t(atomic_t *a)
6459 {
6460         schedule();
6461         return 0;
6462 }
6463
6464 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6465 {
6466         struct btrfs_space_info *space_info = bg->space_info;
6467
6468         ASSERT(bg->ro);
6469
6470         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6471                 return;
6472
6473         /*
6474          * Our block group is read only but before we set it to read only,
6475          * some task might have had allocated an extent from it already, but it
6476          * has not yet created a respective ordered extent (and added it to a
6477          * root's list of ordered extents).
6478          * Therefore wait for any task currently allocating extents, since the
6479          * block group's reservations counter is incremented while a read lock
6480          * on the groups' semaphore is held and decremented after releasing
6481          * the read access on that semaphore and creating the ordered extent.
6482          */
6483         down_write(&space_info->groups_sem);
6484         up_write(&space_info->groups_sem);
6485
6486         wait_on_atomic_t(&bg->reservations,
6487                          btrfs_wait_bg_reservations_atomic_t,
6488                          TASK_UNINTERRUPTIBLE);
6489 }
6490
6491 /**
6492  * btrfs_add_reserved_bytes - update the block_group and space info counters
6493  * @cache:      The cache we are manipulating
6494  * @ram_bytes:  The number of bytes of file content, and will be same to
6495  *              @num_bytes except for the compress path.
6496  * @num_bytes:  The number of bytes in question
6497  * @delalloc:   The blocks are allocated for the delalloc write
6498  *
6499  * This is called by the allocator when it reserves space. Metadata
6500  * reservations should be called with RESERVE_ALLOC so we do the proper
6501  * ENOSPC accounting.  For data we handle the reservation through clearing the
6502  * delalloc bits in the io_tree.  We have to do this since we could end up
6503  * allocating less disk space for the amount of data we have reserved in the
6504  * case of compression.
6505  *
6506  * If this is a reservation and the block group has become read only we cannot
6507  * make the reservation and return -EAGAIN, otherwise this function always
6508  * succeeds.
6509  */
6510 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
6511                                     u64 ram_bytes, u64 num_bytes, int delalloc)
6512 {
6513         struct btrfs_space_info *space_info = cache->space_info;
6514         int ret = 0;
6515
6516         spin_lock(&space_info->lock);
6517         spin_lock(&cache->lock);
6518         if (cache->ro) {
6519                 ret = -EAGAIN;
6520         } else {
6521                 cache->reserved += num_bytes;
6522                 space_info->bytes_reserved += num_bytes;
6523
6524                 trace_btrfs_space_reservation(cache->fs_info,
6525                                 "space_info", space_info->flags,
6526                                 ram_bytes, 0);
6527                 space_info->bytes_may_use -= ram_bytes;
6528                 if (delalloc)
6529                         cache->delalloc_bytes += num_bytes;
6530         }
6531         spin_unlock(&cache->lock);
6532         spin_unlock(&space_info->lock);
6533         return ret;
6534 }
6535
6536 /**
6537  * btrfs_free_reserved_bytes - update the block_group and space info counters
6538  * @cache:      The cache we are manipulating
6539  * @num_bytes:  The number of bytes in question
6540  * @delalloc:   The blocks are allocated for the delalloc write
6541  *
6542  * This is called by somebody who is freeing space that was never actually used
6543  * on disk.  For example if you reserve some space for a new leaf in transaction
6544  * A and before transaction A commits you free that leaf, you call this with
6545  * reserve set to 0 in order to clear the reservation.
6546  */
6547
6548 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
6549                                      u64 num_bytes, int delalloc)
6550 {
6551         struct btrfs_space_info *space_info = cache->space_info;
6552         int ret = 0;
6553
6554         spin_lock(&space_info->lock);
6555         spin_lock(&cache->lock);
6556         if (cache->ro)
6557                 space_info->bytes_readonly += num_bytes;
6558         cache->reserved -= num_bytes;
6559         space_info->bytes_reserved -= num_bytes;
6560
6561         if (delalloc)
6562                 cache->delalloc_bytes -= num_bytes;
6563         spin_unlock(&cache->lock);
6564         spin_unlock(&space_info->lock);
6565         return ret;
6566 }
6567 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6568                                 struct btrfs_root *root)
6569 {
6570         struct btrfs_fs_info *fs_info = root->fs_info;
6571         struct btrfs_caching_control *next;
6572         struct btrfs_caching_control *caching_ctl;
6573         struct btrfs_block_group_cache *cache;
6574
6575         down_write(&fs_info->commit_root_sem);
6576
6577         list_for_each_entry_safe(caching_ctl, next,
6578                                  &fs_info->caching_block_groups, list) {
6579                 cache = caching_ctl->block_group;
6580                 if (block_group_cache_done(cache)) {
6581                         cache->last_byte_to_unpin = (u64)-1;
6582                         list_del_init(&caching_ctl->list);
6583                         put_caching_control(caching_ctl);
6584                 } else {
6585                         cache->last_byte_to_unpin = caching_ctl->progress;
6586                 }
6587         }
6588
6589         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6590                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6591         else
6592                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6593
6594         up_write(&fs_info->commit_root_sem);
6595
6596         update_global_block_rsv(fs_info);
6597 }
6598
6599 /*
6600  * Returns the free cluster for the given space info and sets empty_cluster to
6601  * what it should be based on the mount options.
6602  */
6603 static struct btrfs_free_cluster *
6604 fetch_cluster_info(struct btrfs_root *root, struct btrfs_space_info *space_info,
6605                    u64 *empty_cluster)
6606 {
6607         struct btrfs_free_cluster *ret = NULL;
6608         bool ssd = btrfs_test_opt(root->fs_info, SSD);
6609
6610         *empty_cluster = 0;
6611         if (btrfs_mixed_space_info(space_info))
6612                 return ret;
6613
6614         if (ssd)
6615                 *empty_cluster = SZ_2M;
6616         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6617                 ret = &root->fs_info->meta_alloc_cluster;
6618                 if (!ssd)
6619                         *empty_cluster = SZ_64K;
6620         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) && ssd) {
6621                 ret = &root->fs_info->data_alloc_cluster;
6622         }
6623
6624         return ret;
6625 }
6626
6627 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
6628                               const bool return_free_space)
6629 {
6630         struct btrfs_fs_info *fs_info = root->fs_info;
6631         struct btrfs_block_group_cache *cache = NULL;
6632         struct btrfs_space_info *space_info;
6633         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6634         struct btrfs_free_cluster *cluster = NULL;
6635         u64 len;
6636         u64 total_unpinned = 0;
6637         u64 empty_cluster = 0;
6638         bool readonly;
6639
6640         while (start <= end) {
6641                 readonly = false;
6642                 if (!cache ||
6643                     start >= cache->key.objectid + cache->key.offset) {
6644                         if (cache)
6645                                 btrfs_put_block_group(cache);
6646                         total_unpinned = 0;
6647                         cache = btrfs_lookup_block_group(fs_info, start);
6648                         BUG_ON(!cache); /* Logic error */
6649
6650                         cluster = fetch_cluster_info(root,
6651                                                      cache->space_info,
6652                                                      &empty_cluster);
6653                         empty_cluster <<= 1;
6654                 }
6655
6656                 len = cache->key.objectid + cache->key.offset - start;
6657                 len = min(len, end + 1 - start);
6658
6659                 if (start < cache->last_byte_to_unpin) {
6660                         len = min(len, cache->last_byte_to_unpin - start);
6661                         if (return_free_space)
6662                                 btrfs_add_free_space(cache, start, len);
6663                 }
6664
6665                 start += len;
6666                 total_unpinned += len;
6667                 space_info = cache->space_info;
6668
6669                 /*
6670                  * If this space cluster has been marked as fragmented and we've
6671                  * unpinned enough in this block group to potentially allow a
6672                  * cluster to be created inside of it go ahead and clear the
6673                  * fragmented check.
6674                  */
6675                 if (cluster && cluster->fragmented &&
6676                     total_unpinned > empty_cluster) {
6677                         spin_lock(&cluster->lock);
6678                         cluster->fragmented = 0;
6679                         spin_unlock(&cluster->lock);
6680                 }
6681
6682                 spin_lock(&space_info->lock);
6683                 spin_lock(&cache->lock);
6684                 cache->pinned -= len;
6685                 space_info->bytes_pinned -= len;
6686
6687                 trace_btrfs_space_reservation(fs_info, "pinned",
6688                                               space_info->flags, len, 0);
6689                 space_info->max_extent_size = 0;
6690                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6691                 if (cache->ro) {
6692                         space_info->bytes_readonly += len;
6693                         readonly = true;
6694                 }
6695                 spin_unlock(&cache->lock);
6696                 if (!readonly && return_free_space &&
6697                     global_rsv->space_info == space_info) {
6698                         u64 to_add = len;
6699                         WARN_ON(!return_free_space);
6700                         spin_lock(&global_rsv->lock);
6701                         if (!global_rsv->full) {
6702                                 to_add = min(len, global_rsv->size -
6703                                              global_rsv->reserved);
6704                                 global_rsv->reserved += to_add;
6705                                 space_info->bytes_may_use += to_add;
6706                                 if (global_rsv->reserved >= global_rsv->size)
6707                                         global_rsv->full = 1;
6708                                 trace_btrfs_space_reservation(fs_info,
6709                                                               "space_info",
6710                                                               space_info->flags,
6711                                                               to_add, 1);
6712                                 len -= to_add;
6713                         }
6714                         spin_unlock(&global_rsv->lock);
6715                         /* Add to any tickets we may have */
6716                         if (len)
6717                                 space_info_add_new_bytes(fs_info, space_info,
6718                                                          len);
6719                 }
6720                 spin_unlock(&space_info->lock);
6721         }
6722
6723         if (cache)
6724                 btrfs_put_block_group(cache);
6725         return 0;
6726 }
6727
6728 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6729                                struct btrfs_root *root)
6730 {
6731         struct btrfs_fs_info *fs_info = root->fs_info;
6732         struct btrfs_block_group_cache *block_group, *tmp;
6733         struct list_head *deleted_bgs;
6734         struct extent_io_tree *unpin;
6735         u64 start;
6736         u64 end;
6737         int ret;
6738
6739         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6740                 unpin = &fs_info->freed_extents[1];
6741         else
6742                 unpin = &fs_info->freed_extents[0];
6743
6744         while (!trans->aborted) {
6745                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6746                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6747                                             EXTENT_DIRTY, NULL);
6748                 if (ret) {
6749                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6750                         break;
6751                 }
6752
6753                 if (btrfs_test_opt(root->fs_info, DISCARD))
6754                         ret = btrfs_discard_extent(root, start,
6755                                                    end + 1 - start, NULL);
6756
6757                 clear_extent_dirty(unpin, start, end);
6758                 unpin_extent_range(root, start, end, true);
6759                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6760                 cond_resched();
6761         }
6762
6763         /*
6764          * Transaction is finished.  We don't need the lock anymore.  We
6765          * do need to clean up the block groups in case of a transaction
6766          * abort.
6767          */
6768         deleted_bgs = &trans->transaction->deleted_bgs;
6769         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6770                 u64 trimmed = 0;
6771
6772                 ret = -EROFS;
6773                 if (!trans->aborted)
6774                         ret = btrfs_discard_extent(root,
6775                                                    block_group->key.objectid,
6776                                                    block_group->key.offset,
6777                                                    &trimmed);
6778
6779                 list_del_init(&block_group->bg_list);
6780                 btrfs_put_block_group_trimming(block_group);
6781                 btrfs_put_block_group(block_group);
6782
6783                 if (ret) {
6784                         const char *errstr = btrfs_decode_error(ret);
6785                         btrfs_warn(fs_info,
6786                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6787                                    ret, errstr);
6788                 }
6789         }
6790
6791         return 0;
6792 }
6793
6794 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6795                              u64 owner, u64 root_objectid)
6796 {
6797         struct btrfs_space_info *space_info;
6798         u64 flags;
6799
6800         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6801                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6802                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6803                 else
6804                         flags = BTRFS_BLOCK_GROUP_METADATA;
6805         } else {
6806                 flags = BTRFS_BLOCK_GROUP_DATA;
6807         }
6808
6809         space_info = __find_space_info(fs_info, flags);
6810         BUG_ON(!space_info); /* Logic bug */
6811         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6812 }
6813
6814
6815 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6816                                 struct btrfs_root *root,
6817                                 struct btrfs_delayed_ref_node *node, u64 parent,
6818                                 u64 root_objectid, u64 owner_objectid,
6819                                 u64 owner_offset, int refs_to_drop,
6820                                 struct btrfs_delayed_extent_op *extent_op)
6821 {
6822         struct btrfs_key key;
6823         struct btrfs_path *path;
6824         struct btrfs_fs_info *info = root->fs_info;
6825         struct btrfs_root *extent_root = info->extent_root;
6826         struct extent_buffer *leaf;
6827         struct btrfs_extent_item *ei;
6828         struct btrfs_extent_inline_ref *iref;
6829         int ret;
6830         int is_data;
6831         int extent_slot = 0;
6832         int found_extent = 0;
6833         int num_to_del = 1;
6834         u32 item_size;
6835         u64 refs;
6836         u64 bytenr = node->bytenr;
6837         u64 num_bytes = node->num_bytes;
6838         int last_ref = 0;
6839         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6840                                                  SKINNY_METADATA);
6841
6842         path = btrfs_alloc_path();
6843         if (!path)
6844                 return -ENOMEM;
6845
6846         path->reada = READA_FORWARD;
6847         path->leave_spinning = 1;
6848
6849         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6850         BUG_ON(!is_data && refs_to_drop != 1);
6851
6852         if (is_data)
6853                 skinny_metadata = 0;
6854
6855         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6856                                     bytenr, num_bytes, parent,
6857                                     root_objectid, owner_objectid,
6858                                     owner_offset);
6859         if (ret == 0) {
6860                 extent_slot = path->slots[0];
6861                 while (extent_slot >= 0) {
6862                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6863                                               extent_slot);
6864                         if (key.objectid != bytenr)
6865                                 break;
6866                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6867                             key.offset == num_bytes) {
6868                                 found_extent = 1;
6869                                 break;
6870                         }
6871                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6872                             key.offset == owner_objectid) {
6873                                 found_extent = 1;
6874                                 break;
6875                         }
6876                         if (path->slots[0] - extent_slot > 5)
6877                                 break;
6878                         extent_slot--;
6879                 }
6880 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6881                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6882                 if (found_extent && item_size < sizeof(*ei))
6883                         found_extent = 0;
6884 #endif
6885                 if (!found_extent) {
6886                         BUG_ON(iref);
6887                         ret = remove_extent_backref(trans, extent_root, path,
6888                                                     NULL, refs_to_drop,
6889                                                     is_data, &last_ref);
6890                         if (ret) {
6891                                 btrfs_abort_transaction(trans, ret);
6892                                 goto out;
6893                         }
6894                         btrfs_release_path(path);
6895                         path->leave_spinning = 1;
6896
6897                         key.objectid = bytenr;
6898                         key.type = BTRFS_EXTENT_ITEM_KEY;
6899                         key.offset = num_bytes;
6900
6901                         if (!is_data && skinny_metadata) {
6902                                 key.type = BTRFS_METADATA_ITEM_KEY;
6903                                 key.offset = owner_objectid;
6904                         }
6905
6906                         ret = btrfs_search_slot(trans, extent_root,
6907                                                 &key, path, -1, 1);
6908                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6909                                 /*
6910                                  * Couldn't find our skinny metadata item,
6911                                  * see if we have ye olde extent item.
6912                                  */
6913                                 path->slots[0]--;
6914                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6915                                                       path->slots[0]);
6916                                 if (key.objectid == bytenr &&
6917                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6918                                     key.offset == num_bytes)
6919                                         ret = 0;
6920                         }
6921
6922                         if (ret > 0 && skinny_metadata) {
6923                                 skinny_metadata = false;
6924                                 key.objectid = bytenr;
6925                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6926                                 key.offset = num_bytes;
6927                                 btrfs_release_path(path);
6928                                 ret = btrfs_search_slot(trans, extent_root,
6929                                                         &key, path, -1, 1);
6930                         }
6931
6932                         if (ret) {
6933                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6934                                         ret, bytenr);
6935                                 if (ret > 0)
6936                                         btrfs_print_leaf(extent_root,
6937                                                          path->nodes[0]);
6938                         }
6939                         if (ret < 0) {
6940                                 btrfs_abort_transaction(trans, ret);
6941                                 goto out;
6942                         }
6943                         extent_slot = path->slots[0];
6944                 }
6945         } else if (WARN_ON(ret == -ENOENT)) {
6946                 btrfs_print_leaf(extent_root, path->nodes[0]);
6947                 btrfs_err(info,
6948                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6949                         bytenr, parent, root_objectid, owner_objectid,
6950                         owner_offset);
6951                 btrfs_abort_transaction(trans, ret);
6952                 goto out;
6953         } else {
6954                 btrfs_abort_transaction(trans, ret);
6955                 goto out;
6956         }
6957
6958         leaf = path->nodes[0];
6959         item_size = btrfs_item_size_nr(leaf, extent_slot);
6960 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6961         if (item_size < sizeof(*ei)) {
6962                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6963                 ret = convert_extent_item_v0(trans, extent_root, path,
6964                                              owner_objectid, 0);
6965                 if (ret < 0) {
6966                         btrfs_abort_transaction(trans, ret);
6967                         goto out;
6968                 }
6969
6970                 btrfs_release_path(path);
6971                 path->leave_spinning = 1;
6972
6973                 key.objectid = bytenr;
6974                 key.type = BTRFS_EXTENT_ITEM_KEY;
6975                 key.offset = num_bytes;
6976
6977                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6978                                         -1, 1);
6979                 if (ret) {
6980                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6981                                 ret, bytenr);
6982                         btrfs_print_leaf(extent_root, path->nodes[0]);
6983                 }
6984                 if (ret < 0) {
6985                         btrfs_abort_transaction(trans, ret);
6986                         goto out;
6987                 }
6988
6989                 extent_slot = path->slots[0];
6990                 leaf = path->nodes[0];
6991                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6992         }
6993 #endif
6994         BUG_ON(item_size < sizeof(*ei));
6995         ei = btrfs_item_ptr(leaf, extent_slot,
6996                             struct btrfs_extent_item);
6997         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6998             key.type == BTRFS_EXTENT_ITEM_KEY) {
6999                 struct btrfs_tree_block_info *bi;
7000                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7001                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7002                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7003         }
7004
7005         refs = btrfs_extent_refs(leaf, ei);
7006         if (refs < refs_to_drop) {
7007                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
7008                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
7009                 ret = -EINVAL;
7010                 btrfs_abort_transaction(trans, ret);
7011                 goto out;
7012         }
7013         refs -= refs_to_drop;
7014
7015         if (refs > 0) {
7016                 if (extent_op)
7017                         __run_delayed_extent_op(extent_op, leaf, ei);
7018                 /*
7019                  * In the case of inline back ref, reference count will
7020                  * be updated by remove_extent_backref
7021                  */
7022                 if (iref) {
7023                         BUG_ON(!found_extent);
7024                 } else {
7025                         btrfs_set_extent_refs(leaf, ei, refs);
7026                         btrfs_mark_buffer_dirty(leaf);
7027                 }
7028                 if (found_extent) {
7029                         ret = remove_extent_backref(trans, extent_root, path,
7030                                                     iref, refs_to_drop,
7031                                                     is_data, &last_ref);
7032                         if (ret) {
7033                                 btrfs_abort_transaction(trans, ret);
7034                                 goto out;
7035                         }
7036                 }
7037                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
7038                                  root_objectid);
7039         } else {
7040                 if (found_extent) {
7041                         BUG_ON(is_data && refs_to_drop !=
7042                                extent_data_ref_count(path, iref));
7043                         if (iref) {
7044                                 BUG_ON(path->slots[0] != extent_slot);
7045                         } else {
7046                                 BUG_ON(path->slots[0] != extent_slot + 1);
7047                                 path->slots[0] = extent_slot;
7048                                 num_to_del = 2;
7049                         }
7050                 }
7051
7052                 last_ref = 1;
7053                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7054                                       num_to_del);
7055                 if (ret) {
7056                         btrfs_abort_transaction(trans, ret);
7057                         goto out;
7058                 }
7059                 btrfs_release_path(path);
7060
7061                 if (is_data) {
7062                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
7063                         if (ret) {
7064                                 btrfs_abort_transaction(trans, ret);
7065                                 goto out;
7066                         }
7067                 }
7068
7069                 ret = add_to_free_space_tree(trans, root->fs_info, bytenr,
7070                                              num_bytes);
7071                 if (ret) {
7072                         btrfs_abort_transaction(trans, ret);
7073                         goto out;
7074                 }
7075
7076                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
7077                 if (ret) {
7078                         btrfs_abort_transaction(trans, ret);
7079                         goto out;
7080                 }
7081         }
7082         btrfs_release_path(path);
7083
7084 out:
7085         btrfs_free_path(path);
7086         return ret;
7087 }
7088
7089 /*
7090  * when we free an block, it is possible (and likely) that we free the last
7091  * delayed ref for that extent as well.  This searches the delayed ref tree for
7092  * a given extent, and if there are no other delayed refs to be processed, it
7093  * removes it from the tree.
7094  */
7095 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7096                                       struct btrfs_root *root, u64 bytenr)
7097 {
7098         struct btrfs_delayed_ref_head *head;
7099         struct btrfs_delayed_ref_root *delayed_refs;
7100         int ret = 0;
7101
7102         delayed_refs = &trans->transaction->delayed_refs;
7103         spin_lock(&delayed_refs->lock);
7104         head = btrfs_find_delayed_ref_head(trans, bytenr);
7105         if (!head)
7106                 goto out_delayed_unlock;
7107
7108         spin_lock(&head->lock);
7109         if (!list_empty(&head->ref_list))
7110                 goto out;
7111
7112         if (head->extent_op) {
7113                 if (!head->must_insert_reserved)
7114                         goto out;
7115                 btrfs_free_delayed_extent_op(head->extent_op);
7116                 head->extent_op = NULL;
7117         }
7118
7119         /*
7120          * waiting for the lock here would deadlock.  If someone else has it
7121          * locked they are already in the process of dropping it anyway
7122          */
7123         if (!mutex_trylock(&head->mutex))
7124                 goto out;
7125
7126         /*
7127          * at this point we have a head with no other entries.  Go
7128          * ahead and process it.
7129          */
7130         head->node.in_tree = 0;
7131         rb_erase(&head->href_node, &delayed_refs->href_root);
7132
7133         atomic_dec(&delayed_refs->num_entries);
7134
7135         /*
7136          * we don't take a ref on the node because we're removing it from the
7137          * tree, so we just steal the ref the tree was holding.
7138          */
7139         delayed_refs->num_heads--;
7140         if (head->processing == 0)
7141                 delayed_refs->num_heads_ready--;
7142         head->processing = 0;
7143         spin_unlock(&head->lock);
7144         spin_unlock(&delayed_refs->lock);
7145
7146         BUG_ON(head->extent_op);
7147         if (head->must_insert_reserved)
7148                 ret = 1;
7149
7150         mutex_unlock(&head->mutex);
7151         btrfs_put_delayed_ref(&head->node);
7152         return ret;
7153 out:
7154         spin_unlock(&head->lock);
7155
7156 out_delayed_unlock:
7157         spin_unlock(&delayed_refs->lock);
7158         return 0;
7159 }
7160
7161 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7162                            struct btrfs_root *root,
7163                            struct extent_buffer *buf,
7164                            u64 parent, int last_ref)
7165 {
7166         int pin = 1;
7167         int ret;
7168
7169         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7170                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7171                                         buf->start, buf->len,
7172                                         parent, root->root_key.objectid,
7173                                         btrfs_header_level(buf),
7174                                         BTRFS_DROP_DELAYED_REF, NULL);
7175                 BUG_ON(ret); /* -ENOMEM */
7176         }
7177
7178         if (!last_ref)
7179                 return;
7180
7181         if (btrfs_header_generation(buf) == trans->transid) {
7182                 struct btrfs_block_group_cache *cache;
7183
7184                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7185                         ret = check_ref_cleanup(trans, root, buf->start);
7186                         if (!ret)
7187                                 goto out;
7188                 }
7189
7190                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
7191
7192                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7193                         pin_down_extent(root, cache, buf->start, buf->len, 1);
7194                         btrfs_put_block_group(cache);
7195                         goto out;
7196                 }
7197
7198                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7199
7200                 btrfs_add_free_space(cache, buf->start, buf->len);
7201                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7202                 btrfs_put_block_group(cache);
7203                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
7204                 pin = 0;
7205         }
7206 out:
7207         if (pin)
7208                 add_pinned_bytes(root->fs_info, buf->len,
7209                                  btrfs_header_level(buf),
7210                                  root->root_key.objectid);
7211
7212         /*
7213          * Deleting the buffer, clear the corrupt flag since it doesn't matter
7214          * anymore.
7215          */
7216         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7217 }
7218
7219 /* Can return -ENOMEM */
7220 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7221                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7222                       u64 owner, u64 offset)
7223 {
7224         int ret;
7225         struct btrfs_fs_info *fs_info = root->fs_info;
7226
7227         if (btrfs_is_testing(fs_info))
7228                 return 0;
7229
7230         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
7231
7232         /*
7233          * tree log blocks never actually go into the extent allocation
7234          * tree, just update pinning info and exit early.
7235          */
7236         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7237                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7238                 /* unlocks the pinned mutex */
7239                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
7240                 ret = 0;
7241         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7242                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7243                                         num_bytes,
7244                                         parent, root_objectid, (int)owner,
7245                                         BTRFS_DROP_DELAYED_REF, NULL);
7246         } else {
7247                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7248                                                 num_bytes,
7249                                                 parent, root_objectid, owner,
7250                                                 offset, 0,
7251                                                 BTRFS_DROP_DELAYED_REF, NULL);
7252         }
7253         return ret;
7254 }
7255
7256 /*
7257  * when we wait for progress in the block group caching, its because
7258  * our allocation attempt failed at least once.  So, we must sleep
7259  * and let some progress happen before we try again.
7260  *
7261  * This function will sleep at least once waiting for new free space to
7262  * show up, and then it will check the block group free space numbers
7263  * for our min num_bytes.  Another option is to have it go ahead
7264  * and look in the rbtree for a free extent of a given size, but this
7265  * is a good start.
7266  *
7267  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7268  * any of the information in this block group.
7269  */
7270 static noinline void
7271 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7272                                 u64 num_bytes)
7273 {
7274         struct btrfs_caching_control *caching_ctl;
7275
7276         caching_ctl = get_caching_control(cache);
7277         if (!caching_ctl)
7278                 return;
7279
7280         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7281                    (cache->free_space_ctl->free_space >= num_bytes));
7282
7283         put_caching_control(caching_ctl);
7284 }
7285
7286 static noinline int
7287 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7288 {
7289         struct btrfs_caching_control *caching_ctl;
7290         int ret = 0;
7291
7292         caching_ctl = get_caching_control(cache);
7293         if (!caching_ctl)
7294                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7295
7296         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7297         if (cache->cached == BTRFS_CACHE_ERROR)
7298                 ret = -EIO;
7299         put_caching_control(caching_ctl);
7300         return ret;
7301 }
7302
7303 int __get_raid_index(u64 flags)
7304 {
7305         if (flags & BTRFS_BLOCK_GROUP_RAID10)
7306                 return BTRFS_RAID_RAID10;
7307         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
7308                 return BTRFS_RAID_RAID1;
7309         else if (flags & BTRFS_BLOCK_GROUP_DUP)
7310                 return BTRFS_RAID_DUP;
7311         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
7312                 return BTRFS_RAID_RAID0;
7313         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
7314                 return BTRFS_RAID_RAID5;
7315         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
7316                 return BTRFS_RAID_RAID6;
7317
7318         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
7319 }
7320
7321 int get_block_group_index(struct btrfs_block_group_cache *cache)
7322 {
7323         return __get_raid_index(cache->flags);
7324 }
7325
7326 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7327         [BTRFS_RAID_RAID10]     = "raid10",
7328         [BTRFS_RAID_RAID1]      = "raid1",
7329         [BTRFS_RAID_DUP]        = "dup",
7330         [BTRFS_RAID_RAID0]      = "raid0",
7331         [BTRFS_RAID_SINGLE]     = "single",
7332         [BTRFS_RAID_RAID5]      = "raid5",
7333         [BTRFS_RAID_RAID6]      = "raid6",
7334 };
7335
7336 static const char *get_raid_name(enum btrfs_raid_types type)
7337 {
7338         if (type >= BTRFS_NR_RAID_TYPES)
7339                 return NULL;
7340
7341         return btrfs_raid_type_names[type];
7342 }
7343
7344 enum btrfs_loop_type {
7345         LOOP_CACHING_NOWAIT = 0,
7346         LOOP_CACHING_WAIT = 1,
7347         LOOP_ALLOC_CHUNK = 2,
7348         LOOP_NO_EMPTY_SIZE = 3,
7349 };
7350
7351 static inline void
7352 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7353                        int delalloc)
7354 {
7355         if (delalloc)
7356                 down_read(&cache->data_rwsem);
7357 }
7358
7359 static inline void
7360 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7361                        int delalloc)
7362 {
7363         btrfs_get_block_group(cache);
7364         if (delalloc)
7365                 down_read(&cache->data_rwsem);
7366 }
7367
7368 static struct btrfs_block_group_cache *
7369 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7370                    struct btrfs_free_cluster *cluster,
7371                    int delalloc)
7372 {
7373         struct btrfs_block_group_cache *used_bg = NULL;
7374
7375         spin_lock(&cluster->refill_lock);
7376         while (1) {
7377                 used_bg = cluster->block_group;
7378                 if (!used_bg)
7379                         return NULL;
7380
7381                 if (used_bg == block_group)
7382                         return used_bg;
7383
7384                 btrfs_get_block_group(used_bg);
7385
7386                 if (!delalloc)
7387                         return used_bg;
7388
7389                 if (down_read_trylock(&used_bg->data_rwsem))
7390                         return used_bg;
7391
7392                 spin_unlock(&cluster->refill_lock);
7393
7394                 down_read(&used_bg->data_rwsem);
7395
7396                 spin_lock(&cluster->refill_lock);
7397                 if (used_bg == cluster->block_group)
7398                         return used_bg;
7399
7400                 up_read(&used_bg->data_rwsem);
7401                 btrfs_put_block_group(used_bg);
7402         }
7403 }
7404
7405 static inline void
7406 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7407                          int delalloc)
7408 {
7409         if (delalloc)
7410                 up_read(&cache->data_rwsem);
7411         btrfs_put_block_group(cache);
7412 }
7413
7414 /*
7415  * walks the btree of allocated extents and find a hole of a given size.
7416  * The key ins is changed to record the hole:
7417  * ins->objectid == start position
7418  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7419  * ins->offset == the size of the hole.
7420  * Any available blocks before search_start are skipped.
7421  *
7422  * If there is no suitable free space, we will record the max size of
7423  * the free space extent currently.
7424  */
7425 static noinline int find_free_extent(struct btrfs_root *orig_root,
7426                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7427                                 u64 hint_byte, struct btrfs_key *ins,
7428                                 u64 flags, int delalloc)
7429 {
7430         int ret = 0;
7431         struct btrfs_root *root = orig_root->fs_info->extent_root;
7432         struct btrfs_free_cluster *last_ptr = NULL;
7433         struct btrfs_block_group_cache *block_group = NULL;
7434         u64 search_start = 0;
7435         u64 max_extent_size = 0;
7436         u64 empty_cluster = 0;
7437         struct btrfs_space_info *space_info;
7438         int loop = 0;
7439         int index = __get_raid_index(flags);
7440         bool failed_cluster_refill = false;
7441         bool failed_alloc = false;
7442         bool use_cluster = true;
7443         bool have_caching_bg = false;
7444         bool orig_have_caching_bg = false;
7445         bool full_search = false;
7446
7447         WARN_ON(num_bytes < root->sectorsize);
7448         ins->type = BTRFS_EXTENT_ITEM_KEY;
7449         ins->objectid = 0;
7450         ins->offset = 0;
7451
7452         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
7453
7454         space_info = __find_space_info(root->fs_info, flags);
7455         if (!space_info) {
7456                 btrfs_err(root->fs_info, "No space info for %llu", flags);
7457                 return -ENOSPC;
7458         }
7459
7460         /*
7461          * If our free space is heavily fragmented we may not be able to make
7462          * big contiguous allocations, so instead of doing the expensive search
7463          * for free space, simply return ENOSPC with our max_extent_size so we
7464          * can go ahead and search for a more manageable chunk.
7465          *
7466          * If our max_extent_size is large enough for our allocation simply
7467          * disable clustering since we will likely not be able to find enough
7468          * space to create a cluster and induce latency trying.
7469          */
7470         if (unlikely(space_info->max_extent_size)) {
7471                 spin_lock(&space_info->lock);
7472                 if (space_info->max_extent_size &&
7473                     num_bytes > space_info->max_extent_size) {
7474                         ins->offset = space_info->max_extent_size;
7475                         spin_unlock(&space_info->lock);
7476                         return -ENOSPC;
7477                 } else if (space_info->max_extent_size) {
7478                         use_cluster = false;
7479                 }
7480                 spin_unlock(&space_info->lock);
7481         }
7482
7483         last_ptr = fetch_cluster_info(orig_root, space_info, &empty_cluster);
7484         if (last_ptr) {
7485                 spin_lock(&last_ptr->lock);
7486                 if (last_ptr->block_group)
7487                         hint_byte = last_ptr->window_start;
7488                 if (last_ptr->fragmented) {
7489                         /*
7490                          * We still set window_start so we can keep track of the
7491                          * last place we found an allocation to try and save
7492                          * some time.
7493                          */
7494                         hint_byte = last_ptr->window_start;
7495                         use_cluster = false;
7496                 }
7497                 spin_unlock(&last_ptr->lock);
7498         }
7499
7500         search_start = max(search_start, first_logical_byte(root, 0));
7501         search_start = max(search_start, hint_byte);
7502         if (search_start == hint_byte) {
7503                 block_group = btrfs_lookup_block_group(root->fs_info,
7504                                                        search_start);
7505                 /*
7506                  * we don't want to use the block group if it doesn't match our
7507                  * allocation bits, or if its not cached.
7508                  *
7509                  * However if we are re-searching with an ideal block group
7510                  * picked out then we don't care that the block group is cached.
7511                  */
7512                 if (block_group && block_group_bits(block_group, flags) &&
7513                     block_group->cached != BTRFS_CACHE_NO) {
7514                         down_read(&space_info->groups_sem);
7515                         if (list_empty(&block_group->list) ||
7516                             block_group->ro) {
7517                                 /*
7518                                  * someone is removing this block group,
7519                                  * we can't jump into the have_block_group
7520                                  * target because our list pointers are not
7521                                  * valid
7522                                  */
7523                                 btrfs_put_block_group(block_group);
7524                                 up_read(&space_info->groups_sem);
7525                         } else {
7526                                 index = get_block_group_index(block_group);
7527                                 btrfs_lock_block_group(block_group, delalloc);
7528                                 goto have_block_group;
7529                         }
7530                 } else if (block_group) {
7531                         btrfs_put_block_group(block_group);
7532                 }
7533         }
7534 search:
7535         have_caching_bg = false;
7536         if (index == 0 || index == __get_raid_index(flags))
7537                 full_search = true;
7538         down_read(&space_info->groups_sem);
7539         list_for_each_entry(block_group, &space_info->block_groups[index],
7540                             list) {
7541                 u64 offset;
7542                 int cached;
7543
7544                 btrfs_grab_block_group(block_group, delalloc);
7545                 search_start = block_group->key.objectid;
7546
7547                 /*
7548                  * this can happen if we end up cycling through all the
7549                  * raid types, but we want to make sure we only allocate
7550                  * for the proper type.
7551                  */
7552                 if (!block_group_bits(block_group, flags)) {
7553                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7554                                 BTRFS_BLOCK_GROUP_RAID1 |
7555                                 BTRFS_BLOCK_GROUP_RAID5 |
7556                                 BTRFS_BLOCK_GROUP_RAID6 |
7557                                 BTRFS_BLOCK_GROUP_RAID10;
7558
7559                         /*
7560                          * if they asked for extra copies and this block group
7561                          * doesn't provide them, bail.  This does allow us to
7562                          * fill raid0 from raid1.
7563                          */
7564                         if ((flags & extra) && !(block_group->flags & extra))
7565                                 goto loop;
7566                 }
7567
7568 have_block_group:
7569                 cached = block_group_cache_done(block_group);
7570                 if (unlikely(!cached)) {
7571                         have_caching_bg = true;
7572                         ret = cache_block_group(block_group, 0);
7573                         BUG_ON(ret < 0);
7574                         ret = 0;
7575                 }
7576
7577                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7578                         goto loop;
7579                 if (unlikely(block_group->ro))
7580                         goto loop;
7581
7582                 /*
7583                  * Ok we want to try and use the cluster allocator, so
7584                  * lets look there
7585                  */
7586                 if (last_ptr && use_cluster) {
7587                         struct btrfs_block_group_cache *used_block_group;
7588                         unsigned long aligned_cluster;
7589                         /*
7590                          * the refill lock keeps out other
7591                          * people trying to start a new cluster
7592                          */
7593                         used_block_group = btrfs_lock_cluster(block_group,
7594                                                               last_ptr,
7595                                                               delalloc);
7596                         if (!used_block_group)
7597                                 goto refill_cluster;
7598
7599                         if (used_block_group != block_group &&
7600                             (used_block_group->ro ||
7601                              !block_group_bits(used_block_group, flags)))
7602                                 goto release_cluster;
7603
7604                         offset = btrfs_alloc_from_cluster(used_block_group,
7605                                                 last_ptr,
7606                                                 num_bytes,
7607                                                 used_block_group->key.objectid,
7608                                                 &max_extent_size);
7609                         if (offset) {
7610                                 /* we have a block, we're done */
7611                                 spin_unlock(&last_ptr->refill_lock);
7612                                 trace_btrfs_reserve_extent_cluster(root,
7613                                                 used_block_group,
7614                                                 search_start, num_bytes);
7615                                 if (used_block_group != block_group) {
7616                                         btrfs_release_block_group(block_group,
7617                                                                   delalloc);
7618                                         block_group = used_block_group;
7619                                 }
7620                                 goto checks;
7621                         }
7622
7623                         WARN_ON(last_ptr->block_group != used_block_group);
7624 release_cluster:
7625                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7626                          * set up a new clusters, so lets just skip it
7627                          * and let the allocator find whatever block
7628                          * it can find.  If we reach this point, we
7629                          * will have tried the cluster allocator
7630                          * plenty of times and not have found
7631                          * anything, so we are likely way too
7632                          * fragmented for the clustering stuff to find
7633                          * anything.
7634                          *
7635                          * However, if the cluster is taken from the
7636                          * current block group, release the cluster
7637                          * first, so that we stand a better chance of
7638                          * succeeding in the unclustered
7639                          * allocation.  */
7640                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7641                             used_block_group != block_group) {
7642                                 spin_unlock(&last_ptr->refill_lock);
7643                                 btrfs_release_block_group(used_block_group,
7644                                                           delalloc);
7645                                 goto unclustered_alloc;
7646                         }
7647
7648                         /*
7649                          * this cluster didn't work out, free it and
7650                          * start over
7651                          */
7652                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7653
7654                         if (used_block_group != block_group)
7655                                 btrfs_release_block_group(used_block_group,
7656                                                           delalloc);
7657 refill_cluster:
7658                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7659                                 spin_unlock(&last_ptr->refill_lock);
7660                                 goto unclustered_alloc;
7661                         }
7662
7663                         aligned_cluster = max_t(unsigned long,
7664                                                 empty_cluster + empty_size,
7665                                               block_group->full_stripe_len);
7666
7667                         /* allocate a cluster in this block group */
7668                         ret = btrfs_find_space_cluster(root, block_group,
7669                                                        last_ptr, search_start,
7670                                                        num_bytes,
7671                                                        aligned_cluster);
7672                         if (ret == 0) {
7673                                 /*
7674                                  * now pull our allocation out of this
7675                                  * cluster
7676                                  */
7677                                 offset = btrfs_alloc_from_cluster(block_group,
7678                                                         last_ptr,
7679                                                         num_bytes,
7680                                                         search_start,
7681                                                         &max_extent_size);
7682                                 if (offset) {
7683                                         /* we found one, proceed */
7684                                         spin_unlock(&last_ptr->refill_lock);
7685                                         trace_btrfs_reserve_extent_cluster(root,
7686                                                 block_group, search_start,
7687                                                 num_bytes);
7688                                         goto checks;
7689                                 }
7690                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7691                                    && !failed_cluster_refill) {
7692                                 spin_unlock(&last_ptr->refill_lock);
7693
7694                                 failed_cluster_refill = true;
7695                                 wait_block_group_cache_progress(block_group,
7696                                        num_bytes + empty_cluster + empty_size);
7697                                 goto have_block_group;
7698                         }
7699
7700                         /*
7701                          * at this point we either didn't find a cluster
7702                          * or we weren't able to allocate a block from our
7703                          * cluster.  Free the cluster we've been trying
7704                          * to use, and go to the next block group
7705                          */
7706                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7707                         spin_unlock(&last_ptr->refill_lock);
7708                         goto loop;
7709                 }
7710
7711 unclustered_alloc:
7712                 /*
7713                  * We are doing an unclustered alloc, set the fragmented flag so
7714                  * we don't bother trying to setup a cluster again until we get
7715                  * more space.
7716                  */
7717                 if (unlikely(last_ptr)) {
7718                         spin_lock(&last_ptr->lock);
7719                         last_ptr->fragmented = 1;
7720                         spin_unlock(&last_ptr->lock);
7721                 }
7722                 spin_lock(&block_group->free_space_ctl->tree_lock);
7723                 if (cached &&
7724                     block_group->free_space_ctl->free_space <
7725                     num_bytes + empty_cluster + empty_size) {
7726                         if (block_group->free_space_ctl->free_space >
7727                             max_extent_size)
7728                                 max_extent_size =
7729                                         block_group->free_space_ctl->free_space;
7730                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7731                         goto loop;
7732                 }
7733                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7734
7735                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7736                                                     num_bytes, empty_size,
7737                                                     &max_extent_size);
7738                 /*
7739                  * If we didn't find a chunk, and we haven't failed on this
7740                  * block group before, and this block group is in the middle of
7741                  * caching and we are ok with waiting, then go ahead and wait
7742                  * for progress to be made, and set failed_alloc to true.
7743                  *
7744                  * If failed_alloc is true then we've already waited on this
7745                  * block group once and should move on to the next block group.
7746                  */
7747                 if (!offset && !failed_alloc && !cached &&
7748                     loop > LOOP_CACHING_NOWAIT) {
7749                         wait_block_group_cache_progress(block_group,
7750                                                 num_bytes + empty_size);
7751                         failed_alloc = true;
7752                         goto have_block_group;
7753                 } else if (!offset) {
7754                         goto loop;
7755                 }
7756 checks:
7757                 search_start = ALIGN(offset, root->stripesize);
7758
7759                 /* move on to the next group */
7760                 if (search_start + num_bytes >
7761                     block_group->key.objectid + block_group->key.offset) {
7762                         btrfs_add_free_space(block_group, offset, num_bytes);
7763                         goto loop;
7764                 }
7765
7766                 if (offset < search_start)
7767                         btrfs_add_free_space(block_group, offset,
7768                                              search_start - offset);
7769                 BUG_ON(offset > search_start);
7770
7771                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7772                                 num_bytes, delalloc);
7773                 if (ret == -EAGAIN) {
7774                         btrfs_add_free_space(block_group, offset, num_bytes);
7775                         goto loop;
7776                 }
7777                 btrfs_inc_block_group_reservations(block_group);
7778
7779                 /* we are all good, lets return */
7780                 ins->objectid = search_start;
7781                 ins->offset = num_bytes;
7782
7783                 trace_btrfs_reserve_extent(orig_root, block_group,
7784                                            search_start, num_bytes);
7785                 btrfs_release_block_group(block_group, delalloc);
7786                 break;
7787 loop:
7788                 failed_cluster_refill = false;
7789                 failed_alloc = false;
7790                 BUG_ON(index != get_block_group_index(block_group));
7791                 btrfs_release_block_group(block_group, delalloc);
7792         }
7793         up_read(&space_info->groups_sem);
7794
7795         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7796                 && !orig_have_caching_bg)
7797                 orig_have_caching_bg = true;
7798
7799         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7800                 goto search;
7801
7802         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7803                 goto search;
7804
7805         /*
7806          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7807          *                      caching kthreads as we move along
7808          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7809          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7810          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7811          *                      again
7812          */
7813         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7814                 index = 0;
7815                 if (loop == LOOP_CACHING_NOWAIT) {
7816                         /*
7817                          * We want to skip the LOOP_CACHING_WAIT step if we
7818                          * don't have any uncached bgs and we've already done a
7819                          * full search through.
7820                          */
7821                         if (orig_have_caching_bg || !full_search)
7822                                 loop = LOOP_CACHING_WAIT;
7823                         else
7824                                 loop = LOOP_ALLOC_CHUNK;
7825                 } else {
7826                         loop++;
7827                 }
7828
7829                 if (loop == LOOP_ALLOC_CHUNK) {
7830                         struct btrfs_trans_handle *trans;
7831                         int exist = 0;
7832
7833                         trans = current->journal_info;
7834                         if (trans)
7835                                 exist = 1;
7836                         else
7837                                 trans = btrfs_join_transaction(root);
7838
7839                         if (IS_ERR(trans)) {
7840                                 ret = PTR_ERR(trans);
7841                                 goto out;
7842                         }
7843
7844                         ret = do_chunk_alloc(trans, root, flags,
7845                                              CHUNK_ALLOC_FORCE);
7846
7847                         /*
7848                          * If we can't allocate a new chunk we've already looped
7849                          * through at least once, move on to the NO_EMPTY_SIZE
7850                          * case.
7851                          */
7852                         if (ret == -ENOSPC)
7853                                 loop = LOOP_NO_EMPTY_SIZE;
7854
7855                         /*
7856                          * Do not bail out on ENOSPC since we
7857                          * can do more things.
7858                          */
7859                         if (ret < 0 && ret != -ENOSPC)
7860                                 btrfs_abort_transaction(trans, ret);
7861                         else
7862                                 ret = 0;
7863                         if (!exist)
7864                                 btrfs_end_transaction(trans, root);
7865                         if (ret)
7866                                 goto out;
7867                 }
7868
7869                 if (loop == LOOP_NO_EMPTY_SIZE) {
7870                         /*
7871                          * Don't loop again if we already have no empty_size and
7872                          * no empty_cluster.
7873                          */
7874                         if (empty_size == 0 &&
7875                             empty_cluster == 0) {
7876                                 ret = -ENOSPC;
7877                                 goto out;
7878                         }
7879                         empty_size = 0;
7880                         empty_cluster = 0;
7881                 }
7882
7883                 goto search;
7884         } else if (!ins->objectid) {
7885                 ret = -ENOSPC;
7886         } else if (ins->objectid) {
7887                 if (!use_cluster && last_ptr) {
7888                         spin_lock(&last_ptr->lock);
7889                         last_ptr->window_start = ins->objectid;
7890                         spin_unlock(&last_ptr->lock);
7891                 }
7892                 ret = 0;
7893         }
7894 out:
7895         if (ret == -ENOSPC) {
7896                 spin_lock(&space_info->lock);
7897                 space_info->max_extent_size = max_extent_size;
7898                 spin_unlock(&space_info->lock);
7899                 ins->offset = max_extent_size;
7900         }
7901         return ret;
7902 }
7903
7904 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7905                             int dump_block_groups)
7906 {
7907         struct btrfs_block_group_cache *cache;
7908         int index = 0;
7909
7910         spin_lock(&info->lock);
7911         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7912                info->flags,
7913                info->total_bytes - info->bytes_used - info->bytes_pinned -
7914                info->bytes_reserved - info->bytes_readonly -
7915                info->bytes_may_use, (info->full) ? "" : "not ");
7916         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7917                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7918                info->total_bytes, info->bytes_used, info->bytes_pinned,
7919                info->bytes_reserved, info->bytes_may_use,
7920                info->bytes_readonly);
7921         spin_unlock(&info->lock);
7922
7923         if (!dump_block_groups)
7924                 return;
7925
7926         down_read(&info->groups_sem);
7927 again:
7928         list_for_each_entry(cache, &info->block_groups[index], list) {
7929                 spin_lock(&cache->lock);
7930                 printk(KERN_INFO "BTRFS: "
7931                            "block group %llu has %llu bytes, "
7932                            "%llu used %llu pinned %llu reserved %s\n",
7933                        cache->key.objectid, cache->key.offset,
7934                        btrfs_block_group_used(&cache->item), cache->pinned,
7935                        cache->reserved, cache->ro ? "[readonly]" : "");
7936                 btrfs_dump_free_space(cache, bytes);
7937                 spin_unlock(&cache->lock);
7938         }
7939         if (++index < BTRFS_NR_RAID_TYPES)
7940                 goto again;
7941         up_read(&info->groups_sem);
7942 }
7943
7944 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7945                          u64 num_bytes, u64 min_alloc_size,
7946                          u64 empty_size, u64 hint_byte,
7947                          struct btrfs_key *ins, int is_data, int delalloc)
7948 {
7949         bool final_tried = num_bytes == min_alloc_size;
7950         u64 flags;
7951         int ret;
7952
7953         flags = btrfs_get_alloc_profile(root, is_data);
7954 again:
7955         WARN_ON(num_bytes < root->sectorsize);
7956         ret = find_free_extent(root, ram_bytes, num_bytes, empty_size,
7957                                hint_byte, ins, flags, delalloc);
7958         if (!ret && !is_data) {
7959                 btrfs_dec_block_group_reservations(root->fs_info,
7960                                                    ins->objectid);
7961         } else if (ret == -ENOSPC) {
7962                 if (!final_tried && ins->offset) {
7963                         num_bytes = min(num_bytes >> 1, ins->offset);
7964                         num_bytes = round_down(num_bytes, root->sectorsize);
7965                         num_bytes = max(num_bytes, min_alloc_size);
7966                         ram_bytes = num_bytes;
7967                         if (num_bytes == min_alloc_size)
7968                                 final_tried = true;
7969                         goto again;
7970                 } else if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
7971                         struct btrfs_space_info *sinfo;
7972
7973                         sinfo = __find_space_info(root->fs_info, flags);
7974                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7975                                 flags, num_bytes);
7976                         if (sinfo)
7977                                 dump_space_info(sinfo, num_bytes, 1);
7978                 }
7979         }
7980
7981         return ret;
7982 }
7983
7984 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7985                                         u64 start, u64 len,
7986                                         int pin, int delalloc)
7987 {
7988         struct btrfs_block_group_cache *cache;
7989         int ret = 0;
7990
7991         cache = btrfs_lookup_block_group(root->fs_info, start);
7992         if (!cache) {
7993                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7994                         start);
7995                 return -ENOSPC;
7996         }
7997
7998         if (pin)
7999                 pin_down_extent(root, cache, start, len, 1);
8000         else {
8001                 if (btrfs_test_opt(root->fs_info, DISCARD))
8002                         ret = btrfs_discard_extent(root, start, len, NULL);
8003                 btrfs_add_free_space(cache, start, len);
8004                 btrfs_free_reserved_bytes(cache, len, delalloc);
8005                 trace_btrfs_reserved_extent_free(root, start, len);
8006         }
8007
8008         btrfs_put_block_group(cache);
8009         return ret;
8010 }
8011
8012 int btrfs_free_reserved_extent(struct btrfs_root *root,
8013                                u64 start, u64 len, int delalloc)
8014 {
8015         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
8016 }
8017
8018 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
8019                                        u64 start, u64 len)
8020 {
8021         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
8022 }
8023
8024 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8025                                       struct btrfs_root *root,
8026                                       u64 parent, u64 root_objectid,
8027                                       u64 flags, u64 owner, u64 offset,
8028                                       struct btrfs_key *ins, int ref_mod)
8029 {
8030         int ret;
8031         struct btrfs_fs_info *fs_info = root->fs_info;
8032         struct btrfs_extent_item *extent_item;
8033         struct btrfs_extent_inline_ref *iref;
8034         struct btrfs_path *path;
8035         struct extent_buffer *leaf;
8036         int type;
8037         u32 size;
8038
8039         if (parent > 0)
8040                 type = BTRFS_SHARED_DATA_REF_KEY;
8041         else
8042                 type = BTRFS_EXTENT_DATA_REF_KEY;
8043
8044         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8045
8046         path = btrfs_alloc_path();
8047         if (!path)
8048                 return -ENOMEM;
8049
8050         path->leave_spinning = 1;
8051         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8052                                       ins, size);
8053         if (ret) {
8054                 btrfs_free_path(path);
8055                 return ret;
8056         }
8057
8058         leaf = path->nodes[0];
8059         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8060                                      struct btrfs_extent_item);
8061         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8062         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8063         btrfs_set_extent_flags(leaf, extent_item,
8064                                flags | BTRFS_EXTENT_FLAG_DATA);
8065
8066         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8067         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8068         if (parent > 0) {
8069                 struct btrfs_shared_data_ref *ref;
8070                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8071                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8072                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8073         } else {
8074                 struct btrfs_extent_data_ref *ref;
8075                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8076                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8077                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8078                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8079                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8080         }
8081
8082         btrfs_mark_buffer_dirty(path->nodes[0]);
8083         btrfs_free_path(path);
8084
8085         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8086                                           ins->offset);
8087         if (ret)
8088                 return ret;
8089
8090         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
8091         if (ret) { /* -ENOENT, logic error */
8092                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8093                         ins->objectid, ins->offset);
8094                 BUG();
8095         }
8096         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
8097         return ret;
8098 }
8099
8100 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8101                                      struct btrfs_root *root,
8102                                      u64 parent, u64 root_objectid,
8103                                      u64 flags, struct btrfs_disk_key *key,
8104                                      int level, struct btrfs_key *ins)
8105 {
8106         int ret;
8107         struct btrfs_fs_info *fs_info = root->fs_info;
8108         struct btrfs_extent_item *extent_item;
8109         struct btrfs_tree_block_info *block_info;
8110         struct btrfs_extent_inline_ref *iref;
8111         struct btrfs_path *path;
8112         struct extent_buffer *leaf;
8113         u32 size = sizeof(*extent_item) + sizeof(*iref);
8114         u64 num_bytes = ins->offset;
8115         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8116                                                  SKINNY_METADATA);
8117
8118         if (!skinny_metadata)
8119                 size += sizeof(*block_info);
8120
8121         path = btrfs_alloc_path();
8122         if (!path) {
8123                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
8124                                                    root->nodesize);
8125                 return -ENOMEM;
8126         }
8127
8128         path->leave_spinning = 1;
8129         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8130                                       ins, size);
8131         if (ret) {
8132                 btrfs_free_path(path);
8133                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
8134                                                    root->nodesize);
8135                 return ret;
8136         }
8137
8138         leaf = path->nodes[0];
8139         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8140                                      struct btrfs_extent_item);
8141         btrfs_set_extent_refs(leaf, extent_item, 1);
8142         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8143         btrfs_set_extent_flags(leaf, extent_item,
8144                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8145
8146         if (skinny_metadata) {
8147                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8148                 num_bytes = root->nodesize;
8149         } else {
8150                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8151                 btrfs_set_tree_block_key(leaf, block_info, key);
8152                 btrfs_set_tree_block_level(leaf, block_info, level);
8153                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8154         }
8155
8156         if (parent > 0) {
8157                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8158                 btrfs_set_extent_inline_ref_type(leaf, iref,
8159                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8160                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8161         } else {
8162                 btrfs_set_extent_inline_ref_type(leaf, iref,
8163                                                  BTRFS_TREE_BLOCK_REF_KEY);
8164                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8165         }
8166
8167         btrfs_mark_buffer_dirty(leaf);
8168         btrfs_free_path(path);
8169
8170         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8171                                           num_bytes);
8172         if (ret)
8173                 return ret;
8174
8175         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
8176                                  1);
8177         if (ret) { /* -ENOENT, logic error */
8178                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8179                         ins->objectid, ins->offset);
8180                 BUG();
8181         }
8182
8183         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
8184         return ret;
8185 }
8186
8187 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8188                                      struct btrfs_root *root,
8189                                      u64 root_objectid, u64 owner,
8190                                      u64 offset, u64 ram_bytes,
8191                                      struct btrfs_key *ins)
8192 {
8193         int ret;
8194
8195         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
8196
8197         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
8198                                          ins->offset, 0,
8199                                          root_objectid, owner, offset,
8200                                          ram_bytes, BTRFS_ADD_DELAYED_EXTENT,
8201                                          NULL);
8202         return ret;
8203 }
8204
8205 /*
8206  * this is used by the tree logging recovery code.  It records that
8207  * an extent has been allocated and makes sure to clear the free
8208  * space cache bits as well
8209  */
8210 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8211                                    struct btrfs_root *root,
8212                                    u64 root_objectid, u64 owner, u64 offset,
8213                                    struct btrfs_key *ins)
8214 {
8215         int ret;
8216         struct btrfs_block_group_cache *block_group;
8217         struct btrfs_space_info *space_info;
8218
8219         /*
8220          * Mixed block groups will exclude before processing the log so we only
8221          * need to do the exclude dance if this fs isn't mixed.
8222          */
8223         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
8224                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
8225                 if (ret)
8226                         return ret;
8227         }
8228
8229         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
8230         if (!block_group)
8231                 return -EINVAL;
8232
8233         space_info = block_group->space_info;
8234         spin_lock(&space_info->lock);
8235         spin_lock(&block_group->lock);
8236         space_info->bytes_reserved += ins->offset;
8237         block_group->reserved += ins->offset;
8238         spin_unlock(&block_group->lock);
8239         spin_unlock(&space_info->lock);
8240
8241         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
8242                                          0, owner, offset, ins, 1);
8243         btrfs_put_block_group(block_group);
8244         return ret;
8245 }
8246
8247 static struct extent_buffer *
8248 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8249                       u64 bytenr, int level)
8250 {
8251         struct extent_buffer *buf;
8252
8253         buf = btrfs_find_create_tree_block(root, bytenr);
8254         if (IS_ERR(buf))
8255                 return buf;
8256
8257         btrfs_set_header_generation(buf, trans->transid);
8258         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8259         btrfs_tree_lock(buf);
8260         clean_tree_block(trans, root->fs_info, buf);
8261         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8262
8263         btrfs_set_lock_blocking(buf);
8264         set_extent_buffer_uptodate(buf);
8265
8266         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8267                 buf->log_index = root->log_transid % 2;
8268                 /*
8269                  * we allow two log transactions at a time, use different
8270                  * EXENT bit to differentiate dirty pages.
8271                  */
8272                 if (buf->log_index == 0)
8273                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8274                                         buf->start + buf->len - 1, GFP_NOFS);
8275                 else
8276                         set_extent_new(&root->dirty_log_pages, buf->start,
8277                                         buf->start + buf->len - 1);
8278         } else {
8279                 buf->log_index = -1;
8280                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8281                          buf->start + buf->len - 1, GFP_NOFS);
8282         }
8283         trans->dirty = true;
8284         /* this returns a buffer locked for blocking */
8285         return buf;
8286 }
8287
8288 static struct btrfs_block_rsv *
8289 use_block_rsv(struct btrfs_trans_handle *trans,
8290               struct btrfs_root *root, u32 blocksize)
8291 {
8292         struct btrfs_block_rsv *block_rsv;
8293         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
8294         int ret;
8295         bool global_updated = false;
8296
8297         block_rsv = get_block_rsv(trans, root);
8298
8299         if (unlikely(block_rsv->size == 0))
8300                 goto try_reserve;
8301 again:
8302         ret = block_rsv_use_bytes(block_rsv, blocksize);
8303         if (!ret)
8304                 return block_rsv;
8305
8306         if (block_rsv->failfast)
8307                 return ERR_PTR(ret);
8308
8309         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8310                 global_updated = true;
8311                 update_global_block_rsv(root->fs_info);
8312                 goto again;
8313         }
8314
8315         if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
8316                 static DEFINE_RATELIMIT_STATE(_rs,
8317                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8318                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8319                 if (__ratelimit(&_rs))
8320                         WARN(1, KERN_DEBUG
8321                                 "BTRFS: block rsv returned %d\n", ret);
8322         }
8323 try_reserve:
8324         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8325                                      BTRFS_RESERVE_NO_FLUSH);
8326         if (!ret)
8327                 return block_rsv;
8328         /*
8329          * If we couldn't reserve metadata bytes try and use some from
8330          * the global reserve if its space type is the same as the global
8331          * reservation.
8332          */
8333         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8334             block_rsv->space_info == global_rsv->space_info) {
8335                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8336                 if (!ret)
8337                         return global_rsv;
8338         }
8339         return ERR_PTR(ret);
8340 }
8341
8342 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8343                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8344 {
8345         block_rsv_add_bytes(block_rsv, blocksize, 0);
8346         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8347 }
8348
8349 /*
8350  * finds a free extent and does all the dirty work required for allocation
8351  * returns the tree buffer or an ERR_PTR on error.
8352  */
8353 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8354                                         struct btrfs_root *root,
8355                                         u64 parent, u64 root_objectid,
8356                                         struct btrfs_disk_key *key, int level,
8357                                         u64 hint, u64 empty_size)
8358 {
8359         struct btrfs_key ins;
8360         struct btrfs_block_rsv *block_rsv;
8361         struct extent_buffer *buf;
8362         struct btrfs_delayed_extent_op *extent_op;
8363         u64 flags = 0;
8364         int ret;
8365         u32 blocksize = root->nodesize;
8366         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8367                                                  SKINNY_METADATA);
8368
8369 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8370         if (btrfs_is_testing(root->fs_info)) {
8371                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8372                                             level);
8373                 if (!IS_ERR(buf))
8374                         root->alloc_bytenr += blocksize;
8375                 return buf;
8376         }
8377 #endif
8378
8379         block_rsv = use_block_rsv(trans, root, blocksize);
8380         if (IS_ERR(block_rsv))
8381                 return ERR_CAST(block_rsv);
8382
8383         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8384                                    empty_size, hint, &ins, 0, 0);
8385         if (ret)
8386                 goto out_unuse;
8387
8388         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8389         if (IS_ERR(buf)) {
8390                 ret = PTR_ERR(buf);
8391                 goto out_free_reserved;
8392         }
8393
8394         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8395                 if (parent == 0)
8396                         parent = ins.objectid;
8397                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8398         } else
8399                 BUG_ON(parent > 0);
8400
8401         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8402                 extent_op = btrfs_alloc_delayed_extent_op();
8403                 if (!extent_op) {
8404                         ret = -ENOMEM;
8405                         goto out_free_buf;
8406                 }
8407                 if (key)
8408                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8409                 else
8410                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8411                 extent_op->flags_to_set = flags;
8412                 extent_op->update_key = skinny_metadata ? false : true;
8413                 extent_op->update_flags = true;
8414                 extent_op->is_data = false;
8415                 extent_op->level = level;
8416
8417                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
8418                                                  ins.objectid, ins.offset,
8419                                                  parent, root_objectid, level,
8420                                                  BTRFS_ADD_DELAYED_EXTENT,
8421                                                  extent_op);
8422                 if (ret)
8423                         goto out_free_delayed;
8424         }
8425         return buf;
8426
8427 out_free_delayed:
8428         btrfs_free_delayed_extent_op(extent_op);
8429 out_free_buf:
8430         free_extent_buffer(buf);
8431 out_free_reserved:
8432         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
8433 out_unuse:
8434         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
8435         return ERR_PTR(ret);
8436 }
8437
8438 struct walk_control {
8439         u64 refs[BTRFS_MAX_LEVEL];
8440         u64 flags[BTRFS_MAX_LEVEL];
8441         struct btrfs_key update_progress;
8442         int stage;
8443         int level;
8444         int shared_level;
8445         int update_ref;
8446         int keep_locks;
8447         int reada_slot;
8448         int reada_count;
8449         int for_reloc;
8450 };
8451
8452 #define DROP_REFERENCE  1
8453 #define UPDATE_BACKREF  2
8454
8455 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8456                                      struct btrfs_root *root,
8457                                      struct walk_control *wc,
8458                                      struct btrfs_path *path)
8459 {
8460         u64 bytenr;
8461         u64 generation;
8462         u64 refs;
8463         u64 flags;
8464         u32 nritems;
8465         u32 blocksize;
8466         struct btrfs_key key;
8467         struct extent_buffer *eb;
8468         int ret;
8469         int slot;
8470         int nread = 0;
8471
8472         if (path->slots[wc->level] < wc->reada_slot) {
8473                 wc->reada_count = wc->reada_count * 2 / 3;
8474                 wc->reada_count = max(wc->reada_count, 2);
8475         } else {
8476                 wc->reada_count = wc->reada_count * 3 / 2;
8477                 wc->reada_count = min_t(int, wc->reada_count,
8478                                         BTRFS_NODEPTRS_PER_BLOCK(root));
8479         }
8480
8481         eb = path->nodes[wc->level];
8482         nritems = btrfs_header_nritems(eb);
8483         blocksize = root->nodesize;
8484
8485         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8486                 if (nread >= wc->reada_count)
8487                         break;
8488
8489                 cond_resched();
8490                 bytenr = btrfs_node_blockptr(eb, slot);
8491                 generation = btrfs_node_ptr_generation(eb, slot);
8492
8493                 if (slot == path->slots[wc->level])
8494                         goto reada;
8495
8496                 if (wc->stage == UPDATE_BACKREF &&
8497                     generation <= root->root_key.offset)
8498                         continue;
8499
8500                 /* We don't lock the tree block, it's OK to be racy here */
8501                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
8502                                                wc->level - 1, 1, &refs,
8503                                                &flags);
8504                 /* We don't care about errors in readahead. */
8505                 if (ret < 0)
8506                         continue;
8507                 BUG_ON(refs == 0);
8508
8509                 if (wc->stage == DROP_REFERENCE) {
8510                         if (refs == 1)
8511                                 goto reada;
8512
8513                         if (wc->level == 1 &&
8514                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8515                                 continue;
8516                         if (!wc->update_ref ||
8517                             generation <= root->root_key.offset)
8518                                 continue;
8519                         btrfs_node_key_to_cpu(eb, &key, slot);
8520                         ret = btrfs_comp_cpu_keys(&key,
8521                                                   &wc->update_progress);
8522                         if (ret < 0)
8523                                 continue;
8524                 } else {
8525                         if (wc->level == 1 &&
8526                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8527                                 continue;
8528                 }
8529 reada:
8530                 readahead_tree_block(root, bytenr);
8531                 nread++;
8532         }
8533         wc->reada_slot = slot;
8534 }
8535
8536 static int account_leaf_items(struct btrfs_trans_handle *trans,
8537                               struct btrfs_root *root,
8538                               struct extent_buffer *eb)
8539 {
8540         int nr = btrfs_header_nritems(eb);
8541         int i, extent_type, ret;
8542         struct btrfs_key key;
8543         struct btrfs_file_extent_item *fi;
8544         u64 bytenr, num_bytes;
8545
8546         /* We can be called directly from walk_up_proc() */
8547         if (!root->fs_info->quota_enabled)
8548                 return 0;
8549
8550         for (i = 0; i < nr; i++) {
8551                 btrfs_item_key_to_cpu(eb, &key, i);
8552
8553                 if (key.type != BTRFS_EXTENT_DATA_KEY)
8554                         continue;
8555
8556                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
8557                 /* filter out non qgroup-accountable extents  */
8558                 extent_type = btrfs_file_extent_type(eb, fi);
8559
8560                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
8561                         continue;
8562
8563                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
8564                 if (!bytenr)
8565                         continue;
8566
8567                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
8568
8569                 ret = btrfs_qgroup_insert_dirty_extent(trans, root->fs_info,
8570                                 bytenr, num_bytes, GFP_NOFS);
8571                 if (ret)
8572                         return ret;
8573         }
8574         return 0;
8575 }
8576
8577 /*
8578  * Walk up the tree from the bottom, freeing leaves and any interior
8579  * nodes which have had all slots visited. If a node (leaf or
8580  * interior) is freed, the node above it will have it's slot
8581  * incremented. The root node will never be freed.
8582  *
8583  * At the end of this function, we should have a path which has all
8584  * slots incremented to the next position for a search. If we need to
8585  * read a new node it will be NULL and the node above it will have the
8586  * correct slot selected for a later read.
8587  *
8588  * If we increment the root nodes slot counter past the number of
8589  * elements, 1 is returned to signal completion of the search.
8590  */
8591 static int adjust_slots_upwards(struct btrfs_root *root,
8592                                 struct btrfs_path *path, int root_level)
8593 {
8594         int level = 0;
8595         int nr, slot;
8596         struct extent_buffer *eb;
8597
8598         if (root_level == 0)
8599                 return 1;
8600
8601         while (level <= root_level) {
8602                 eb = path->nodes[level];
8603                 nr = btrfs_header_nritems(eb);
8604                 path->slots[level]++;
8605                 slot = path->slots[level];
8606                 if (slot >= nr || level == 0) {
8607                         /*
8608                          * Don't free the root -  we will detect this
8609                          * condition after our loop and return a
8610                          * positive value for caller to stop walking the tree.
8611                          */
8612                         if (level != root_level) {
8613                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8614                                 path->locks[level] = 0;
8615
8616                                 free_extent_buffer(eb);
8617                                 path->nodes[level] = NULL;
8618                                 path->slots[level] = 0;
8619                         }
8620                 } else {
8621                         /*
8622                          * We have a valid slot to walk back down
8623                          * from. Stop here so caller can process these
8624                          * new nodes.
8625                          */
8626                         break;
8627                 }
8628
8629                 level++;
8630         }
8631
8632         eb = path->nodes[root_level];
8633         if (path->slots[root_level] >= btrfs_header_nritems(eb))
8634                 return 1;
8635
8636         return 0;
8637 }
8638
8639 /*
8640  * root_eb is the subtree root and is locked before this function is called.
8641  */
8642 static int account_shared_subtree(struct btrfs_trans_handle *trans,
8643                                   struct btrfs_root *root,
8644                                   struct extent_buffer *root_eb,
8645                                   u64 root_gen,
8646                                   int root_level)
8647 {
8648         int ret = 0;
8649         int level;
8650         struct extent_buffer *eb = root_eb;
8651         struct btrfs_path *path = NULL;
8652
8653         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
8654         BUG_ON(root_eb == NULL);
8655
8656         if (!root->fs_info->quota_enabled)
8657                 return 0;
8658
8659         if (!extent_buffer_uptodate(root_eb)) {
8660                 ret = btrfs_read_buffer(root_eb, root_gen);
8661                 if (ret)
8662                         goto out;
8663         }
8664
8665         if (root_level == 0) {
8666                 ret = account_leaf_items(trans, root, root_eb);
8667                 goto out;
8668         }
8669
8670         path = btrfs_alloc_path();
8671         if (!path)
8672                 return -ENOMEM;
8673
8674         /*
8675          * Walk down the tree.  Missing extent blocks are filled in as
8676          * we go. Metadata is accounted every time we read a new
8677          * extent block.
8678          *
8679          * When we reach a leaf, we account for file extent items in it,
8680          * walk back up the tree (adjusting slot pointers as we go)
8681          * and restart the search process.
8682          */
8683         extent_buffer_get(root_eb); /* For path */
8684         path->nodes[root_level] = root_eb;
8685         path->slots[root_level] = 0;
8686         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
8687 walk_down:
8688         level = root_level;
8689         while (level >= 0) {
8690                 if (path->nodes[level] == NULL) {
8691                         int parent_slot;
8692                         u64 child_gen;
8693                         u64 child_bytenr;
8694
8695                         /* We need to get child blockptr/gen from
8696                          * parent before we can read it. */
8697                         eb = path->nodes[level + 1];
8698                         parent_slot = path->slots[level + 1];
8699                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8700                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8701
8702                         eb = read_tree_block(root, child_bytenr, child_gen);
8703                         if (IS_ERR(eb)) {
8704                                 ret = PTR_ERR(eb);
8705                                 goto out;
8706                         } else if (!extent_buffer_uptodate(eb)) {
8707                                 free_extent_buffer(eb);
8708                                 ret = -EIO;
8709                                 goto out;
8710                         }
8711
8712                         path->nodes[level] = eb;
8713                         path->slots[level] = 0;
8714
8715                         btrfs_tree_read_lock(eb);
8716                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8717                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8718
8719                         ret = btrfs_qgroup_insert_dirty_extent(trans,
8720                                         root->fs_info, child_bytenr,
8721                                         root->nodesize, GFP_NOFS);
8722                         if (ret)
8723                                 goto out;
8724                 }
8725
8726                 if (level == 0) {
8727                         ret = account_leaf_items(trans, root, path->nodes[level]);
8728                         if (ret)
8729                                 goto out;
8730
8731                         /* Nonzero return here means we completed our search */
8732                         ret = adjust_slots_upwards(root, path, root_level);
8733                         if (ret)
8734                                 break;
8735
8736                         /* Restart search with new slots */
8737                         goto walk_down;
8738                 }
8739
8740                 level--;
8741         }
8742
8743         ret = 0;
8744 out:
8745         btrfs_free_path(path);
8746
8747         return ret;
8748 }
8749
8750 /*
8751  * helper to process tree block while walking down the tree.
8752  *
8753  * when wc->stage == UPDATE_BACKREF, this function updates
8754  * back refs for pointers in the block.
8755  *
8756  * NOTE: return value 1 means we should stop walking down.
8757  */
8758 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8759                                    struct btrfs_root *root,
8760                                    struct btrfs_path *path,
8761                                    struct walk_control *wc, int lookup_info)
8762 {
8763         int level = wc->level;
8764         struct extent_buffer *eb = path->nodes[level];
8765         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8766         int ret;
8767
8768         if (wc->stage == UPDATE_BACKREF &&
8769             btrfs_header_owner(eb) != root->root_key.objectid)
8770                 return 1;
8771
8772         /*
8773          * when reference count of tree block is 1, it won't increase
8774          * again. once full backref flag is set, we never clear it.
8775          */
8776         if (lookup_info &&
8777             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8778              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8779                 BUG_ON(!path->locks[level]);
8780                 ret = btrfs_lookup_extent_info(trans, root,
8781                                                eb->start, level, 1,
8782                                                &wc->refs[level],
8783                                                &wc->flags[level]);
8784                 BUG_ON(ret == -ENOMEM);
8785                 if (ret)
8786                         return ret;
8787                 BUG_ON(wc->refs[level] == 0);
8788         }
8789
8790         if (wc->stage == DROP_REFERENCE) {
8791                 if (wc->refs[level] > 1)
8792                         return 1;
8793
8794                 if (path->locks[level] && !wc->keep_locks) {
8795                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8796                         path->locks[level] = 0;
8797                 }
8798                 return 0;
8799         }
8800
8801         /* wc->stage == UPDATE_BACKREF */
8802         if (!(wc->flags[level] & flag)) {
8803                 BUG_ON(!path->locks[level]);
8804                 ret = btrfs_inc_ref(trans, root, eb, 1);
8805                 BUG_ON(ret); /* -ENOMEM */
8806                 ret = btrfs_dec_ref(trans, root, eb, 0);
8807                 BUG_ON(ret); /* -ENOMEM */
8808                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8809                                                   eb->len, flag,
8810                                                   btrfs_header_level(eb), 0);
8811                 BUG_ON(ret); /* -ENOMEM */
8812                 wc->flags[level] |= flag;
8813         }
8814
8815         /*
8816          * the block is shared by multiple trees, so it's not good to
8817          * keep the tree lock
8818          */
8819         if (path->locks[level] && level > 0) {
8820                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8821                 path->locks[level] = 0;
8822         }
8823         return 0;
8824 }
8825
8826 /*
8827  * helper to process tree block pointer.
8828  *
8829  * when wc->stage == DROP_REFERENCE, this function checks
8830  * reference count of the block pointed to. if the block
8831  * is shared and we need update back refs for the subtree
8832  * rooted at the block, this function changes wc->stage to
8833  * UPDATE_BACKREF. if the block is shared and there is no
8834  * need to update back, this function drops the reference
8835  * to the block.
8836  *
8837  * NOTE: return value 1 means we should stop walking down.
8838  */
8839 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8840                                  struct btrfs_root *root,
8841                                  struct btrfs_path *path,
8842                                  struct walk_control *wc, int *lookup_info)
8843 {
8844         u64 bytenr;
8845         u64 generation;
8846         u64 parent;
8847         u32 blocksize;
8848         struct btrfs_key key;
8849         struct extent_buffer *next;
8850         int level = wc->level;
8851         int reada = 0;
8852         int ret = 0;
8853         bool need_account = false;
8854
8855         generation = btrfs_node_ptr_generation(path->nodes[level],
8856                                                path->slots[level]);
8857         /*
8858          * if the lower level block was created before the snapshot
8859          * was created, we know there is no need to update back refs
8860          * for the subtree
8861          */
8862         if (wc->stage == UPDATE_BACKREF &&
8863             generation <= root->root_key.offset) {
8864                 *lookup_info = 1;
8865                 return 1;
8866         }
8867
8868         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8869         blocksize = root->nodesize;
8870
8871         next = btrfs_find_tree_block(root->fs_info, bytenr);
8872         if (!next) {
8873                 next = btrfs_find_create_tree_block(root, bytenr);
8874                 if (IS_ERR(next))
8875                         return PTR_ERR(next);
8876
8877                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8878                                                level - 1);
8879                 reada = 1;
8880         }
8881         btrfs_tree_lock(next);
8882         btrfs_set_lock_blocking(next);
8883
8884         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8885                                        &wc->refs[level - 1],
8886                                        &wc->flags[level - 1]);
8887         if (ret < 0) {
8888                 btrfs_tree_unlock(next);
8889                 return ret;
8890         }
8891
8892         if (unlikely(wc->refs[level - 1] == 0)) {
8893                 btrfs_err(root->fs_info, "Missing references.");
8894                 BUG();
8895         }
8896         *lookup_info = 0;
8897
8898         if (wc->stage == DROP_REFERENCE) {
8899                 if (wc->refs[level - 1] > 1) {
8900                         need_account = true;
8901                         if (level == 1 &&
8902                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8903                                 goto skip;
8904
8905                         if (!wc->update_ref ||
8906                             generation <= root->root_key.offset)
8907                                 goto skip;
8908
8909                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8910                                               path->slots[level]);
8911                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8912                         if (ret < 0)
8913                                 goto skip;
8914
8915                         wc->stage = UPDATE_BACKREF;
8916                         wc->shared_level = level - 1;
8917                 }
8918         } else {
8919                 if (level == 1 &&
8920                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8921                         goto skip;
8922         }
8923
8924         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8925                 btrfs_tree_unlock(next);
8926                 free_extent_buffer(next);
8927                 next = NULL;
8928                 *lookup_info = 1;
8929         }
8930
8931         if (!next) {
8932                 if (reada && level == 1)
8933                         reada_walk_down(trans, root, wc, path);
8934                 next = read_tree_block(root, bytenr, generation);
8935                 if (IS_ERR(next)) {
8936                         return PTR_ERR(next);
8937                 } else if (!extent_buffer_uptodate(next)) {
8938                         free_extent_buffer(next);
8939                         return -EIO;
8940                 }
8941                 btrfs_tree_lock(next);
8942                 btrfs_set_lock_blocking(next);
8943         }
8944
8945         level--;
8946         BUG_ON(level != btrfs_header_level(next));
8947         path->nodes[level] = next;
8948         path->slots[level] = 0;
8949         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8950         wc->level = level;
8951         if (wc->level == 1)
8952                 wc->reada_slot = 0;
8953         return 0;
8954 skip:
8955         wc->refs[level - 1] = 0;
8956         wc->flags[level - 1] = 0;
8957         if (wc->stage == DROP_REFERENCE) {
8958                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8959                         parent = path->nodes[level]->start;
8960                 } else {
8961                         BUG_ON(root->root_key.objectid !=
8962                                btrfs_header_owner(path->nodes[level]));
8963                         parent = 0;
8964                 }
8965
8966                 if (need_account) {
8967                         ret = account_shared_subtree(trans, root, next,
8968                                                      generation, level - 1);
8969                         if (ret) {
8970                                 btrfs_err_rl(root->fs_info,
8971                                         "Error "
8972                                         "%d accounting shared subtree. Quota "
8973                                         "is out of sync, rescan required.",
8974                                         ret);
8975                         }
8976                 }
8977                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8978                                 root->root_key.objectid, level - 1, 0);
8979                 BUG_ON(ret); /* -ENOMEM */
8980         }
8981         btrfs_tree_unlock(next);
8982         free_extent_buffer(next);
8983         *lookup_info = 1;
8984         return 1;
8985 }
8986
8987 /*
8988  * helper to process tree block while walking up the tree.
8989  *
8990  * when wc->stage == DROP_REFERENCE, this function drops
8991  * reference count on the block.
8992  *
8993  * when wc->stage == UPDATE_BACKREF, this function changes
8994  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8995  * to UPDATE_BACKREF previously while processing the block.
8996  *
8997  * NOTE: return value 1 means we should stop walking up.
8998  */
8999 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9000                                  struct btrfs_root *root,
9001                                  struct btrfs_path *path,
9002                                  struct walk_control *wc)
9003 {
9004         int ret;
9005         int level = wc->level;
9006         struct extent_buffer *eb = path->nodes[level];
9007         u64 parent = 0;
9008
9009         if (wc->stage == UPDATE_BACKREF) {
9010                 BUG_ON(wc->shared_level < level);
9011                 if (level < wc->shared_level)
9012                         goto out;
9013
9014                 ret = find_next_key(path, level + 1, &wc->update_progress);
9015                 if (ret > 0)
9016                         wc->update_ref = 0;
9017
9018                 wc->stage = DROP_REFERENCE;
9019                 wc->shared_level = -1;
9020                 path->slots[level] = 0;
9021
9022                 /*
9023                  * check reference count again if the block isn't locked.
9024                  * we should start walking down the tree again if reference
9025                  * count is one.
9026                  */
9027                 if (!path->locks[level]) {
9028                         BUG_ON(level == 0);
9029                         btrfs_tree_lock(eb);
9030                         btrfs_set_lock_blocking(eb);
9031                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9032
9033                         ret = btrfs_lookup_extent_info(trans, root,
9034                                                        eb->start, level, 1,
9035                                                        &wc->refs[level],
9036                                                        &wc->flags[level]);
9037                         if (ret < 0) {
9038                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9039                                 path->locks[level] = 0;
9040                                 return ret;
9041                         }
9042                         BUG_ON(wc->refs[level] == 0);
9043                         if (wc->refs[level] == 1) {
9044                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9045                                 path->locks[level] = 0;
9046                                 return 1;
9047                         }
9048                 }
9049         }
9050
9051         /* wc->stage == DROP_REFERENCE */
9052         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9053
9054         if (wc->refs[level] == 1) {
9055                 if (level == 0) {
9056                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9057                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9058                         else
9059                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9060                         BUG_ON(ret); /* -ENOMEM */
9061                         ret = account_leaf_items(trans, root, eb);
9062                         if (ret) {
9063                                 btrfs_err_rl(root->fs_info,
9064                                         "error "
9065                                         "%d accounting leaf items. Quota "
9066                                         "is out of sync, rescan required.",
9067                                         ret);
9068                         }
9069                 }
9070                 /* make block locked assertion in clean_tree_block happy */
9071                 if (!path->locks[level] &&
9072                     btrfs_header_generation(eb) == trans->transid) {
9073                         btrfs_tree_lock(eb);
9074                         btrfs_set_lock_blocking(eb);
9075                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9076                 }
9077                 clean_tree_block(trans, root->fs_info, eb);
9078         }
9079
9080         if (eb == root->node) {
9081                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9082                         parent = eb->start;
9083                 else
9084                         BUG_ON(root->root_key.objectid !=
9085                                btrfs_header_owner(eb));
9086         } else {
9087                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9088                         parent = path->nodes[level + 1]->start;
9089                 else
9090                         BUG_ON(root->root_key.objectid !=
9091                                btrfs_header_owner(path->nodes[level + 1]));
9092         }
9093
9094         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9095 out:
9096         wc->refs[level] = 0;
9097         wc->flags[level] = 0;
9098         return 0;
9099 }
9100
9101 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9102                                    struct btrfs_root *root,
9103                                    struct btrfs_path *path,
9104                                    struct walk_control *wc)
9105 {
9106         int level = wc->level;
9107         int lookup_info = 1;
9108         int ret;
9109
9110         while (level >= 0) {
9111                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9112                 if (ret > 0)
9113                         break;
9114
9115                 if (level == 0)
9116                         break;
9117
9118                 if (path->slots[level] >=
9119                     btrfs_header_nritems(path->nodes[level]))
9120                         break;
9121
9122                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9123                 if (ret > 0) {
9124                         path->slots[level]++;
9125                         continue;
9126                 } else if (ret < 0)
9127                         return ret;
9128                 level = wc->level;
9129         }
9130         return 0;
9131 }
9132
9133 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9134                                  struct btrfs_root *root,
9135                                  struct btrfs_path *path,
9136                                  struct walk_control *wc, int max_level)
9137 {
9138         int level = wc->level;
9139         int ret;
9140
9141         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9142         while (level < max_level && path->nodes[level]) {
9143                 wc->level = level;
9144                 if (path->slots[level] + 1 <
9145                     btrfs_header_nritems(path->nodes[level])) {
9146                         path->slots[level]++;
9147                         return 0;
9148                 } else {
9149                         ret = walk_up_proc(trans, root, path, wc);
9150                         if (ret > 0)
9151                                 return 0;
9152
9153                         if (path->locks[level]) {
9154                                 btrfs_tree_unlock_rw(path->nodes[level],
9155                                                      path->locks[level]);
9156                                 path->locks[level] = 0;
9157                         }
9158                         free_extent_buffer(path->nodes[level]);
9159                         path->nodes[level] = NULL;
9160                         level++;
9161                 }
9162         }
9163         return 1;
9164 }
9165
9166 /*
9167  * drop a subvolume tree.
9168  *
9169  * this function traverses the tree freeing any blocks that only
9170  * referenced by the tree.
9171  *
9172  * when a shared tree block is found. this function decreases its
9173  * reference count by one. if update_ref is true, this function
9174  * also make sure backrefs for the shared block and all lower level
9175  * blocks are properly updated.
9176  *
9177  * If called with for_reloc == 0, may exit early with -EAGAIN
9178  */
9179 int btrfs_drop_snapshot(struct btrfs_root *root,
9180                          struct btrfs_block_rsv *block_rsv, int update_ref,
9181                          int for_reloc)
9182 {
9183         struct btrfs_path *path;
9184         struct btrfs_trans_handle *trans;
9185         struct btrfs_root *tree_root = root->fs_info->tree_root;
9186         struct btrfs_root_item *root_item = &root->root_item;
9187         struct walk_control *wc;
9188         struct btrfs_key key;
9189         int err = 0;
9190         int ret;
9191         int level;
9192         bool root_dropped = false;
9193
9194         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
9195
9196         path = btrfs_alloc_path();
9197         if (!path) {
9198                 err = -ENOMEM;
9199                 goto out;
9200         }
9201
9202         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9203         if (!wc) {
9204                 btrfs_free_path(path);
9205                 err = -ENOMEM;
9206                 goto out;
9207         }
9208
9209         trans = btrfs_start_transaction(tree_root, 0);
9210         if (IS_ERR(trans)) {
9211                 err = PTR_ERR(trans);
9212                 goto out_free;
9213         }
9214
9215         if (block_rsv)
9216                 trans->block_rsv = block_rsv;
9217
9218         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9219                 level = btrfs_header_level(root->node);
9220                 path->nodes[level] = btrfs_lock_root_node(root);
9221                 btrfs_set_lock_blocking(path->nodes[level]);
9222                 path->slots[level] = 0;
9223                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9224                 memset(&wc->update_progress, 0,
9225                        sizeof(wc->update_progress));
9226         } else {
9227                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9228                 memcpy(&wc->update_progress, &key,
9229                        sizeof(wc->update_progress));
9230
9231                 level = root_item->drop_level;
9232                 BUG_ON(level == 0);
9233                 path->lowest_level = level;
9234                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9235                 path->lowest_level = 0;
9236                 if (ret < 0) {
9237                         err = ret;
9238                         goto out_end_trans;
9239                 }
9240                 WARN_ON(ret > 0);
9241
9242                 /*
9243                  * unlock our path, this is safe because only this
9244                  * function is allowed to delete this snapshot
9245                  */
9246                 btrfs_unlock_up_safe(path, 0);
9247
9248                 level = btrfs_header_level(root->node);
9249                 while (1) {
9250                         btrfs_tree_lock(path->nodes[level]);
9251                         btrfs_set_lock_blocking(path->nodes[level]);
9252                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9253
9254                         ret = btrfs_lookup_extent_info(trans, root,
9255                                                 path->nodes[level]->start,
9256                                                 level, 1, &wc->refs[level],
9257                                                 &wc->flags[level]);
9258                         if (ret < 0) {
9259                                 err = ret;
9260                                 goto out_end_trans;
9261                         }
9262                         BUG_ON(wc->refs[level] == 0);
9263
9264                         if (level == root_item->drop_level)
9265                                 break;
9266
9267                         btrfs_tree_unlock(path->nodes[level]);
9268                         path->locks[level] = 0;
9269                         WARN_ON(wc->refs[level] != 1);
9270                         level--;
9271                 }
9272         }
9273
9274         wc->level = level;
9275         wc->shared_level = -1;
9276         wc->stage = DROP_REFERENCE;
9277         wc->update_ref = update_ref;
9278         wc->keep_locks = 0;
9279         wc->for_reloc = for_reloc;
9280         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9281
9282         while (1) {
9283
9284                 ret = walk_down_tree(trans, root, path, wc);
9285                 if (ret < 0) {
9286                         err = ret;
9287                         break;
9288                 }
9289
9290                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9291                 if (ret < 0) {
9292                         err = ret;
9293                         break;
9294                 }
9295
9296                 if (ret > 0) {
9297                         BUG_ON(wc->stage != DROP_REFERENCE);
9298                         break;
9299                 }
9300
9301                 if (wc->stage == DROP_REFERENCE) {
9302                         level = wc->level;
9303                         btrfs_node_key(path->nodes[level],
9304                                        &root_item->drop_progress,
9305                                        path->slots[level]);
9306                         root_item->drop_level = level;
9307                 }
9308
9309                 BUG_ON(wc->level == 0);
9310                 if (btrfs_should_end_transaction(trans, tree_root) ||
9311                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
9312                         ret = btrfs_update_root(trans, tree_root,
9313                                                 &root->root_key,
9314                                                 root_item);
9315                         if (ret) {
9316                                 btrfs_abort_transaction(trans, ret);
9317                                 err = ret;
9318                                 goto out_end_trans;
9319                         }
9320
9321                         btrfs_end_transaction_throttle(trans, tree_root);
9322                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
9323                                 pr_debug("BTRFS: drop snapshot early exit\n");
9324                                 err = -EAGAIN;
9325                                 goto out_free;
9326                         }
9327
9328                         trans = btrfs_start_transaction(tree_root, 0);
9329                         if (IS_ERR(trans)) {
9330                                 err = PTR_ERR(trans);
9331                                 goto out_free;
9332                         }
9333                         if (block_rsv)
9334                                 trans->block_rsv = block_rsv;
9335                 }
9336         }
9337         btrfs_release_path(path);
9338         if (err)
9339                 goto out_end_trans;
9340
9341         ret = btrfs_del_root(trans, tree_root, &root->root_key);
9342         if (ret) {
9343                 btrfs_abort_transaction(trans, ret);
9344                 goto out_end_trans;
9345         }
9346
9347         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9348                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9349                                       NULL, NULL);
9350                 if (ret < 0) {
9351                         btrfs_abort_transaction(trans, ret);
9352                         err = ret;
9353                         goto out_end_trans;
9354                 } else if (ret > 0) {
9355                         /* if we fail to delete the orphan item this time
9356                          * around, it'll get picked up the next time.
9357                          *
9358                          * The most common failure here is just -ENOENT.
9359                          */
9360                         btrfs_del_orphan_item(trans, tree_root,
9361                                               root->root_key.objectid);
9362                 }
9363         }
9364
9365         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9366                 btrfs_add_dropped_root(trans, root);
9367         } else {
9368                 free_extent_buffer(root->node);
9369                 free_extent_buffer(root->commit_root);
9370                 btrfs_put_fs_root(root);
9371         }
9372         root_dropped = true;
9373 out_end_trans:
9374         btrfs_end_transaction_throttle(trans, tree_root);
9375 out_free:
9376         kfree(wc);
9377         btrfs_free_path(path);
9378 out:
9379         /*
9380          * So if we need to stop dropping the snapshot for whatever reason we
9381          * need to make sure to add it back to the dead root list so that we
9382          * keep trying to do the work later.  This also cleans up roots if we
9383          * don't have it in the radix (like when we recover after a power fail
9384          * or unmount) so we don't leak memory.
9385          */
9386         if (!for_reloc && root_dropped == false)
9387                 btrfs_add_dead_root(root);
9388         if (err && err != -EAGAIN)
9389                 btrfs_handle_fs_error(root->fs_info, err, NULL);
9390         return err;
9391 }
9392
9393 /*
9394  * drop subtree rooted at tree block 'node'.
9395  *
9396  * NOTE: this function will unlock and release tree block 'node'
9397  * only used by relocation code
9398  */
9399 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9400                         struct btrfs_root *root,
9401                         struct extent_buffer *node,
9402                         struct extent_buffer *parent)
9403 {
9404         struct btrfs_path *path;
9405         struct walk_control *wc;
9406         int level;
9407         int parent_level;
9408         int ret = 0;
9409         int wret;
9410
9411         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9412
9413         path = btrfs_alloc_path();
9414         if (!path)
9415                 return -ENOMEM;
9416
9417         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9418         if (!wc) {
9419                 btrfs_free_path(path);
9420                 return -ENOMEM;
9421         }
9422
9423         btrfs_assert_tree_locked(parent);
9424         parent_level = btrfs_header_level(parent);
9425         extent_buffer_get(parent);
9426         path->nodes[parent_level] = parent;
9427         path->slots[parent_level] = btrfs_header_nritems(parent);
9428
9429         btrfs_assert_tree_locked(node);
9430         level = btrfs_header_level(node);
9431         path->nodes[level] = node;
9432         path->slots[level] = 0;
9433         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9434
9435         wc->refs[parent_level] = 1;
9436         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9437         wc->level = level;
9438         wc->shared_level = -1;
9439         wc->stage = DROP_REFERENCE;
9440         wc->update_ref = 0;
9441         wc->keep_locks = 1;
9442         wc->for_reloc = 1;
9443         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9444
9445         while (1) {
9446                 wret = walk_down_tree(trans, root, path, wc);
9447                 if (wret < 0) {
9448                         ret = wret;
9449                         break;
9450                 }
9451
9452                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9453                 if (wret < 0)
9454                         ret = wret;
9455                 if (wret != 0)
9456                         break;
9457         }
9458
9459         kfree(wc);
9460         btrfs_free_path(path);
9461         return ret;
9462 }
9463
9464 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
9465 {
9466         u64 num_devices;
9467         u64 stripped;
9468
9469         /*
9470          * if restripe for this chunk_type is on pick target profile and
9471          * return, otherwise do the usual balance
9472          */
9473         stripped = get_restripe_target(root->fs_info, flags);
9474         if (stripped)
9475                 return extended_to_chunk(stripped);
9476
9477         num_devices = root->fs_info->fs_devices->rw_devices;
9478
9479         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9480                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9481                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9482
9483         if (num_devices == 1) {
9484                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9485                 stripped = flags & ~stripped;
9486
9487                 /* turn raid0 into single device chunks */
9488                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9489                         return stripped;
9490
9491                 /* turn mirroring into duplication */
9492                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9493                              BTRFS_BLOCK_GROUP_RAID10))
9494                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9495         } else {
9496                 /* they already had raid on here, just return */
9497                 if (flags & stripped)
9498                         return flags;
9499
9500                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9501                 stripped = flags & ~stripped;
9502
9503                 /* switch duplicated blocks with raid1 */
9504                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9505                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9506
9507                 /* this is drive concat, leave it alone */
9508         }
9509
9510         return flags;
9511 }
9512
9513 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9514 {
9515         struct btrfs_space_info *sinfo = cache->space_info;
9516         u64 num_bytes;
9517         u64 min_allocable_bytes;
9518         int ret = -ENOSPC;
9519
9520         /*
9521          * We need some metadata space and system metadata space for
9522          * allocating chunks in some corner cases until we force to set
9523          * it to be readonly.
9524          */
9525         if ((sinfo->flags &
9526              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9527             !force)
9528                 min_allocable_bytes = SZ_1M;
9529         else
9530                 min_allocable_bytes = 0;
9531
9532         spin_lock(&sinfo->lock);
9533         spin_lock(&cache->lock);
9534
9535         if (cache->ro) {
9536                 cache->ro++;
9537                 ret = 0;
9538                 goto out;
9539         }
9540
9541         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9542                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9543
9544         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
9545             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
9546             min_allocable_bytes <= sinfo->total_bytes) {
9547                 sinfo->bytes_readonly += num_bytes;
9548                 cache->ro++;
9549                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9550                 ret = 0;
9551         }
9552 out:
9553         spin_unlock(&cache->lock);
9554         spin_unlock(&sinfo->lock);
9555         return ret;
9556 }
9557
9558 int btrfs_inc_block_group_ro(struct btrfs_root *root,
9559                              struct btrfs_block_group_cache *cache)
9560
9561 {
9562         struct btrfs_trans_handle *trans;
9563         u64 alloc_flags;
9564         int ret;
9565
9566 again:
9567         trans = btrfs_join_transaction(root);
9568         if (IS_ERR(trans))
9569                 return PTR_ERR(trans);
9570
9571         /*
9572          * we're not allowed to set block groups readonly after the dirty
9573          * block groups cache has started writing.  If it already started,
9574          * back off and let this transaction commit
9575          */
9576         mutex_lock(&root->fs_info->ro_block_group_mutex);
9577         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9578                 u64 transid = trans->transid;
9579
9580                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
9581                 btrfs_end_transaction(trans, root);
9582
9583                 ret = btrfs_wait_for_commit(root, transid);
9584                 if (ret)
9585                         return ret;
9586                 goto again;
9587         }
9588
9589         /*
9590          * if we are changing raid levels, try to allocate a corresponding
9591          * block group with the new raid level.
9592          */
9593         alloc_flags = update_block_group_flags(root, cache->flags);
9594         if (alloc_flags != cache->flags) {
9595                 ret = do_chunk_alloc(trans, root, alloc_flags,
9596                                      CHUNK_ALLOC_FORCE);
9597                 /*
9598                  * ENOSPC is allowed here, we may have enough space
9599                  * already allocated at the new raid level to
9600                  * carry on
9601                  */
9602                 if (ret == -ENOSPC)
9603                         ret = 0;
9604                 if (ret < 0)
9605                         goto out;
9606         }
9607
9608         ret = inc_block_group_ro(cache, 0);
9609         if (!ret)
9610                 goto out;
9611         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
9612         ret = do_chunk_alloc(trans, root, alloc_flags,
9613                              CHUNK_ALLOC_FORCE);
9614         if (ret < 0)
9615                 goto out;
9616         ret = inc_block_group_ro(cache, 0);
9617 out:
9618         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9619                 alloc_flags = update_block_group_flags(root, cache->flags);
9620                 lock_chunks(root->fs_info->chunk_root);
9621                 check_system_chunk(trans, root, alloc_flags);
9622                 unlock_chunks(root->fs_info->chunk_root);
9623         }
9624         mutex_unlock(&root->fs_info->ro_block_group_mutex);
9625
9626         btrfs_end_transaction(trans, root);
9627         return ret;
9628 }
9629
9630 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9631                             struct btrfs_root *root, u64 type)
9632 {
9633         u64 alloc_flags = get_alloc_profile(root, type);
9634         return do_chunk_alloc(trans, root, alloc_flags,
9635                               CHUNK_ALLOC_FORCE);
9636 }
9637
9638 /*
9639  * helper to account the unused space of all the readonly block group in the
9640  * space_info. takes mirrors into account.
9641  */
9642 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9643 {
9644         struct btrfs_block_group_cache *block_group;
9645         u64 free_bytes = 0;
9646         int factor;
9647
9648         /* It's df, we don't care if it's racy */
9649         if (list_empty(&sinfo->ro_bgs))
9650                 return 0;
9651
9652         spin_lock(&sinfo->lock);
9653         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9654                 spin_lock(&block_group->lock);
9655
9656                 if (!block_group->ro) {
9657                         spin_unlock(&block_group->lock);
9658                         continue;
9659                 }
9660
9661                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9662                                           BTRFS_BLOCK_GROUP_RAID10 |
9663                                           BTRFS_BLOCK_GROUP_DUP))
9664                         factor = 2;
9665                 else
9666                         factor = 1;
9667
9668                 free_bytes += (block_group->key.offset -
9669                                btrfs_block_group_used(&block_group->item)) *
9670                                factor;
9671
9672                 spin_unlock(&block_group->lock);
9673         }
9674         spin_unlock(&sinfo->lock);
9675
9676         return free_bytes;
9677 }
9678
9679 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9680                               struct btrfs_block_group_cache *cache)
9681 {
9682         struct btrfs_space_info *sinfo = cache->space_info;
9683         u64 num_bytes;
9684
9685         BUG_ON(!cache->ro);
9686
9687         spin_lock(&sinfo->lock);
9688         spin_lock(&cache->lock);
9689         if (!--cache->ro) {
9690                 num_bytes = cache->key.offset - cache->reserved -
9691                             cache->pinned - cache->bytes_super -
9692                             btrfs_block_group_used(&cache->item);
9693                 sinfo->bytes_readonly -= num_bytes;
9694                 list_del_init(&cache->ro_list);
9695         }
9696         spin_unlock(&cache->lock);
9697         spin_unlock(&sinfo->lock);
9698 }
9699
9700 /*
9701  * checks to see if its even possible to relocate this block group.
9702  *
9703  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9704  * ok to go ahead and try.
9705  */
9706 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9707 {
9708         struct btrfs_block_group_cache *block_group;
9709         struct btrfs_space_info *space_info;
9710         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9711         struct btrfs_device *device;
9712         struct btrfs_trans_handle *trans;
9713         u64 min_free;
9714         u64 dev_min = 1;
9715         u64 dev_nr = 0;
9716         u64 target;
9717         int debug;
9718         int index;
9719         int full = 0;
9720         int ret = 0;
9721
9722         debug = btrfs_test_opt(root->fs_info, ENOSPC_DEBUG);
9723
9724         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9725
9726         /* odd, couldn't find the block group, leave it alone */
9727         if (!block_group) {
9728                 if (debug)
9729                         btrfs_warn(root->fs_info,
9730                                    "can't find block group for bytenr %llu",
9731                                    bytenr);
9732                 return -1;
9733         }
9734
9735         min_free = btrfs_block_group_used(&block_group->item);
9736
9737         /* no bytes used, we're good */
9738         if (!min_free)
9739                 goto out;
9740
9741         space_info = block_group->space_info;
9742         spin_lock(&space_info->lock);
9743
9744         full = space_info->full;
9745
9746         /*
9747          * if this is the last block group we have in this space, we can't
9748          * relocate it unless we're able to allocate a new chunk below.
9749          *
9750          * Otherwise, we need to make sure we have room in the space to handle
9751          * all of the extents from this block group.  If we can, we're good
9752          */
9753         if ((space_info->total_bytes != block_group->key.offset) &&
9754             (space_info->bytes_used + space_info->bytes_reserved +
9755              space_info->bytes_pinned + space_info->bytes_readonly +
9756              min_free < space_info->total_bytes)) {
9757                 spin_unlock(&space_info->lock);
9758                 goto out;
9759         }
9760         spin_unlock(&space_info->lock);
9761
9762         /*
9763          * ok we don't have enough space, but maybe we have free space on our
9764          * devices to allocate new chunks for relocation, so loop through our
9765          * alloc devices and guess if we have enough space.  if this block
9766          * group is going to be restriped, run checks against the target
9767          * profile instead of the current one.
9768          */
9769         ret = -1;
9770
9771         /*
9772          * index:
9773          *      0: raid10
9774          *      1: raid1
9775          *      2: dup
9776          *      3: raid0
9777          *      4: single
9778          */
9779         target = get_restripe_target(root->fs_info, block_group->flags);
9780         if (target) {
9781                 index = __get_raid_index(extended_to_chunk(target));
9782         } else {
9783                 /*
9784                  * this is just a balance, so if we were marked as full
9785                  * we know there is no space for a new chunk
9786                  */
9787                 if (full) {
9788                         if (debug)
9789                                 btrfs_warn(root->fs_info,
9790                                         "no space to alloc new chunk for block group %llu",
9791                                         block_group->key.objectid);
9792                         goto out;
9793                 }
9794
9795                 index = get_block_group_index(block_group);
9796         }
9797
9798         if (index == BTRFS_RAID_RAID10) {
9799                 dev_min = 4;
9800                 /* Divide by 2 */
9801                 min_free >>= 1;
9802         } else if (index == BTRFS_RAID_RAID1) {
9803                 dev_min = 2;
9804         } else if (index == BTRFS_RAID_DUP) {
9805                 /* Multiply by 2 */
9806                 min_free <<= 1;
9807         } else if (index == BTRFS_RAID_RAID0) {
9808                 dev_min = fs_devices->rw_devices;
9809                 min_free = div64_u64(min_free, dev_min);
9810         }
9811
9812         /* We need to do this so that we can look at pending chunks */
9813         trans = btrfs_join_transaction(root);
9814         if (IS_ERR(trans)) {
9815                 ret = PTR_ERR(trans);
9816                 goto out;
9817         }
9818
9819         mutex_lock(&root->fs_info->chunk_mutex);
9820         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9821                 u64 dev_offset;
9822
9823                 /*
9824                  * check to make sure we can actually find a chunk with enough
9825                  * space to fit our block group in.
9826                  */
9827                 if (device->total_bytes > device->bytes_used + min_free &&
9828                     !device->is_tgtdev_for_dev_replace) {
9829                         ret = find_free_dev_extent(trans, device, min_free,
9830                                                    &dev_offset, NULL);
9831                         if (!ret)
9832                                 dev_nr++;
9833
9834                         if (dev_nr >= dev_min)
9835                                 break;
9836
9837                         ret = -1;
9838                 }
9839         }
9840         if (debug && ret == -1)
9841                 btrfs_warn(root->fs_info,
9842                         "no space to allocate a new chunk for block group %llu",
9843                         block_group->key.objectid);
9844         mutex_unlock(&root->fs_info->chunk_mutex);
9845         btrfs_end_transaction(trans, root);
9846 out:
9847         btrfs_put_block_group(block_group);
9848         return ret;
9849 }
9850
9851 static int find_first_block_group(struct btrfs_root *root,
9852                 struct btrfs_path *path, struct btrfs_key *key)
9853 {
9854         int ret = 0;
9855         struct btrfs_key found_key;
9856         struct extent_buffer *leaf;
9857         int slot;
9858
9859         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9860         if (ret < 0)
9861                 goto out;
9862
9863         while (1) {
9864                 slot = path->slots[0];
9865                 leaf = path->nodes[0];
9866                 if (slot >= btrfs_header_nritems(leaf)) {
9867                         ret = btrfs_next_leaf(root, path);
9868                         if (ret == 0)
9869                                 continue;
9870                         if (ret < 0)
9871                                 goto out;
9872                         break;
9873                 }
9874                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9875
9876                 if (found_key.objectid >= key->objectid &&
9877                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9878                         struct extent_map_tree *em_tree;
9879                         struct extent_map *em;
9880
9881                         em_tree = &root->fs_info->mapping_tree.map_tree;
9882                         read_lock(&em_tree->lock);
9883                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9884                                                    found_key.offset);
9885                         read_unlock(&em_tree->lock);
9886                         if (!em) {
9887                                 btrfs_err(root->fs_info,
9888                         "logical %llu len %llu found bg but no related chunk",
9889                                           found_key.objectid, found_key.offset);
9890                                 ret = -ENOENT;
9891                         } else {
9892                                 ret = 0;
9893                         }
9894                         free_extent_map(em);
9895                         goto out;
9896                 }
9897                 path->slots[0]++;
9898         }
9899 out:
9900         return ret;
9901 }
9902
9903 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9904 {
9905         struct btrfs_block_group_cache *block_group;
9906         u64 last = 0;
9907
9908         while (1) {
9909                 struct inode *inode;
9910
9911                 block_group = btrfs_lookup_first_block_group(info, last);
9912                 while (block_group) {
9913                         spin_lock(&block_group->lock);
9914                         if (block_group->iref)
9915                                 break;
9916                         spin_unlock(&block_group->lock);
9917                         block_group = next_block_group(info->tree_root,
9918                                                        block_group);
9919                 }
9920                 if (!block_group) {
9921                         if (last == 0)
9922                                 break;
9923                         last = 0;
9924                         continue;
9925                 }
9926
9927                 inode = block_group->inode;
9928                 block_group->iref = 0;
9929                 block_group->inode = NULL;
9930                 spin_unlock(&block_group->lock);
9931                 ASSERT(block_group->io_ctl.inode == NULL);
9932                 iput(inode);
9933                 last = block_group->key.objectid + block_group->key.offset;
9934                 btrfs_put_block_group(block_group);
9935         }
9936 }
9937
9938 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9939 {
9940         struct btrfs_block_group_cache *block_group;
9941         struct btrfs_space_info *space_info;
9942         struct btrfs_caching_control *caching_ctl;
9943         struct rb_node *n;
9944
9945         down_write(&info->commit_root_sem);
9946         while (!list_empty(&info->caching_block_groups)) {
9947                 caching_ctl = list_entry(info->caching_block_groups.next,
9948                                          struct btrfs_caching_control, list);
9949                 list_del(&caching_ctl->list);
9950                 put_caching_control(caching_ctl);
9951         }
9952         up_write(&info->commit_root_sem);
9953
9954         spin_lock(&info->unused_bgs_lock);
9955         while (!list_empty(&info->unused_bgs)) {
9956                 block_group = list_first_entry(&info->unused_bgs,
9957                                                struct btrfs_block_group_cache,
9958                                                bg_list);
9959                 list_del_init(&block_group->bg_list);
9960                 btrfs_put_block_group(block_group);
9961         }
9962         spin_unlock(&info->unused_bgs_lock);
9963
9964         spin_lock(&info->block_group_cache_lock);
9965         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9966                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9967                                        cache_node);
9968                 rb_erase(&block_group->cache_node,
9969                          &info->block_group_cache_tree);
9970                 RB_CLEAR_NODE(&block_group->cache_node);
9971                 spin_unlock(&info->block_group_cache_lock);
9972
9973                 down_write(&block_group->space_info->groups_sem);
9974                 list_del(&block_group->list);
9975                 up_write(&block_group->space_info->groups_sem);
9976
9977                 if (block_group->cached == BTRFS_CACHE_STARTED)
9978                         wait_block_group_cache_done(block_group);
9979
9980                 /*
9981                  * We haven't cached this block group, which means we could
9982                  * possibly have excluded extents on this block group.
9983                  */
9984                 if (block_group->cached == BTRFS_CACHE_NO ||
9985                     block_group->cached == BTRFS_CACHE_ERROR)
9986                         free_excluded_extents(info->extent_root, block_group);
9987
9988                 btrfs_remove_free_space_cache(block_group);
9989                 ASSERT(list_empty(&block_group->dirty_list));
9990                 ASSERT(list_empty(&block_group->io_list));
9991                 ASSERT(list_empty(&block_group->bg_list));
9992                 ASSERT(atomic_read(&block_group->count) == 1);
9993                 btrfs_put_block_group(block_group);
9994
9995                 spin_lock(&info->block_group_cache_lock);
9996         }
9997         spin_unlock(&info->block_group_cache_lock);
9998
9999         /* now that all the block groups are freed, go through and
10000          * free all the space_info structs.  This is only called during
10001          * the final stages of unmount, and so we know nobody is
10002          * using them.  We call synchronize_rcu() once before we start,
10003          * just to be on the safe side.
10004          */
10005         synchronize_rcu();
10006
10007         release_global_block_rsv(info);
10008
10009         while (!list_empty(&info->space_info)) {
10010                 int i;
10011
10012                 space_info = list_entry(info->space_info.next,
10013                                         struct btrfs_space_info,
10014                                         list);
10015
10016                 /*
10017                  * Do not hide this behind enospc_debug, this is actually
10018                  * important and indicates a real bug if this happens.
10019                  */
10020                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10021                             space_info->bytes_reserved > 0 ||
10022                             space_info->bytes_may_use > 0))
10023                         dump_space_info(space_info, 0, 0);
10024                 list_del(&space_info->list);
10025                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10026                         struct kobject *kobj;
10027                         kobj = space_info->block_group_kobjs[i];
10028                         space_info->block_group_kobjs[i] = NULL;
10029                         if (kobj) {
10030                                 kobject_del(kobj);
10031                                 kobject_put(kobj);
10032                         }
10033                 }
10034                 kobject_del(&space_info->kobj);
10035                 kobject_put(&space_info->kobj);
10036         }
10037         return 0;
10038 }
10039
10040 static void __link_block_group(struct btrfs_space_info *space_info,
10041                                struct btrfs_block_group_cache *cache)
10042 {
10043         int index = get_block_group_index(cache);
10044         bool first = false;
10045
10046         down_write(&space_info->groups_sem);
10047         if (list_empty(&space_info->block_groups[index]))
10048                 first = true;
10049         list_add_tail(&cache->list, &space_info->block_groups[index]);
10050         up_write(&space_info->groups_sem);
10051
10052         if (first) {
10053                 struct raid_kobject *rkobj;
10054                 int ret;
10055
10056                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10057                 if (!rkobj)
10058                         goto out_err;
10059                 rkobj->raid_type = index;
10060                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10061                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10062                                   "%s", get_raid_name(index));
10063                 if (ret) {
10064                         kobject_put(&rkobj->kobj);
10065                         goto out_err;
10066                 }
10067                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10068         }
10069
10070         return;
10071 out_err:
10072         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
10073 }
10074
10075 static struct btrfs_block_group_cache *
10076 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
10077 {
10078         struct btrfs_block_group_cache *cache;
10079
10080         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10081         if (!cache)
10082                 return NULL;
10083
10084         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10085                                         GFP_NOFS);
10086         if (!cache->free_space_ctl) {
10087                 kfree(cache);
10088                 return NULL;
10089         }
10090
10091         cache->key.objectid = start;
10092         cache->key.offset = size;
10093         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10094
10095         cache->sectorsize = root->sectorsize;
10096         cache->fs_info = root->fs_info;
10097         cache->full_stripe_len = btrfs_full_stripe_len(root,
10098                                                &root->fs_info->mapping_tree,
10099                                                start);
10100         set_free_space_tree_thresholds(cache);
10101
10102         atomic_set(&cache->count, 1);
10103         spin_lock_init(&cache->lock);
10104         init_rwsem(&cache->data_rwsem);
10105         INIT_LIST_HEAD(&cache->list);
10106         INIT_LIST_HEAD(&cache->cluster_list);
10107         INIT_LIST_HEAD(&cache->bg_list);
10108         INIT_LIST_HEAD(&cache->ro_list);
10109         INIT_LIST_HEAD(&cache->dirty_list);
10110         INIT_LIST_HEAD(&cache->io_list);
10111         btrfs_init_free_space_ctl(cache);
10112         atomic_set(&cache->trimming, 0);
10113         mutex_init(&cache->free_space_lock);
10114
10115         return cache;
10116 }
10117
10118 int btrfs_read_block_groups(struct btrfs_root *root)
10119 {
10120         struct btrfs_path *path;
10121         int ret;
10122         struct btrfs_block_group_cache *cache;
10123         struct btrfs_fs_info *info = root->fs_info;
10124         struct btrfs_space_info *space_info;
10125         struct btrfs_key key;
10126         struct btrfs_key found_key;
10127         struct extent_buffer *leaf;
10128         int need_clear = 0;
10129         u64 cache_gen;
10130
10131         root = info->extent_root;
10132         key.objectid = 0;
10133         key.offset = 0;
10134         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10135         path = btrfs_alloc_path();
10136         if (!path)
10137                 return -ENOMEM;
10138         path->reada = READA_FORWARD;
10139
10140         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
10141         if (btrfs_test_opt(root->fs_info, SPACE_CACHE) &&
10142             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
10143                 need_clear = 1;
10144         if (btrfs_test_opt(root->fs_info, CLEAR_CACHE))
10145                 need_clear = 1;
10146
10147         while (1) {
10148                 ret = find_first_block_group(root, path, &key);
10149                 if (ret > 0)
10150                         break;
10151                 if (ret != 0)
10152                         goto error;
10153
10154                 leaf = path->nodes[0];
10155                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10156
10157                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
10158                                                        found_key.offset);
10159                 if (!cache) {
10160                         ret = -ENOMEM;
10161                         goto error;
10162                 }
10163
10164                 if (need_clear) {
10165                         /*
10166                          * When we mount with old space cache, we need to
10167                          * set BTRFS_DC_CLEAR and set dirty flag.
10168                          *
10169                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10170                          *    truncate the old free space cache inode and
10171                          *    setup a new one.
10172                          * b) Setting 'dirty flag' makes sure that we flush
10173                          *    the new space cache info onto disk.
10174                          */
10175                         if (btrfs_test_opt(root->fs_info, SPACE_CACHE))
10176                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10177                 }
10178
10179                 read_extent_buffer(leaf, &cache->item,
10180                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10181                                    sizeof(cache->item));
10182                 cache->flags = btrfs_block_group_flags(&cache->item);
10183
10184                 key.objectid = found_key.objectid + found_key.offset;
10185                 btrfs_release_path(path);
10186
10187                 /*
10188                  * We need to exclude the super stripes now so that the space
10189                  * info has super bytes accounted for, otherwise we'll think
10190                  * we have more space than we actually do.
10191                  */
10192                 ret = exclude_super_stripes(root, cache);
10193                 if (ret) {
10194                         /*
10195                          * We may have excluded something, so call this just in
10196                          * case.
10197                          */
10198                         free_excluded_extents(root, cache);
10199                         btrfs_put_block_group(cache);
10200                         goto error;
10201                 }
10202
10203                 /*
10204                  * check for two cases, either we are full, and therefore
10205                  * don't need to bother with the caching work since we won't
10206                  * find any space, or we are empty, and we can just add all
10207                  * the space in and be done with it.  This saves us _alot_ of
10208                  * time, particularly in the full case.
10209                  */
10210                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10211                         cache->last_byte_to_unpin = (u64)-1;
10212                         cache->cached = BTRFS_CACHE_FINISHED;
10213                         free_excluded_extents(root, cache);
10214                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10215                         cache->last_byte_to_unpin = (u64)-1;
10216                         cache->cached = BTRFS_CACHE_FINISHED;
10217                         add_new_free_space(cache, root->fs_info,
10218                                            found_key.objectid,
10219                                            found_key.objectid +
10220                                            found_key.offset);
10221                         free_excluded_extents(root, cache);
10222                 }
10223
10224                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
10225                 if (ret) {
10226                         btrfs_remove_free_space_cache(cache);
10227                         btrfs_put_block_group(cache);
10228                         goto error;
10229                 }
10230
10231                 trace_btrfs_add_block_group(root->fs_info, cache, 0);
10232                 ret = update_space_info(info, cache->flags, found_key.offset,
10233                                         btrfs_block_group_used(&cache->item),
10234                                         cache->bytes_super, &space_info);
10235                 if (ret) {
10236                         btrfs_remove_free_space_cache(cache);
10237                         spin_lock(&info->block_group_cache_lock);
10238                         rb_erase(&cache->cache_node,
10239                                  &info->block_group_cache_tree);
10240                         RB_CLEAR_NODE(&cache->cache_node);
10241                         spin_unlock(&info->block_group_cache_lock);
10242                         btrfs_put_block_group(cache);
10243                         goto error;
10244                 }
10245
10246                 cache->space_info = space_info;
10247
10248                 __link_block_group(space_info, cache);
10249
10250                 set_avail_alloc_bits(root->fs_info, cache->flags);
10251                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
10252                         inc_block_group_ro(cache, 1);
10253                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10254                         spin_lock(&info->unused_bgs_lock);
10255                         /* Should always be true but just in case. */
10256                         if (list_empty(&cache->bg_list)) {
10257                                 btrfs_get_block_group(cache);
10258                                 list_add_tail(&cache->bg_list,
10259                                               &info->unused_bgs);
10260                         }
10261                         spin_unlock(&info->unused_bgs_lock);
10262                 }
10263         }
10264
10265         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
10266                 if (!(get_alloc_profile(root, space_info->flags) &
10267                       (BTRFS_BLOCK_GROUP_RAID10 |
10268                        BTRFS_BLOCK_GROUP_RAID1 |
10269                        BTRFS_BLOCK_GROUP_RAID5 |
10270                        BTRFS_BLOCK_GROUP_RAID6 |
10271                        BTRFS_BLOCK_GROUP_DUP)))
10272                         continue;
10273                 /*
10274                  * avoid allocating from un-mirrored block group if there are
10275                  * mirrored block groups.
10276                  */
10277                 list_for_each_entry(cache,
10278                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10279                                 list)
10280                         inc_block_group_ro(cache, 1);
10281                 list_for_each_entry(cache,
10282                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10283                                 list)
10284                         inc_block_group_ro(cache, 1);
10285         }
10286
10287         init_global_block_rsv(info);
10288         ret = 0;
10289 error:
10290         btrfs_free_path(path);
10291         return ret;
10292 }
10293
10294 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
10295                                        struct btrfs_root *root)
10296 {
10297         struct btrfs_block_group_cache *block_group, *tmp;
10298         struct btrfs_root *extent_root = root->fs_info->extent_root;
10299         struct btrfs_block_group_item item;
10300         struct btrfs_key key;
10301         int ret = 0;
10302         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10303
10304         trans->can_flush_pending_bgs = false;
10305         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10306                 if (ret)
10307                         goto next;
10308
10309                 spin_lock(&block_group->lock);
10310                 memcpy(&item, &block_group->item, sizeof(item));
10311                 memcpy(&key, &block_group->key, sizeof(key));
10312                 spin_unlock(&block_group->lock);
10313
10314                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10315                                         sizeof(item));
10316                 if (ret)
10317                         btrfs_abort_transaction(trans, ret);
10318                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
10319                                                key.objectid, key.offset);
10320                 if (ret)
10321                         btrfs_abort_transaction(trans, ret);
10322                 add_block_group_free_space(trans, root->fs_info, block_group);
10323                 /* already aborted the transaction if it failed. */
10324 next:
10325                 list_del_init(&block_group->bg_list);
10326         }
10327         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10328 }
10329
10330 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10331                            struct btrfs_root *root, u64 bytes_used,
10332                            u64 type, u64 chunk_objectid, u64 chunk_offset,
10333                            u64 size)
10334 {
10335         int ret;
10336         struct btrfs_root *extent_root;
10337         struct btrfs_block_group_cache *cache;
10338         extent_root = root->fs_info->extent_root;
10339
10340         btrfs_set_log_full_commit(root->fs_info, trans);
10341
10342         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
10343         if (!cache)
10344                 return -ENOMEM;
10345
10346         btrfs_set_block_group_used(&cache->item, bytes_used);
10347         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
10348         btrfs_set_block_group_flags(&cache->item, type);
10349
10350         cache->flags = type;
10351         cache->last_byte_to_unpin = (u64)-1;
10352         cache->cached = BTRFS_CACHE_FINISHED;
10353         cache->needs_free_space = 1;
10354         ret = exclude_super_stripes(root, cache);
10355         if (ret) {
10356                 /*
10357                  * We may have excluded something, so call this just in
10358                  * case.
10359                  */
10360                 free_excluded_extents(root, cache);
10361                 btrfs_put_block_group(cache);
10362                 return ret;
10363         }
10364
10365         add_new_free_space(cache, root->fs_info, chunk_offset,
10366                            chunk_offset + size);
10367
10368         free_excluded_extents(root, cache);
10369
10370 #ifdef CONFIG_BTRFS_DEBUG
10371         if (btrfs_should_fragment_free_space(root, cache)) {
10372                 u64 new_bytes_used = size - bytes_used;
10373
10374                 bytes_used += new_bytes_used >> 1;
10375                 fragment_free_space(root, cache);
10376         }
10377 #endif
10378         /*
10379          * Call to ensure the corresponding space_info object is created and
10380          * assigned to our block group, but don't update its counters just yet.
10381          * We want our bg to be added to the rbtree with its ->space_info set.
10382          */
10383         ret = update_space_info(root->fs_info, cache->flags, 0, 0, 0,
10384                                 &cache->space_info);
10385         if (ret) {
10386                 btrfs_remove_free_space_cache(cache);
10387                 btrfs_put_block_group(cache);
10388                 return ret;
10389         }
10390
10391         ret = btrfs_add_block_group_cache(root->fs_info, cache);
10392         if (ret) {
10393                 btrfs_remove_free_space_cache(cache);
10394                 btrfs_put_block_group(cache);
10395                 return ret;
10396         }
10397
10398         /*
10399          * Now that our block group has its ->space_info set and is inserted in
10400          * the rbtree, update the space info's counters.
10401          */
10402         trace_btrfs_add_block_group(root->fs_info, cache, 1);
10403         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
10404                                 cache->bytes_super, &cache->space_info);
10405         if (ret) {
10406                 btrfs_remove_free_space_cache(cache);
10407                 spin_lock(&root->fs_info->block_group_cache_lock);
10408                 rb_erase(&cache->cache_node,
10409                          &root->fs_info->block_group_cache_tree);
10410                 RB_CLEAR_NODE(&cache->cache_node);
10411                 spin_unlock(&root->fs_info->block_group_cache_lock);
10412                 btrfs_put_block_group(cache);
10413                 return ret;
10414         }
10415         update_global_block_rsv(root->fs_info);
10416
10417         __link_block_group(cache->space_info, cache);
10418
10419         list_add_tail(&cache->bg_list, &trans->new_bgs);
10420
10421         set_avail_alloc_bits(extent_root->fs_info, type);
10422         return 0;
10423 }
10424
10425 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10426 {
10427         u64 extra_flags = chunk_to_extended(flags) &
10428                                 BTRFS_EXTENDED_PROFILE_MASK;
10429
10430         write_seqlock(&fs_info->profiles_lock);
10431         if (flags & BTRFS_BLOCK_GROUP_DATA)
10432                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10433         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10434                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10435         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10436                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10437         write_sequnlock(&fs_info->profiles_lock);
10438 }
10439
10440 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10441                              struct btrfs_root *root, u64 group_start,
10442                              struct extent_map *em)
10443 {
10444         struct btrfs_path *path;
10445         struct btrfs_block_group_cache *block_group;
10446         struct btrfs_free_cluster *cluster;
10447         struct btrfs_root *tree_root = root->fs_info->tree_root;
10448         struct btrfs_key key;
10449         struct inode *inode;
10450         struct kobject *kobj = NULL;
10451         int ret;
10452         int index;
10453         int factor;
10454         struct btrfs_caching_control *caching_ctl = NULL;
10455         bool remove_em;
10456
10457         root = root->fs_info->extent_root;
10458
10459         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
10460         BUG_ON(!block_group);
10461         BUG_ON(!block_group->ro);
10462
10463         /*
10464          * Free the reserved super bytes from this block group before
10465          * remove it.
10466          */
10467         free_excluded_extents(root, block_group);
10468
10469         memcpy(&key, &block_group->key, sizeof(key));
10470         index = get_block_group_index(block_group);
10471         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10472                                   BTRFS_BLOCK_GROUP_RAID1 |
10473                                   BTRFS_BLOCK_GROUP_RAID10))
10474                 factor = 2;
10475         else
10476                 factor = 1;
10477
10478         /* make sure this block group isn't part of an allocation cluster */
10479         cluster = &root->fs_info->data_alloc_cluster;
10480         spin_lock(&cluster->refill_lock);
10481         btrfs_return_cluster_to_free_space(block_group, cluster);
10482         spin_unlock(&cluster->refill_lock);
10483
10484         /*
10485          * make sure this block group isn't part of a metadata
10486          * allocation cluster
10487          */
10488         cluster = &root->fs_info->meta_alloc_cluster;
10489         spin_lock(&cluster->refill_lock);
10490         btrfs_return_cluster_to_free_space(block_group, cluster);
10491         spin_unlock(&cluster->refill_lock);
10492
10493         path = btrfs_alloc_path();
10494         if (!path) {
10495                 ret = -ENOMEM;
10496                 goto out;
10497         }
10498
10499         /*
10500          * get the inode first so any iput calls done for the io_list
10501          * aren't the final iput (no unlinks allowed now)
10502          */
10503         inode = lookup_free_space_inode(tree_root, block_group, path);
10504
10505         mutex_lock(&trans->transaction->cache_write_mutex);
10506         /*
10507          * make sure our free spache cache IO is done before remove the
10508          * free space inode
10509          */
10510         spin_lock(&trans->transaction->dirty_bgs_lock);
10511         if (!list_empty(&block_group->io_list)) {
10512                 list_del_init(&block_group->io_list);
10513
10514                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10515
10516                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10517                 btrfs_wait_cache_io(root, trans, block_group,
10518                                     &block_group->io_ctl, path,
10519                                     block_group->key.objectid);
10520                 btrfs_put_block_group(block_group);
10521                 spin_lock(&trans->transaction->dirty_bgs_lock);
10522         }
10523
10524         if (!list_empty(&block_group->dirty_list)) {
10525                 list_del_init(&block_group->dirty_list);
10526                 btrfs_put_block_group(block_group);
10527         }
10528         spin_unlock(&trans->transaction->dirty_bgs_lock);
10529         mutex_unlock(&trans->transaction->cache_write_mutex);
10530
10531         if (!IS_ERR(inode)) {
10532                 ret = btrfs_orphan_add(trans, inode);
10533                 if (ret) {
10534                         btrfs_add_delayed_iput(inode);
10535                         goto out;
10536                 }
10537                 clear_nlink(inode);
10538                 /* One for the block groups ref */
10539                 spin_lock(&block_group->lock);
10540                 if (block_group->iref) {
10541                         block_group->iref = 0;
10542                         block_group->inode = NULL;
10543                         spin_unlock(&block_group->lock);
10544                         iput(inode);
10545                 } else {
10546                         spin_unlock(&block_group->lock);
10547                 }
10548                 /* One for our lookup ref */
10549                 btrfs_add_delayed_iput(inode);
10550         }
10551
10552         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10553         key.offset = block_group->key.objectid;
10554         key.type = 0;
10555
10556         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10557         if (ret < 0)
10558                 goto out;
10559         if (ret > 0)
10560                 btrfs_release_path(path);
10561         if (ret == 0) {
10562                 ret = btrfs_del_item(trans, tree_root, path);
10563                 if (ret)
10564                         goto out;
10565                 btrfs_release_path(path);
10566         }
10567
10568         spin_lock(&root->fs_info->block_group_cache_lock);
10569         rb_erase(&block_group->cache_node,
10570                  &root->fs_info->block_group_cache_tree);
10571         RB_CLEAR_NODE(&block_group->cache_node);
10572
10573         if (root->fs_info->first_logical_byte == block_group->key.objectid)
10574                 root->fs_info->first_logical_byte = (u64)-1;
10575         spin_unlock(&root->fs_info->block_group_cache_lock);
10576
10577         down_write(&block_group->space_info->groups_sem);
10578         /*
10579          * we must use list_del_init so people can check to see if they
10580          * are still on the list after taking the semaphore
10581          */
10582         list_del_init(&block_group->list);
10583         if (list_empty(&block_group->space_info->block_groups[index])) {
10584                 kobj = block_group->space_info->block_group_kobjs[index];
10585                 block_group->space_info->block_group_kobjs[index] = NULL;
10586                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
10587         }
10588         up_write(&block_group->space_info->groups_sem);
10589         if (kobj) {
10590                 kobject_del(kobj);
10591                 kobject_put(kobj);
10592         }
10593
10594         if (block_group->has_caching_ctl)
10595                 caching_ctl = get_caching_control(block_group);
10596         if (block_group->cached == BTRFS_CACHE_STARTED)
10597                 wait_block_group_cache_done(block_group);
10598         if (block_group->has_caching_ctl) {
10599                 down_write(&root->fs_info->commit_root_sem);
10600                 if (!caching_ctl) {
10601                         struct btrfs_caching_control *ctl;
10602
10603                         list_for_each_entry(ctl,
10604                                     &root->fs_info->caching_block_groups, list)
10605                                 if (ctl->block_group == block_group) {
10606                                         caching_ctl = ctl;
10607                                         atomic_inc(&caching_ctl->count);
10608                                         break;
10609                                 }
10610                 }
10611                 if (caching_ctl)
10612                         list_del_init(&caching_ctl->list);
10613                 up_write(&root->fs_info->commit_root_sem);
10614                 if (caching_ctl) {
10615                         /* Once for the caching bgs list and once for us. */
10616                         put_caching_control(caching_ctl);
10617                         put_caching_control(caching_ctl);
10618                 }
10619         }
10620
10621         spin_lock(&trans->transaction->dirty_bgs_lock);
10622         if (!list_empty(&block_group->dirty_list)) {
10623                 WARN_ON(1);
10624         }
10625         if (!list_empty(&block_group->io_list)) {
10626                 WARN_ON(1);
10627         }
10628         spin_unlock(&trans->transaction->dirty_bgs_lock);
10629         btrfs_remove_free_space_cache(block_group);
10630
10631         spin_lock(&block_group->space_info->lock);
10632         list_del_init(&block_group->ro_list);
10633
10634         if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
10635                 WARN_ON(block_group->space_info->total_bytes
10636                         < block_group->key.offset);
10637                 WARN_ON(block_group->space_info->bytes_readonly
10638                         < block_group->key.offset);
10639                 WARN_ON(block_group->space_info->disk_total
10640                         < block_group->key.offset * factor);
10641         }
10642         block_group->space_info->total_bytes -= block_group->key.offset;
10643         block_group->space_info->bytes_readonly -= block_group->key.offset;
10644         block_group->space_info->disk_total -= block_group->key.offset * factor;
10645
10646         spin_unlock(&block_group->space_info->lock);
10647
10648         memcpy(&key, &block_group->key, sizeof(key));
10649
10650         lock_chunks(root);
10651         if (!list_empty(&em->list)) {
10652                 /* We're in the transaction->pending_chunks list. */
10653                 free_extent_map(em);
10654         }
10655         spin_lock(&block_group->lock);
10656         block_group->removed = 1;
10657         /*
10658          * At this point trimming can't start on this block group, because we
10659          * removed the block group from the tree fs_info->block_group_cache_tree
10660          * so no one can't find it anymore and even if someone already got this
10661          * block group before we removed it from the rbtree, they have already
10662          * incremented block_group->trimming - if they didn't, they won't find
10663          * any free space entries because we already removed them all when we
10664          * called btrfs_remove_free_space_cache().
10665          *
10666          * And we must not remove the extent map from the fs_info->mapping_tree
10667          * to prevent the same logical address range and physical device space
10668          * ranges from being reused for a new block group. This is because our
10669          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10670          * completely transactionless, so while it is trimming a range the
10671          * currently running transaction might finish and a new one start,
10672          * allowing for new block groups to be created that can reuse the same
10673          * physical device locations unless we take this special care.
10674          *
10675          * There may also be an implicit trim operation if the file system
10676          * is mounted with -odiscard. The same protections must remain
10677          * in place until the extents have been discarded completely when
10678          * the transaction commit has completed.
10679          */
10680         remove_em = (atomic_read(&block_group->trimming) == 0);
10681         /*
10682          * Make sure a trimmer task always sees the em in the pinned_chunks list
10683          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10684          * before checking block_group->removed).
10685          */
10686         if (!remove_em) {
10687                 /*
10688                  * Our em might be in trans->transaction->pending_chunks which
10689                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10690                  * and so is the fs_info->pinned_chunks list.
10691                  *
10692                  * So at this point we must be holding the chunk_mutex to avoid
10693                  * any races with chunk allocation (more specifically at
10694                  * volumes.c:contains_pending_extent()), to ensure it always
10695                  * sees the em, either in the pending_chunks list or in the
10696                  * pinned_chunks list.
10697                  */
10698                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10699         }
10700         spin_unlock(&block_group->lock);
10701
10702         if (remove_em) {
10703                 struct extent_map_tree *em_tree;
10704
10705                 em_tree = &root->fs_info->mapping_tree.map_tree;
10706                 write_lock(&em_tree->lock);
10707                 /*
10708                  * The em might be in the pending_chunks list, so make sure the
10709                  * chunk mutex is locked, since remove_extent_mapping() will
10710                  * delete us from that list.
10711                  */
10712                 remove_extent_mapping(em_tree, em);
10713                 write_unlock(&em_tree->lock);
10714                 /* once for the tree */
10715                 free_extent_map(em);
10716         }
10717
10718         unlock_chunks(root);
10719
10720         ret = remove_block_group_free_space(trans, root->fs_info, block_group);
10721         if (ret)
10722                 goto out;
10723
10724         btrfs_put_block_group(block_group);
10725         btrfs_put_block_group(block_group);
10726
10727         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10728         if (ret > 0)
10729                 ret = -EIO;
10730         if (ret < 0)
10731                 goto out;
10732
10733         ret = btrfs_del_item(trans, root, path);
10734 out:
10735         btrfs_free_path(path);
10736         return ret;
10737 }
10738
10739 struct btrfs_trans_handle *
10740 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10741                                      const u64 chunk_offset)
10742 {
10743         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10744         struct extent_map *em;
10745         struct map_lookup *map;
10746         unsigned int num_items;
10747
10748         read_lock(&em_tree->lock);
10749         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10750         read_unlock(&em_tree->lock);
10751         ASSERT(em && em->start == chunk_offset);
10752
10753         /*
10754          * We need to reserve 3 + N units from the metadata space info in order
10755          * to remove a block group (done at btrfs_remove_chunk() and at
10756          * btrfs_remove_block_group()), which are used for:
10757          *
10758          * 1 unit for adding the free space inode's orphan (located in the tree
10759          * of tree roots).
10760          * 1 unit for deleting the block group item (located in the extent
10761          * tree).
10762          * 1 unit for deleting the free space item (located in tree of tree
10763          * roots).
10764          * N units for deleting N device extent items corresponding to each
10765          * stripe (located in the device tree).
10766          *
10767          * In order to remove a block group we also need to reserve units in the
10768          * system space info in order to update the chunk tree (update one or
10769          * more device items and remove one chunk item), but this is done at
10770          * btrfs_remove_chunk() through a call to check_system_chunk().
10771          */
10772         map = em->map_lookup;
10773         num_items = 3 + map->num_stripes;
10774         free_extent_map(em);
10775
10776         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10777                                                            num_items, 1);
10778 }
10779
10780 /*
10781  * Process the unused_bgs list and remove any that don't have any allocated
10782  * space inside of them.
10783  */
10784 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10785 {
10786         struct btrfs_block_group_cache *block_group;
10787         struct btrfs_space_info *space_info;
10788         struct btrfs_root *root = fs_info->extent_root;
10789         struct btrfs_trans_handle *trans;
10790         int ret = 0;
10791
10792         if (!fs_info->open)
10793                 return;
10794
10795         spin_lock(&fs_info->unused_bgs_lock);
10796         while (!list_empty(&fs_info->unused_bgs)) {
10797                 u64 start, end;
10798                 int trimming;
10799
10800                 block_group = list_first_entry(&fs_info->unused_bgs,
10801                                                struct btrfs_block_group_cache,
10802                                                bg_list);
10803                 list_del_init(&block_group->bg_list);
10804
10805                 space_info = block_group->space_info;
10806
10807                 if (ret || btrfs_mixed_space_info(space_info)) {
10808                         btrfs_put_block_group(block_group);
10809                         continue;
10810                 }
10811                 spin_unlock(&fs_info->unused_bgs_lock);
10812
10813                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10814
10815                 /* Don't want to race with allocators so take the groups_sem */
10816                 down_write(&space_info->groups_sem);
10817                 spin_lock(&block_group->lock);
10818                 if (block_group->reserved ||
10819                     btrfs_block_group_used(&block_group->item) ||
10820                     block_group->ro ||
10821                     list_is_singular(&block_group->list)) {
10822                         /*
10823                          * We want to bail if we made new allocations or have
10824                          * outstanding allocations in this block group.  We do
10825                          * the ro check in case balance is currently acting on
10826                          * this block group.
10827                          */
10828                         spin_unlock(&block_group->lock);
10829                         up_write(&space_info->groups_sem);
10830                         goto next;
10831                 }
10832                 spin_unlock(&block_group->lock);
10833
10834                 /* We don't want to force the issue, only flip if it's ok. */
10835                 ret = inc_block_group_ro(block_group, 0);
10836                 up_write(&space_info->groups_sem);
10837                 if (ret < 0) {
10838                         ret = 0;
10839                         goto next;
10840                 }
10841
10842                 /*
10843                  * Want to do this before we do anything else so we can recover
10844                  * properly if we fail to join the transaction.
10845                  */
10846                 trans = btrfs_start_trans_remove_block_group(fs_info,
10847                                                      block_group->key.objectid);
10848                 if (IS_ERR(trans)) {
10849                         btrfs_dec_block_group_ro(root, block_group);
10850                         ret = PTR_ERR(trans);
10851                         goto next;
10852                 }
10853
10854                 /*
10855                  * We could have pending pinned extents for this block group,
10856                  * just delete them, we don't care about them anymore.
10857                  */
10858                 start = block_group->key.objectid;
10859                 end = start + block_group->key.offset - 1;
10860                 /*
10861                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10862                  * btrfs_finish_extent_commit(). If we are at transaction N,
10863                  * another task might be running finish_extent_commit() for the
10864                  * previous transaction N - 1, and have seen a range belonging
10865                  * to the block group in freed_extents[] before we were able to
10866                  * clear the whole block group range from freed_extents[]. This
10867                  * means that task can lookup for the block group after we
10868                  * unpinned it from freed_extents[] and removed it, leading to
10869                  * a BUG_ON() at btrfs_unpin_extent_range().
10870                  */
10871                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10872                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10873                                   EXTENT_DIRTY);
10874                 if (ret) {
10875                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10876                         btrfs_dec_block_group_ro(root, block_group);
10877                         goto end_trans;
10878                 }
10879                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10880                                   EXTENT_DIRTY);
10881                 if (ret) {
10882                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10883                         btrfs_dec_block_group_ro(root, block_group);
10884                         goto end_trans;
10885                 }
10886                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10887
10888                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10889                 spin_lock(&space_info->lock);
10890                 spin_lock(&block_group->lock);
10891
10892                 space_info->bytes_pinned -= block_group->pinned;
10893                 space_info->bytes_readonly += block_group->pinned;
10894                 percpu_counter_add(&space_info->total_bytes_pinned,
10895                                    -block_group->pinned);
10896                 block_group->pinned = 0;
10897
10898                 spin_unlock(&block_group->lock);
10899                 spin_unlock(&space_info->lock);
10900
10901                 /* DISCARD can flip during remount */
10902                 trimming = btrfs_test_opt(root->fs_info, DISCARD);
10903
10904                 /* Implicit trim during transaction commit. */
10905                 if (trimming)
10906                         btrfs_get_block_group_trimming(block_group);
10907
10908                 /*
10909                  * Btrfs_remove_chunk will abort the transaction if things go
10910                  * horribly wrong.
10911                  */
10912                 ret = btrfs_remove_chunk(trans, root,
10913                                          block_group->key.objectid);
10914
10915                 if (ret) {
10916                         if (trimming)
10917                                 btrfs_put_block_group_trimming(block_group);
10918                         goto end_trans;
10919                 }
10920
10921                 /*
10922                  * If we're not mounted with -odiscard, we can just forget
10923                  * about this block group. Otherwise we'll need to wait
10924                  * until transaction commit to do the actual discard.
10925                  */
10926                 if (trimming) {
10927                         spin_lock(&fs_info->unused_bgs_lock);
10928                         /*
10929                          * A concurrent scrub might have added us to the list
10930                          * fs_info->unused_bgs, so use a list_move operation
10931                          * to add the block group to the deleted_bgs list.
10932                          */
10933                         list_move(&block_group->bg_list,
10934                                   &trans->transaction->deleted_bgs);
10935                         spin_unlock(&fs_info->unused_bgs_lock);
10936                         btrfs_get_block_group(block_group);
10937                 }
10938 end_trans:
10939                 btrfs_end_transaction(trans, root);
10940 next:
10941                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10942                 btrfs_put_block_group(block_group);
10943                 spin_lock(&fs_info->unused_bgs_lock);
10944         }
10945         spin_unlock(&fs_info->unused_bgs_lock);
10946 }
10947
10948 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10949 {
10950         struct btrfs_space_info *space_info;
10951         struct btrfs_super_block *disk_super;
10952         u64 features;
10953         u64 flags;
10954         int mixed = 0;
10955         int ret;
10956
10957         disk_super = fs_info->super_copy;
10958         if (!btrfs_super_root(disk_super))
10959                 return -EINVAL;
10960
10961         features = btrfs_super_incompat_flags(disk_super);
10962         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10963                 mixed = 1;
10964
10965         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10966         ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10967         if (ret)
10968                 goto out;
10969
10970         if (mixed) {
10971                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10972                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10973         } else {
10974                 flags = BTRFS_BLOCK_GROUP_METADATA;
10975                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10976                 if (ret)
10977                         goto out;
10978
10979                 flags = BTRFS_BLOCK_GROUP_DATA;
10980                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10981         }
10982 out:
10983         return ret;
10984 }
10985
10986 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10987 {
10988         return unpin_extent_range(root, start, end, false);
10989 }
10990
10991 /*
10992  * It used to be that old block groups would be left around forever.
10993  * Iterating over them would be enough to trim unused space.  Since we
10994  * now automatically remove them, we also need to iterate over unallocated
10995  * space.
10996  *
10997  * We don't want a transaction for this since the discard may take a
10998  * substantial amount of time.  We don't require that a transaction be
10999  * running, but we do need to take a running transaction into account
11000  * to ensure that we're not discarding chunks that were released in
11001  * the current transaction.
11002  *
11003  * Holding the chunks lock will prevent other threads from allocating
11004  * or releasing chunks, but it won't prevent a running transaction
11005  * from committing and releasing the memory that the pending chunks
11006  * list head uses.  For that, we need to take a reference to the
11007  * transaction.
11008  */
11009 static int btrfs_trim_free_extents(struct btrfs_device *device,
11010                                    u64 minlen, u64 *trimmed)
11011 {
11012         u64 start = 0, len = 0;
11013         int ret;
11014
11015         *trimmed = 0;
11016
11017         /* Not writeable = nothing to do. */
11018         if (!device->writeable)
11019                 return 0;
11020
11021         /* No free space = nothing to do. */
11022         if (device->total_bytes <= device->bytes_used)
11023                 return 0;
11024
11025         ret = 0;
11026
11027         while (1) {
11028                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
11029                 struct btrfs_transaction *trans;
11030                 u64 bytes;
11031
11032                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11033                 if (ret)
11034                         return ret;
11035
11036                 down_read(&fs_info->commit_root_sem);
11037
11038                 spin_lock(&fs_info->trans_lock);
11039                 trans = fs_info->running_transaction;
11040                 if (trans)
11041                         atomic_inc(&trans->use_count);
11042                 spin_unlock(&fs_info->trans_lock);
11043
11044                 ret = find_free_dev_extent_start(trans, device, minlen, start,
11045                                                  &start, &len);
11046                 if (trans)
11047                         btrfs_put_transaction(trans);
11048
11049                 if (ret) {
11050                         up_read(&fs_info->commit_root_sem);
11051                         mutex_unlock(&fs_info->chunk_mutex);
11052                         if (ret == -ENOSPC)
11053                                 ret = 0;
11054                         break;
11055                 }
11056
11057                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
11058                 up_read(&fs_info->commit_root_sem);
11059                 mutex_unlock(&fs_info->chunk_mutex);
11060
11061                 if (ret)
11062                         break;
11063
11064                 start += len;
11065                 *trimmed += bytes;
11066
11067                 if (fatal_signal_pending(current)) {
11068                         ret = -ERESTARTSYS;
11069                         break;
11070                 }
11071
11072                 cond_resched();
11073         }
11074
11075         return ret;
11076 }
11077
11078 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
11079 {
11080         struct btrfs_fs_info *fs_info = root->fs_info;
11081         struct btrfs_block_group_cache *cache = NULL;
11082         struct btrfs_device *device;
11083         struct list_head *devices;
11084         u64 group_trimmed;
11085         u64 start;
11086         u64 end;
11087         u64 trimmed = 0;
11088         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
11089         int ret = 0;
11090
11091         /*
11092          * try to trim all FS space, our block group may start from non-zero.
11093          */
11094         if (range->len == total_bytes)
11095                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
11096         else
11097                 cache = btrfs_lookup_block_group(fs_info, range->start);
11098
11099         while (cache) {
11100                 if (cache->key.objectid >= (range->start + range->len)) {
11101                         btrfs_put_block_group(cache);
11102                         break;
11103                 }
11104
11105                 start = max(range->start, cache->key.objectid);
11106                 end = min(range->start + range->len,
11107                                 cache->key.objectid + cache->key.offset);
11108
11109                 if (end - start >= range->minlen) {
11110                         if (!block_group_cache_done(cache)) {
11111                                 ret = cache_block_group(cache, 0);
11112                                 if (ret) {
11113                                         btrfs_put_block_group(cache);
11114                                         break;
11115                                 }
11116                                 ret = wait_block_group_cache_done(cache);
11117                                 if (ret) {
11118                                         btrfs_put_block_group(cache);
11119                                         break;
11120                                 }
11121                         }
11122                         ret = btrfs_trim_block_group(cache,
11123                                                      &group_trimmed,
11124                                                      start,
11125                                                      end,
11126                                                      range->minlen);
11127
11128                         trimmed += group_trimmed;
11129                         if (ret) {
11130                                 btrfs_put_block_group(cache);
11131                                 break;
11132                         }
11133                 }
11134
11135                 cache = next_block_group(fs_info->tree_root, cache);
11136         }
11137
11138         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
11139         devices = &root->fs_info->fs_devices->alloc_list;
11140         list_for_each_entry(device, devices, dev_alloc_list) {
11141                 ret = btrfs_trim_free_extents(device, range->minlen,
11142                                               &group_trimmed);
11143                 if (ret)
11144                         break;
11145
11146                 trimmed += group_trimmed;
11147         }
11148         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
11149
11150         range->len = trimmed;
11151         return ret;
11152 }
11153
11154 /*
11155  * btrfs_{start,end}_write_no_snapshoting() are similar to
11156  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11157  * data into the page cache through nocow before the subvolume is snapshoted,
11158  * but flush the data into disk after the snapshot creation, or to prevent
11159  * operations while snapshoting is ongoing and that cause the snapshot to be
11160  * inconsistent (writes followed by expanding truncates for example).
11161  */
11162 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
11163 {
11164         percpu_counter_dec(&root->subv_writers->counter);
11165         /*
11166          * Make sure counter is updated before we wake up waiters.
11167          */
11168         smp_mb();
11169         if (waitqueue_active(&root->subv_writers->wait))
11170                 wake_up(&root->subv_writers->wait);
11171 }
11172
11173 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
11174 {
11175         if (atomic_read(&root->will_be_snapshoted))
11176                 return 0;
11177
11178         percpu_counter_inc(&root->subv_writers->counter);
11179         /*
11180          * Make sure counter is updated before we check for snapshot creation.
11181          */
11182         smp_mb();
11183         if (atomic_read(&root->will_be_snapshoted)) {
11184                 btrfs_end_write_no_snapshoting(root);
11185                 return 0;
11186         }
11187         return 1;
11188 }
11189
11190 static int wait_snapshoting_atomic_t(atomic_t *a)
11191 {
11192         schedule();
11193         return 0;
11194 }
11195
11196 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11197 {
11198         while (true) {
11199                 int ret;
11200
11201                 ret = btrfs_start_write_no_snapshoting(root);
11202                 if (ret)
11203                         break;
11204                 wait_on_atomic_t(&root->will_be_snapshoted,
11205                                  wait_snapshoting_atomic_t,
11206                                  TASK_UNINTERRUPTIBLE);
11207         }
11208 }