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_fs_info *fs_info,
  91                             struct btrfs_space_info *info, u64 bytes,
  92                             int dump_block_groups);
  93 static int btrfs_add_reserved_bytes(struct btrfs_block_group_cache *cache,
  94                                     u64 ram_bytes, u64 num_bytes, int delalloc);
  95 static int btrfs_free_reserved_bytes(struct btrfs_block_group_cache *cache,
  96                                      u64 num_bytes, int delalloc);
  97 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
  98                                u64 num_bytes);
  99 int btrfs_pin_extent(struct btrfs_root *root,
 100                      u64 bytenr, u64 num_bytes, int reserved);
 101 static int __reserve_metadata_bytes(struct btrfs_root *root,
 102                                     struct btrfs_space_info *space_info,
 103                                     u64 orig_bytes,
 104                                     enum btrfs_reserve_flush_enum flush);
 105 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
 106                                      struct btrfs_space_info *space_info,
 107                                      u64 num_bytes);
 108 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 109                                      struct btrfs_space_info *space_info,
 110                                      u64 num_bytes);
 111
 112 static noinline int
 113 block_group_cache_done(struct btrfs_block_group_cache *cache)
 114 {
 115         smp_mb();
 116         return cache->cached == BTRFS_CACHE_FINISHED ||
 117                 cache->cached == BTRFS_CACHE_ERROR;
 118 }
 119
 120 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 121 {
 122         return (cache->flags & bits) == bits;
 123 }
 124
 125 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 126 {
 127         atomic_inc(&cache->count);
 128 }
 129
 130 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 131 {
 132         if (atomic_dec_and_test(&cache->count)) {
 133                 WARN_ON(cache->pinned > 0);
 134                 WARN_ON(cache->reserved > 0);
 135                 kfree(cache->free_space_ctl);
 136                 kfree(cache);
 137         }
 138 }
 139
 140 /*
 141  * this adds the block group to the fs_info rb tree for the block group
 142  * cache
 143  */
 144 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 145                                 struct btrfs_block_group_cache *block_group)
 146 {
 147         struct rb_node **p;
 148         struct rb_node *parent = NULL;
 149         struct btrfs_block_group_cache *cache;
 150
 151         spin_lock(&info->block_group_cache_lock);
 152         p = &info->block_group_cache_tree.rb_node;
 153
 154         while (*p) {
 155                 parent = *p;
 156                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 157                                  cache_node);
 158                 if (block_group->key.objectid < cache->key.objectid) {
 159                         p = &(*p)->rb_left;
 160                 } else if (block_group->key.objectid > cache->key.objectid) {
 161                         p = &(*p)->rb_right;
 162                 } else {
 163                         spin_unlock(&info->block_group_cache_lock);
 164                         return -EEXIST;
 165                 }
 166         }
 167
 168         rb_link_node(&block_group->cache_node, parent, p);
 169         rb_insert_color(&block_group->cache_node,
 170                         &info->block_group_cache_tree);
 171
 172         if (info->first_logical_byte > block_group->key.objectid)
 173                 info->first_logical_byte = block_group->key.objectid;
 174
 175         spin_unlock(&info->block_group_cache_lock);
 176
 177         return 0;
 178 }
 179
 180 /*
 181  * This will return the block group at or after bytenr if contains is 0, else
 182  * it will return the block group that contains the bytenr
 183  */
 184 static struct btrfs_block_group_cache *
 185 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 186                               int contains)
 187 {
 188         struct btrfs_block_group_cache *cache, *ret = NULL;
 189         struct rb_node *n;
 190         u64 end, start;
 191
 192         spin_lock(&info->block_group_cache_lock);
 193         n = info->block_group_cache_tree.rb_node;
 194
 195         while (n) {
 196                 cache = rb_entry(n, struct btrfs_block_group_cache,
 197                                  cache_node);
 198                 end = cache->key.objectid + cache->key.offset - 1;
 199                 start = cache->key.objectid;
 200
 201                 if (bytenr < start) {
 202                         if (!contains && (!ret || start < ret->key.objectid))
 203                                 ret = cache;
 204                         n = n->rb_left;
 205                 } else if (bytenr > start) {
 206                         if (contains && bytenr <= end) {
 207                                 ret = cache;
 208                                 break;
 209                         }
 210                         n = n->rb_right;
 211                 } else {
 212                         ret = cache;
 213                         break;
 214                 }
 215         }
 216         if (ret) {
 217                 btrfs_get_block_group(ret);
 218                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 219                         info->first_logical_byte = ret->key.objectid;
 220         }
 221         spin_unlock(&info->block_group_cache_lock);
 222
 223         return ret;
 224 }
 225
 226 static int add_excluded_extent(struct btrfs_root *root,
 227                                u64 start, u64 num_bytes)
 228 {
 229         u64 end = start + num_bytes - 1;
 230         set_extent_bits(&root->fs_info->freed_extents[0],
 231                         start, end, EXTENT_UPTODATE);
 232         set_extent_bits(&root->fs_info->freed_extents[1],
 233                         start, end, EXTENT_UPTODATE);
 234         return 0;
 235 }
 236
 237 static void free_excluded_extents(struct btrfs_root *root,
 238                                   struct btrfs_block_group_cache *cache)
 239 {
 240         u64 start, end;
 241
 242         start = cache->key.objectid;
 243         end = start + cache->key.offset - 1;
 244
 245         clear_extent_bits(&root->fs_info->freed_extents[0],
 246                           start, end, EXTENT_UPTODATE);
 247         clear_extent_bits(&root->fs_info->freed_extents[1],
 248                           start, end, EXTENT_UPTODATE);
 249 }
 250
 251 static int exclude_super_stripes(struct btrfs_root *root,
 252                                  struct btrfs_block_group_cache *cache)
 253 {
 254         u64 bytenr;
 255         u64 *logical;
 256         int stripe_len;
 257         int i, nr, ret;
 258
 259         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 260                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 261                 cache->bytes_super += stripe_len;
 262                 ret = add_excluded_extent(root, cache->key.objectid,
 263                                           stripe_len);
 264                 if (ret)
 265                         return ret;
 266         }
 267
 268         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 269                 bytenr = btrfs_sb_offset(i);
 270                 ret = btrfs_rmap_block(root->fs_info, cache->key.objectid,
 271                                        bytenr, 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         return block_group_cache_tree_search(info, bytenr, 0);
 734 }
 735
 736 /*
 737  * return the block group that contains the given bytenr
 738  */
 739 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 740                                                  struct btrfs_fs_info *info,
 741                                                  u64 bytenr)
 742 {
 743         return block_group_cache_tree_search(info, bytenr, 1);
 744 }
 745
 746 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 747                                                   u64 flags)
 748 {
 749         struct list_head *head = &info->space_info;
 750         struct btrfs_space_info *found;
 751
 752         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 753
 754         rcu_read_lock();
 755         list_for_each_entry_rcu(found, head, list) {
 756                 if (found->flags & flags) {
 757                         rcu_read_unlock();
 758                         return found;
 759                 }
 760         }
 761         rcu_read_unlock();
 762         return NULL;
 763 }
 764
 765 /*
 766  * after adding space to the filesystem, we need to clear the full flags
 767  * on all the space infos.
 768  */
 769 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 770 {
 771         struct list_head *head = &info->space_info;
 772         struct btrfs_space_info *found;
 773
 774         rcu_read_lock();
 775         list_for_each_entry_rcu(found, head, list)
 776                 found->full = 0;
 777         rcu_read_unlock();
 778 }
 779
 780 /* simple helper to search for an existing data extent at a given offset */
 781 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 782 {
 783         int ret;
 784         struct btrfs_key key;
 785         struct btrfs_path *path;
 786
 787         path = btrfs_alloc_path();
 788         if (!path)
 789                 return -ENOMEM;
 790
 791         key.objectid = start;
 792         key.offset = len;
 793         key.type = BTRFS_EXTENT_ITEM_KEY;
 794         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 795                                 0, 0);
 796         btrfs_free_path(path);
 797         return ret;
 798 }
 799
 800 /*
 801  * helper function to lookup reference count and flags of a tree block.
 802  *
 803  * the head node for delayed ref is used to store the sum of all the
 804  * reference count modifications queued up in the rbtree. the head
 805  * node may also store the extent flags to set. This way you can check
 806  * to see what the reference count and extent flags would be if all of
 807  * the delayed refs are not processed.
 808  */
 809 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 810                              struct btrfs_root *root, u64 bytenr,
 811                              u64 offset, int metadata, u64 *refs, u64 *flags)
 812 {
 813         struct btrfs_delayed_ref_head *head;
 814         struct btrfs_delayed_ref_root *delayed_refs;
 815         struct btrfs_path *path;
 816         struct btrfs_extent_item *ei;
 817         struct extent_buffer *leaf;
 818         struct btrfs_key key;
 819         u32 item_size;
 820         u64 num_refs;
 821         u64 extent_flags;
 822         int ret;
 823
 824         /*
 825          * If we don't have skinny metadata, don't bother doing anything
 826          * different
 827          */
 828         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 829                 offset = root->nodesize;
 830                 metadata = 0;
 831         }
 832
 833         path = btrfs_alloc_path();
 834         if (!path)
 835                 return -ENOMEM;
 836
 837         if (!trans) {
 838                 path->skip_locking = 1;
 839                 path->search_commit_root = 1;
 840         }
 841
 842 search_again:
 843         key.objectid = bytenr;
 844         key.offset = offset;
 845         if (metadata)
 846                 key.type = BTRFS_METADATA_ITEM_KEY;
 847         else
 848                 key.type = BTRFS_EXTENT_ITEM_KEY;
 849
 850         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 851                                 &key, path, 0, 0);
 852         if (ret < 0)
 853                 goto out_free;
 854
 855         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 856                 if (path->slots[0]) {
 857                         path->slots[0]--;
 858                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 859                                               path->slots[0]);
 860                         if (key.objectid == bytenr &&
 861                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 862                             key.offset == root->nodesize)
 863                                 ret = 0;
 864                 }
 865         }
 866
 867         if (ret == 0) {
 868                 leaf = path->nodes[0];
 869                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 870                 if (item_size >= sizeof(*ei)) {
 871                         ei = btrfs_item_ptr(leaf, path->slots[0],
 872                                             struct btrfs_extent_item);
 873                         num_refs = btrfs_extent_refs(leaf, ei);
 874                         extent_flags = btrfs_extent_flags(leaf, ei);
 875                 } else {
 876 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 877                         struct btrfs_extent_item_v0 *ei0;
 878                         BUG_ON(item_size != sizeof(*ei0));
 879                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 880                                              struct btrfs_extent_item_v0);
 881                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 882                         /* FIXME: this isn't correct for data */
 883                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 884 #else
 885                         BUG();
 886 #endif
 887                 }
 888                 BUG_ON(num_refs == 0);
 889         } else {
 890                 num_refs = 0;
 891                 extent_flags = 0;
 892                 ret = 0;
 893         }
 894
 895         if (!trans)
 896                 goto out;
 897
 898         delayed_refs = &trans->transaction->delayed_refs;
 899         spin_lock(&delayed_refs->lock);
 900         head = btrfs_find_delayed_ref_head(trans, bytenr);
 901         if (head) {
 902                 if (!mutex_trylock(&head->mutex)) {
 903                         atomic_inc(&head->node.refs);
 904                         spin_unlock(&delayed_refs->lock);
 905
 906                         btrfs_release_path(path);
 907
 908                         /*
 909                          * Mutex was contended, block until it's released and try
 910                          * again
 911                          */
 912                         mutex_lock(&head->mutex);
 913                         mutex_unlock(&head->mutex);
 914                         btrfs_put_delayed_ref(&head->node);
 915                         goto search_again;
 916                 }
 917                 spin_lock(&head->lock);
 918                 if (head->extent_op && head->extent_op->update_flags)
 919                         extent_flags |= head->extent_op->flags_to_set;
 920                 else
 921                         BUG_ON(num_refs == 0);
 922
 923                 num_refs += head->node.ref_mod;
 924                 spin_unlock(&head->lock);
 925                 mutex_unlock(&head->mutex);
 926         }
 927         spin_unlock(&delayed_refs->lock);
 928 out:
 929         WARN_ON(num_refs == 0);
 930         if (refs)
 931                 *refs = num_refs;
 932         if (flags)
 933                 *flags = extent_flags;
 934 out_free:
 935         btrfs_free_path(path);
 936         return ret;
 937 }
 938
 939 /*
 940  * Back reference rules.  Back refs have three main goals:
 941  *
 942  * 1) differentiate between all holders of references to an extent so that
 943  *    when a reference is dropped we can make sure it was a valid reference
 944  *    before freeing the extent.
 945  *
 946  * 2) Provide enough information to quickly find the holders of an extent
 947  *    if we notice a given block is corrupted or bad.
 948  *
 949  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 950  *    maintenance.  This is actually the same as #2, but with a slightly
 951  *    different use case.
 952  *
 953  * There are two kinds of back refs. The implicit back refs is optimized
 954  * for pointers in non-shared tree blocks. For a given pointer in a block,
 955  * back refs of this kind provide information about the block's owner tree
 956  * and the pointer's key. These information allow us to find the block by
 957  * b-tree searching. The full back refs is for pointers in tree blocks not
 958  * referenced by their owner trees. The location of tree block is recorded
 959  * in the back refs. Actually the full back refs is generic, and can be
 960  * used in all cases the implicit back refs is used. The major shortcoming
 961  * of the full back refs is its overhead. Every time a tree block gets
 962  * COWed, we have to update back refs entry for all pointers in it.
 963  *
 964  * For a newly allocated tree block, we use implicit back refs for
 965  * pointers in it. This means most tree related operations only involve
 966  * implicit back refs. For a tree block created in old transaction, the
 967  * only way to drop a reference to it is COW it. So we can detect the
 968  * event that tree block loses its owner tree's reference and do the
 969  * back refs conversion.
 970  *
 971  * When a tree block is COWed through a tree, there are four cases:
 972  *
 973  * The reference count of the block is one and the tree is the block's
 974  * owner tree. Nothing to do in this case.
 975  *
 976  * The reference count of the block is one and the tree is not the
 977  * block's owner tree. In this case, full back refs is used for pointers
 978  * in the block. Remove these full back refs, add implicit back refs for
 979  * every pointers in the new block.
 980  *
 981  * The reference count of the block is greater than one and the tree is
 982  * the block's owner tree. In this case, implicit back refs is used for
 983  * pointers in the block. Add full back refs for every pointers in the
 984  * block, increase lower level extents' reference counts. The original
 985  * implicit back refs are entailed to the new block.
 986  *
 987  * The reference count of the block is greater than one and the tree is
 988  * not the block's owner tree. Add implicit back refs for every pointer in
 989  * the new block, increase lower level extents' reference count.
 990  *
 991  * Back Reference Key composing:
 992  *
 993  * The key objectid corresponds to the first byte in the extent,
 994  * The key type is used to differentiate between types of back refs.
 995  * There are different meanings of the key offset for different types
 996  * of back refs.
 997  *
 998  * File extents can be referenced by:
 999  *
1000  * - multiple snapshots, subvolumes, or different generations in one subvol
1001  * - different files inside a single subvolume
1002  * - different offsets inside a file (bookend extents in file.c)
1003  *
1004  * The extent ref structure for the implicit back refs has fields for:
1005  *
1006  * - Objectid of the subvolume root
1007  * - objectid of the file holding the reference
1008  * - original offset in the file
1009  * - how many bookend extents
1010  *
1011  * The key offset for the implicit back refs is hash of the first
1012  * three fields.
1013  *
1014  * The extent ref structure for the full back refs has field for:
1015  *
1016  * - number of pointers in the tree leaf
1017  *
1018  * The key offset for the implicit back refs is the first byte of
1019  * the tree leaf
1020  *
1021  * When a file extent is allocated, The implicit back refs is used.
1022  * the fields are filled in:
1023  *
1024  *     (root_key.objectid, inode objectid, offset in file, 1)
1025  *
1026  * When a file extent is removed file truncation, we find the
1027  * corresponding implicit back refs and check the following fields:
1028  *
1029  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1030  *
1031  * Btree extents can be referenced by:
1032  *
1033  * - Different subvolumes
1034  *
1035  * Both the implicit back refs and the full back refs for tree blocks
1036  * only consist of key. The key offset for the implicit back refs is
1037  * objectid of block's owner tree. The key offset for the full back refs
1038  * is the first byte of parent block.
1039  *
1040  * When implicit back refs is used, information about the lowest key and
1041  * level of the tree block are required. These information are stored in
1042  * tree block info structure.
1043  */
1044
1045 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1046 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1047                                   struct btrfs_root *root,
1048                                   struct btrfs_path *path,
1049                                   u64 owner, u32 extra_size)
1050 {
1051         struct btrfs_extent_item *item;
1052         struct btrfs_extent_item_v0 *ei0;
1053         struct btrfs_extent_ref_v0 *ref0;
1054         struct btrfs_tree_block_info *bi;
1055         struct extent_buffer *leaf;
1056         struct btrfs_key key;
1057         struct btrfs_key found_key;
1058         u32 new_size = sizeof(*item);
1059         u64 refs;
1060         int ret;
1061
1062         leaf = path->nodes[0];
1063         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1064
1065         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1066         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1067                              struct btrfs_extent_item_v0);
1068         refs = btrfs_extent_refs_v0(leaf, ei0);
1069
1070         if (owner == (u64)-1) {
1071                 while (1) {
1072                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1073                                 ret = btrfs_next_leaf(root, path);
1074                                 if (ret < 0)
1075                                         return ret;
1076                                 BUG_ON(ret > 0); /* Corruption */
1077                                 leaf = path->nodes[0];
1078                         }
1079                         btrfs_item_key_to_cpu(leaf, &found_key,
1080                                               path->slots[0]);
1081                         BUG_ON(key.objectid != found_key.objectid);
1082                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1083                                 path->slots[0]++;
1084                                 continue;
1085                         }
1086                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1087                                               struct btrfs_extent_ref_v0);
1088                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1089                         break;
1090                 }
1091         }
1092         btrfs_release_path(path);
1093
1094         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1095                 new_size += sizeof(*bi);
1096
1097         new_size -= sizeof(*ei0);
1098         ret = btrfs_search_slot(trans, root, &key, path,
1099                                 new_size + extra_size, 1);
1100         if (ret < 0)
1101                 return ret;
1102         BUG_ON(ret); /* Corruption */
1103
1104         btrfs_extend_item(root, path, new_size);
1105
1106         leaf = path->nodes[0];
1107         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1108         btrfs_set_extent_refs(leaf, item, refs);
1109         /* FIXME: get real generation */
1110         btrfs_set_extent_generation(leaf, item, 0);
1111         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1112                 btrfs_set_extent_flags(leaf, item,
1113                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1114                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1115                 bi = (struct btrfs_tree_block_info *)(item + 1);
1116                 /* FIXME: get first key of the block */
1117                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1118                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1119         } else {
1120                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1121         }
1122         btrfs_mark_buffer_dirty(leaf);
1123         return 0;
1124 }
1125 #endif
1126
1127 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1128 {
1129         u32 high_crc = ~(u32)0;
1130         u32 low_crc = ~(u32)0;
1131         __le64 lenum;
1132
1133         lenum = cpu_to_le64(root_objectid);
1134         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1135         lenum = cpu_to_le64(owner);
1136         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1137         lenum = cpu_to_le64(offset);
1138         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1139
1140         return ((u64)high_crc << 31) ^ (u64)low_crc;
1141 }
1142
1143 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1144                                      struct btrfs_extent_data_ref *ref)
1145 {
1146         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1147                                     btrfs_extent_data_ref_objectid(leaf, ref),
1148                                     btrfs_extent_data_ref_offset(leaf, ref));
1149 }
1150
1151 static int match_extent_data_ref(struct extent_buffer *leaf,
1152                                  struct btrfs_extent_data_ref *ref,
1153                                  u64 root_objectid, u64 owner, u64 offset)
1154 {
1155         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1156             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1157             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1158                 return 0;
1159         return 1;
1160 }
1161
1162 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1163                                            struct btrfs_root *root,
1164                                            struct btrfs_path *path,
1165                                            u64 bytenr, u64 parent,
1166                                            u64 root_objectid,
1167                                            u64 owner, u64 offset)
1168 {
1169         struct btrfs_key key;
1170         struct btrfs_extent_data_ref *ref;
1171         struct extent_buffer *leaf;
1172         u32 nritems;
1173         int ret;
1174         int recow;
1175         int err = -ENOENT;
1176
1177         key.objectid = bytenr;
1178         if (parent) {
1179                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1180                 key.offset = parent;
1181         } else {
1182                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1183                 key.offset = hash_extent_data_ref(root_objectid,
1184                                                   owner, offset);
1185         }
1186 again:
1187         recow = 0;
1188         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1189         if (ret < 0) {
1190                 err = ret;
1191                 goto fail;
1192         }
1193
1194         if (parent) {
1195                 if (!ret)
1196                         return 0;
1197 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1198                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1199                 btrfs_release_path(path);
1200                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1201                 if (ret < 0) {
1202                         err = ret;
1203                         goto fail;
1204                 }
1205                 if (!ret)
1206                         return 0;
1207 #endif
1208                 goto fail;
1209         }
1210
1211         leaf = path->nodes[0];
1212         nritems = btrfs_header_nritems(leaf);
1213         while (1) {
1214                 if (path->slots[0] >= nritems) {
1215                         ret = btrfs_next_leaf(root, path);
1216                         if (ret < 0)
1217                                 err = ret;
1218                         if (ret)
1219                                 goto fail;
1220
1221                         leaf = path->nodes[0];
1222                         nritems = btrfs_header_nritems(leaf);
1223                         recow = 1;
1224                 }
1225
1226                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1227                 if (key.objectid != bytenr ||
1228                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1229                         goto fail;
1230
1231                 ref = btrfs_item_ptr(leaf, path->slots[0],
1232                                      struct btrfs_extent_data_ref);
1233
1234                 if (match_extent_data_ref(leaf, ref, root_objectid,
1235                                           owner, offset)) {
1236                         if (recow) {
1237                                 btrfs_release_path(path);
1238                                 goto again;
1239                         }
1240                         err = 0;
1241                         break;
1242                 }
1243                 path->slots[0]++;
1244         }
1245 fail:
1246         return err;
1247 }
1248
1249 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1250                                            struct btrfs_root *root,
1251                                            struct btrfs_path *path,
1252                                            u64 bytenr, u64 parent,
1253                                            u64 root_objectid, u64 owner,
1254                                            u64 offset, int refs_to_add)
1255 {
1256         struct btrfs_key key;
1257         struct extent_buffer *leaf;
1258         u32 size;
1259         u32 num_refs;
1260         int ret;
1261
1262         key.objectid = bytenr;
1263         if (parent) {
1264                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1265                 key.offset = parent;
1266                 size = sizeof(struct btrfs_shared_data_ref);
1267         } else {
1268                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1269                 key.offset = hash_extent_data_ref(root_objectid,
1270                                                   owner, offset);
1271                 size = sizeof(struct btrfs_extent_data_ref);
1272         }
1273
1274         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1275         if (ret && ret != -EEXIST)
1276                 goto fail;
1277
1278         leaf = path->nodes[0];
1279         if (parent) {
1280                 struct btrfs_shared_data_ref *ref;
1281                 ref = btrfs_item_ptr(leaf, path->slots[0],
1282                                      struct btrfs_shared_data_ref);
1283                 if (ret == 0) {
1284                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1285                 } else {
1286                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1287                         num_refs += refs_to_add;
1288                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1289                 }
1290         } else {
1291                 struct btrfs_extent_data_ref *ref;
1292                 while (ret == -EEXIST) {
1293                         ref = btrfs_item_ptr(leaf, path->slots[0],
1294                                              struct btrfs_extent_data_ref);
1295                         if (match_extent_data_ref(leaf, ref, root_objectid,
1296                                                   owner, offset))
1297                                 break;
1298                         btrfs_release_path(path);
1299                         key.offset++;
1300                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1301                                                       size);
1302                         if (ret && ret != -EEXIST)
1303                                 goto fail;
1304
1305                         leaf = path->nodes[0];
1306                 }
1307                 ref = btrfs_item_ptr(leaf, path->slots[0],
1308                                      struct btrfs_extent_data_ref);
1309                 if (ret == 0) {
1310                         btrfs_set_extent_data_ref_root(leaf, ref,
1311                                                        root_objectid);
1312                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1313                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1314                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1315                 } else {
1316                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1317                         num_refs += refs_to_add;
1318                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1319                 }
1320         }
1321         btrfs_mark_buffer_dirty(leaf);
1322         ret = 0;
1323 fail:
1324         btrfs_release_path(path);
1325         return ret;
1326 }
1327
1328 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1329                                            struct btrfs_root *root,
1330                                            struct btrfs_path *path,
1331                                            int refs_to_drop, int *last_ref)
1332 {
1333         struct btrfs_key key;
1334         struct btrfs_extent_data_ref *ref1 = NULL;
1335         struct btrfs_shared_data_ref *ref2 = NULL;
1336         struct extent_buffer *leaf;
1337         u32 num_refs = 0;
1338         int ret = 0;
1339
1340         leaf = path->nodes[0];
1341         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1342
1343         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1344                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1345                                       struct btrfs_extent_data_ref);
1346                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1347         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1348                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1349                                       struct btrfs_shared_data_ref);
1350                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1351 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1352         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1353                 struct btrfs_extent_ref_v0 *ref0;
1354                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1355                                       struct btrfs_extent_ref_v0);
1356                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1357 #endif
1358         } else {
1359                 BUG();
1360         }
1361
1362         BUG_ON(num_refs < refs_to_drop);
1363         num_refs -= refs_to_drop;
1364
1365         if (num_refs == 0) {
1366                 ret = btrfs_del_item(trans, root, path);
1367                 *last_ref = 1;
1368         } else {
1369                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1370                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1371                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1372                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1373 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1374                 else {
1375                         struct btrfs_extent_ref_v0 *ref0;
1376                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1377                                         struct btrfs_extent_ref_v0);
1378                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1379                 }
1380 #endif
1381                 btrfs_mark_buffer_dirty(leaf);
1382         }
1383         return ret;
1384 }
1385
1386 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1387                                           struct btrfs_extent_inline_ref *iref)
1388 {
1389         struct btrfs_key key;
1390         struct extent_buffer *leaf;
1391         struct btrfs_extent_data_ref *ref1;
1392         struct btrfs_shared_data_ref *ref2;
1393         u32 num_refs = 0;
1394
1395         leaf = path->nodes[0];
1396         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1397         if (iref) {
1398                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1399                     BTRFS_EXTENT_DATA_REF_KEY) {
1400                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1401                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1402                 } else {
1403                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1404                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1405                 }
1406         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1407                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1408                                       struct btrfs_extent_data_ref);
1409                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1410         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1411                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1412                                       struct btrfs_shared_data_ref);
1413                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1414 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1415         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1416                 struct btrfs_extent_ref_v0 *ref0;
1417                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1418                                       struct btrfs_extent_ref_v0);
1419                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1420 #endif
1421         } else {
1422                 WARN_ON(1);
1423         }
1424         return num_refs;
1425 }
1426
1427 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1428                                           struct btrfs_root *root,
1429                                           struct btrfs_path *path,
1430                                           u64 bytenr, u64 parent,
1431                                           u64 root_objectid)
1432 {
1433         struct btrfs_key key;
1434         int ret;
1435
1436         key.objectid = bytenr;
1437         if (parent) {
1438                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1439                 key.offset = parent;
1440         } else {
1441                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1442                 key.offset = root_objectid;
1443         }
1444
1445         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1446         if (ret > 0)
1447                 ret = -ENOENT;
1448 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1449         if (ret == -ENOENT && parent) {
1450                 btrfs_release_path(path);
1451                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1452                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1453                 if (ret > 0)
1454                         ret = -ENOENT;
1455         }
1456 #endif
1457         return ret;
1458 }
1459
1460 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1461                                           struct btrfs_root *root,
1462                                           struct btrfs_path *path,
1463                                           u64 bytenr, u64 parent,
1464                                           u64 root_objectid)
1465 {
1466         struct btrfs_key key;
1467         int ret;
1468
1469         key.objectid = bytenr;
1470         if (parent) {
1471                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1472                 key.offset = parent;
1473         } else {
1474                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1475                 key.offset = root_objectid;
1476         }
1477
1478         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1479         btrfs_release_path(path);
1480         return ret;
1481 }
1482
1483 static inline int extent_ref_type(u64 parent, u64 owner)
1484 {
1485         int type;
1486         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1487                 if (parent > 0)
1488                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1489                 else
1490                         type = BTRFS_TREE_BLOCK_REF_KEY;
1491         } else {
1492                 if (parent > 0)
1493                         type = BTRFS_SHARED_DATA_REF_KEY;
1494                 else
1495                         type = BTRFS_EXTENT_DATA_REF_KEY;
1496         }
1497         return type;
1498 }
1499
1500 static int find_next_key(struct btrfs_path *path, int level,
1501                          struct btrfs_key *key)
1502
1503 {
1504         for (; level < BTRFS_MAX_LEVEL; level++) {
1505                 if (!path->nodes[level])
1506                         break;
1507                 if (path->slots[level] + 1 >=
1508                     btrfs_header_nritems(path->nodes[level]))
1509                         continue;
1510                 if (level == 0)
1511                         btrfs_item_key_to_cpu(path->nodes[level], key,
1512                                               path->slots[level] + 1);
1513                 else
1514                         btrfs_node_key_to_cpu(path->nodes[level], key,
1515                                               path->slots[level] + 1);
1516                 return 0;
1517         }
1518         return 1;
1519 }
1520
1521 /*
1522  * look for inline back ref. if back ref is found, *ref_ret is set
1523  * to the address of inline back ref, and 0 is returned.
1524  *
1525  * if back ref isn't found, *ref_ret is set to the address where it
1526  * should be inserted, and -ENOENT is returned.
1527  *
1528  * if insert is true and there are too many inline back refs, the path
1529  * points to the extent item, and -EAGAIN is returned.
1530  *
1531  * NOTE: inline back refs are ordered in the same way that back ref
1532  *       items in the tree are ordered.
1533  */
1534 static noinline_for_stack
1535 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1536                                  struct btrfs_root *root,
1537                                  struct btrfs_path *path,
1538                                  struct btrfs_extent_inline_ref **ref_ret,
1539                                  u64 bytenr, u64 num_bytes,
1540                                  u64 parent, u64 root_objectid,
1541                                  u64 owner, u64 offset, int insert)
1542 {
1543         struct btrfs_key key;
1544         struct extent_buffer *leaf;
1545         struct btrfs_extent_item *ei;
1546         struct btrfs_extent_inline_ref *iref;
1547         u64 flags;
1548         u64 item_size;
1549         unsigned long ptr;
1550         unsigned long end;
1551         int extra_size;
1552         int type;
1553         int want;
1554         int ret;
1555         int err = 0;
1556         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1557                                                  SKINNY_METADATA);
1558
1559         key.objectid = bytenr;
1560         key.type = BTRFS_EXTENT_ITEM_KEY;
1561         key.offset = num_bytes;
1562
1563         want = extent_ref_type(parent, owner);
1564         if (insert) {
1565                 extra_size = btrfs_extent_inline_ref_size(want);
1566                 path->keep_locks = 1;
1567         } else
1568                 extra_size = -1;
1569
1570         /*
1571          * Owner is our parent level, so we can just add one to get the level
1572          * for the block we are interested in.
1573          */
1574         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1575                 key.type = BTRFS_METADATA_ITEM_KEY;
1576                 key.offset = owner;
1577         }
1578
1579 again:
1580         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1581         if (ret < 0) {
1582                 err = ret;
1583                 goto out;
1584         }
1585
1586         /*
1587          * We may be a newly converted file system which still has the old fat
1588          * extent entries for metadata, so try and see if we have one of those.
1589          */
1590         if (ret > 0 && skinny_metadata) {
1591                 skinny_metadata = false;
1592                 if (path->slots[0]) {
1593                         path->slots[0]--;
1594                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1595                                               path->slots[0]);
1596                         if (key.objectid == bytenr &&
1597                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1598                             key.offset == num_bytes)
1599                                 ret = 0;
1600                 }
1601                 if (ret) {
1602                         key.objectid = bytenr;
1603                         key.type = BTRFS_EXTENT_ITEM_KEY;
1604                         key.offset = num_bytes;
1605                         btrfs_release_path(path);
1606                         goto again;
1607                 }
1608         }
1609
1610         if (ret && !insert) {
1611                 err = -ENOENT;
1612                 goto out;
1613         } else if (WARN_ON(ret)) {
1614                 err = -EIO;
1615                 goto out;
1616         }
1617
1618         leaf = path->nodes[0];
1619         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1620 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1621         if (item_size < sizeof(*ei)) {
1622                 if (!insert) {
1623                         err = -ENOENT;
1624                         goto out;
1625                 }
1626                 ret = convert_extent_item_v0(trans, root, path, owner,
1627                                              extra_size);
1628                 if (ret < 0) {
1629                         err = ret;
1630                         goto out;
1631                 }
1632                 leaf = path->nodes[0];
1633                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1634         }
1635 #endif
1636         BUG_ON(item_size < sizeof(*ei));
1637
1638         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1639         flags = btrfs_extent_flags(leaf, ei);
1640
1641         ptr = (unsigned long)(ei + 1);
1642         end = (unsigned long)ei + item_size;
1643
1644         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1645                 ptr += sizeof(struct btrfs_tree_block_info);
1646                 BUG_ON(ptr > end);
1647         }
1648
1649         err = -ENOENT;
1650         while (1) {
1651                 if (ptr >= end) {
1652                         WARN_ON(ptr > end);
1653                         break;
1654                 }
1655                 iref = (struct btrfs_extent_inline_ref *)ptr;
1656                 type = btrfs_extent_inline_ref_type(leaf, iref);
1657                 if (want < type)
1658                         break;
1659                 if (want > type) {
1660                         ptr += btrfs_extent_inline_ref_size(type);
1661                         continue;
1662                 }
1663
1664                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1665                         struct btrfs_extent_data_ref *dref;
1666                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1667                         if (match_extent_data_ref(leaf, dref, root_objectid,
1668                                                   owner, offset)) {
1669                                 err = 0;
1670                                 break;
1671                         }
1672                         if (hash_extent_data_ref_item(leaf, dref) <
1673                             hash_extent_data_ref(root_objectid, owner, offset))
1674                                 break;
1675                 } else {
1676                         u64 ref_offset;
1677                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1678                         if (parent > 0) {
1679                                 if (parent == ref_offset) {
1680                                         err = 0;
1681                                         break;
1682                                 }
1683                                 if (ref_offset < parent)
1684                                         break;
1685                         } else {
1686                                 if (root_objectid == ref_offset) {
1687                                         err = 0;
1688                                         break;
1689                                 }
1690                                 if (ref_offset < root_objectid)
1691                                         break;
1692                         }
1693                 }
1694                 ptr += btrfs_extent_inline_ref_size(type);
1695         }
1696         if (err == -ENOENT && insert) {
1697                 if (item_size + extra_size >=
1698                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1699                         err = -EAGAIN;
1700                         goto out;
1701                 }
1702                 /*
1703                  * To add new inline back ref, we have to make sure
1704                  * there is no corresponding back ref item.
1705                  * For simplicity, we just do not add new inline back
1706                  * ref if there is any kind of item for this block
1707                  */
1708                 if (find_next_key(path, 0, &key) == 0 &&
1709                     key.objectid == bytenr &&
1710                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1711                         err = -EAGAIN;
1712                         goto out;
1713                 }
1714         }
1715         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1716 out:
1717         if (insert) {
1718                 path->keep_locks = 0;
1719                 btrfs_unlock_up_safe(path, 1);
1720         }
1721         return err;
1722 }
1723
1724 /*
1725  * helper to add new inline back ref
1726  */
1727 static noinline_for_stack
1728 void setup_inline_extent_backref(struct btrfs_root *root,
1729                                  struct btrfs_path *path,
1730                                  struct btrfs_extent_inline_ref *iref,
1731                                  u64 parent, u64 root_objectid,
1732                                  u64 owner, u64 offset, int refs_to_add,
1733                                  struct btrfs_delayed_extent_op *extent_op)
1734 {
1735         struct extent_buffer *leaf;
1736         struct btrfs_extent_item *ei;
1737         unsigned long ptr;
1738         unsigned long end;
1739         unsigned long item_offset;
1740         u64 refs;
1741         int size;
1742         int type;
1743
1744         leaf = path->nodes[0];
1745         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1746         item_offset = (unsigned long)iref - (unsigned long)ei;
1747
1748         type = extent_ref_type(parent, owner);
1749         size = btrfs_extent_inline_ref_size(type);
1750
1751         btrfs_extend_item(root, path, size);
1752
1753         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1754         refs = btrfs_extent_refs(leaf, ei);
1755         refs += refs_to_add;
1756         btrfs_set_extent_refs(leaf, ei, refs);
1757         if (extent_op)
1758                 __run_delayed_extent_op(extent_op, leaf, ei);
1759
1760         ptr = (unsigned long)ei + item_offset;
1761         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1762         if (ptr < end - size)
1763                 memmove_extent_buffer(leaf, ptr + size, ptr,
1764                                       end - size - ptr);
1765
1766         iref = (struct btrfs_extent_inline_ref *)ptr;
1767         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1768         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1769                 struct btrfs_extent_data_ref *dref;
1770                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1771                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1772                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1773                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1774                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1775         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1776                 struct btrfs_shared_data_ref *sref;
1777                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1778                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1779                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1780         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1781                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1782         } else {
1783                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1784         }
1785         btrfs_mark_buffer_dirty(leaf);
1786 }
1787
1788 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1789                                  struct btrfs_root *root,
1790                                  struct btrfs_path *path,
1791                                  struct btrfs_extent_inline_ref **ref_ret,
1792                                  u64 bytenr, u64 num_bytes, u64 parent,
1793                                  u64 root_objectid, u64 owner, u64 offset)
1794 {
1795         int ret;
1796
1797         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1798                                            bytenr, num_bytes, parent,
1799                                            root_objectid, owner, offset, 0);
1800         if (ret != -ENOENT)
1801                 return ret;
1802
1803         btrfs_release_path(path);
1804         *ref_ret = NULL;
1805
1806         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1807                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1808                                             root_objectid);
1809         } else {
1810                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1811                                              root_objectid, owner, offset);
1812         }
1813         return ret;
1814 }
1815
1816 /*
1817  * helper to update/remove inline back ref
1818  */
1819 static noinline_for_stack
1820 void update_inline_extent_backref(struct btrfs_root *root,
1821                                   struct btrfs_path *path,
1822                                   struct btrfs_extent_inline_ref *iref,
1823                                   int refs_to_mod,
1824                                   struct btrfs_delayed_extent_op *extent_op,
1825                                   int *last_ref)
1826 {
1827         struct extent_buffer *leaf;
1828         struct btrfs_extent_item *ei;
1829         struct btrfs_extent_data_ref *dref = NULL;
1830         struct btrfs_shared_data_ref *sref = NULL;
1831         unsigned long ptr;
1832         unsigned long end;
1833         u32 item_size;
1834         int size;
1835         int type;
1836         u64 refs;
1837
1838         leaf = path->nodes[0];
1839         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1840         refs = btrfs_extent_refs(leaf, ei);
1841         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1842         refs += refs_to_mod;
1843         btrfs_set_extent_refs(leaf, ei, refs);
1844         if (extent_op)
1845                 __run_delayed_extent_op(extent_op, leaf, ei);
1846
1847         type = btrfs_extent_inline_ref_type(leaf, iref);
1848
1849         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1850                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1851                 refs = btrfs_extent_data_ref_count(leaf, dref);
1852         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1853                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1854                 refs = btrfs_shared_data_ref_count(leaf, sref);
1855         } else {
1856                 refs = 1;
1857                 BUG_ON(refs_to_mod != -1);
1858         }
1859
1860         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1861         refs += refs_to_mod;
1862
1863         if (refs > 0) {
1864                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1865                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1866                 else
1867                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1868         } else {
1869                 *last_ref = 1;
1870                 size =  btrfs_extent_inline_ref_size(type);
1871                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1872                 ptr = (unsigned long)iref;
1873                 end = (unsigned long)ei + item_size;
1874                 if (ptr + size < end)
1875                         memmove_extent_buffer(leaf, ptr, ptr + size,
1876                                               end - ptr - size);
1877                 item_size -= size;
1878                 btrfs_truncate_item(root, path, item_size, 1);
1879         }
1880         btrfs_mark_buffer_dirty(leaf);
1881 }
1882
1883 static noinline_for_stack
1884 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1885                                  struct btrfs_root *root,
1886                                  struct btrfs_path *path,
1887                                  u64 bytenr, u64 num_bytes, u64 parent,
1888                                  u64 root_objectid, u64 owner,
1889                                  u64 offset, int refs_to_add,
1890                                  struct btrfs_delayed_extent_op *extent_op)
1891 {
1892         struct btrfs_extent_inline_ref *iref;
1893         int ret;
1894
1895         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1896                                            bytenr, num_bytes, parent,
1897                                            root_objectid, owner, offset, 1);
1898         if (ret == 0) {
1899                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1900                 update_inline_extent_backref(root, path, iref,
1901                                              refs_to_add, extent_op, NULL);
1902         } else if (ret == -ENOENT) {
1903                 setup_inline_extent_backref(root, path, iref, parent,
1904                                             root_objectid, owner, offset,
1905                                             refs_to_add, extent_op);
1906                 ret = 0;
1907         }
1908         return ret;
1909 }
1910
1911 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1912                                  struct btrfs_root *root,
1913                                  struct btrfs_path *path,
1914                                  u64 bytenr, u64 parent, u64 root_objectid,
1915                                  u64 owner, u64 offset, int refs_to_add)
1916 {
1917         int ret;
1918         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1919                 BUG_ON(refs_to_add != 1);
1920                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1921                                             parent, root_objectid);
1922         } else {
1923                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1924                                              parent, root_objectid,
1925                                              owner, offset, refs_to_add);
1926         }
1927         return ret;
1928 }
1929
1930 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1931                                  struct btrfs_root *root,
1932                                  struct btrfs_path *path,
1933                                  struct btrfs_extent_inline_ref *iref,
1934                                  int refs_to_drop, int is_data, int *last_ref)
1935 {
1936         int ret = 0;
1937
1938         BUG_ON(!is_data && refs_to_drop != 1);
1939         if (iref) {
1940                 update_inline_extent_backref(root, path, iref,
1941                                              -refs_to_drop, NULL, last_ref);
1942         } else if (is_data) {
1943                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1944                                              last_ref);
1945         } else {
1946                 *last_ref = 1;
1947                 ret = btrfs_del_item(trans, root, path);
1948         }
1949         return ret;
1950 }
1951
1952 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1953 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1954                                u64 *discarded_bytes)
1955 {
1956         int j, ret = 0;
1957         u64 bytes_left, end;
1958         u64 aligned_start = ALIGN(start, 1 << 9);
1959
1960         if (WARN_ON(start != aligned_start)) {
1961                 len -= aligned_start - start;
1962                 len = round_down(len, 1 << 9);
1963                 start = aligned_start;
1964         }
1965
1966         *discarded_bytes = 0;
1967
1968         if (!len)
1969                 return 0;
1970
1971         end = start + len;
1972         bytes_left = len;
1973
1974         /* Skip any superblocks on this device. */
1975         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1976                 u64 sb_start = btrfs_sb_offset(j);
1977                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1978                 u64 size = sb_start - start;
1979
1980                 if (!in_range(sb_start, start, bytes_left) &&
1981                     !in_range(sb_end, start, bytes_left) &&
1982                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1983                         continue;
1984
1985                 /*
1986                  * Superblock spans beginning of range.  Adjust start and
1987                  * try again.
1988                  */
1989                 if (sb_start <= start) {
1990                         start += sb_end - start;
1991                         if (start > end) {
1992                                 bytes_left = 0;
1993                                 break;
1994                         }
1995                         bytes_left = end - start;
1996                         continue;
1997                 }
1998
1999                 if (size) {
2000                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2001                                                    GFP_NOFS, 0);
2002                         if (!ret)
2003                                 *discarded_bytes += size;
2004                         else if (ret != -EOPNOTSUPP)
2005                                 return ret;
2006                 }
2007
2008                 start = sb_end;
2009                 if (start > end) {
2010                         bytes_left = 0;
2011                         break;
2012                 }
2013                 bytes_left = end - start;
2014         }
2015
2016         if (bytes_left) {
2017                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2018                                            GFP_NOFS, 0);
2019                 if (!ret)
2020                         *discarded_bytes += bytes_left;
2021         }
2022         return ret;
2023 }
2024
2025 int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
2026                          u64 num_bytes, u64 *actual_bytes)
2027 {
2028         int ret;
2029         u64 discarded_bytes = 0;
2030         struct btrfs_bio *bbio = NULL;
2031
2032
2033         /*
2034          * Avoid races with device replace and make sure our bbio has devices
2035          * associated to its stripes that don't go away while we are discarding.
2036          */
2037         btrfs_bio_counter_inc_blocked(root->fs_info);
2038         /* Tell the block device(s) that the sectors can be discarded */
2039         ret = btrfs_map_block(root->fs_info, REQ_OP_DISCARD,
2040                               bytenr, &num_bytes, &bbio, 0);
2041         /* Error condition is -ENOMEM */
2042         if (!ret) {
2043                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2044                 int i;
2045
2046
2047                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2048                         u64 bytes;
2049                         if (!stripe->dev->can_discard)
2050                                 continue;
2051
2052                         ret = btrfs_issue_discard(stripe->dev->bdev,
2053                                                   stripe->physical,
2054                                                   stripe->length,
2055                                                   &bytes);
2056                         if (!ret)
2057                                 discarded_bytes += bytes;
2058                         else if (ret != -EOPNOTSUPP)
2059                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2060
2061                         /*
2062                          * Just in case we get back EOPNOTSUPP for some reason,
2063                          * just ignore the return value so we don't screw up
2064                          * people calling discard_extent.
2065                          */
2066                         ret = 0;
2067                 }
2068                 btrfs_put_bbio(bbio);
2069         }
2070         btrfs_bio_counter_dec(root->fs_info);
2071
2072         if (actual_bytes)
2073                 *actual_bytes = discarded_bytes;
2074
2075
2076         if (ret == -EOPNOTSUPP)
2077                 ret = 0;
2078         return ret;
2079 }
2080
2081 /* Can return -ENOMEM */
2082 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2083                          struct btrfs_root *root,
2084                          u64 bytenr, u64 num_bytes, u64 parent,
2085                          u64 root_objectid, u64 owner, u64 offset)
2086 {
2087         int ret;
2088         struct btrfs_fs_info *fs_info = root->fs_info;
2089
2090         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2091                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2092
2093         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2094                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2095                                         num_bytes,
2096                                         parent, root_objectid, (int)owner,
2097                                         BTRFS_ADD_DELAYED_REF, NULL);
2098         } else {
2099                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2100                                         num_bytes, parent, root_objectid,
2101                                         owner, offset, 0,
2102                                         BTRFS_ADD_DELAYED_REF, NULL);
2103         }
2104         return ret;
2105 }
2106
2107 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2108                                   struct btrfs_root *root,
2109                                   struct btrfs_delayed_ref_node *node,
2110                                   u64 parent, u64 root_objectid,
2111                                   u64 owner, u64 offset, int refs_to_add,
2112                                   struct btrfs_delayed_extent_op *extent_op)
2113 {
2114         struct btrfs_fs_info *fs_info = root->fs_info;
2115         struct btrfs_path *path;
2116         struct extent_buffer *leaf;
2117         struct btrfs_extent_item *item;
2118         struct btrfs_key key;
2119         u64 bytenr = node->bytenr;
2120         u64 num_bytes = node->num_bytes;
2121         u64 refs;
2122         int ret;
2123
2124         path = btrfs_alloc_path();
2125         if (!path)
2126                 return -ENOMEM;
2127
2128         path->reada = READA_FORWARD;
2129         path->leave_spinning = 1;
2130         /* this will setup the path even if it fails to insert the back ref */
2131         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2132                                            bytenr, num_bytes, parent,
2133                                            root_objectid, owner, offset,
2134                                            refs_to_add, extent_op);
2135         if ((ret < 0 && ret != -EAGAIN) || !ret)
2136                 goto out;
2137
2138         /*
2139          * Ok we had -EAGAIN which means we didn't have space to insert and
2140          * inline extent ref, so just update the reference count and add a
2141          * normal backref.
2142          */
2143         leaf = path->nodes[0];
2144         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2145         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2146         refs = btrfs_extent_refs(leaf, item);
2147         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2148         if (extent_op)
2149                 __run_delayed_extent_op(extent_op, leaf, item);
2150
2151         btrfs_mark_buffer_dirty(leaf);
2152         btrfs_release_path(path);
2153
2154         path->reada = READA_FORWARD;
2155         path->leave_spinning = 1;
2156         /* now insert the actual backref */
2157         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2158                                     path, bytenr, parent, root_objectid,
2159                                     owner, offset, refs_to_add);
2160         if (ret)
2161                 btrfs_abort_transaction(trans, ret);
2162 out:
2163         btrfs_free_path(path);
2164         return ret;
2165 }
2166
2167 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2168                                 struct btrfs_root *root,
2169                                 struct btrfs_delayed_ref_node *node,
2170                                 struct btrfs_delayed_extent_op *extent_op,
2171                                 int insert_reserved)
2172 {
2173         int ret = 0;
2174         struct btrfs_delayed_data_ref *ref;
2175         struct btrfs_key ins;
2176         u64 parent = 0;
2177         u64 ref_root = 0;
2178         u64 flags = 0;
2179
2180         ins.objectid = node->bytenr;
2181         ins.offset = node->num_bytes;
2182         ins.type = BTRFS_EXTENT_ITEM_KEY;
2183
2184         ref = btrfs_delayed_node_to_data_ref(node);
2185         trace_run_delayed_data_ref(root->fs_info, node, ref, node->action);
2186
2187         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2188                 parent = ref->parent;
2189         ref_root = ref->root;
2190
2191         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2192                 if (extent_op)
2193                         flags |= extent_op->flags_to_set;
2194                 ret = alloc_reserved_file_extent(trans, root,
2195                                                  parent, ref_root, flags,
2196                                                  ref->objectid, ref->offset,
2197                                                  &ins, node->ref_mod);
2198         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2199                 ret = __btrfs_inc_extent_ref(trans, root, node, parent,
2200                                              ref_root, ref->objectid,
2201                                              ref->offset, node->ref_mod,
2202                                              extent_op);
2203         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2204                 ret = __btrfs_free_extent(trans, root, node, parent,
2205                                           ref_root, ref->objectid,
2206                                           ref->offset, node->ref_mod,
2207                                           extent_op);
2208         } else {
2209                 BUG();
2210         }
2211         return ret;
2212 }
2213
2214 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2215                                     struct extent_buffer *leaf,
2216                                     struct btrfs_extent_item *ei)
2217 {
2218         u64 flags = btrfs_extent_flags(leaf, ei);
2219         if (extent_op->update_flags) {
2220                 flags |= extent_op->flags_to_set;
2221                 btrfs_set_extent_flags(leaf, ei, flags);
2222         }
2223
2224         if (extent_op->update_key) {
2225                 struct btrfs_tree_block_info *bi;
2226                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2227                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2228                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2229         }
2230 }
2231
2232 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2233                                  struct btrfs_root *root,
2234                                  struct btrfs_delayed_ref_node *node,
2235                                  struct btrfs_delayed_extent_op *extent_op)
2236 {
2237         struct btrfs_key key;
2238         struct btrfs_path *path;
2239         struct btrfs_extent_item *ei;
2240         struct extent_buffer *leaf;
2241         u32 item_size;
2242         int ret;
2243         int err = 0;
2244         int metadata = !extent_op->is_data;
2245
2246         if (trans->aborted)
2247                 return 0;
2248
2249         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2250                 metadata = 0;
2251
2252         path = btrfs_alloc_path();
2253         if (!path)
2254                 return -ENOMEM;
2255
2256         key.objectid = node->bytenr;
2257
2258         if (metadata) {
2259                 key.type = BTRFS_METADATA_ITEM_KEY;
2260                 key.offset = extent_op->level;
2261         } else {
2262                 key.type = BTRFS_EXTENT_ITEM_KEY;
2263                 key.offset = node->num_bytes;
2264         }
2265
2266 again:
2267         path->reada = READA_FORWARD;
2268         path->leave_spinning = 1;
2269         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2270                                 path, 0, 1);
2271         if (ret < 0) {
2272                 err = ret;
2273                 goto out;
2274         }
2275         if (ret > 0) {
2276                 if (metadata) {
2277                         if (path->slots[0] > 0) {
2278                                 path->slots[0]--;
2279                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2280                                                       path->slots[0]);
2281                                 if (key.objectid == node->bytenr &&
2282                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2283                                     key.offset == node->num_bytes)
2284                                         ret = 0;
2285                         }
2286                         if (ret > 0) {
2287                                 btrfs_release_path(path);
2288                                 metadata = 0;
2289
2290                                 key.objectid = node->bytenr;
2291                                 key.offset = node->num_bytes;
2292                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2293                                 goto again;
2294                         }
2295                 } else {
2296                         err = -EIO;
2297                         goto out;
2298                 }
2299         }
2300
2301         leaf = path->nodes[0];
2302         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2303 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2304         if (item_size < sizeof(*ei)) {
2305                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2306                                              path, (u64)-1, 0);
2307                 if (ret < 0) {
2308                         err = ret;
2309                         goto out;
2310                 }
2311                 leaf = path->nodes[0];
2312                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2313         }
2314 #endif
2315         BUG_ON(item_size < sizeof(*ei));
2316         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2317         __run_delayed_extent_op(extent_op, leaf, ei);
2318
2319         btrfs_mark_buffer_dirty(leaf);
2320 out:
2321         btrfs_free_path(path);
2322         return err;
2323 }
2324
2325 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2326                                 struct btrfs_root *root,
2327                                 struct btrfs_delayed_ref_node *node,
2328                                 struct btrfs_delayed_extent_op *extent_op,
2329                                 int insert_reserved)
2330 {
2331         int ret = 0;
2332         struct btrfs_delayed_tree_ref *ref;
2333         struct btrfs_key ins;
2334         u64 parent = 0;
2335         u64 ref_root = 0;
2336         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2337                                                  SKINNY_METADATA);
2338
2339         ref = btrfs_delayed_node_to_tree_ref(node);
2340         trace_run_delayed_tree_ref(root->fs_info, node, ref, node->action);
2341
2342         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2343                 parent = ref->parent;
2344         ref_root = ref->root;
2345
2346         ins.objectid = node->bytenr;
2347         if (skinny_metadata) {
2348                 ins.offset = ref->level;
2349                 ins.type = BTRFS_METADATA_ITEM_KEY;
2350         } else {
2351                 ins.offset = node->num_bytes;
2352                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2353         }
2354
2355         if (node->ref_mod != 1) {
2356                 btrfs_err(root->fs_info,
2357         "btree block(%llu) has %d references rather than 1: action %d ref_root %llu parent %llu",
2358                           node->bytenr, node->ref_mod, node->action, ref_root,
2359                           parent);
2360                 return -EIO;
2361         }
2362         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2363                 BUG_ON(!extent_op || !extent_op->update_flags);
2364                 ret = alloc_reserved_tree_block(trans, root,
2365                                                 parent, ref_root,
2366                                                 extent_op->flags_to_set,
2367                                                 &extent_op->key,
2368                                                 ref->level, &ins);
2369         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2370                 ret = __btrfs_inc_extent_ref(trans, root, node,
2371                                              parent, ref_root,
2372                                              ref->level, 0, 1,
2373                                              extent_op);
2374         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2375                 ret = __btrfs_free_extent(trans, root, node,
2376                                           parent, ref_root,
2377                                           ref->level, 0, 1, extent_op);
2378         } else {
2379                 BUG();
2380         }
2381         return ret;
2382 }
2383
2384 /* helper function to actually process a single delayed ref entry */
2385 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2386                                struct btrfs_root *root,
2387                                struct btrfs_delayed_ref_node *node,
2388                                struct btrfs_delayed_extent_op *extent_op,
2389                                int insert_reserved)
2390 {
2391         int ret = 0;
2392
2393         if (trans->aborted) {
2394                 if (insert_reserved)
2395                         btrfs_pin_extent(root, node->bytenr,
2396                                          node->num_bytes, 1);
2397                 return 0;
2398         }
2399
2400         if (btrfs_delayed_ref_is_head(node)) {
2401                 struct btrfs_delayed_ref_head *head;
2402                 /*
2403                  * we've hit the end of the chain and we were supposed
2404                  * to insert this extent into the tree.  But, it got
2405                  * deleted before we ever needed to insert it, so all
2406                  * we have to do is clean up the accounting
2407                  */
2408                 BUG_ON(extent_op);
2409                 head = btrfs_delayed_node_to_head(node);
2410                 trace_run_delayed_ref_head(root->fs_info, node, head,
2411                                            node->action);
2412
2413                 if (insert_reserved) {
2414                         btrfs_pin_extent(root, node->bytenr,
2415                                          node->num_bytes, 1);
2416                         if (head->is_data) {
2417                                 ret = btrfs_del_csums(trans, root,
2418                                                       node->bytenr,
2419                                                       node->num_bytes);
2420                         }
2421                 }
2422
2423                 /* Also free its reserved qgroup space */
2424                 btrfs_qgroup_free_delayed_ref(root->fs_info,
2425                                               head->qgroup_ref_root,
2426                                               head->qgroup_reserved);
2427                 return ret;
2428         }
2429
2430         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2431             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2432                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2433                                            insert_reserved);
2434         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2435                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2436                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2437                                            insert_reserved);
2438         else
2439                 BUG();
2440         return ret;
2441 }
2442
2443 static inline struct btrfs_delayed_ref_node *
2444 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2445 {
2446         struct btrfs_delayed_ref_node *ref;
2447
2448         if (list_empty(&head->ref_list))
2449                 return NULL;
2450
2451         /*
2452          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2453          * This is to prevent a ref count from going down to zero, which deletes
2454          * the extent item from the extent tree, when there still are references
2455          * to add, which would fail because they would not find the extent item.
2456          */
2457         list_for_each_entry(ref, &head->ref_list, list) {
2458                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2459                         return ref;
2460         }
2461
2462         return list_entry(head->ref_list.next, struct btrfs_delayed_ref_node,
2463                           list);
2464 }
2465
2466 /*
2467  * Returns 0 on success or if called with an already aborted transaction.
2468  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2469  */
2470 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2471                                              struct btrfs_root *root,
2472                                              unsigned long nr)
2473 {
2474         struct btrfs_delayed_ref_root *delayed_refs;
2475         struct btrfs_delayed_ref_node *ref;
2476         struct btrfs_delayed_ref_head *locked_ref = NULL;
2477         struct btrfs_delayed_extent_op *extent_op;
2478         struct btrfs_fs_info *fs_info = root->fs_info;
2479         ktime_t start = ktime_get();
2480         int ret;
2481         unsigned long count = 0;
2482         unsigned long actual_count = 0;
2483         int must_insert_reserved = 0;
2484
2485         delayed_refs = &trans->transaction->delayed_refs;
2486         while (1) {
2487                 if (!locked_ref) {
2488                         if (count >= nr)
2489                                 break;
2490
2491                         spin_lock(&delayed_refs->lock);
2492                         locked_ref = btrfs_select_ref_head(trans);
2493                         if (!locked_ref) {
2494                                 spin_unlock(&delayed_refs->lock);
2495                                 break;
2496                         }
2497
2498                         /* grab the lock that says we are going to process
2499                          * all the refs for this head */
2500                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2501                         spin_unlock(&delayed_refs->lock);
2502                         /*
2503                          * we may have dropped the spin lock to get the head
2504                          * mutex lock, and that might have given someone else
2505                          * time to free the head.  If that's true, it has been
2506                          * removed from our list and we can move on.
2507                          */
2508                         if (ret == -EAGAIN) {
2509                                 locked_ref = NULL;
2510                                 count++;
2511                                 continue;
2512                         }
2513                 }
2514
2515                 /*
2516                  * We need to try and merge add/drops of the same ref since we
2517                  * can run into issues with relocate dropping the implicit ref
2518                  * and then it being added back again before the drop can
2519                  * finish.  If we merged anything we need to re-loop so we can
2520                  * get a good ref.
2521                  * Or we can get node references of the same type that weren't
2522                  * merged when created due to bumps in the tree mod seq, and
2523                  * we need to merge them to prevent adding an inline extent
2524                  * backref before dropping it (triggering a BUG_ON at
2525                  * insert_inline_extent_backref()).
2526                  */
2527                 spin_lock(&locked_ref->lock);
2528                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2529                                          locked_ref);
2530
2531                 /*
2532                  * locked_ref is the head node, so we have to go one
2533                  * node back for any delayed ref updates
2534                  */
2535                 ref = select_delayed_ref(locked_ref);
2536
2537                 if (ref && ref->seq &&
2538                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2539                         spin_unlock(&locked_ref->lock);
2540                         btrfs_delayed_ref_unlock(locked_ref);
2541                         spin_lock(&delayed_refs->lock);
2542                         locked_ref->processing = 0;
2543                         delayed_refs->num_heads_ready++;
2544                         spin_unlock(&delayed_refs->lock);
2545                         locked_ref = NULL;
2546                         cond_resched();
2547                         count++;
2548                         continue;
2549                 }
2550
2551                 /*
2552                  * record the must insert reserved flag before we
2553                  * drop the spin lock.
2554                  */
2555                 must_insert_reserved = locked_ref->must_insert_reserved;
2556                 locked_ref->must_insert_reserved = 0;
2557
2558                 extent_op = locked_ref->extent_op;
2559                 locked_ref->extent_op = NULL;
2560
2561                 if (!ref) {
2562
2563
2564                         /* All delayed refs have been processed, Go ahead
2565                          * and send the head node to run_one_delayed_ref,
2566                          * so that any accounting fixes can happen
2567                          */
2568                         ref = &locked_ref->node;
2569
2570                         if (extent_op && must_insert_reserved) {
2571                                 btrfs_free_delayed_extent_op(extent_op);
2572                                 extent_op = NULL;
2573                         }
2574
2575                         if (extent_op) {
2576                                 spin_unlock(&locked_ref->lock);
2577                                 ret = run_delayed_extent_op(trans, root,
2578                                                             ref, extent_op);
2579                                 btrfs_free_delayed_extent_op(extent_op);
2580
2581                                 if (ret) {
2582                                         /*
2583                                          * Need to reset must_insert_reserved if
2584                                          * there was an error so the abort stuff
2585                                          * can cleanup the reserved space
2586                                          * properly.
2587                                          */
2588                                         if (must_insert_reserved)
2589                                                 locked_ref->must_insert_reserved = 1;
2590                                         locked_ref->processing = 0;
2591                                         btrfs_debug(fs_info,
2592                                                     "run_delayed_extent_op returned %d",
2593                                                     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",
2654                                     ret);
2655                         return ret;
2656                 }
2657
2658                 /*
2659                  * If this node is a head, that means all the refs in this head
2660                  * have been dealt with, and we will pick the next head to deal
2661                  * with, so we must unlock the head and drop it from the cluster
2662                  * list before we release it.
2663                  */
2664                 if (btrfs_delayed_ref_is_head(ref)) {
2665                         if (locked_ref->is_data &&
2666                             locked_ref->total_ref_mod < 0) {
2667                                 spin_lock(&delayed_refs->lock);
2668                                 delayed_refs->pending_csums -= ref->num_bytes;
2669                                 spin_unlock(&delayed_refs->lock);
2670                         }
2671                         btrfs_delayed_ref_unlock(locked_ref);
2672                         locked_ref = NULL;
2673                 }
2674                 btrfs_put_delayed_ref(ref);
2675                 count++;
2676                 cond_resched();
2677         }
2678
2679         /*
2680          * We don't want to include ref heads since we can have empty ref heads
2681          * and those will drastically skew our runtime down since we just do
2682          * accounting, no actual extent tree updates.
2683          */
2684         if (actual_count > 0) {
2685                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2686                 u64 avg;
2687
2688                 /*
2689                  * We weigh the current average higher than our current runtime
2690                  * to avoid large swings in the average.
2691                  */
2692                 spin_lock(&delayed_refs->lock);
2693                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2694                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2695                 spin_unlock(&delayed_refs->lock);
2696         }
2697         return 0;
2698 }
2699
2700 #ifdef SCRAMBLE_DELAYED_REFS
2701 /*
2702  * Normally delayed refs get processed in ascending bytenr order. This
2703  * correlates in most cases to the order added. To expose dependencies on this
2704  * order, we start to process the tree in the middle instead of the beginning
2705  */
2706 static u64 find_middle(struct rb_root *root)
2707 {
2708         struct rb_node *n = root->rb_node;
2709         struct btrfs_delayed_ref_node *entry;
2710         int alt = 1;
2711         u64 middle;
2712         u64 first = 0, last = 0;
2713
2714         n = rb_first(root);
2715         if (n) {
2716                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2717                 first = entry->bytenr;
2718         }
2719         n = rb_last(root);
2720         if (n) {
2721                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2722                 last = entry->bytenr;
2723         }
2724         n = root->rb_node;
2725
2726         while (n) {
2727                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2728                 WARN_ON(!entry->in_tree);
2729
2730                 middle = entry->bytenr;
2731
2732                 if (alt)
2733                         n = n->rb_left;
2734                 else
2735                         n = n->rb_right;
2736
2737                 alt = 1 - alt;
2738         }
2739         return middle;
2740 }
2741 #endif
2742
2743 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2744 {
2745         u64 num_bytes;
2746
2747         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2748                              sizeof(struct btrfs_extent_inline_ref));
2749         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2750                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2751
2752         /*
2753          * We don't ever fill up leaves all the way so multiply by 2 just to be
2754          * closer to what we're really going to want to use.
2755          */
2756         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2757 }
2758
2759 /*
2760  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2761  * would require to store the csums for that many bytes.
2762  */
2763 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2764 {
2765         u64 csum_size;
2766         u64 num_csums_per_leaf;
2767         u64 num_csums;
2768
2769         csum_size = BTRFS_MAX_ITEM_SIZE(root);
2770         num_csums_per_leaf = div64_u64(csum_size,
2771                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2772         num_csums = div64_u64(csum_bytes, root->sectorsize);
2773         num_csums += num_csums_per_leaf - 1;
2774         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2775         return num_csums;
2776 }
2777
2778 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2779                                        struct btrfs_root *root)
2780 {
2781         struct btrfs_block_rsv *global_rsv;
2782         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2783         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2784         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2785         u64 num_bytes, num_dirty_bgs_bytes;
2786         int ret = 0;
2787
2788         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2789         num_heads = heads_to_leaves(root, num_heads);
2790         if (num_heads > 1)
2791                 num_bytes += (num_heads - 1) * root->nodesize;
2792         num_bytes <<= 1;
2793         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2794         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2795                                                              num_dirty_bgs);
2796         global_rsv = &root->fs_info->global_block_rsv;
2797
2798         /*
2799          * If we can't allocate any more chunks lets make sure we have _lots_ of
2800          * wiggle room since running delayed refs can create more delayed refs.
2801          */
2802         if (global_rsv->space_info->full) {
2803                 num_dirty_bgs_bytes <<= 1;
2804                 num_bytes <<= 1;
2805         }
2806
2807         spin_lock(&global_rsv->lock);
2808         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2809                 ret = 1;
2810         spin_unlock(&global_rsv->lock);
2811         return ret;
2812 }
2813
2814 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2815                                        struct btrfs_root *root)
2816 {
2817         struct btrfs_fs_info *fs_info = root->fs_info;
2818         u64 num_entries =
2819                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2820         u64 avg_runtime;
2821         u64 val;
2822
2823         smp_mb();
2824         avg_runtime = fs_info->avg_delayed_ref_runtime;
2825         val = num_entries * avg_runtime;
2826         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2827                 return 1;
2828         if (val >= NSEC_PER_SEC / 2)
2829                 return 2;
2830
2831         return btrfs_check_space_for_delayed_refs(trans, root);
2832 }
2833
2834 struct async_delayed_refs {
2835         struct btrfs_root *root;
2836         u64 transid;
2837         int count;
2838         int error;
2839         int sync;
2840         struct completion wait;
2841         struct btrfs_work work;
2842 };
2843
2844 static void delayed_ref_async_start(struct btrfs_work *work)
2845 {
2846         struct async_delayed_refs *async;
2847         struct btrfs_trans_handle *trans;
2848         int ret;
2849
2850         async = container_of(work, struct async_delayed_refs, work);
2851
2852         /* if the commit is already started, we don't need to wait here */
2853         if (btrfs_transaction_blocked(async->root->fs_info))
2854                 goto done;
2855
2856         trans = btrfs_join_transaction(async->root);
2857         if (IS_ERR(trans)) {
2858                 async->error = PTR_ERR(trans);
2859                 goto done;
2860         }
2861
2862         /*
2863          * trans->sync means that when we call end_transaction, we won't
2864          * wait on delayed refs
2865          */
2866         trans->sync = true;
2867
2868         /* Don't bother flushing if we got into a different transaction */
2869         if (trans->transid > async->transid)
2870                 goto end;
2871
2872         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2873         if (ret)
2874                 async->error = ret;
2875 end:
2876         ret = btrfs_end_transaction(trans, async->root);
2877         if (ret && !async->error)
2878                 async->error = ret;
2879 done:
2880         if (async->sync)
2881                 complete(&async->wait);
2882         else
2883                 kfree(async);
2884 }
2885
2886 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2887                                  unsigned long count, u64 transid, int wait)
2888 {
2889         struct async_delayed_refs *async;
2890         int ret;
2891
2892         async = kmalloc(sizeof(*async), GFP_NOFS);
2893         if (!async)
2894                 return -ENOMEM;
2895
2896         async->root = root->fs_info->tree_root;
2897         async->count = count;
2898         async->error = 0;
2899         async->transid = transid;
2900         if (wait)
2901                 async->sync = 1;
2902         else
2903                 async->sync = 0;
2904         init_completion(&async->wait);
2905
2906         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2907                         delayed_ref_async_start, NULL, NULL);
2908
2909         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2910
2911         if (wait) {
2912                 wait_for_completion(&async->wait);
2913                 ret = async->error;
2914                 kfree(async);
2915                 return ret;
2916         }
2917         return 0;
2918 }
2919
2920 /*
2921  * this starts processing the delayed reference count updates and
2922  * extent insertions we have queued up so far.  count can be
2923  * 0, which means to process everything in the tree at the start
2924  * of the run (but not newly added entries), or it can be some target
2925  * number you'd like to process.
2926  *
2927  * Returns 0 on success or if called with an aborted transaction
2928  * Returns <0 on error and aborts the transaction
2929  */
2930 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2931                            struct btrfs_root *root, unsigned long count)
2932 {
2933         struct rb_node *node;
2934         struct btrfs_delayed_ref_root *delayed_refs;
2935         struct btrfs_delayed_ref_head *head;
2936         int ret;
2937         int run_all = count == (unsigned long)-1;
2938         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2939
2940         /* We'll clean this up in btrfs_cleanup_transaction */
2941         if (trans->aborted)
2942                 return 0;
2943
2944         if (test_bit(BTRFS_FS_CREATING_FREE_SPACE_TREE, &root->fs_info->flags))
2945                 return 0;
2946
2947         if (root == root->fs_info->extent_root)
2948                 root = root->fs_info->tree_root;
2949
2950         delayed_refs = &trans->transaction->delayed_refs;
2951         if (count == 0)
2952                 count = atomic_read(&delayed_refs->num_entries) * 2;
2953
2954 again:
2955 #ifdef SCRAMBLE_DELAYED_REFS
2956         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2957 #endif
2958         trans->can_flush_pending_bgs = false;
2959         ret = __btrfs_run_delayed_refs(trans, root, count);
2960         if (ret < 0) {
2961                 btrfs_abort_transaction(trans, ret);
2962                 return ret;
2963         }
2964
2965         if (run_all) {
2966                 if (!list_empty(&trans->new_bgs))
2967                         btrfs_create_pending_block_groups(trans, root);
2968
2969                 spin_lock(&delayed_refs->lock);
2970                 node = rb_first(&delayed_refs->href_root);
2971                 if (!node) {
2972                         spin_unlock(&delayed_refs->lock);
2973                         goto out;
2974                 }
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         } else if (ret < 0) {
3698                 btrfs_cleanup_dirty_bgs(cur_trans, root);
3699         }
3700
3701         btrfs_free_path(path);
3702         return ret;
3703 }
3704
3705 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3706                                    struct btrfs_root *root)
3707 {
3708         struct btrfs_block_group_cache *cache;
3709         struct btrfs_transaction *cur_trans = trans->transaction;
3710         int ret = 0;
3711         int should_put;
3712         struct btrfs_path *path;
3713         struct list_head *io = &cur_trans->io_bgs;
3714         int num_started = 0;
3715
3716         path = btrfs_alloc_path();
3717         if (!path)
3718                 return -ENOMEM;
3719
3720         /*
3721          * Even though we are in the critical section of the transaction commit,
3722          * we can still have concurrent tasks adding elements to this
3723          * transaction's list of dirty block groups. These tasks correspond to
3724          * endio free space workers started when writeback finishes for a
3725          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3726          * allocate new block groups as a result of COWing nodes of the root
3727          * tree when updating the free space inode. The writeback for the space
3728          * caches is triggered by an earlier call to
3729          * btrfs_start_dirty_block_groups() and iterations of the following
3730          * loop.
3731          * Also we want to do the cache_save_setup first and then run the
3732          * delayed refs to make sure we have the best chance at doing this all
3733          * in one shot.
3734          */
3735         spin_lock(&cur_trans->dirty_bgs_lock);
3736         while (!list_empty(&cur_trans->dirty_bgs)) {
3737                 cache = list_first_entry(&cur_trans->dirty_bgs,
3738                                          struct btrfs_block_group_cache,
3739                                          dirty_list);
3740
3741                 /*
3742                  * this can happen if cache_save_setup re-dirties a block
3743                  * group that is already under IO.  Just wait for it to
3744                  * finish and then do it all again
3745                  */
3746                 if (!list_empty(&cache->io_list)) {
3747                         spin_unlock(&cur_trans->dirty_bgs_lock);
3748                         list_del_init(&cache->io_list);
3749                         btrfs_wait_cache_io(root, trans, cache,
3750                                             &cache->io_ctl, path,
3751                                             cache->key.objectid);
3752                         btrfs_put_block_group(cache);
3753                         spin_lock(&cur_trans->dirty_bgs_lock);
3754                 }
3755
3756                 /*
3757                  * don't remove from the dirty list until after we've waited
3758                  * on any pending IO
3759                  */
3760                 list_del_init(&cache->dirty_list);
3761                 spin_unlock(&cur_trans->dirty_bgs_lock);
3762                 should_put = 1;
3763
3764                 cache_save_setup(cache, trans, path);
3765
3766                 if (!ret)
3767                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3768
3769                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3770                         cache->io_ctl.inode = NULL;
3771                         ret = btrfs_write_out_cache(root, trans, cache, path);
3772                         if (ret == 0 && cache->io_ctl.inode) {
3773                                 num_started++;
3774                                 should_put = 0;
3775                                 list_add_tail(&cache->io_list, io);
3776                         } else {
3777                                 /*
3778                                  * if we failed to write the cache, the
3779                                  * generation will be bad and life goes on
3780                                  */
3781                                 ret = 0;
3782                         }
3783                 }
3784                 if (!ret) {
3785                         ret = write_one_cache_group(trans, root, path, cache);
3786                         /*
3787                          * One of the free space endio workers might have
3788                          * created a new block group while updating a free space
3789                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3790                          * and hasn't released its transaction handle yet, in
3791                          * which case the new block group is still attached to
3792                          * its transaction handle and its creation has not
3793                          * finished yet (no block group item in the extent tree
3794                          * yet, etc). If this is the case, wait for all free
3795                          * space endio workers to finish and retry. This is a
3796                          * a very rare case so no need for a more efficient and
3797                          * complex approach.
3798                          */
3799                         if (ret == -ENOENT) {
3800                                 wait_event(cur_trans->writer_wait,
3801                                    atomic_read(&cur_trans->num_writers) == 1);
3802                                 ret = write_one_cache_group(trans, root, path,
3803                                                             cache);
3804                         }
3805                         if (ret)
3806                                 btrfs_abort_transaction(trans, ret);
3807                 }
3808
3809                 /* if its not on the io list, we need to put the block group */
3810                 if (should_put)
3811                         btrfs_put_block_group(cache);
3812                 spin_lock(&cur_trans->dirty_bgs_lock);
3813         }
3814         spin_unlock(&cur_trans->dirty_bgs_lock);
3815
3816         while (!list_empty(io)) {
3817                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3818                                          io_list);
3819                 list_del_init(&cache->io_list);
3820                 btrfs_wait_cache_io(root, trans, cache,
3821                                     &cache->io_ctl, path, cache->key.objectid);
3822                 btrfs_put_block_group(cache);
3823         }
3824
3825         btrfs_free_path(path);
3826         return ret;
3827 }
3828
3829 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3830 {
3831         struct btrfs_block_group_cache *block_group;
3832         int readonly = 0;
3833
3834         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3835         if (!block_group || block_group->ro)
3836                 readonly = 1;
3837         if (block_group)
3838                 btrfs_put_block_group(block_group);
3839         return readonly;
3840 }
3841
3842 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3843 {
3844         struct btrfs_block_group_cache *bg;
3845         bool ret = true;
3846
3847         bg = btrfs_lookup_block_group(fs_info, bytenr);
3848         if (!bg)
3849                 return false;
3850
3851         spin_lock(&bg->lock);
3852         if (bg->ro)
3853                 ret = false;
3854         else
3855                 atomic_inc(&bg->nocow_writers);
3856         spin_unlock(&bg->lock);
3857
3858         /* no put on block group, done by btrfs_dec_nocow_writers */
3859         if (!ret)
3860                 btrfs_put_block_group(bg);
3861
3862         return ret;
3863
3864 }
3865
3866 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3867 {
3868         struct btrfs_block_group_cache *bg;
3869
3870         bg = btrfs_lookup_block_group(fs_info, bytenr);
3871         ASSERT(bg);
3872         if (atomic_dec_and_test(&bg->nocow_writers))
3873                 wake_up_atomic_t(&bg->nocow_writers);
3874         /*
3875          * Once for our lookup and once for the lookup done by a previous call
3876          * to btrfs_inc_nocow_writers()
3877          */
3878         btrfs_put_block_group(bg);
3879         btrfs_put_block_group(bg);
3880 }
3881
3882 static int btrfs_wait_nocow_writers_atomic_t(atomic_t *a)
3883 {
3884         schedule();
3885         return 0;
3886 }
3887
3888 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3889 {
3890         wait_on_atomic_t(&bg->nocow_writers,
3891                          btrfs_wait_nocow_writers_atomic_t,
3892                          TASK_UNINTERRUPTIBLE);
3893 }
3894
3895 static const char *alloc_name(u64 flags)
3896 {
3897         switch (flags) {
3898         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3899                 return "mixed";
3900         case BTRFS_BLOCK_GROUP_METADATA:
3901                 return "metadata";
3902         case BTRFS_BLOCK_GROUP_DATA:
3903                 return "data";
3904         case BTRFS_BLOCK_GROUP_SYSTEM:
3905                 return "system";
3906         default:
3907                 WARN_ON(1);
3908                 return "invalid-combination";
3909         };
3910 }
3911
3912 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3913                              u64 total_bytes, u64 bytes_used,
3914                              u64 bytes_readonly,
3915                              struct btrfs_space_info **space_info)
3916 {
3917         struct btrfs_space_info *found;
3918         int i;
3919         int factor;
3920         int ret;
3921
3922         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3923                      BTRFS_BLOCK_GROUP_RAID10))
3924                 factor = 2;
3925         else
3926                 factor = 1;
3927
3928         found = __find_space_info(info, flags);
3929         if (found) {
3930                 spin_lock(&found->lock);
3931                 found->total_bytes += total_bytes;
3932                 found->disk_total += total_bytes * factor;
3933                 found->bytes_used += bytes_used;
3934                 found->disk_used += bytes_used * factor;
3935                 found->bytes_readonly += bytes_readonly;
3936                 if (total_bytes > 0)
3937                         found->full = 0;
3938                 space_info_add_new_bytes(info, found, total_bytes -
3939                                          bytes_used - bytes_readonly);
3940                 spin_unlock(&found->lock);
3941                 *space_info = found;
3942                 return 0;
3943         }
3944         found = kzalloc(sizeof(*found), GFP_NOFS);
3945         if (!found)
3946                 return -ENOMEM;
3947
3948         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3949         if (ret) {
3950                 kfree(found);
3951                 return ret;
3952         }
3953
3954         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3955                 INIT_LIST_HEAD(&found->block_groups[i]);
3956         init_rwsem(&found->groups_sem);
3957         spin_lock_init(&found->lock);
3958         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3959         found->total_bytes = total_bytes;
3960         found->disk_total = total_bytes * factor;
3961         found->bytes_used = bytes_used;
3962         found->disk_used = bytes_used * factor;
3963         found->bytes_pinned = 0;
3964         found->bytes_reserved = 0;
3965         found->bytes_readonly = bytes_readonly;
3966         found->bytes_may_use = 0;
3967         found->full = 0;
3968         found->max_extent_size = 0;
3969         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3970         found->chunk_alloc = 0;
3971         found->flush = 0;
3972         init_waitqueue_head(&found->wait);
3973         INIT_LIST_HEAD(&found->ro_bgs);
3974         INIT_LIST_HEAD(&found->tickets);
3975         INIT_LIST_HEAD(&found->priority_tickets);
3976
3977         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3978                                     info->space_info_kobj, "%s",
3979                                     alloc_name(found->flags));
3980         if (ret) {
3981                 kfree(found);
3982                 return ret;
3983         }
3984
3985         *space_info = found;
3986         list_add_rcu(&found->list, &info->space_info);
3987         if (flags & BTRFS_BLOCK_GROUP_DATA)
3988                 info->data_sinfo = found;
3989
3990         return ret;
3991 }
3992
3993 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3994 {
3995         u64 extra_flags = chunk_to_extended(flags) &
3996                                 BTRFS_EXTENDED_PROFILE_MASK;
3997
3998         write_seqlock(&fs_info->profiles_lock);
3999         if (flags & BTRFS_BLOCK_GROUP_DATA)
4000                 fs_info->avail_data_alloc_bits |= extra_flags;
4001         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4002                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4003         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4004                 fs_info->avail_system_alloc_bits |= extra_flags;
4005         write_sequnlock(&fs_info->profiles_lock);
4006 }
4007
4008 /*
4009  * returns target flags in extended format or 0 if restripe for this
4010  * chunk_type is not in progress
4011  *
4012  * should be called with either volume_mutex or balance_lock held
4013  */
4014 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4015 {
4016         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4017         u64 target = 0;
4018
4019         if (!bctl)
4020                 return 0;
4021
4022         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4023             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4024                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4025         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4026                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4027                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4028         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4029                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4030                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4031         }
4032
4033         return target;
4034 }
4035
4036 /*
4037  * @flags: available profiles in extended format (see ctree.h)
4038  *
4039  * Returns reduced profile in chunk format.  If profile changing is in
4040  * progress (either running or paused) picks the target profile (if it's
4041  * already available), otherwise falls back to plain reducing.
4042  */
4043 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
4044 {
4045         u64 num_devices = root->fs_info->fs_devices->rw_devices;
4046         u64 target;
4047         u64 raid_type;
4048         u64 allowed = 0;
4049
4050         /*
4051          * see if restripe for this chunk_type is in progress, if so
4052          * try to reduce to the target profile
4053          */
4054         spin_lock(&root->fs_info->balance_lock);
4055         target = get_restripe_target(root->fs_info, flags);
4056         if (target) {
4057                 /* pick target profile only if it's already available */
4058                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4059                         spin_unlock(&root->fs_info->balance_lock);
4060                         return extended_to_chunk(target);
4061                 }
4062         }
4063         spin_unlock(&root->fs_info->balance_lock);
4064
4065         /* First, mask out the RAID levels which aren't possible */
4066         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4067                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4068                         allowed |= btrfs_raid_group[raid_type];
4069         }
4070         allowed &= flags;
4071
4072         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4073                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4074         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4075                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4076         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4077                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4078         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4079                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4080         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4081                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4082
4083         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4084
4085         return extended_to_chunk(flags | allowed);
4086 }
4087
4088 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
4089 {
4090         unsigned seq;
4091         u64 flags;
4092
4093         do {
4094                 flags = orig_flags;
4095                 seq = read_seqbegin(&root->fs_info->profiles_lock);
4096
4097                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4098                         flags |= root->fs_info->avail_data_alloc_bits;
4099                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4100                         flags |= root->fs_info->avail_system_alloc_bits;
4101                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4102                         flags |= root->fs_info->avail_metadata_alloc_bits;
4103         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
4104
4105         return btrfs_reduce_alloc_profile(root, flags);
4106 }
4107
4108 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
4109 {
4110         u64 flags;
4111         u64 ret;
4112
4113         if (data)
4114                 flags = BTRFS_BLOCK_GROUP_DATA;
4115         else if (root == root->fs_info->chunk_root)
4116                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4117         else
4118                 flags = BTRFS_BLOCK_GROUP_METADATA;
4119
4120         ret = get_alloc_profile(root, flags);
4121         return ret;
4122 }
4123
4124 int btrfs_alloc_data_chunk_ondemand(struct inode *inode, u64 bytes)
4125 {
4126         struct btrfs_space_info *data_sinfo;
4127         struct btrfs_root *root = BTRFS_I(inode)->root;
4128         struct btrfs_fs_info *fs_info = root->fs_info;
4129         u64 used;
4130         int ret = 0;
4131         int need_commit = 2;
4132         int have_pinned_space;
4133
4134         /* make sure bytes are sectorsize aligned */
4135         bytes = ALIGN(bytes, root->sectorsize);
4136
4137         if (btrfs_is_free_space_inode(inode)) {
4138                 need_commit = 0;
4139                 ASSERT(current->journal_info);
4140         }
4141
4142         data_sinfo = fs_info->data_sinfo;
4143         if (!data_sinfo)
4144                 goto alloc;
4145
4146 again:
4147         /* make sure we have enough space to handle the data first */
4148         spin_lock(&data_sinfo->lock);
4149         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
4150                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
4151                 data_sinfo->bytes_may_use;
4152
4153         if (used + bytes > data_sinfo->total_bytes) {
4154                 struct btrfs_trans_handle *trans;
4155
4156                 /*
4157                  * if we don't have enough free bytes in this space then we need
4158                  * to alloc a new chunk.
4159                  */
4160                 if (!data_sinfo->full) {
4161                         u64 alloc_target;
4162
4163                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4164                         spin_unlock(&data_sinfo->lock);
4165 alloc:
4166                         alloc_target = btrfs_get_alloc_profile(root, 1);
4167                         /*
4168                          * It is ugly that we don't call nolock join
4169                          * transaction for the free space inode case here.
4170                          * But it is safe because we only do the data space
4171                          * reservation for the free space cache in the
4172                          * transaction context, the common join transaction
4173                          * just increase the counter of the current transaction
4174                          * handler, doesn't try to acquire the trans_lock of
4175                          * the fs.
4176                          */
4177                         trans = btrfs_join_transaction(root);
4178                         if (IS_ERR(trans))
4179                                 return PTR_ERR(trans);
4180
4181                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4182                                              alloc_target,
4183                                              CHUNK_ALLOC_NO_FORCE);
4184                         btrfs_end_transaction(trans, root);
4185                         if (ret < 0) {
4186                                 if (ret != -ENOSPC)
4187                                         return ret;
4188                                 else {
4189                                         have_pinned_space = 1;
4190                                         goto commit_trans;
4191                                 }
4192                         }
4193
4194                         if (!data_sinfo)
4195                                 data_sinfo = fs_info->data_sinfo;
4196
4197                         goto again;
4198                 }
4199
4200                 /*
4201                  * If we don't have enough pinned space to deal with this
4202                  * allocation, and no removed chunk in current transaction,
4203                  * don't bother committing the transaction.
4204                  */
4205                 have_pinned_space = percpu_counter_compare(
4206                         &data_sinfo->total_bytes_pinned,
4207                         used + bytes - data_sinfo->total_bytes);
4208                 spin_unlock(&data_sinfo->lock);
4209
4210                 /* commit the current transaction and try again */
4211 commit_trans:
4212                 if (need_commit &&
4213                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
4214                         need_commit--;
4215
4216                         if (need_commit > 0) {
4217                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4218                                 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
4219                         }
4220
4221                         trans = btrfs_join_transaction(root);
4222                         if (IS_ERR(trans))
4223                                 return PTR_ERR(trans);
4224                         if (have_pinned_space >= 0 ||
4225                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4226                                      &trans->transaction->flags) ||
4227                             need_commit > 0) {
4228                                 ret = btrfs_commit_transaction(trans, root);
4229                                 if (ret)
4230                                         return ret;
4231                                 /*
4232                                  * The cleaner kthread might still be doing iput
4233                                  * operations. Wait for it to finish so that
4234                                  * more space is released.
4235                                  */
4236                                 mutex_lock(&root->fs_info->cleaner_delayed_iput_mutex);
4237                                 mutex_unlock(&root->fs_info->cleaner_delayed_iput_mutex);
4238                                 goto again;
4239                         } else {
4240                                 btrfs_end_transaction(trans, root);
4241                         }
4242                 }
4243
4244                 trace_btrfs_space_reservation(root->fs_info,
4245                                               "space_info:enospc",
4246                                               data_sinfo->flags, bytes, 1);
4247                 return -ENOSPC;
4248         }
4249         data_sinfo->bytes_may_use += bytes;
4250         trace_btrfs_space_reservation(root->fs_info, "space_info",
4251                                       data_sinfo->flags, bytes, 1);
4252         spin_unlock(&data_sinfo->lock);
4253
4254         return ret;
4255 }
4256
4257 /*
4258  * New check_data_free_space() with ability for precious data reservation
4259  * Will replace old btrfs_check_data_free_space(), but for patch split,
4260  * add a new function first and then replace it.
4261  */
4262 int btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4263 {
4264         struct btrfs_root *root = BTRFS_I(inode)->root;
4265         int ret;
4266
4267         /* align the range */
4268         len = round_up(start + len, root->sectorsize) -
4269               round_down(start, root->sectorsize);
4270         start = round_down(start, root->sectorsize);
4271
4272         ret = btrfs_alloc_data_chunk_ondemand(inode, len);
4273         if (ret < 0)
4274                 return ret;
4275
4276         /* Use new btrfs_qgroup_reserve_data to reserve precious data space. */
4277         ret = btrfs_qgroup_reserve_data(inode, start, len);
4278         if (ret)
4279                 btrfs_free_reserved_data_space_noquota(inode, start, len);
4280         return ret;
4281 }
4282
4283 /*
4284  * Called if we need to clear a data reservation for this inode
4285  * Normally in a error case.
4286  *
4287  * This one will *NOT* use accurate qgroup reserved space API, just for case
4288  * which we can't sleep and is sure it won't affect qgroup reserved space.
4289  * Like clear_bit_hook().
4290  */
4291 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4292                                             u64 len)
4293 {
4294         struct btrfs_root *root = BTRFS_I(inode)->root;
4295         struct btrfs_space_info *data_sinfo;
4296
4297         /* Make sure the range is aligned to sectorsize */
4298         len = round_up(start + len, root->sectorsize) -
4299               round_down(start, root->sectorsize);
4300         start = round_down(start, root->sectorsize);
4301
4302         data_sinfo = root->fs_info->data_sinfo;
4303         spin_lock(&data_sinfo->lock);
4304         if (WARN_ON(data_sinfo->bytes_may_use < len))
4305                 data_sinfo->bytes_may_use = 0;
4306         else
4307                 data_sinfo->bytes_may_use -= len;
4308         trace_btrfs_space_reservation(root->fs_info, "space_info",
4309                                       data_sinfo->flags, len, 0);
4310         spin_unlock(&data_sinfo->lock);
4311 }
4312
4313 /*
4314  * Called if we need to clear a data reservation for this inode
4315  * Normally in a error case.
4316  *
4317  * This one will handle the per-inode data rsv map for accurate reserved
4318  * space framework.
4319  */
4320 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4321 {
4322         btrfs_free_reserved_data_space_noquota(inode, start, len);
4323         btrfs_qgroup_free_data(inode, start, len);
4324 }
4325
4326 static void force_metadata_allocation(struct btrfs_fs_info *info)
4327 {
4328         struct list_head *head = &info->space_info;
4329         struct btrfs_space_info *found;
4330
4331         rcu_read_lock();
4332         list_for_each_entry_rcu(found, head, list) {
4333                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4334                         found->force_alloc = CHUNK_ALLOC_FORCE;
4335         }
4336         rcu_read_unlock();
4337 }
4338
4339 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4340 {
4341         return (global->size << 1);
4342 }
4343
4344 static int should_alloc_chunk(struct btrfs_root *root,
4345                               struct btrfs_space_info *sinfo, int force)
4346 {
4347         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4348         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4349         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4350         u64 thresh;
4351
4352         if (force == CHUNK_ALLOC_FORCE)
4353                 return 1;
4354
4355         /*
4356          * We need to take into account the global rsv because for all intents
4357          * and purposes it's used space.  Don't worry about locking the
4358          * global_rsv, it doesn't change except when the transaction commits.
4359          */
4360         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4361                 num_allocated += calc_global_rsv_need_space(global_rsv);
4362
4363         /*
4364          * in limited mode, we want to have some free space up to
4365          * about 1% of the FS size.
4366          */
4367         if (force == CHUNK_ALLOC_LIMITED) {
4368                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4369                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4370
4371                 if (num_bytes - num_allocated < thresh)
4372                         return 1;
4373         }
4374
4375         if (num_allocated + SZ_2M < div_factor(num_bytes, 8))
4376                 return 0;
4377         return 1;
4378 }
4379
4380 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4381 {
4382         u64 num_dev;
4383
4384         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4385                     BTRFS_BLOCK_GROUP_RAID0 |
4386                     BTRFS_BLOCK_GROUP_RAID5 |
4387                     BTRFS_BLOCK_GROUP_RAID6))
4388                 num_dev = root->fs_info->fs_devices->rw_devices;
4389         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4390                 num_dev = 2;
4391         else
4392                 num_dev = 1;    /* DUP or single */
4393
4394         return num_dev;
4395 }
4396
4397 /*
4398  * If @is_allocation is true, reserve space in the system space info necessary
4399  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4400  * removing a chunk.
4401  */
4402 void check_system_chunk(struct btrfs_trans_handle *trans,
4403                         struct btrfs_root *root,
4404                         u64 type)
4405 {
4406         struct btrfs_space_info *info;
4407         u64 left;
4408         u64 thresh;
4409         int ret = 0;
4410         u64 num_devs;
4411
4412         /*
4413          * Needed because we can end up allocating a system chunk and for an
4414          * atomic and race free space reservation in the chunk block reserve.
4415          */
4416         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4417
4418         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4419         spin_lock(&info->lock);
4420         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4421                 info->bytes_reserved - info->bytes_readonly -
4422                 info->bytes_may_use;
4423         spin_unlock(&info->lock);
4424
4425         num_devs = get_profile_num_devs(root, type);
4426
4427         /* num_devs device items to update and 1 chunk item to add or remove */
4428         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4429                 btrfs_calc_trans_metadata_size(root, 1);
4430
4431         if (left < thresh && btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
4432                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4433                         left, thresh, type);
4434                 dump_space_info(root->fs_info, info, 0, 0);
4435         }
4436
4437         if (left < thresh) {
4438                 u64 flags;
4439
4440                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4441                 /*
4442                  * Ignore failure to create system chunk. We might end up not
4443                  * needing it, as we might not need to COW all nodes/leafs from
4444                  * the paths we visit in the chunk tree (they were already COWed
4445                  * or created in the current transaction for example).
4446                  */
4447                 ret = btrfs_alloc_chunk(trans, root, flags);
4448         }
4449
4450         if (!ret) {
4451                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4452                                           &root->fs_info->chunk_block_rsv,
4453                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4454                 if (!ret)
4455                         trans->chunk_bytes_reserved += thresh;
4456         }
4457 }
4458
4459 /*
4460  * If force is CHUNK_ALLOC_FORCE:
4461  *    - return 1 if it successfully allocates a chunk,
4462  *    - return errors including -ENOSPC otherwise.
4463  * If force is NOT CHUNK_ALLOC_FORCE:
4464  *    - return 0 if it doesn't need to allocate a new chunk,
4465  *    - return 1 if it successfully allocates a chunk,
4466  *    - return errors including -ENOSPC otherwise.
4467  */
4468 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4469                           struct btrfs_root *extent_root, u64 flags, int force)
4470 {
4471         struct btrfs_space_info *space_info;
4472         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4473         int wait_for_alloc = 0;
4474         int ret = 0;
4475
4476         /* Don't re-enter if we're already allocating a chunk */
4477         if (trans->allocating_chunk)
4478                 return -ENOSPC;
4479
4480         space_info = __find_space_info(extent_root->fs_info, flags);
4481         if (!space_info) {
4482                 ret = update_space_info(extent_root->fs_info, flags,
4483                                         0, 0, 0, &space_info);
4484                 BUG_ON(ret); /* -ENOMEM */
4485         }
4486         BUG_ON(!space_info); /* Logic error */
4487
4488 again:
4489         spin_lock(&space_info->lock);
4490         if (force < space_info->force_alloc)
4491                 force = space_info->force_alloc;
4492         if (space_info->full) {
4493                 if (should_alloc_chunk(extent_root, space_info, force))
4494                         ret = -ENOSPC;
4495                 else
4496                         ret = 0;
4497                 spin_unlock(&space_info->lock);
4498                 return ret;
4499         }
4500
4501         if (!should_alloc_chunk(extent_root, space_info, force)) {
4502                 spin_unlock(&space_info->lock);
4503                 return 0;
4504         } else if (space_info->chunk_alloc) {
4505                 wait_for_alloc = 1;
4506         } else {
4507                 space_info->chunk_alloc = 1;
4508         }
4509
4510         spin_unlock(&space_info->lock);
4511
4512         mutex_lock(&fs_info->chunk_mutex);
4513
4514         /*
4515          * The chunk_mutex is held throughout the entirety of a chunk
4516          * allocation, so once we've acquired the chunk_mutex we know that the
4517          * other guy is done and we need to recheck and see if we should
4518          * allocate.
4519          */
4520         if (wait_for_alloc) {
4521                 mutex_unlock(&fs_info->chunk_mutex);
4522                 wait_for_alloc = 0;
4523                 goto again;
4524         }
4525
4526         trans->allocating_chunk = true;
4527
4528         /*
4529          * If we have mixed data/metadata chunks we want to make sure we keep
4530          * allocating mixed chunks instead of individual chunks.
4531          */
4532         if (btrfs_mixed_space_info(space_info))
4533                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4534
4535         /*
4536          * if we're doing a data chunk, go ahead and make sure that
4537          * we keep a reasonable number of metadata chunks allocated in the
4538          * FS as well.
4539          */
4540         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4541                 fs_info->data_chunk_allocations++;
4542                 if (!(fs_info->data_chunk_allocations %
4543                       fs_info->metadata_ratio))
4544                         force_metadata_allocation(fs_info);
4545         }
4546
4547         /*
4548          * Check if we have enough space in SYSTEM chunk because we may need
4549          * to update devices.
4550          */
4551         check_system_chunk(trans, extent_root, flags);
4552
4553         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4554         trans->allocating_chunk = false;
4555
4556         spin_lock(&space_info->lock);
4557         if (ret < 0 && ret != -ENOSPC)
4558                 goto out;
4559         if (ret)
4560                 space_info->full = 1;
4561         else
4562                 ret = 1;
4563
4564         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4565 out:
4566         space_info->chunk_alloc = 0;
4567         spin_unlock(&space_info->lock);
4568         mutex_unlock(&fs_info->chunk_mutex);
4569         /*
4570          * When we allocate a new chunk we reserve space in the chunk block
4571          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4572          * add new nodes/leafs to it if we end up needing to do it when
4573          * inserting the chunk item and updating device items as part of the
4574          * second phase of chunk allocation, performed by
4575          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4576          * large number of new block groups to create in our transaction
4577          * handle's new_bgs list to avoid exhausting the chunk block reserve
4578          * in extreme cases - like having a single transaction create many new
4579          * block groups when starting to write out the free space caches of all
4580          * the block groups that were made dirty during the lifetime of the
4581          * transaction.
4582          */
4583         if (trans->can_flush_pending_bgs &&
4584             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4585                 btrfs_create_pending_block_groups(trans, extent_root);
4586                 btrfs_trans_release_chunk_metadata(trans);
4587         }
4588         return ret;
4589 }
4590
4591 static int can_overcommit(struct btrfs_root *root,
4592                           struct btrfs_space_info *space_info, u64 bytes,
4593                           enum btrfs_reserve_flush_enum flush)
4594 {
4595         struct btrfs_block_rsv *global_rsv;
4596         u64 profile;
4597         u64 space_size;
4598         u64 avail;
4599         u64 used;
4600
4601         /* Don't overcommit when in mixed mode. */
4602         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4603                 return 0;
4604
4605         BUG_ON(root->fs_info == NULL);
4606         global_rsv = &root->fs_info->global_block_rsv;
4607         profile = btrfs_get_alloc_profile(root, 0);
4608         used = space_info->bytes_used + space_info->bytes_reserved +
4609                 space_info->bytes_pinned + space_info->bytes_readonly;
4610
4611         /*
4612          * We only want to allow over committing if we have lots of actual space
4613          * free, but if we don't have enough space to handle the global reserve
4614          * space then we could end up having a real enospc problem when trying
4615          * to allocate a chunk or some other such important allocation.
4616          */
4617         spin_lock(&global_rsv->lock);
4618         space_size = calc_global_rsv_need_space(global_rsv);
4619         spin_unlock(&global_rsv->lock);
4620         if (used + space_size >= space_info->total_bytes)
4621                 return 0;
4622
4623         used += space_info->bytes_may_use;
4624
4625         spin_lock(&root->fs_info->free_chunk_lock);
4626         avail = root->fs_info->free_chunk_space;
4627         spin_unlock(&root->fs_info->free_chunk_lock);
4628
4629         /*
4630          * If we have dup, raid1 or raid10 then only half of the free
4631          * space is actually useable.  For raid56, the space info used
4632          * doesn't include the parity drive, so we don't have to
4633          * change the math
4634          */
4635         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4636                        BTRFS_BLOCK_GROUP_RAID1 |
4637                        BTRFS_BLOCK_GROUP_RAID10))
4638                 avail >>= 1;
4639
4640         /*
4641          * If we aren't flushing all things, let us overcommit up to
4642          * 1/2th of the space. If we can flush, don't let us overcommit
4643          * too much, let it overcommit up to 1/8 of the space.
4644          */
4645         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4646                 avail >>= 3;
4647         else
4648                 avail >>= 1;
4649
4650         if (used + bytes < space_info->total_bytes + avail)
4651                 return 1;
4652         return 0;
4653 }
4654
4655 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4656                                          unsigned long nr_pages, int nr_items)
4657 {
4658         struct super_block *sb = root->fs_info->sb;
4659
4660         if (down_read_trylock(&sb->s_umount)) {
4661                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4662                 up_read(&sb->s_umount);
4663         } else {
4664                 /*
4665                  * We needn't worry the filesystem going from r/w to r/o though
4666                  * we don't acquire ->s_umount mutex, because the filesystem
4667                  * should guarantee the delalloc inodes list be empty after
4668                  * the filesystem is readonly(all dirty pages are written to
4669                  * the disk).
4670                  */
4671                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4672                 if (!current->journal_info)
4673                         btrfs_wait_ordered_roots(root->fs_info, nr_items,
4674                                                  0, (u64)-1);
4675         }
4676 }
4677
4678 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4679 {
4680         u64 bytes;
4681         int nr;
4682
4683         bytes = btrfs_calc_trans_metadata_size(root, 1);
4684         nr = (int)div64_u64(to_reclaim, bytes);
4685         if (!nr)
4686                 nr = 1;
4687         return nr;
4688 }
4689
4690 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4691
4692 /*
4693  * shrink metadata reservation for delalloc
4694  */
4695 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4696                             bool wait_ordered)
4697 {
4698         struct btrfs_block_rsv *block_rsv;
4699         struct btrfs_space_info *space_info;
4700         struct btrfs_trans_handle *trans;
4701         u64 delalloc_bytes;
4702         u64 max_reclaim;
4703         long time_left;
4704         unsigned long nr_pages;
4705         int loops;
4706         int items;
4707         enum btrfs_reserve_flush_enum flush;
4708
4709         /* Calc the number of the pages we need flush for space reservation */
4710         items = calc_reclaim_items_nr(root, to_reclaim);
4711         to_reclaim = (u64)items * EXTENT_SIZE_PER_ITEM;
4712
4713         trans = (struct btrfs_trans_handle *)current->journal_info;
4714         block_rsv = &root->fs_info->delalloc_block_rsv;
4715         space_info = block_rsv->space_info;
4716
4717         delalloc_bytes = percpu_counter_sum_positive(
4718                                                 &root->fs_info->delalloc_bytes);
4719         if (delalloc_bytes == 0) {
4720                 if (trans)
4721                         return;
4722                 if (wait_ordered)
4723                         btrfs_wait_ordered_roots(root->fs_info, items,
4724                                                  0, (u64)-1);
4725                 return;
4726         }
4727
4728         loops = 0;
4729         while (delalloc_bytes && loops < 3) {
4730                 max_reclaim = min(delalloc_bytes, to_reclaim);
4731                 nr_pages = max_reclaim >> PAGE_SHIFT;
4732                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4733                 /*
4734                  * We need to wait for the async pages to actually start before
4735                  * we do anything.
4736                  */
4737                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4738                 if (!max_reclaim)
4739                         goto skip_async;
4740
4741                 if (max_reclaim <= nr_pages)
4742                         max_reclaim = 0;
4743                 else
4744                         max_reclaim -= nr_pages;
4745
4746                 wait_event(root->fs_info->async_submit_wait,
4747                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4748                            (int)max_reclaim);
4749 skip_async:
4750                 if (!trans)
4751                         flush = BTRFS_RESERVE_FLUSH_ALL;
4752                 else
4753                         flush = BTRFS_RESERVE_NO_FLUSH;
4754                 spin_lock(&space_info->lock);
4755                 if (can_overcommit(root, space_info, orig, flush)) {
4756                         spin_unlock(&space_info->lock);
4757                         break;
4758                 }
4759                 if (list_empty(&space_info->tickets) &&
4760                     list_empty(&space_info->priority_tickets)) {
4761                         spin_unlock(&space_info->lock);
4762                         break;
4763                 }
4764                 spin_unlock(&space_info->lock);
4765
4766                 loops++;
4767                 if (wait_ordered && !trans) {
4768                         btrfs_wait_ordered_roots(root->fs_info, items,
4769                                                  0, (u64)-1);
4770                 } else {
4771                         time_left = schedule_timeout_killable(1);
4772                         if (time_left)
4773                                 break;
4774                 }
4775                 delalloc_bytes = percpu_counter_sum_positive(
4776                                                 &root->fs_info->delalloc_bytes);
4777         }
4778 }
4779
4780 /**
4781  * maybe_commit_transaction - possibly commit the transaction if its ok to
4782  * @root - the root we're allocating for
4783  * @bytes - the number of bytes we want to reserve
4784  * @force - force the commit
4785  *
4786  * This will check to make sure that committing the transaction will actually
4787  * get us somewhere and then commit the transaction if it does.  Otherwise it
4788  * will return -ENOSPC.
4789  */
4790 static int may_commit_transaction(struct btrfs_root *root,
4791                                   struct btrfs_space_info *space_info,
4792                                   u64 bytes, int force)
4793 {
4794         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4795         struct btrfs_trans_handle *trans;
4796
4797         trans = (struct btrfs_trans_handle *)current->journal_info;
4798         if (trans)
4799                 return -EAGAIN;
4800
4801         if (force)
4802                 goto commit;
4803
4804         /* See if there is enough pinned space to make this reservation */
4805         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4806                                    bytes) >= 0)
4807                 goto commit;
4808
4809         /*
4810          * See if there is some space in the delayed insertion reservation for
4811          * this reservation.
4812          */
4813         if (space_info != delayed_rsv->space_info)
4814                 return -ENOSPC;
4815
4816         spin_lock(&delayed_rsv->lock);
4817         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4818                                    bytes - delayed_rsv->size) >= 0) {
4819                 spin_unlock(&delayed_rsv->lock);
4820                 return -ENOSPC;
4821         }
4822         spin_unlock(&delayed_rsv->lock);
4823
4824 commit:
4825         trans = btrfs_join_transaction(root);
4826         if (IS_ERR(trans))
4827                 return -ENOSPC;
4828
4829         return btrfs_commit_transaction(trans, root);
4830 }
4831
4832 struct reserve_ticket {
4833         u64 bytes;
4834         int error;
4835         struct list_head list;
4836         wait_queue_head_t wait;
4837 };
4838
4839 static int flush_space(struct btrfs_root *root,
4840                        struct btrfs_space_info *space_info, u64 num_bytes,
4841                        u64 orig_bytes, int state)
4842 {
4843         struct btrfs_trans_handle *trans;
4844         int nr;
4845         int ret = 0;
4846
4847         switch (state) {
4848         case FLUSH_DELAYED_ITEMS_NR:
4849         case FLUSH_DELAYED_ITEMS:
4850                 if (state == FLUSH_DELAYED_ITEMS_NR)
4851                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4852                 else
4853                         nr = -1;
4854
4855                 trans = btrfs_join_transaction(root);
4856                 if (IS_ERR(trans)) {
4857                         ret = PTR_ERR(trans);
4858                         break;
4859                 }
4860                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4861                 btrfs_end_transaction(trans, root);
4862                 break;
4863         case FLUSH_DELALLOC:
4864         case FLUSH_DELALLOC_WAIT:
4865                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4866                                 state == FLUSH_DELALLOC_WAIT);
4867                 break;
4868         case ALLOC_CHUNK:
4869                 trans = btrfs_join_transaction(root);
4870                 if (IS_ERR(trans)) {
4871                         ret = PTR_ERR(trans);
4872                         break;
4873                 }
4874                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4875                                      btrfs_get_alloc_profile(root, 0),
4876                                      CHUNK_ALLOC_NO_FORCE);
4877                 btrfs_end_transaction(trans, root);
4878                 if (ret > 0 || ret == -ENOSPC)
4879                         ret = 0;
4880                 break;
4881         case COMMIT_TRANS:
4882                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4883                 break;
4884         default:
4885                 ret = -ENOSPC;
4886                 break;
4887         }
4888
4889         trace_btrfs_flush_space(root->fs_info, space_info->flags, num_bytes,
4890                                 orig_bytes, state, ret);
4891         return ret;
4892 }
4893
4894 static inline u64
4895 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4896                                  struct btrfs_space_info *space_info)
4897 {
4898         struct reserve_ticket *ticket;
4899         u64 used;
4900         u64 expected;
4901         u64 to_reclaim = 0;
4902
4903         list_for_each_entry(ticket, &space_info->tickets, list)
4904                 to_reclaim += ticket->bytes;
4905         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4906                 to_reclaim += ticket->bytes;
4907         if (to_reclaim)
4908                 return to_reclaim;
4909
4910         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
4911         if (can_overcommit(root, space_info, to_reclaim,
4912                            BTRFS_RESERVE_FLUSH_ALL))
4913                 return 0;
4914
4915         used = space_info->bytes_used + space_info->bytes_reserved +
4916                space_info->bytes_pinned + space_info->bytes_readonly +
4917                space_info->bytes_may_use;
4918         if (can_overcommit(root, space_info, SZ_1M, BTRFS_RESERVE_FLUSH_ALL))
4919                 expected = div_factor_fine(space_info->total_bytes, 95);
4920         else
4921                 expected = div_factor_fine(space_info->total_bytes, 90);
4922
4923         if (used > expected)
4924                 to_reclaim = used - expected;
4925         else
4926                 to_reclaim = 0;
4927         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4928                                      space_info->bytes_reserved);
4929         return to_reclaim;
4930 }
4931
4932 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4933                                         struct btrfs_root *root, u64 used)
4934 {
4935         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4936
4937         /* If we're just plain full then async reclaim just slows us down. */
4938         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4939                 return 0;
4940
4941         if (!btrfs_calc_reclaim_metadata_size(root, space_info))
4942                 return 0;
4943
4944         return (used >= thresh && !btrfs_fs_closing(root->fs_info) &&
4945                 !test_bit(BTRFS_FS_STATE_REMOUNTING,
4946                           &root->fs_info->fs_state));
4947 }
4948
4949 static void wake_all_tickets(struct list_head *head)
4950 {
4951         struct reserve_ticket *ticket;
4952
4953         while (!list_empty(head)) {
4954                 ticket = list_first_entry(head, struct reserve_ticket, list);
4955                 list_del_init(&ticket->list);
4956                 ticket->error = -ENOSPC;
4957                 wake_up(&ticket->wait);
4958         }
4959 }
4960
4961 /*
4962  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4963  * will loop and continuously try to flush as long as we are making progress.
4964  * We count progress as clearing off tickets each time we have to loop.
4965  */
4966 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4967 {
4968         struct btrfs_fs_info *fs_info;
4969         struct btrfs_space_info *space_info;
4970         u64 to_reclaim;
4971         int flush_state;
4972         int commit_cycles = 0;
4973         u64 last_tickets_id;
4974
4975         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4976         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4977
4978         spin_lock(&space_info->lock);
4979         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4980                                                       space_info);
4981         if (!to_reclaim) {
4982                 space_info->flush = 0;
4983                 spin_unlock(&space_info->lock);
4984                 return;
4985         }
4986         last_tickets_id = space_info->tickets_id;
4987         spin_unlock(&space_info->lock);
4988
4989         flush_state = FLUSH_DELAYED_ITEMS_NR;
4990         do {
4991                 struct reserve_ticket *ticket;
4992                 int ret;
4993
4994                 ret = flush_space(fs_info->fs_root, space_info, to_reclaim,
4995                             to_reclaim, flush_state);
4996                 spin_lock(&space_info->lock);
4997                 if (list_empty(&space_info->tickets)) {
4998                         space_info->flush = 0;
4999                         spin_unlock(&space_info->lock);
5000                         return;
5001                 }
5002                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5003                                                               space_info);
5004                 ticket = list_first_entry(&space_info->tickets,
5005                                           struct reserve_ticket, list);
5006                 if (last_tickets_id == space_info->tickets_id) {
5007                         flush_state++;
5008                 } else {
5009                         last_tickets_id = space_info->tickets_id;
5010                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5011                         if (commit_cycles)
5012                                 commit_cycles--;
5013                 }
5014
5015                 if (flush_state > COMMIT_TRANS) {
5016                         commit_cycles++;
5017                         if (commit_cycles > 2) {
5018                                 wake_all_tickets(&space_info->tickets);
5019                                 space_info->flush = 0;
5020                         } else {
5021                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5022                         }
5023                 }
5024                 spin_unlock(&space_info->lock);
5025         } while (flush_state <= COMMIT_TRANS);
5026 }
5027
5028 void btrfs_init_async_reclaim_work(struct work_struct *work)
5029 {
5030         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5031 }
5032
5033 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5034                                             struct btrfs_space_info *space_info,
5035                                             struct reserve_ticket *ticket)
5036 {
5037         u64 to_reclaim;
5038         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5039
5040         spin_lock(&space_info->lock);
5041         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5042                                                       space_info);
5043         if (!to_reclaim) {
5044                 spin_unlock(&space_info->lock);
5045                 return;
5046         }
5047         spin_unlock(&space_info->lock);
5048
5049         do {
5050                 flush_space(fs_info->fs_root, space_info, to_reclaim,
5051                             to_reclaim, flush_state);
5052                 flush_state++;
5053                 spin_lock(&space_info->lock);
5054                 if (ticket->bytes == 0) {
5055                         spin_unlock(&space_info->lock);
5056                         return;
5057                 }
5058                 spin_unlock(&space_info->lock);
5059
5060                 /*
5061                  * Priority flushers can't wait on delalloc without
5062                  * deadlocking.
5063                  */
5064                 if (flush_state == FLUSH_DELALLOC ||
5065                     flush_state == FLUSH_DELALLOC_WAIT)
5066                         flush_state = ALLOC_CHUNK;
5067         } while (flush_state < COMMIT_TRANS);
5068 }
5069
5070 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5071                                struct btrfs_space_info *space_info,
5072                                struct reserve_ticket *ticket, u64 orig_bytes)
5073
5074 {
5075         DEFINE_WAIT(wait);
5076         int ret = 0;
5077
5078         spin_lock(&space_info->lock);
5079         while (ticket->bytes > 0 && ticket->error == 0) {
5080                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5081                 if (ret) {
5082                         ret = -EINTR;
5083                         break;
5084                 }
5085                 spin_unlock(&space_info->lock);
5086
5087                 schedule();
5088
5089                 finish_wait(&ticket->wait, &wait);
5090                 spin_lock(&space_info->lock);
5091         }
5092         if (!ret)
5093                 ret = ticket->error;
5094         if (!list_empty(&ticket->list))
5095                 list_del_init(&ticket->list);
5096         if (ticket->bytes && ticket->bytes < orig_bytes) {
5097                 u64 num_bytes = orig_bytes - ticket->bytes;
5098                 space_info->bytes_may_use -= num_bytes;
5099                 trace_btrfs_space_reservation(fs_info, "space_info",
5100                                               space_info->flags, num_bytes, 0);
5101         }
5102         spin_unlock(&space_info->lock);
5103
5104         return ret;
5105 }
5106
5107 /**
5108  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5109  * @root - the root we're allocating for
5110  * @space_info - the space info we want to allocate from
5111  * @orig_bytes - the number of bytes we want
5112  * @flush - whether or not we can flush to make our reservation
5113  *
5114  * This will reserve orig_bytes number of bytes from the space info associated
5115  * with the block_rsv.  If there is not enough space it will make an attempt to
5116  * flush out space to make room.  It will do this by flushing delalloc if
5117  * possible or committing the transaction.  If flush is 0 then no attempts to
5118  * regain reservations will be made and this will fail if there is not enough
5119  * space already.
5120  */
5121 static int __reserve_metadata_bytes(struct btrfs_root *root,
5122                                     struct btrfs_space_info *space_info,
5123                                     u64 orig_bytes,
5124                                     enum btrfs_reserve_flush_enum flush)
5125 {
5126         struct reserve_ticket ticket;
5127         u64 used;
5128         int ret = 0;
5129
5130         ASSERT(orig_bytes);
5131         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5132
5133         spin_lock(&space_info->lock);
5134         ret = -ENOSPC;
5135         used = space_info->bytes_used + space_info->bytes_reserved +
5136                 space_info->bytes_pinned + space_info->bytes_readonly +
5137                 space_info->bytes_may_use;
5138
5139         /*
5140          * If we have enough space then hooray, make our reservation and carry
5141          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5142          * If not things get more complicated.
5143          */
5144         if (used + orig_bytes <= space_info->total_bytes) {
5145                 space_info->bytes_may_use += orig_bytes;
5146                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5147                                               space_info->flags, orig_bytes,
5148                                               1);
5149                 ret = 0;
5150         } else if (can_overcommit(root, space_info, orig_bytes, flush)) {
5151                 space_info->bytes_may_use += orig_bytes;
5152                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5153                                               space_info->flags, orig_bytes,
5154                                               1);
5155                 ret = 0;
5156         }
5157
5158         /*
5159          * If we couldn't make a reservation then setup our reservation ticket
5160          * and kick the async worker if it's not already running.
5161          *
5162          * If we are a priority flusher then we just need to add our ticket to
5163          * the list and we will do our own flushing further down.
5164          */
5165         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5166                 ticket.bytes = orig_bytes;
5167                 ticket.error = 0;
5168                 init_waitqueue_head(&ticket.wait);
5169                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5170                         list_add_tail(&ticket.list, &space_info->tickets);
5171                         if (!space_info->flush) {
5172                                 space_info->flush = 1;
5173                                 trace_btrfs_trigger_flush(root->fs_info,
5174                                                           space_info->flags,
5175                                                           orig_bytes, flush,
5176                                                           "enospc");
5177                                 queue_work(system_unbound_wq,
5178                                            &root->fs_info->async_reclaim_work);
5179                         }
5180                 } else {
5181                         list_add_tail(&ticket.list,
5182                                       &space_info->priority_tickets);
5183                 }
5184         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5185                 used += orig_bytes;
5186                 /*
5187                  * We will do the space reservation dance during log replay,
5188                  * which means we won't have fs_info->fs_root set, so don't do
5189                  * the async reclaim as we will panic.
5190                  */
5191                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags) &&
5192                     need_do_async_reclaim(space_info, root, used) &&
5193                     !work_busy(&root->fs_info->async_reclaim_work)) {
5194                         trace_btrfs_trigger_flush(root->fs_info,
5195                                                   space_info->flags,
5196                                                   orig_bytes, flush,
5197                                                   "preempt");
5198                         queue_work(system_unbound_wq,
5199                                    &root->fs_info->async_reclaim_work);
5200                 }
5201         }
5202         spin_unlock(&space_info->lock);
5203         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5204                 return ret;
5205
5206         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5207                 return wait_reserve_ticket(root->fs_info, space_info, &ticket,
5208                                            orig_bytes);
5209
5210         ret = 0;
5211         priority_reclaim_metadata_space(root->fs_info, space_info, &ticket);
5212         spin_lock(&space_info->lock);
5213         if (ticket.bytes) {
5214                 if (ticket.bytes < orig_bytes) {
5215                         u64 num_bytes = orig_bytes - ticket.bytes;
5216                         space_info->bytes_may_use -= num_bytes;
5217                         trace_btrfs_space_reservation(root->fs_info,
5218                                         "space_info", space_info->flags,
5219                                         num_bytes, 0);
5220
5221                 }
5222                 list_del_init(&ticket.list);
5223                 ret = -ENOSPC;
5224         }
5225         spin_unlock(&space_info->lock);
5226         ASSERT(list_empty(&ticket.list));
5227         return ret;
5228 }
5229
5230 /**
5231  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5232  * @root - the root we're allocating for
5233  * @block_rsv - the block_rsv we're allocating for
5234  * @orig_bytes - the number of bytes we want
5235  * @flush - whether or not we can flush to make our reservation
5236  *
5237  * This will reserve orgi_bytes number of bytes from the space info associated
5238  * with the block_rsv.  If there is not enough space it will make an attempt to
5239  * flush out space to make room.  It will do this by flushing delalloc if
5240  * possible or committing the transaction.  If flush is 0 then no attempts to
5241  * regain reservations will be made and this will fail if there is not enough
5242  * space already.
5243  */
5244 static int reserve_metadata_bytes(struct btrfs_root *root,
5245                                   struct btrfs_block_rsv *block_rsv,
5246                                   u64 orig_bytes,
5247                                   enum btrfs_reserve_flush_enum flush)
5248 {
5249         int ret;
5250
5251         ret = __reserve_metadata_bytes(root, block_rsv->space_info, orig_bytes,
5252                                        flush);
5253         if (ret == -ENOSPC &&
5254             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5255                 struct btrfs_block_rsv *global_rsv =
5256                         &root->fs_info->global_block_rsv;
5257
5258                 if (block_rsv != global_rsv &&
5259                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5260                         ret = 0;
5261         }
5262         if (ret == -ENOSPC)
5263                 trace_btrfs_space_reservation(root->fs_info,
5264                                               "space_info:enospc",
5265                                               block_rsv->space_info->flags,
5266                                               orig_bytes, 1);
5267         return ret;
5268 }
5269
5270 static struct btrfs_block_rsv *get_block_rsv(
5271                                         const struct btrfs_trans_handle *trans,
5272                                         const struct btrfs_root *root)
5273 {
5274         struct btrfs_block_rsv *block_rsv = NULL;
5275
5276         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5277             (root == root->fs_info->csum_root && trans->adding_csums) ||
5278              (root == root->fs_info->uuid_root))
5279                 block_rsv = trans->block_rsv;
5280
5281         if (!block_rsv)
5282                 block_rsv = root->block_rsv;
5283
5284         if (!block_rsv)
5285                 block_rsv = &root->fs_info->empty_block_rsv;
5286
5287         return block_rsv;
5288 }
5289
5290 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5291                                u64 num_bytes)
5292 {
5293         int ret = -ENOSPC;
5294         spin_lock(&block_rsv->lock);
5295         if (block_rsv->reserved >= num_bytes) {
5296                 block_rsv->reserved -= num_bytes;
5297                 if (block_rsv->reserved < block_rsv->size)
5298                         block_rsv->full = 0;
5299                 ret = 0;
5300         }
5301         spin_unlock(&block_rsv->lock);
5302         return ret;
5303 }
5304
5305 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5306                                 u64 num_bytes, int update_size)
5307 {
5308         spin_lock(&block_rsv->lock);
5309         block_rsv->reserved += num_bytes;
5310         if (update_size)
5311                 block_rsv->size += num_bytes;
5312         else if (block_rsv->reserved >= block_rsv->size)
5313                 block_rsv->full = 1;
5314         spin_unlock(&block_rsv->lock);
5315 }
5316
5317 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5318                              struct btrfs_block_rsv *dest, u64 num_bytes,
5319                              int min_factor)
5320 {
5321         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5322         u64 min_bytes;
5323
5324         if (global_rsv->space_info != dest->space_info)
5325                 return -ENOSPC;
5326
5327         spin_lock(&global_rsv->lock);
5328         min_bytes = div_factor(global_rsv->size, min_factor);
5329         if (global_rsv->reserved < min_bytes + num_bytes) {
5330                 spin_unlock(&global_rsv->lock);
5331                 return -ENOSPC;
5332         }
5333         global_rsv->reserved -= num_bytes;
5334         if (global_rsv->reserved < global_rsv->size)
5335                 global_rsv->full = 0;
5336         spin_unlock(&global_rsv->lock);
5337
5338         block_rsv_add_bytes(dest, num_bytes, 1);
5339         return 0;
5340 }
5341
5342 /*
5343  * This is for space we already have accounted in space_info->bytes_may_use, so
5344  * basically when we're returning space from block_rsv's.
5345  */
5346 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5347                                      struct btrfs_space_info *space_info,
5348                                      u64 num_bytes)
5349 {
5350         struct reserve_ticket *ticket;
5351         struct list_head *head;
5352         u64 used;
5353         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5354         bool check_overcommit = false;
5355
5356         spin_lock(&space_info->lock);
5357         head = &space_info->priority_tickets;
5358
5359         /*
5360          * If we are over our limit then we need to check and see if we can
5361          * overcommit, and if we can't then we just need to free up our space
5362          * and not satisfy any requests.
5363          */
5364         used = space_info->bytes_used + space_info->bytes_reserved +
5365                 space_info->bytes_pinned + space_info->bytes_readonly +
5366                 space_info->bytes_may_use;
5367         if (used - num_bytes >= space_info->total_bytes)
5368                 check_overcommit = true;
5369 again:
5370         while (!list_empty(head) && num_bytes) {
5371                 ticket = list_first_entry(head, struct reserve_ticket,
5372                                           list);
5373                 /*
5374                  * We use 0 bytes because this space is already reserved, so
5375                  * adding the ticket space would be a double count.
5376                  */
5377                 if (check_overcommit &&
5378                     !can_overcommit(fs_info->extent_root, space_info, 0,
5379                                     flush))
5380                         break;
5381                 if (num_bytes >= ticket->bytes) {
5382                         list_del_init(&ticket->list);
5383                         num_bytes -= ticket->bytes;
5384                         ticket->bytes = 0;
5385                         space_info->tickets_id++;
5386                         wake_up(&ticket->wait);
5387                 } else {
5388                         ticket->bytes -= num_bytes;
5389                         num_bytes = 0;
5390                 }
5391         }
5392
5393         if (num_bytes && head == &space_info->priority_tickets) {
5394                 head = &space_info->tickets;
5395                 flush = BTRFS_RESERVE_FLUSH_ALL;
5396                 goto again;
5397         }
5398         space_info->bytes_may_use -= num_bytes;
5399         trace_btrfs_space_reservation(fs_info, "space_info",
5400                                       space_info->flags, num_bytes, 0);
5401         spin_unlock(&space_info->lock);
5402 }
5403
5404 /*
5405  * This is for newly allocated space that isn't accounted in
5406  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5407  * we use this helper.
5408  */
5409 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5410                                      struct btrfs_space_info *space_info,
5411                                      u64 num_bytes)
5412 {
5413         struct reserve_ticket *ticket;
5414         struct list_head *head = &space_info->priority_tickets;
5415
5416 again:
5417         while (!list_empty(head) && num_bytes) {
5418                 ticket = list_first_entry(head, struct reserve_ticket,
5419                                           list);
5420                 if (num_bytes >= ticket->bytes) {
5421                         trace_btrfs_space_reservation(fs_info, "space_info",
5422                                                       space_info->flags,
5423                                                       ticket->bytes, 1);
5424                         list_del_init(&ticket->list);
5425                         num_bytes -= ticket->bytes;
5426                         space_info->bytes_may_use += ticket->bytes;
5427                         ticket->bytes = 0;
5428                         space_info->tickets_id++;
5429                         wake_up(&ticket->wait);
5430                 } else {
5431                         trace_btrfs_space_reservation(fs_info, "space_info",
5432                                                       space_info->flags,
5433                                                       num_bytes, 1);
5434                         space_info->bytes_may_use += num_bytes;
5435                         ticket->bytes -= num_bytes;
5436                         num_bytes = 0;
5437                 }
5438         }
5439
5440         if (num_bytes && head == &space_info->priority_tickets) {
5441                 head = &space_info->tickets;
5442                 goto again;
5443         }
5444 }
5445
5446 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5447                                     struct btrfs_block_rsv *block_rsv,
5448                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5449 {
5450         struct btrfs_space_info *space_info = block_rsv->space_info;
5451
5452         spin_lock(&block_rsv->lock);
5453         if (num_bytes == (u64)-1)
5454                 num_bytes = block_rsv->size;
5455         block_rsv->size -= num_bytes;
5456         if (block_rsv->reserved >= block_rsv->size) {
5457                 num_bytes = block_rsv->reserved - block_rsv->size;
5458                 block_rsv->reserved = block_rsv->size;
5459                 block_rsv->full = 1;
5460         } else {
5461                 num_bytes = 0;
5462         }
5463         spin_unlock(&block_rsv->lock);
5464
5465         if (num_bytes > 0) {
5466                 if (dest) {
5467                         spin_lock(&dest->lock);
5468                         if (!dest->full) {
5469                                 u64 bytes_to_add;
5470
5471                                 bytes_to_add = dest->size - dest->reserved;
5472                                 bytes_to_add = min(num_bytes, bytes_to_add);
5473                                 dest->reserved += bytes_to_add;
5474                                 if (dest->reserved >= dest->size)
5475                                         dest->full = 1;
5476                                 num_bytes -= bytes_to_add;
5477                         }
5478                         spin_unlock(&dest->lock);
5479                 }
5480                 if (num_bytes)
5481                         space_info_add_old_bytes(fs_info, space_info,
5482                                                  num_bytes);
5483         }
5484 }
5485
5486 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5487                             struct btrfs_block_rsv *dst, u64 num_bytes,
5488                             int update_size)
5489 {
5490         int ret;
5491
5492         ret = block_rsv_use_bytes(src, num_bytes);
5493         if (ret)
5494                 return ret;
5495
5496         block_rsv_add_bytes(dst, num_bytes, update_size);
5497         return 0;
5498 }
5499
5500 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5501 {
5502         memset(rsv, 0, sizeof(*rsv));
5503         spin_lock_init(&rsv->lock);
5504         rsv->type = type;
5505 }
5506
5507 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5508                                               unsigned short type)
5509 {
5510         struct btrfs_block_rsv *block_rsv;
5511         struct btrfs_fs_info *fs_info = root->fs_info;
5512
5513         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5514         if (!block_rsv)
5515                 return NULL;
5516
5517         btrfs_init_block_rsv(block_rsv, type);
5518         block_rsv->space_info = __find_space_info(fs_info,
5519                                                   BTRFS_BLOCK_GROUP_METADATA);
5520         return block_rsv;
5521 }
5522
5523 void btrfs_free_block_rsv(struct btrfs_root *root,
5524                           struct btrfs_block_rsv *rsv)
5525 {
5526         if (!rsv)
5527                 return;
5528         btrfs_block_rsv_release(root, rsv, (u64)-1);
5529         kfree(rsv);
5530 }
5531
5532 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5533 {
5534         kfree(rsv);
5535 }
5536
5537 int btrfs_block_rsv_add(struct btrfs_root *root,
5538                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5539                         enum btrfs_reserve_flush_enum flush)
5540 {
5541         int ret;
5542
5543         if (num_bytes == 0)
5544                 return 0;
5545
5546         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5547         if (!ret) {
5548                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5549                 return 0;
5550         }
5551
5552         return ret;
5553 }
5554
5555 int btrfs_block_rsv_check(struct btrfs_root *root,
5556                           struct btrfs_block_rsv *block_rsv, int min_factor)
5557 {
5558         u64 num_bytes = 0;
5559         int ret = -ENOSPC;
5560
5561         if (!block_rsv)
5562                 return 0;
5563
5564         spin_lock(&block_rsv->lock);
5565         num_bytes = div_factor(block_rsv->size, min_factor);
5566         if (block_rsv->reserved >= num_bytes)
5567                 ret = 0;
5568         spin_unlock(&block_rsv->lock);
5569
5570         return ret;
5571 }
5572
5573 int btrfs_block_rsv_refill(struct btrfs_root *root,
5574                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5575                            enum btrfs_reserve_flush_enum flush)
5576 {
5577         u64 num_bytes = 0;
5578         int ret = -ENOSPC;
5579
5580         if (!block_rsv)
5581                 return 0;
5582
5583         spin_lock(&block_rsv->lock);
5584         num_bytes = min_reserved;
5585         if (block_rsv->reserved >= num_bytes)
5586                 ret = 0;
5587         else
5588                 num_bytes -= block_rsv->reserved;
5589         spin_unlock(&block_rsv->lock);
5590
5591         if (!ret)
5592                 return 0;
5593
5594         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5595         if (!ret) {
5596                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5597                 return 0;
5598         }
5599
5600         return ret;
5601 }
5602
5603 void btrfs_block_rsv_release(struct btrfs_root *root,
5604                              struct btrfs_block_rsv *block_rsv,
5605                              u64 num_bytes)
5606 {
5607         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5608         if (global_rsv == block_rsv ||
5609             block_rsv->space_info != global_rsv->space_info)
5610                 global_rsv = NULL;
5611         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5612                                 num_bytes);
5613 }
5614
5615 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5616 {
5617         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5618         struct btrfs_space_info *sinfo = block_rsv->space_info;
5619         u64 num_bytes;
5620
5621         /*
5622          * The global block rsv is based on the size of the extent tree, the
5623          * checksum tree and the root tree.  If the fs is empty we want to set
5624          * it to a minimal amount for safety.
5625          */
5626         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5627                 btrfs_root_used(&fs_info->csum_root->root_item) +
5628                 btrfs_root_used(&fs_info->tree_root->root_item);
5629         num_bytes = max_t(u64, num_bytes, SZ_16M);
5630
5631         spin_lock(&sinfo->lock);
5632         spin_lock(&block_rsv->lock);
5633
5634         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5635
5636         if (block_rsv->reserved < block_rsv->size) {
5637                 num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5638                         sinfo->bytes_reserved + sinfo->bytes_readonly +
5639                         sinfo->bytes_may_use;
5640                 if (sinfo->total_bytes > num_bytes) {
5641                         num_bytes = sinfo->total_bytes - num_bytes;
5642                         num_bytes = min(num_bytes,
5643                                         block_rsv->size - block_rsv->reserved);
5644                         block_rsv->reserved += num_bytes;
5645                         sinfo->bytes_may_use += num_bytes;
5646                         trace_btrfs_space_reservation(fs_info, "space_info",
5647                                                       sinfo->flags, num_bytes,
5648                                                       1);
5649                 }
5650         } else if (block_rsv->reserved > block_rsv->size) {
5651                 num_bytes = block_rsv->reserved - block_rsv->size;
5652                 sinfo->bytes_may_use -= num_bytes;
5653                 trace_btrfs_space_reservation(fs_info, "space_info",
5654                                       sinfo->flags, num_bytes, 0);
5655                 block_rsv->reserved = block_rsv->size;
5656         }
5657
5658         if (block_rsv->reserved == block_rsv->size)
5659                 block_rsv->full = 1;
5660         else
5661                 block_rsv->full = 0;
5662
5663         spin_unlock(&block_rsv->lock);
5664         spin_unlock(&sinfo->lock);
5665 }
5666
5667 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5668 {
5669         struct btrfs_space_info *space_info;
5670
5671         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5672         fs_info->chunk_block_rsv.space_info = space_info;
5673
5674         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5675         fs_info->global_block_rsv.space_info = space_info;
5676         fs_info->delalloc_block_rsv.space_info = space_info;
5677         fs_info->trans_block_rsv.space_info = space_info;
5678         fs_info->empty_block_rsv.space_info = space_info;
5679         fs_info->delayed_block_rsv.space_info = space_info;
5680
5681         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5682         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5683         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5684         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5685         if (fs_info->quota_root)
5686                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5687         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5688
5689         update_global_block_rsv(fs_info);
5690 }
5691
5692 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5693 {
5694         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5695                                 (u64)-1);
5696         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5697         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5698         WARN_ON(fs_info->trans_block_rsv.size > 0);
5699         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5700         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5701         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5702         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5703         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5704 }
5705
5706 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5707                                   struct btrfs_root *root)
5708 {
5709         if (!trans->block_rsv)
5710                 return;
5711
5712         if (!trans->bytes_reserved)
5713                 return;
5714
5715         trace_btrfs_space_reservation(root->fs_info, "transaction",
5716                                       trans->transid, trans->bytes_reserved, 0);
5717         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5718         trans->bytes_reserved = 0;
5719 }
5720
5721 /*
5722  * To be called after all the new block groups attached to the transaction
5723  * handle have been created (btrfs_create_pending_block_groups()).
5724  */
5725 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5726 {
5727         struct btrfs_fs_info *fs_info = trans->fs_info;
5728
5729         if (!trans->chunk_bytes_reserved)
5730                 return;
5731
5732         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5733
5734         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5735                                 trans->chunk_bytes_reserved);
5736         trans->chunk_bytes_reserved = 0;
5737 }
5738
5739 /* Can only return 0 or -ENOSPC */
5740 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5741                                   struct inode *inode)
5742 {
5743         struct btrfs_root *root = BTRFS_I(inode)->root;
5744         /*
5745          * We always use trans->block_rsv here as we will have reserved space
5746          * for our orphan when starting the transaction, using get_block_rsv()
5747          * here will sometimes make us choose the wrong block rsv as we could be
5748          * doing a reloc inode for a non refcounted root.
5749          */
5750         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5751         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5752
5753         /*
5754          * We need to hold space in order to delete our orphan item once we've
5755          * added it, so this takes the reservation so we can release it later
5756          * when we are truly done with the orphan item.
5757          */
5758         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5759         trace_btrfs_space_reservation(root->fs_info, "orphan",
5760                                       btrfs_ino(inode), num_bytes, 1);
5761         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5762 }
5763
5764 void btrfs_orphan_release_metadata(struct inode *inode)
5765 {
5766         struct btrfs_root *root = BTRFS_I(inode)->root;
5767         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5768         trace_btrfs_space_reservation(root->fs_info, "orphan",
5769                                       btrfs_ino(inode), num_bytes, 0);
5770         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5771 }
5772
5773 /*
5774  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5775  * root: the root of the parent directory
5776  * rsv: block reservation
5777  * items: the number of items that we need do reservation
5778  * qgroup_reserved: used to return the reserved size in qgroup
5779  *
5780  * This function is used to reserve the space for snapshot/subvolume
5781  * creation and deletion. Those operations are different with the
5782  * common file/directory operations, they change two fs/file trees
5783  * and root tree, the number of items that the qgroup reserves is
5784  * different with the free space reservation. So we can not use
5785  * the space reservation mechanism in start_transaction().
5786  */
5787 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5788                                      struct btrfs_block_rsv *rsv,
5789                                      int items,
5790                                      u64 *qgroup_reserved,
5791                                      bool use_global_rsv)
5792 {
5793         u64 num_bytes;
5794         int ret;
5795         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5796
5797         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags)) {
5798                 /* One for parent inode, two for dir entries */
5799                 num_bytes = 3 * root->nodesize;
5800                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5801                 if (ret)
5802                         return ret;
5803         } else {
5804                 num_bytes = 0;
5805         }
5806
5807         *qgroup_reserved = num_bytes;
5808
5809         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5810         rsv->space_info = __find_space_info(root->fs_info,
5811                                             BTRFS_BLOCK_GROUP_METADATA);
5812         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5813                                   BTRFS_RESERVE_FLUSH_ALL);
5814
5815         if (ret == -ENOSPC && use_global_rsv)
5816                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5817
5818         if (ret && *qgroup_reserved)
5819                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5820
5821         return ret;
5822 }
5823
5824 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5825                                       struct btrfs_block_rsv *rsv,
5826                                       u64 qgroup_reserved)
5827 {
5828         btrfs_block_rsv_release(root, rsv, (u64)-1);
5829 }
5830
5831 /**
5832  * drop_outstanding_extent - drop an outstanding extent
5833  * @inode: the inode we're dropping the extent for
5834  * @num_bytes: the number of bytes we're releasing.
5835  *
5836  * This is called when we are freeing up an outstanding extent, either called
5837  * after an error or after an extent is written.  This will return the number of
5838  * reserved extents that need to be freed.  This must be called with
5839  * BTRFS_I(inode)->lock held.
5840  */
5841 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5842 {
5843         unsigned drop_inode_space = 0;
5844         unsigned dropped_extents = 0;
5845         unsigned num_extents = 0;
5846
5847         num_extents = (unsigned)div64_u64(num_bytes +
5848                                           BTRFS_MAX_EXTENT_SIZE - 1,
5849                                           BTRFS_MAX_EXTENT_SIZE);
5850         ASSERT(num_extents);
5851         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5852         BTRFS_I(inode)->outstanding_extents -= num_extents;
5853
5854         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5855             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5856                                &BTRFS_I(inode)->runtime_flags))
5857                 drop_inode_space = 1;
5858
5859         /*
5860          * If we have more or the same amount of outstanding extents than we have
5861          * reserved then we need to leave the reserved extents count alone.
5862          */
5863         if (BTRFS_I(inode)->outstanding_extents >=
5864             BTRFS_I(inode)->reserved_extents)
5865                 return drop_inode_space;
5866
5867         dropped_extents = BTRFS_I(inode)->reserved_extents -
5868                 BTRFS_I(inode)->outstanding_extents;
5869         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5870         return dropped_extents + drop_inode_space;
5871 }
5872
5873 /**
5874  * calc_csum_metadata_size - return the amount of metadata space that must be
5875  *      reserved/freed for the given bytes.
5876  * @inode: the inode we're manipulating
5877  * @num_bytes: the number of bytes in question
5878  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5879  *
5880  * This adjusts the number of csum_bytes in the inode and then returns the
5881  * correct amount of metadata that must either be reserved or freed.  We
5882  * calculate how many checksums we can fit into one leaf and then divide the
5883  * number of bytes that will need to be checksumed by this value to figure out
5884  * how many checksums will be required.  If we are adding bytes then the number
5885  * may go up and we will return the number of additional bytes that must be
5886  * reserved.  If it is going down we will return the number of bytes that must
5887  * be freed.
5888  *
5889  * This must be called with BTRFS_I(inode)->lock held.
5890  */
5891 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5892                                    int reserve)
5893 {
5894         struct btrfs_root *root = BTRFS_I(inode)->root;
5895         u64 old_csums, num_csums;
5896
5897         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5898             BTRFS_I(inode)->csum_bytes == 0)
5899                 return 0;
5900
5901         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5902         if (reserve)
5903                 BTRFS_I(inode)->csum_bytes += num_bytes;
5904         else
5905                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5906         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5907
5908         /* No change, no need to reserve more */
5909         if (old_csums == num_csums)
5910                 return 0;
5911
5912         if (reserve)
5913                 return btrfs_calc_trans_metadata_size(root,
5914                                                       num_csums - old_csums);
5915
5916         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5917 }
5918
5919 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5920 {
5921         struct btrfs_root *root = BTRFS_I(inode)->root;
5922         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5923         u64 to_reserve = 0;
5924         u64 csum_bytes;
5925         unsigned nr_extents = 0;
5926         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5927         int ret = 0;
5928         bool delalloc_lock = true;
5929         u64 to_free = 0;
5930         unsigned dropped;
5931         bool release_extra = false;
5932
5933         /* If we are a free space inode we need to not flush since we will be in
5934          * the middle of a transaction commit.  We also don't need the delalloc
5935          * mutex since we won't race with anybody.  We need this mostly to make
5936          * lockdep shut its filthy mouth.
5937          *
5938          * If we have a transaction open (can happen if we call truncate_block
5939          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5940          */
5941         if (btrfs_is_free_space_inode(inode)) {
5942                 flush = BTRFS_RESERVE_NO_FLUSH;
5943                 delalloc_lock = false;
5944         } else if (current->journal_info) {
5945                 flush = BTRFS_RESERVE_FLUSH_LIMIT;
5946         }
5947
5948         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5949             btrfs_transaction_in_commit(root->fs_info))
5950                 schedule_timeout(1);
5951
5952         if (delalloc_lock)
5953                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5954
5955         num_bytes = ALIGN(num_bytes, root->sectorsize);
5956
5957         spin_lock(&BTRFS_I(inode)->lock);
5958         nr_extents = (unsigned)div64_u64(num_bytes +
5959                                          BTRFS_MAX_EXTENT_SIZE - 1,
5960                                          BTRFS_MAX_EXTENT_SIZE);
5961         BTRFS_I(inode)->outstanding_extents += nr_extents;
5962
5963         nr_extents = 0;
5964         if (BTRFS_I(inode)->outstanding_extents >
5965             BTRFS_I(inode)->reserved_extents)
5966                 nr_extents += BTRFS_I(inode)->outstanding_extents -
5967                         BTRFS_I(inode)->reserved_extents;
5968
5969         /* We always want to reserve a slot for updating the inode. */
5970         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents + 1);
5971         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5972         csum_bytes = BTRFS_I(inode)->csum_bytes;
5973         spin_unlock(&BTRFS_I(inode)->lock);
5974
5975         if (test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags)) {
5976                 ret = btrfs_qgroup_reserve_meta(root,
5977                                 nr_extents * root->nodesize);
5978                 if (ret)
5979                         goto out_fail;
5980         }
5981
5982         ret = btrfs_block_rsv_add(root, block_rsv, to_reserve, flush);
5983         if (unlikely(ret)) {
5984                 btrfs_qgroup_free_meta(root, nr_extents * root->nodesize);
5985                 goto out_fail;
5986         }
5987
5988         spin_lock(&BTRFS_I(inode)->lock);
5989         if (test_and_set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5990                              &BTRFS_I(inode)->runtime_flags)) {
5991                 to_reserve -= btrfs_calc_trans_metadata_size(root, 1);
5992                 release_extra = true;
5993         }
5994         BTRFS_I(inode)->reserved_extents += nr_extents;
5995         spin_unlock(&BTRFS_I(inode)->lock);
5996
5997         if (delalloc_lock)
5998                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5999
6000         if (to_reserve)
6001                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
6002                                               btrfs_ino(inode), to_reserve, 1);
6003         if (release_extra)
6004                 btrfs_block_rsv_release(root, block_rsv,
6005                                         btrfs_calc_trans_metadata_size(root,
6006                                                                        1));
6007         return 0;
6008
6009 out_fail:
6010         spin_lock(&BTRFS_I(inode)->lock);
6011         dropped = drop_outstanding_extent(inode, num_bytes);
6012         /*
6013          * If the inodes csum_bytes is the same as the original
6014          * csum_bytes then we know we haven't raced with any free()ers
6015          * so we can just reduce our inodes csum bytes and carry on.
6016          */
6017         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
6018                 calc_csum_metadata_size(inode, num_bytes, 0);
6019         } else {
6020                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
6021                 u64 bytes;
6022
6023                 /*
6024                  * This is tricky, but first we need to figure out how much we
6025                  * freed from any free-ers that occurred during this
6026                  * reservation, so we reset ->csum_bytes to the csum_bytes
6027                  * before we dropped our lock, and then call the free for the
6028                  * number of bytes that were freed while we were trying our
6029                  * reservation.
6030                  */
6031                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
6032                 BTRFS_I(inode)->csum_bytes = csum_bytes;
6033                 to_free = calc_csum_metadata_size(inode, bytes, 0);
6034
6035
6036                 /*
6037                  * Now we need to see how much we would have freed had we not
6038                  * been making this reservation and our ->csum_bytes were not
6039                  * artificially inflated.
6040                  */
6041                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
6042                 bytes = csum_bytes - orig_csum_bytes;
6043                 bytes = calc_csum_metadata_size(inode, bytes, 0);
6044
6045                 /*
6046                  * Now reset ->csum_bytes to what it should be.  If bytes is
6047                  * more than to_free then we would have freed more space had we
6048                  * not had an artificially high ->csum_bytes, so we need to free
6049                  * the remainder.  If bytes is the same or less then we don't
6050                  * need to do anything, the other free-ers did the correct
6051                  * thing.
6052                  */
6053                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
6054                 if (bytes > to_free)
6055                         to_free = bytes - to_free;
6056                 else
6057                         to_free = 0;
6058         }
6059         spin_unlock(&BTRFS_I(inode)->lock);
6060         if (dropped)
6061                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
6062
6063         if (to_free) {
6064                 btrfs_block_rsv_release(root, block_rsv, to_free);
6065                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
6066                                               btrfs_ino(inode), to_free, 0);
6067         }
6068         if (delalloc_lock)
6069                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6070         return ret;
6071 }
6072
6073 /**
6074  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6075  * @inode: the inode to release the reservation for
6076  * @num_bytes: the number of bytes we're releasing
6077  *
6078  * This will release the metadata reservation for an inode.  This can be called
6079  * once we complete IO for a given set of bytes to release their metadata
6080  * reservations.
6081  */
6082 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
6083 {
6084         struct btrfs_root *root = BTRFS_I(inode)->root;
6085         u64 to_free = 0;
6086         unsigned dropped;
6087
6088         num_bytes = ALIGN(num_bytes, root->sectorsize);
6089         spin_lock(&BTRFS_I(inode)->lock);
6090         dropped = drop_outstanding_extent(inode, num_bytes);
6091
6092         if (num_bytes)
6093                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
6094         spin_unlock(&BTRFS_I(inode)->lock);
6095         if (dropped > 0)
6096                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
6097
6098         if (btrfs_is_testing(root->fs_info))
6099                 return;
6100
6101         trace_btrfs_space_reservation(root->fs_info, "delalloc",
6102                                       btrfs_ino(inode), to_free, 0);
6103
6104         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
6105                                 to_free);
6106 }
6107
6108 /**
6109  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6110  * delalloc
6111  * @inode: inode we're writing to
6112  * @start: start range we are writing to
6113  * @len: how long the range we are writing to
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,
6934                                           "umm, got %d back from search, was looking for %llu",
6935                                           ret, bytenr);
6936                                 if (ret > 0)
6937                                         btrfs_print_leaf(extent_root,
6938                                                          path->nodes[0]);
6939                         }
6940                         if (ret < 0) {
6941                                 btrfs_abort_transaction(trans, ret);
6942                                 goto out;
6943                         }
6944                         extent_slot = path->slots[0];
6945                 }
6946         } else if (WARN_ON(ret == -ENOENT)) {
6947                 btrfs_print_leaf(extent_root, path->nodes[0]);
6948                 btrfs_err(info,
6949                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6950                         bytenr, parent, root_objectid, owner_objectid,
6951                         owner_offset);
6952                 btrfs_abort_transaction(trans, ret);
6953                 goto out;
6954         } else {
6955                 btrfs_abort_transaction(trans, ret);
6956                 goto out;
6957         }
6958
6959         leaf = path->nodes[0];
6960         item_size = btrfs_item_size_nr(leaf, extent_slot);
6961 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6962         if (item_size < sizeof(*ei)) {
6963                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6964                 ret = convert_extent_item_v0(trans, extent_root, path,
6965                                              owner_objectid, 0);
6966                 if (ret < 0) {
6967                         btrfs_abort_transaction(trans, ret);
6968                         goto out;
6969                 }
6970
6971                 btrfs_release_path(path);
6972                 path->leave_spinning = 1;
6973
6974                 key.objectid = bytenr;
6975                 key.type = BTRFS_EXTENT_ITEM_KEY;
6976                 key.offset = num_bytes;
6977
6978                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6979                                         -1, 1);
6980                 if (ret) {
6981                         btrfs_err(info,
6982                                   "umm, got %d back from search, was looking for %llu",
6983                                 ret, bytenr);
6984                         btrfs_print_leaf(extent_root, path->nodes[0]);
6985                 }
6986                 if (ret < 0) {
6987                         btrfs_abort_transaction(trans, ret);
6988                         goto out;
6989                 }
6990
6991                 extent_slot = path->slots[0];
6992                 leaf = path->nodes[0];
6993                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6994         }
6995 #endif
6996         BUG_ON(item_size < sizeof(*ei));
6997         ei = btrfs_item_ptr(leaf, extent_slot,
6998                             struct btrfs_extent_item);
6999         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
7000             key.type == BTRFS_EXTENT_ITEM_KEY) {
7001                 struct btrfs_tree_block_info *bi;
7002                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
7003                 bi = (struct btrfs_tree_block_info *)(ei + 1);
7004                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
7005         }
7006
7007         refs = btrfs_extent_refs(leaf, ei);
7008         if (refs < refs_to_drop) {
7009                 btrfs_err(info,
7010                           "trying to drop %d refs but we only have %Lu for bytenr %Lu",
7011                           refs_to_drop, refs, bytenr);
7012                 ret = -EINVAL;
7013                 btrfs_abort_transaction(trans, ret);
7014                 goto out;
7015         }
7016         refs -= refs_to_drop;
7017
7018         if (refs > 0) {
7019                 if (extent_op)
7020                         __run_delayed_extent_op(extent_op, leaf, ei);
7021                 /*
7022                  * In the case of inline back ref, reference count will
7023                  * be updated by remove_extent_backref
7024                  */
7025                 if (iref) {
7026                         BUG_ON(!found_extent);
7027                 } else {
7028                         btrfs_set_extent_refs(leaf, ei, refs);
7029                         btrfs_mark_buffer_dirty(leaf);
7030                 }
7031                 if (found_extent) {
7032                         ret = remove_extent_backref(trans, extent_root, path,
7033                                                     iref, refs_to_drop,
7034                                                     is_data, &last_ref);
7035                         if (ret) {
7036                                 btrfs_abort_transaction(trans, ret);
7037                                 goto out;
7038                         }
7039                 }
7040                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
7041                                  root_objectid);
7042         } else {
7043                 if (found_extent) {
7044                         BUG_ON(is_data && refs_to_drop !=
7045                                extent_data_ref_count(path, iref));
7046                         if (iref) {
7047                                 BUG_ON(path->slots[0] != extent_slot);
7048                         } else {
7049                                 BUG_ON(path->slots[0] != extent_slot + 1);
7050                                 path->slots[0] = extent_slot;
7051                                 num_to_del = 2;
7052                         }
7053                 }
7054
7055                 last_ref = 1;
7056                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7057                                       num_to_del);
7058                 if (ret) {
7059                         btrfs_abort_transaction(trans, ret);
7060                         goto out;
7061                 }
7062                 btrfs_release_path(path);
7063
7064                 if (is_data) {
7065                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
7066                         if (ret) {
7067                                 btrfs_abort_transaction(trans, ret);
7068                                 goto out;
7069                         }
7070                 }
7071
7072                 ret = add_to_free_space_tree(trans, root->fs_info, bytenr,
7073                                              num_bytes);
7074                 if (ret) {
7075                         btrfs_abort_transaction(trans, ret);
7076                         goto out;
7077                 }
7078
7079                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
7080                 if (ret) {
7081                         btrfs_abort_transaction(trans, ret);
7082                         goto out;
7083                 }
7084         }
7085         btrfs_release_path(path);
7086
7087 out:
7088         btrfs_free_path(path);
7089         return ret;
7090 }
7091
7092 /*
7093  * when we free an block, it is possible (and likely) that we free the last
7094  * delayed ref for that extent as well.  This searches the delayed ref tree for
7095  * a given extent, and if there are no other delayed refs to be processed, it
7096  * removes it from the tree.
7097  */
7098 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7099                                       struct btrfs_root *root, u64 bytenr)
7100 {
7101         struct btrfs_delayed_ref_head *head;
7102         struct btrfs_delayed_ref_root *delayed_refs;
7103         int ret = 0;
7104
7105         delayed_refs = &trans->transaction->delayed_refs;
7106         spin_lock(&delayed_refs->lock);
7107         head = btrfs_find_delayed_ref_head(trans, bytenr);
7108         if (!head)
7109                 goto out_delayed_unlock;
7110
7111         spin_lock(&head->lock);
7112         if (!list_empty(&head->ref_list))
7113                 goto out;
7114
7115         if (head->extent_op) {
7116                 if (!head->must_insert_reserved)
7117                         goto out;
7118                 btrfs_free_delayed_extent_op(head->extent_op);
7119                 head->extent_op = NULL;
7120         }
7121
7122         /*
7123          * waiting for the lock here would deadlock.  If someone else has it
7124          * locked they are already in the process of dropping it anyway
7125          */
7126         if (!mutex_trylock(&head->mutex))
7127                 goto out;
7128
7129         /*
7130          * at this point we have a head with no other entries.  Go
7131          * ahead and process it.
7132          */
7133         head->node.in_tree = 0;
7134         rb_erase(&head->href_node, &delayed_refs->href_root);
7135
7136         atomic_dec(&delayed_refs->num_entries);
7137
7138         /*
7139          * we don't take a ref on the node because we're removing it from the
7140          * tree, so we just steal the ref the tree was holding.
7141          */
7142         delayed_refs->num_heads--;
7143         if (head->processing == 0)
7144                 delayed_refs->num_heads_ready--;
7145         head->processing = 0;
7146         spin_unlock(&head->lock);
7147         spin_unlock(&delayed_refs->lock);
7148
7149         BUG_ON(head->extent_op);
7150         if (head->must_insert_reserved)
7151                 ret = 1;
7152
7153         mutex_unlock(&head->mutex);
7154         btrfs_put_delayed_ref(&head->node);
7155         return ret;
7156 out:
7157         spin_unlock(&head->lock);
7158
7159 out_delayed_unlock:
7160         spin_unlock(&delayed_refs->lock);
7161         return 0;
7162 }
7163
7164 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7165                            struct btrfs_root *root,
7166                            struct extent_buffer *buf,
7167                            u64 parent, int last_ref)
7168 {
7169         int pin = 1;
7170         int ret;
7171
7172         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7173                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7174                                         buf->start, buf->len,
7175                                         parent, root->root_key.objectid,
7176                                         btrfs_header_level(buf),
7177                                         BTRFS_DROP_DELAYED_REF, NULL);
7178                 BUG_ON(ret); /* -ENOMEM */
7179         }
7180
7181         if (!last_ref)
7182                 return;
7183
7184         if (btrfs_header_generation(buf) == trans->transid) {
7185                 struct btrfs_block_group_cache *cache;
7186
7187                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7188                         ret = check_ref_cleanup(trans, root, buf->start);
7189                         if (!ret)
7190                                 goto out;
7191                 }
7192
7193                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
7194
7195                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7196                         pin_down_extent(root, cache, buf->start, buf->len, 1);
7197                         btrfs_put_block_group(cache);
7198                         goto out;
7199                 }
7200
7201                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7202
7203                 btrfs_add_free_space(cache, buf->start, buf->len);
7204                 btrfs_free_reserved_bytes(cache, buf->len, 0);
7205                 btrfs_put_block_group(cache);
7206                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
7207                 pin = 0;
7208         }
7209 out:
7210         if (pin)
7211                 add_pinned_bytes(root->fs_info, buf->len,
7212                                  btrfs_header_level(buf),
7213                                  root->root_key.objectid);
7214
7215         /*
7216          * Deleting the buffer, clear the corrupt flag since it doesn't matter
7217          * anymore.
7218          */
7219         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7220 }
7221
7222 /* Can return -ENOMEM */
7223 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7224                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7225                       u64 owner, u64 offset)
7226 {
7227         int ret;
7228         struct btrfs_fs_info *fs_info = root->fs_info;
7229
7230         if (btrfs_is_testing(fs_info))
7231                 return 0;
7232
7233         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
7234
7235         /*
7236          * tree log blocks never actually go into the extent allocation
7237          * tree, just update pinning info and exit early.
7238          */
7239         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7240                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7241                 /* unlocks the pinned mutex */
7242                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
7243                 ret = 0;
7244         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7245                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7246                                         num_bytes,
7247                                         parent, root_objectid, (int)owner,
7248                                         BTRFS_DROP_DELAYED_REF, NULL);
7249         } else {
7250                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7251                                                 num_bytes,
7252                                                 parent, root_objectid, owner,
7253                                                 offset, 0,
7254                                                 BTRFS_DROP_DELAYED_REF, NULL);
7255         }
7256         return ret;
7257 }
7258
7259 /*
7260  * when we wait for progress in the block group caching, its because
7261  * our allocation attempt failed at least once.  So, we must sleep
7262  * and let some progress happen before we try again.
7263  *
7264  * This function will sleep at least once waiting for new free space to
7265  * show up, and then it will check the block group free space numbers
7266  * for our min num_bytes.  Another option is to have it go ahead
7267  * and look in the rbtree for a free extent of a given size, but this
7268  * is a good start.
7269  *
7270  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7271  * any of the information in this block group.
7272  */
7273 static noinline void
7274 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7275                                 u64 num_bytes)
7276 {
7277         struct btrfs_caching_control *caching_ctl;
7278
7279         caching_ctl = get_caching_control(cache);
7280         if (!caching_ctl)
7281                 return;
7282
7283         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7284                    (cache->free_space_ctl->free_space >= num_bytes));
7285
7286         put_caching_control(caching_ctl);
7287 }
7288
7289 static noinline int
7290 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7291 {
7292         struct btrfs_caching_control *caching_ctl;
7293         int ret = 0;
7294
7295         caching_ctl = get_caching_control(cache);
7296         if (!caching_ctl)
7297                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7298
7299         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7300         if (cache->cached == BTRFS_CACHE_ERROR)
7301                 ret = -EIO;
7302         put_caching_control(caching_ctl);
7303         return ret;
7304 }
7305
7306 int __get_raid_index(u64 flags)
7307 {
7308         if (flags & BTRFS_BLOCK_GROUP_RAID10)
7309                 return BTRFS_RAID_RAID10;
7310         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
7311                 return BTRFS_RAID_RAID1;
7312         else if (flags & BTRFS_BLOCK_GROUP_DUP)
7313                 return BTRFS_RAID_DUP;
7314         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
7315                 return BTRFS_RAID_RAID0;
7316         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
7317                 return BTRFS_RAID_RAID5;
7318         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
7319                 return BTRFS_RAID_RAID6;
7320
7321         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
7322 }
7323
7324 int get_block_group_index(struct btrfs_block_group_cache *cache)
7325 {
7326         return __get_raid_index(cache->flags);
7327 }
7328
7329 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7330         [BTRFS_RAID_RAID10]     = "raid10",
7331         [BTRFS_RAID_RAID1]      = "raid1",
7332         [BTRFS_RAID_DUP]        = "dup",
7333         [BTRFS_RAID_RAID0]      = "raid0",
7334         [BTRFS_RAID_SINGLE]     = "single",
7335         [BTRFS_RAID_RAID5]      = "raid5",
7336         [BTRFS_RAID_RAID6]      = "raid6",
7337 };
7338
7339 static const char *get_raid_name(enum btrfs_raid_types type)
7340 {
7341         if (type >= BTRFS_NR_RAID_TYPES)
7342                 return NULL;
7343
7344         return btrfs_raid_type_names[type];
7345 }
7346
7347 enum btrfs_loop_type {
7348         LOOP_CACHING_NOWAIT = 0,
7349         LOOP_CACHING_WAIT = 1,
7350         LOOP_ALLOC_CHUNK = 2,
7351         LOOP_NO_EMPTY_SIZE = 3,
7352 };
7353
7354 static inline void
7355 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7356                        int delalloc)
7357 {
7358         if (delalloc)
7359                 down_read(&cache->data_rwsem);
7360 }
7361
7362 static inline void
7363 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7364                        int delalloc)
7365 {
7366         btrfs_get_block_group(cache);
7367         if (delalloc)
7368                 down_read(&cache->data_rwsem);
7369 }
7370
7371 static struct btrfs_block_group_cache *
7372 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7373                    struct btrfs_free_cluster *cluster,
7374                    int delalloc)
7375 {
7376         struct btrfs_block_group_cache *used_bg = NULL;
7377
7378         spin_lock(&cluster->refill_lock);
7379         while (1) {
7380                 used_bg = cluster->block_group;
7381                 if (!used_bg)
7382                         return NULL;
7383
7384                 if (used_bg == block_group)
7385                         return used_bg;
7386
7387                 btrfs_get_block_group(used_bg);
7388
7389                 if (!delalloc)
7390                         return used_bg;
7391
7392                 if (down_read_trylock(&used_bg->data_rwsem))
7393                         return used_bg;
7394
7395                 spin_unlock(&cluster->refill_lock);
7396
7397                 down_read(&used_bg->data_rwsem);
7398
7399                 spin_lock(&cluster->refill_lock);
7400                 if (used_bg == cluster->block_group)
7401                         return used_bg;
7402
7403                 up_read(&used_bg->data_rwsem);
7404                 btrfs_put_block_group(used_bg);
7405         }
7406 }
7407
7408 static inline void
7409 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7410                          int delalloc)
7411 {
7412         if (delalloc)
7413                 up_read(&cache->data_rwsem);
7414         btrfs_put_block_group(cache);
7415 }
7416
7417 /*
7418  * walks the btree of allocated extents and find a hole of a given size.
7419  * The key ins is changed to record the hole:
7420  * ins->objectid == start position
7421  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7422  * ins->offset == the size of the hole.
7423  * Any available blocks before search_start are skipped.
7424  *
7425  * If there is no suitable free space, we will record the max size of
7426  * the free space extent currently.
7427  */
7428 static noinline int find_free_extent(struct btrfs_root *orig_root,
7429                                 u64 ram_bytes, u64 num_bytes, u64 empty_size,
7430                                 u64 hint_byte, struct btrfs_key *ins,
7431                                 u64 flags, int delalloc)
7432 {
7433         int ret = 0;
7434         struct btrfs_root *root = orig_root->fs_info->extent_root;
7435         struct btrfs_free_cluster *last_ptr = NULL;
7436         struct btrfs_block_group_cache *block_group = NULL;
7437         u64 search_start = 0;
7438         u64 max_extent_size = 0;
7439         u64 empty_cluster = 0;
7440         struct btrfs_space_info *space_info;
7441         int loop = 0;
7442         int index = __get_raid_index(flags);
7443         bool failed_cluster_refill = false;
7444         bool failed_alloc = false;
7445         bool use_cluster = true;
7446         bool have_caching_bg = false;
7447         bool orig_have_caching_bg = false;
7448         bool full_search = false;
7449
7450         WARN_ON(num_bytes < root->sectorsize);
7451         ins->type = BTRFS_EXTENT_ITEM_KEY;
7452         ins->objectid = 0;
7453         ins->offset = 0;
7454
7455         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
7456
7457         space_info = __find_space_info(root->fs_info, flags);
7458         if (!space_info) {
7459                 btrfs_err(root->fs_info, "No space info for %llu", flags);
7460                 return -ENOSPC;
7461         }
7462
7463         /*
7464          * If our free space is heavily fragmented we may not be able to make
7465          * big contiguous allocations, so instead of doing the expensive search
7466          * for free space, simply return ENOSPC with our max_extent_size so we
7467          * can go ahead and search for a more manageable chunk.
7468          *
7469          * If our max_extent_size is large enough for our allocation simply
7470          * disable clustering since we will likely not be able to find enough
7471          * space to create a cluster and induce latency trying.
7472          */
7473         if (unlikely(space_info->max_extent_size)) {
7474                 spin_lock(&space_info->lock);
7475                 if (space_info->max_extent_size &&
7476                     num_bytes > space_info->max_extent_size) {
7477                         ins->offset = space_info->max_extent_size;
7478                         spin_unlock(&space_info->lock);
7479                         return -ENOSPC;
7480                 } else if (space_info->max_extent_size) {
7481                         use_cluster = false;
7482                 }
7483                 spin_unlock(&space_info->lock);
7484         }
7485
7486         last_ptr = fetch_cluster_info(orig_root, space_info, &empty_cluster);
7487         if (last_ptr) {
7488                 spin_lock(&last_ptr->lock);
7489                 if (last_ptr->block_group)
7490                         hint_byte = last_ptr->window_start;
7491                 if (last_ptr->fragmented) {
7492                         /*
7493                          * We still set window_start so we can keep track of the
7494                          * last place we found an allocation to try and save
7495                          * some time.
7496                          */
7497                         hint_byte = last_ptr->window_start;
7498                         use_cluster = false;
7499                 }
7500                 spin_unlock(&last_ptr->lock);
7501         }
7502
7503         search_start = max(search_start, first_logical_byte(root, 0));
7504         search_start = max(search_start, hint_byte);
7505         if (search_start == hint_byte) {
7506                 block_group = btrfs_lookup_block_group(root->fs_info,
7507                                                        search_start);
7508                 /*
7509                  * we don't want to use the block group if it doesn't match our
7510                  * allocation bits, or if its not cached.
7511                  *
7512                  * However if we are re-searching with an ideal block group
7513                  * picked out then we don't care that the block group is cached.
7514                  */
7515                 if (block_group && block_group_bits(block_group, flags) &&
7516                     block_group->cached != BTRFS_CACHE_NO) {
7517                         down_read(&space_info->groups_sem);
7518                         if (list_empty(&block_group->list) ||
7519                             block_group->ro) {
7520                                 /*
7521                                  * someone is removing this block group,
7522                                  * we can't jump into the have_block_group
7523                                  * target because our list pointers are not
7524                                  * valid
7525                                  */
7526                                 btrfs_put_block_group(block_group);
7527                                 up_read(&space_info->groups_sem);
7528                         } else {
7529                                 index = get_block_group_index(block_group);
7530                                 btrfs_lock_block_group(block_group, delalloc);
7531                                 goto have_block_group;
7532                         }
7533                 } else if (block_group) {
7534                         btrfs_put_block_group(block_group);
7535                 }
7536         }
7537 search:
7538         have_caching_bg = false;
7539         if (index == 0 || index == __get_raid_index(flags))
7540                 full_search = true;
7541         down_read(&space_info->groups_sem);
7542         list_for_each_entry(block_group, &space_info->block_groups[index],
7543                             list) {
7544                 u64 offset;
7545                 int cached;
7546
7547                 btrfs_grab_block_group(block_group, delalloc);
7548                 search_start = block_group->key.objectid;
7549
7550                 /*
7551                  * this can happen if we end up cycling through all the
7552                  * raid types, but we want to make sure we only allocate
7553                  * for the proper type.
7554                  */
7555                 if (!block_group_bits(block_group, flags)) {
7556                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7557                                 BTRFS_BLOCK_GROUP_RAID1 |
7558                                 BTRFS_BLOCK_GROUP_RAID5 |
7559                                 BTRFS_BLOCK_GROUP_RAID6 |
7560                                 BTRFS_BLOCK_GROUP_RAID10;
7561
7562                         /*
7563                          * if they asked for extra copies and this block group
7564                          * doesn't provide them, bail.  This does allow us to
7565                          * fill raid0 from raid1.
7566                          */
7567                         if ((flags & extra) && !(block_group->flags & extra))
7568                                 goto loop;
7569                 }
7570
7571 have_block_group:
7572                 cached = block_group_cache_done(block_group);
7573                 if (unlikely(!cached)) {
7574                         have_caching_bg = true;
7575                         ret = cache_block_group(block_group, 0);
7576                         BUG_ON(ret < 0);
7577                         ret = 0;
7578                 }
7579
7580                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7581                         goto loop;
7582                 if (unlikely(block_group->ro))
7583                         goto loop;
7584
7585                 /*
7586                  * Ok we want to try and use the cluster allocator, so
7587                  * lets look there
7588                  */
7589                 if (last_ptr && use_cluster) {
7590                         struct btrfs_block_group_cache *used_block_group;
7591                         unsigned long aligned_cluster;
7592                         /*
7593                          * the refill lock keeps out other
7594                          * people trying to start a new cluster
7595                          */
7596                         used_block_group = btrfs_lock_cluster(block_group,
7597                                                               last_ptr,
7598                                                               delalloc);
7599                         if (!used_block_group)
7600                                 goto refill_cluster;
7601
7602                         if (used_block_group != block_group &&
7603                             (used_block_group->ro ||
7604                              !block_group_bits(used_block_group, flags)))
7605                                 goto release_cluster;
7606
7607                         offset = btrfs_alloc_from_cluster(used_block_group,
7608                                                 last_ptr,
7609                                                 num_bytes,
7610                                                 used_block_group->key.objectid,
7611                                                 &max_extent_size);
7612                         if (offset) {
7613                                 /* we have a block, we're done */
7614                                 spin_unlock(&last_ptr->refill_lock);
7615                                 trace_btrfs_reserve_extent_cluster(root,
7616                                                 used_block_group,
7617                                                 search_start, num_bytes);
7618                                 if (used_block_group != block_group) {
7619                                         btrfs_release_block_group(block_group,
7620                                                                   delalloc);
7621                                         block_group = used_block_group;
7622                                 }
7623                                 goto checks;
7624                         }
7625
7626                         WARN_ON(last_ptr->block_group != used_block_group);
7627 release_cluster:
7628                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7629                          * set up a new clusters, so lets just skip it
7630                          * and let the allocator find whatever block
7631                          * it can find.  If we reach this point, we
7632                          * will have tried the cluster allocator
7633                          * plenty of times and not have found
7634                          * anything, so we are likely way too
7635                          * fragmented for the clustering stuff to find
7636                          * anything.
7637                          *
7638                          * However, if the cluster is taken from the
7639                          * current block group, release the cluster
7640                          * first, so that we stand a better chance of
7641                          * succeeding in the unclustered
7642                          * allocation.  */
7643                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7644                             used_block_group != block_group) {
7645                                 spin_unlock(&last_ptr->refill_lock);
7646                                 btrfs_release_block_group(used_block_group,
7647                                                           delalloc);
7648                                 goto unclustered_alloc;
7649                         }
7650
7651                         /*
7652                          * this cluster didn't work out, free it and
7653                          * start over
7654                          */
7655                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7656
7657                         if (used_block_group != block_group)
7658                                 btrfs_release_block_group(used_block_group,
7659                                                           delalloc);
7660 refill_cluster:
7661                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7662                                 spin_unlock(&last_ptr->refill_lock);
7663                                 goto unclustered_alloc;
7664                         }
7665
7666                         aligned_cluster = max_t(unsigned long,
7667                                                 empty_cluster + empty_size,
7668                                               block_group->full_stripe_len);
7669
7670                         /* allocate a cluster in this block group */
7671                         ret = btrfs_find_space_cluster(root, block_group,
7672                                                        last_ptr, search_start,
7673                                                        num_bytes,
7674                                                        aligned_cluster);
7675                         if (ret == 0) {
7676                                 /*
7677                                  * now pull our allocation out of this
7678                                  * cluster
7679                                  */
7680                                 offset = btrfs_alloc_from_cluster(block_group,
7681                                                         last_ptr,
7682                                                         num_bytes,
7683                                                         search_start,
7684                                                         &max_extent_size);
7685                                 if (offset) {
7686                                         /* we found one, proceed */
7687                                         spin_unlock(&last_ptr->refill_lock);
7688                                         trace_btrfs_reserve_extent_cluster(root,
7689                                                 block_group, search_start,
7690                                                 num_bytes);
7691                                         goto checks;
7692                                 }
7693                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7694                                    && !failed_cluster_refill) {
7695                                 spin_unlock(&last_ptr->refill_lock);
7696
7697                                 failed_cluster_refill = true;
7698                                 wait_block_group_cache_progress(block_group,
7699                                        num_bytes + empty_cluster + empty_size);
7700                                 goto have_block_group;
7701                         }
7702
7703                         /*
7704                          * at this point we either didn't find a cluster
7705                          * or we weren't able to allocate a block from our
7706                          * cluster.  Free the cluster we've been trying
7707                          * to use, and go to the next block group
7708                          */
7709                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7710                         spin_unlock(&last_ptr->refill_lock);
7711                         goto loop;
7712                 }
7713
7714 unclustered_alloc:
7715                 /*
7716                  * We are doing an unclustered alloc, set the fragmented flag so
7717                  * we don't bother trying to setup a cluster again until we get
7718                  * more space.
7719                  */
7720                 if (unlikely(last_ptr)) {
7721                         spin_lock(&last_ptr->lock);
7722                         last_ptr->fragmented = 1;
7723                         spin_unlock(&last_ptr->lock);
7724                 }
7725                 spin_lock(&block_group->free_space_ctl->tree_lock);
7726                 if (cached &&
7727                     block_group->free_space_ctl->free_space <
7728                     num_bytes + empty_cluster + empty_size) {
7729                         if (block_group->free_space_ctl->free_space >
7730                             max_extent_size)
7731                                 max_extent_size =
7732                                         block_group->free_space_ctl->free_space;
7733                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7734                         goto loop;
7735                 }
7736                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7737
7738                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7739                                                     num_bytes, empty_size,
7740                                                     &max_extent_size);
7741                 /*
7742                  * If we didn't find a chunk, and we haven't failed on this
7743                  * block group before, and this block group is in the middle of
7744                  * caching and we are ok with waiting, then go ahead and wait
7745                  * for progress to be made, and set failed_alloc to true.
7746                  *
7747                  * If failed_alloc is true then we've already waited on this
7748                  * block group once and should move on to the next block group.
7749                  */
7750                 if (!offset && !failed_alloc && !cached &&
7751                     loop > LOOP_CACHING_NOWAIT) {
7752                         wait_block_group_cache_progress(block_group,
7753                                                 num_bytes + empty_size);
7754                         failed_alloc = true;
7755                         goto have_block_group;
7756                 } else if (!offset) {
7757                         goto loop;
7758                 }
7759 checks:
7760                 search_start = ALIGN(offset, root->stripesize);
7761
7762                 /* move on to the next group */
7763                 if (search_start + num_bytes >
7764                     block_group->key.objectid + block_group->key.offset) {
7765                         btrfs_add_free_space(block_group, offset, num_bytes);
7766                         goto loop;
7767                 }
7768
7769                 if (offset < search_start)
7770                         btrfs_add_free_space(block_group, offset,
7771                                              search_start - offset);
7772                 BUG_ON(offset > search_start);
7773
7774                 ret = btrfs_add_reserved_bytes(block_group, ram_bytes,
7775                                 num_bytes, delalloc);
7776                 if (ret == -EAGAIN) {
7777                         btrfs_add_free_space(block_group, offset, num_bytes);
7778                         goto loop;
7779                 }
7780                 btrfs_inc_block_group_reservations(block_group);
7781
7782                 /* we are all good, lets return */
7783                 ins->objectid = search_start;
7784                 ins->offset = num_bytes;
7785
7786                 trace_btrfs_reserve_extent(orig_root, block_group,
7787                                            search_start, num_bytes);
7788                 btrfs_release_block_group(block_group, delalloc);
7789                 break;
7790 loop:
7791                 failed_cluster_refill = false;
7792                 failed_alloc = false;
7793                 BUG_ON(index != get_block_group_index(block_group));
7794                 btrfs_release_block_group(block_group, delalloc);
7795         }
7796         up_read(&space_info->groups_sem);
7797
7798         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7799                 && !orig_have_caching_bg)
7800                 orig_have_caching_bg = true;
7801
7802         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7803                 goto search;
7804
7805         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7806                 goto search;
7807
7808         /*
7809          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7810          *                      caching kthreads as we move along
7811          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7812          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7813          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7814          *                      again
7815          */
7816         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7817                 index = 0;
7818                 if (loop == LOOP_CACHING_NOWAIT) {
7819                         /*
7820                          * We want to skip the LOOP_CACHING_WAIT step if we
7821                          * don't have any uncached bgs and we've already done a
7822                          * full search through.
7823                          */
7824                         if (orig_have_caching_bg || !full_search)
7825                                 loop = LOOP_CACHING_WAIT;
7826                         else
7827                                 loop = LOOP_ALLOC_CHUNK;
7828                 } else {
7829                         loop++;
7830                 }
7831
7832                 if (loop == LOOP_ALLOC_CHUNK) {
7833                         struct btrfs_trans_handle *trans;
7834                         int exist = 0;
7835
7836                         trans = current->journal_info;
7837                         if (trans)
7838                                 exist = 1;
7839                         else
7840                                 trans = btrfs_join_transaction(root);
7841
7842                         if (IS_ERR(trans)) {
7843                                 ret = PTR_ERR(trans);
7844                                 goto out;
7845                         }
7846
7847                         ret = do_chunk_alloc(trans, root, flags,
7848                                              CHUNK_ALLOC_FORCE);
7849
7850                         /*
7851                          * If we can't allocate a new chunk we've already looped
7852                          * through at least once, move on to the NO_EMPTY_SIZE
7853                          * case.
7854                          */
7855                         if (ret == -ENOSPC)
7856                                 loop = LOOP_NO_EMPTY_SIZE;
7857
7858                         /*
7859                          * Do not bail out on ENOSPC since we
7860                          * can do more things.
7861                          */
7862                         if (ret < 0 && ret != -ENOSPC)
7863                                 btrfs_abort_transaction(trans, ret);
7864                         else
7865                                 ret = 0;
7866                         if (!exist)
7867                                 btrfs_end_transaction(trans, root);
7868                         if (ret)
7869                                 goto out;
7870                 }
7871
7872                 if (loop == LOOP_NO_EMPTY_SIZE) {
7873                         /*
7874                          * Don't loop again if we already have no empty_size and
7875                          * no empty_cluster.
7876                          */
7877                         if (empty_size == 0 &&
7878                             empty_cluster == 0) {
7879                                 ret = -ENOSPC;
7880                                 goto out;
7881                         }
7882                         empty_size = 0;
7883                         empty_cluster = 0;
7884                 }
7885
7886                 goto search;
7887         } else if (!ins->objectid) {
7888                 ret = -ENOSPC;
7889         } else if (ins->objectid) {
7890                 if (!use_cluster && last_ptr) {
7891                         spin_lock(&last_ptr->lock);
7892                         last_ptr->window_start = ins->objectid;
7893                         spin_unlock(&last_ptr->lock);
7894                 }
7895                 ret = 0;
7896         }
7897 out:
7898         if (ret == -ENOSPC) {
7899                 spin_lock(&space_info->lock);
7900                 space_info->max_extent_size = max_extent_size;
7901                 spin_unlock(&space_info->lock);
7902                 ins->offset = max_extent_size;
7903         }
7904         return ret;
7905 }
7906
7907 static void dump_space_info(struct btrfs_fs_info *fs_info,
7908                             struct btrfs_space_info *info, u64 bytes,
7909                             int dump_block_groups)
7910 {
7911         struct btrfs_block_group_cache *cache;
7912         int index = 0;
7913
7914         spin_lock(&info->lock);
7915         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
7916                    info->flags,
7917                    info->total_bytes - info->bytes_used - info->bytes_pinned -
7918                    info->bytes_reserved - info->bytes_readonly -
7919                    info->bytes_may_use, (info->full) ? "" : "not ");
7920         btrfs_info(fs_info,
7921                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
7922                 info->total_bytes, info->bytes_used, info->bytes_pinned,
7923                 info->bytes_reserved, info->bytes_may_use,
7924                 info->bytes_readonly);
7925         spin_unlock(&info->lock);
7926
7927         if (!dump_block_groups)
7928                 return;
7929
7930         down_read(&info->groups_sem);
7931 again:
7932         list_for_each_entry(cache, &info->block_groups[index], list) {
7933                 spin_lock(&cache->lock);
7934                 btrfs_info(fs_info,
7935                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
7936                         cache->key.objectid, cache->key.offset,
7937                         btrfs_block_group_used(&cache->item), cache->pinned,
7938                         cache->reserved, cache->ro ? "[readonly]" : "");
7939                 btrfs_dump_free_space(cache, bytes);
7940                 spin_unlock(&cache->lock);
7941         }
7942         if (++index < BTRFS_NR_RAID_TYPES)
7943                 goto again;
7944         up_read(&info->groups_sem);
7945 }
7946
7947 int btrfs_reserve_extent(struct btrfs_root *root, u64 ram_bytes,
7948                          u64 num_bytes, u64 min_alloc_size,
7949                          u64 empty_size, u64 hint_byte,
7950                          struct btrfs_key *ins, int is_data, int delalloc)
7951 {
7952         struct btrfs_fs_info *fs_info = root->fs_info;
7953         bool final_tried = num_bytes == min_alloc_size;
7954         u64 flags;
7955         int ret;
7956
7957         flags = btrfs_get_alloc_profile(root, is_data);
7958 again:
7959         WARN_ON(num_bytes < root->sectorsize);
7960         ret = find_free_extent(root, ram_bytes, num_bytes, empty_size,
7961                                hint_byte, ins, flags, delalloc);
7962         if (!ret && !is_data) {
7963                 btrfs_dec_block_group_reservations(fs_info, ins->objectid);
7964         } else if (ret == -ENOSPC) {
7965                 if (!final_tried && ins->offset) {
7966                         num_bytes = min(num_bytes >> 1, ins->offset);
7967                         num_bytes = round_down(num_bytes, root->sectorsize);
7968                         num_bytes = max(num_bytes, min_alloc_size);
7969                         ram_bytes = num_bytes;
7970                         if (num_bytes == min_alloc_size)
7971                                 final_tried = true;
7972                         goto again;
7973                 } else if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
7974                         struct btrfs_space_info *sinfo;
7975
7976                         sinfo = __find_space_info(fs_info, flags);
7977                         btrfs_err(root->fs_info,
7978                                   "allocation failed flags %llu, wanted %llu",
7979                                   flags, num_bytes);
7980                         if (sinfo)
7981                                 dump_space_info(fs_info, sinfo, num_bytes, 1);
7982                 }
7983         }
7984
7985         return ret;
7986 }
7987
7988 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7989                                         u64 start, u64 len,
7990                                         int pin, int delalloc)
7991 {
7992         struct btrfs_block_group_cache *cache;
7993         int ret = 0;
7994
7995         cache = btrfs_lookup_block_group(root->fs_info, start);
7996         if (!cache) {
7997                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7998                         start);
7999                 return -ENOSPC;
8000         }
8001
8002         if (pin)
8003                 pin_down_extent(root, cache, start, len, 1);
8004         else {
8005                 if (btrfs_test_opt(root->fs_info, DISCARD))
8006                         ret = btrfs_discard_extent(root, start, len, NULL);
8007                 btrfs_add_free_space(cache, start, len);
8008                 btrfs_free_reserved_bytes(cache, len, delalloc);
8009                 trace_btrfs_reserved_extent_free(root, start, len);
8010         }
8011
8012         btrfs_put_block_group(cache);
8013         return ret;
8014 }
8015
8016 int btrfs_free_reserved_extent(struct btrfs_root *root,
8017                                u64 start, u64 len, int delalloc)
8018 {
8019         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
8020 }
8021
8022 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
8023                                        u64 start, u64 len)
8024 {
8025         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
8026 }
8027
8028 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8029                                       struct btrfs_root *root,
8030                                       u64 parent, u64 root_objectid,
8031                                       u64 flags, u64 owner, u64 offset,
8032                                       struct btrfs_key *ins, int ref_mod)
8033 {
8034         int ret;
8035         struct btrfs_fs_info *fs_info = root->fs_info;
8036         struct btrfs_extent_item *extent_item;
8037         struct btrfs_extent_inline_ref *iref;
8038         struct btrfs_path *path;
8039         struct extent_buffer *leaf;
8040         int type;
8041         u32 size;
8042
8043         if (parent > 0)
8044                 type = BTRFS_SHARED_DATA_REF_KEY;
8045         else
8046                 type = BTRFS_EXTENT_DATA_REF_KEY;
8047
8048         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8049
8050         path = btrfs_alloc_path();
8051         if (!path)
8052                 return -ENOMEM;
8053
8054         path->leave_spinning = 1;
8055         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8056                                       ins, size);
8057         if (ret) {
8058                 btrfs_free_path(path);
8059                 return ret;
8060         }
8061
8062         leaf = path->nodes[0];
8063         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8064                                      struct btrfs_extent_item);
8065         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8066         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8067         btrfs_set_extent_flags(leaf, extent_item,
8068                                flags | BTRFS_EXTENT_FLAG_DATA);
8069
8070         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8071         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8072         if (parent > 0) {
8073                 struct btrfs_shared_data_ref *ref;
8074                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8075                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8076                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8077         } else {
8078                 struct btrfs_extent_data_ref *ref;
8079                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8080                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8081                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8082                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8083                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8084         }
8085
8086         btrfs_mark_buffer_dirty(path->nodes[0]);
8087         btrfs_free_path(path);
8088
8089         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8090                                           ins->offset);
8091         if (ret)
8092                 return ret;
8093
8094         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
8095         if (ret) { /* -ENOENT, logic error */
8096                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8097                         ins->objectid, ins->offset);
8098                 BUG();
8099         }
8100         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
8101         return ret;
8102 }
8103
8104 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8105                                      struct btrfs_root *root,
8106                                      u64 parent, u64 root_objectid,
8107                                      u64 flags, struct btrfs_disk_key *key,
8108                                      int level, struct btrfs_key *ins)
8109 {
8110         int ret;
8111         struct btrfs_fs_info *fs_info = root->fs_info;
8112         struct btrfs_extent_item *extent_item;
8113         struct btrfs_tree_block_info *block_info;
8114         struct btrfs_extent_inline_ref *iref;
8115         struct btrfs_path *path;
8116         struct extent_buffer *leaf;
8117         u32 size = sizeof(*extent_item) + sizeof(*iref);
8118         u64 num_bytes = ins->offset;
8119         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8120                                                  SKINNY_METADATA);
8121
8122         if (!skinny_metadata)
8123                 size += sizeof(*block_info);
8124
8125         path = btrfs_alloc_path();
8126         if (!path) {
8127                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
8128                                                    root->nodesize);
8129                 return -ENOMEM;
8130         }
8131
8132         path->leave_spinning = 1;
8133         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8134                                       ins, size);
8135         if (ret) {
8136                 btrfs_free_path(path);
8137                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
8138                                                    root->nodesize);
8139                 return ret;
8140         }
8141
8142         leaf = path->nodes[0];
8143         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8144                                      struct btrfs_extent_item);
8145         btrfs_set_extent_refs(leaf, extent_item, 1);
8146         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8147         btrfs_set_extent_flags(leaf, extent_item,
8148                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8149
8150         if (skinny_metadata) {
8151                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8152                 num_bytes = root->nodesize;
8153         } else {
8154                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8155                 btrfs_set_tree_block_key(leaf, block_info, key);
8156                 btrfs_set_tree_block_level(leaf, block_info, level);
8157                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8158         }
8159
8160         if (parent > 0) {
8161                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8162                 btrfs_set_extent_inline_ref_type(leaf, iref,
8163                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8164                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8165         } else {
8166                 btrfs_set_extent_inline_ref_type(leaf, iref,
8167                                                  BTRFS_TREE_BLOCK_REF_KEY);
8168                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8169         }
8170
8171         btrfs_mark_buffer_dirty(leaf);
8172         btrfs_free_path(path);
8173
8174         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8175                                           num_bytes);
8176         if (ret)
8177                 return ret;
8178
8179         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
8180                                  1);
8181         if (ret) { /* -ENOENT, logic error */
8182                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8183                         ins->objectid, ins->offset);
8184                 BUG();
8185         }
8186
8187         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
8188         return ret;
8189 }
8190
8191 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8192                                      struct btrfs_root *root,
8193                                      u64 root_objectid, u64 owner,
8194                                      u64 offset, u64 ram_bytes,
8195                                      struct btrfs_key *ins)
8196 {
8197         int ret;
8198
8199         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
8200
8201         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
8202                                          ins->offset, 0,
8203                                          root_objectid, owner, offset,
8204                                          ram_bytes, BTRFS_ADD_DELAYED_EXTENT,
8205                                          NULL);
8206         return ret;
8207 }
8208
8209 /*
8210  * this is used by the tree logging recovery code.  It records that
8211  * an extent has been allocated and makes sure to clear the free
8212  * space cache bits as well
8213  */
8214 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8215                                    struct btrfs_root *root,
8216                                    u64 root_objectid, u64 owner, u64 offset,
8217                                    struct btrfs_key *ins)
8218 {
8219         int ret;
8220         struct btrfs_block_group_cache *block_group;
8221         struct btrfs_space_info *space_info;
8222
8223         /*
8224          * Mixed block groups will exclude before processing the log so we only
8225          * need to do the exclude dance if this fs isn't mixed.
8226          */
8227         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
8228                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
8229                 if (ret)
8230                         return ret;
8231         }
8232
8233         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
8234         if (!block_group)
8235                 return -EINVAL;
8236
8237         space_info = block_group->space_info;
8238         spin_lock(&space_info->lock);
8239         spin_lock(&block_group->lock);
8240         space_info->bytes_reserved += ins->offset;
8241         block_group->reserved += ins->offset;
8242         spin_unlock(&block_group->lock);
8243         spin_unlock(&space_info->lock);
8244
8245         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
8246                                          0, owner, offset, ins, 1);
8247         btrfs_put_block_group(block_group);
8248         return ret;
8249 }
8250
8251 static struct extent_buffer *
8252 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8253                       u64 bytenr, int level)
8254 {
8255         struct extent_buffer *buf;
8256
8257         buf = btrfs_find_create_tree_block(root, bytenr);
8258         if (IS_ERR(buf))
8259                 return buf;
8260
8261         btrfs_set_header_generation(buf, trans->transid);
8262         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8263         btrfs_tree_lock(buf);
8264         clean_tree_block(trans, root->fs_info, buf);
8265         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8266
8267         btrfs_set_lock_blocking(buf);
8268         set_extent_buffer_uptodate(buf);
8269
8270         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8271                 buf->log_index = root->log_transid % 2;
8272                 /*
8273                  * we allow two log transactions at a time, use different
8274                  * EXENT bit to differentiate dirty pages.
8275                  */
8276                 if (buf->log_index == 0)
8277                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8278                                         buf->start + buf->len - 1, GFP_NOFS);
8279                 else
8280                         set_extent_new(&root->dirty_log_pages, buf->start,
8281                                         buf->start + buf->len - 1);
8282         } else {
8283                 buf->log_index = -1;
8284                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8285                          buf->start + buf->len - 1, GFP_NOFS);
8286         }
8287         trans->dirty = true;
8288         /* this returns a buffer locked for blocking */
8289         return buf;
8290 }
8291
8292 static struct btrfs_block_rsv *
8293 use_block_rsv(struct btrfs_trans_handle *trans,
8294               struct btrfs_root *root, u32 blocksize)
8295 {
8296         struct btrfs_block_rsv *block_rsv;
8297         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
8298         int ret;
8299         bool global_updated = false;
8300
8301         block_rsv = get_block_rsv(trans, root);
8302
8303         if (unlikely(block_rsv->size == 0))
8304                 goto try_reserve;
8305 again:
8306         ret = block_rsv_use_bytes(block_rsv, blocksize);
8307         if (!ret)
8308                 return block_rsv;
8309
8310         if (block_rsv->failfast)
8311                 return ERR_PTR(ret);
8312
8313         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8314                 global_updated = true;
8315                 update_global_block_rsv(root->fs_info);
8316                 goto again;
8317         }
8318
8319         if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
8320                 static DEFINE_RATELIMIT_STATE(_rs,
8321                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8322                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8323                 if (__ratelimit(&_rs))
8324                         WARN(1, KERN_DEBUG
8325                                 "BTRFS: block rsv returned %d\n", ret);
8326         }
8327 try_reserve:
8328         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8329                                      BTRFS_RESERVE_NO_FLUSH);
8330         if (!ret)
8331                 return block_rsv;
8332         /*
8333          * If we couldn't reserve metadata bytes try and use some from
8334          * the global reserve if its space type is the same as the global
8335          * reservation.
8336          */
8337         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8338             block_rsv->space_info == global_rsv->space_info) {
8339                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8340                 if (!ret)
8341                         return global_rsv;
8342         }
8343         return ERR_PTR(ret);
8344 }
8345
8346 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8347                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8348 {
8349         block_rsv_add_bytes(block_rsv, blocksize, 0);
8350         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8351 }
8352
8353 /*
8354  * finds a free extent and does all the dirty work required for allocation
8355  * returns the tree buffer or an ERR_PTR on error.
8356  */
8357 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8358                                         struct btrfs_root *root,
8359                                         u64 parent, u64 root_objectid,
8360                                         struct btrfs_disk_key *key, int level,
8361                                         u64 hint, u64 empty_size)
8362 {
8363         struct btrfs_key ins;
8364         struct btrfs_block_rsv *block_rsv;
8365         struct extent_buffer *buf;
8366         struct btrfs_delayed_extent_op *extent_op;
8367         u64 flags = 0;
8368         int ret;
8369         u32 blocksize = root->nodesize;
8370         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8371                                                  SKINNY_METADATA);
8372
8373 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8374         if (btrfs_is_testing(root->fs_info)) {
8375                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8376                                             level);
8377                 if (!IS_ERR(buf))
8378                         root->alloc_bytenr += blocksize;
8379                 return buf;
8380         }
8381 #endif
8382
8383         block_rsv = use_block_rsv(trans, root, blocksize);
8384         if (IS_ERR(block_rsv))
8385                 return ERR_CAST(block_rsv);
8386
8387         ret = btrfs_reserve_extent(root, blocksize, blocksize, blocksize,
8388                                    empty_size, hint, &ins, 0, 0);
8389         if (ret)
8390                 goto out_unuse;
8391
8392         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8393         if (IS_ERR(buf)) {
8394                 ret = PTR_ERR(buf);
8395                 goto out_free_reserved;
8396         }
8397
8398         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8399                 if (parent == 0)
8400                         parent = ins.objectid;
8401                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8402         } else
8403                 BUG_ON(parent > 0);
8404
8405         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8406                 extent_op = btrfs_alloc_delayed_extent_op();
8407                 if (!extent_op) {
8408                         ret = -ENOMEM;
8409                         goto out_free_buf;
8410                 }
8411                 if (key)
8412                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8413                 else
8414                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8415                 extent_op->flags_to_set = flags;
8416                 extent_op->update_key = skinny_metadata ? false : true;
8417                 extent_op->update_flags = true;
8418                 extent_op->is_data = false;
8419                 extent_op->level = level;
8420
8421                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
8422                                                  ins.objectid, ins.offset,
8423                                                  parent, root_objectid, level,
8424                                                  BTRFS_ADD_DELAYED_EXTENT,
8425                                                  extent_op);
8426                 if (ret)
8427                         goto out_free_delayed;
8428         }
8429         return buf;
8430
8431 out_free_delayed:
8432         btrfs_free_delayed_extent_op(extent_op);
8433 out_free_buf:
8434         free_extent_buffer(buf);
8435 out_free_reserved:
8436         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
8437 out_unuse:
8438         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
8439         return ERR_PTR(ret);
8440 }
8441
8442 struct walk_control {
8443         u64 refs[BTRFS_MAX_LEVEL];
8444         u64 flags[BTRFS_MAX_LEVEL];
8445         struct btrfs_key update_progress;
8446         int stage;
8447         int level;
8448         int shared_level;
8449         int update_ref;
8450         int keep_locks;
8451         int reada_slot;
8452         int reada_count;
8453         int for_reloc;
8454 };
8455
8456 #define DROP_REFERENCE  1
8457 #define UPDATE_BACKREF  2
8458
8459 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8460                                      struct btrfs_root *root,
8461                                      struct walk_control *wc,
8462                                      struct btrfs_path *path)
8463 {
8464         u64 bytenr;
8465         u64 generation;
8466         u64 refs;
8467         u64 flags;
8468         u32 nritems;
8469         struct btrfs_key key;
8470         struct extent_buffer *eb;
8471         int ret;
8472         int slot;
8473         int nread = 0;
8474
8475         if (path->slots[wc->level] < wc->reada_slot) {
8476                 wc->reada_count = wc->reada_count * 2 / 3;
8477                 wc->reada_count = max(wc->reada_count, 2);
8478         } else {
8479                 wc->reada_count = wc->reada_count * 3 / 2;
8480                 wc->reada_count = min_t(int, wc->reada_count,
8481                                         BTRFS_NODEPTRS_PER_BLOCK(root));
8482         }
8483
8484         eb = path->nodes[wc->level];
8485         nritems = btrfs_header_nritems(eb);
8486
8487         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8488                 if (nread >= wc->reada_count)
8489                         break;
8490
8491                 cond_resched();
8492                 bytenr = btrfs_node_blockptr(eb, slot);
8493                 generation = btrfs_node_ptr_generation(eb, slot);
8494
8495                 if (slot == path->slots[wc->level])
8496                         goto reada;
8497
8498                 if (wc->stage == UPDATE_BACKREF &&
8499                     generation <= root->root_key.offset)
8500                         continue;
8501
8502                 /* We don't lock the tree block, it's OK to be racy here */
8503                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
8504                                                wc->level - 1, 1, &refs,
8505                                                &flags);
8506                 /* We don't care about errors in readahead. */
8507                 if (ret < 0)
8508                         continue;
8509                 BUG_ON(refs == 0);
8510
8511                 if (wc->stage == DROP_REFERENCE) {
8512                         if (refs == 1)
8513                                 goto reada;
8514
8515                         if (wc->level == 1 &&
8516                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8517                                 continue;
8518                         if (!wc->update_ref ||
8519                             generation <= root->root_key.offset)
8520                                 continue;
8521                         btrfs_node_key_to_cpu(eb, &key, slot);
8522                         ret = btrfs_comp_cpu_keys(&key,
8523                                                   &wc->update_progress);
8524                         if (ret < 0)
8525                                 continue;
8526                 } else {
8527                         if (wc->level == 1 &&
8528                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8529                                 continue;
8530                 }
8531 reada:
8532                 readahead_tree_block(root, bytenr);
8533                 nread++;
8534         }
8535         wc->reada_slot = slot;
8536 }
8537
8538 static int account_leaf_items(struct btrfs_trans_handle *trans,
8539                               struct btrfs_root *root,
8540                               struct extent_buffer *eb)
8541 {
8542         int nr = btrfs_header_nritems(eb);
8543         int i, extent_type, ret;
8544         struct btrfs_key key;
8545         struct btrfs_file_extent_item *fi;
8546         u64 bytenr, num_bytes;
8547
8548         /* We can be called directly from walk_up_proc() */
8549         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags))
8550                 return 0;
8551
8552         for (i = 0; i < nr; i++) {
8553                 btrfs_item_key_to_cpu(eb, &key, i);
8554
8555                 if (key.type != BTRFS_EXTENT_DATA_KEY)
8556                         continue;
8557
8558                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
8559                 /* filter out non qgroup-accountable extents  */
8560                 extent_type = btrfs_file_extent_type(eb, fi);
8561
8562                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
8563                         continue;
8564
8565                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
8566                 if (!bytenr)
8567                         continue;
8568
8569                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
8570
8571                 ret = btrfs_qgroup_insert_dirty_extent(trans, root->fs_info,
8572                                 bytenr, num_bytes, GFP_NOFS);
8573                 if (ret)
8574                         return ret;
8575         }
8576         return 0;
8577 }
8578
8579 /*
8580  * Walk up the tree from the bottom, freeing leaves and any interior
8581  * nodes which have had all slots visited. If a node (leaf or
8582  * interior) is freed, the node above it will have it's slot
8583  * incremented. The root node will never be freed.
8584  *
8585  * At the end of this function, we should have a path which has all
8586  * slots incremented to the next position for a search. If we need to
8587  * read a new node it will be NULL and the node above it will have the
8588  * correct slot selected for a later read.
8589  *
8590  * If we increment the root nodes slot counter past the number of
8591  * elements, 1 is returned to signal completion of the search.
8592  */
8593 static int adjust_slots_upwards(struct btrfs_root *root,
8594                                 struct btrfs_path *path, int root_level)
8595 {
8596         int level = 0;
8597         int nr, slot;
8598         struct extent_buffer *eb;
8599
8600         if (root_level == 0)
8601                 return 1;
8602
8603         while (level <= root_level) {
8604                 eb = path->nodes[level];
8605                 nr = btrfs_header_nritems(eb);
8606                 path->slots[level]++;
8607                 slot = path->slots[level];
8608                 if (slot >= nr || level == 0) {
8609                         /*
8610                          * Don't free the root -  we will detect this
8611                          * condition after our loop and return a
8612                          * positive value for caller to stop walking the tree.
8613                          */
8614                         if (level != root_level) {
8615                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8616                                 path->locks[level] = 0;
8617
8618                                 free_extent_buffer(eb);
8619                                 path->nodes[level] = NULL;
8620                                 path->slots[level] = 0;
8621                         }
8622                 } else {
8623                         /*
8624                          * We have a valid slot to walk back down
8625                          * from. Stop here so caller can process these
8626                          * new nodes.
8627                          */
8628                         break;
8629                 }
8630
8631                 level++;
8632         }
8633
8634         eb = path->nodes[root_level];
8635         if (path->slots[root_level] >= btrfs_header_nritems(eb))
8636                 return 1;
8637
8638         return 0;
8639 }
8640
8641 /*
8642  * root_eb is the subtree root and is locked before this function is called.
8643  */
8644 static int account_shared_subtree(struct btrfs_trans_handle *trans,
8645                                   struct btrfs_root *root,
8646                                   struct extent_buffer *root_eb,
8647                                   u64 root_gen,
8648                                   int root_level)
8649 {
8650         int ret = 0;
8651         int level;
8652         struct extent_buffer *eb = root_eb;
8653         struct btrfs_path *path = NULL;
8654
8655         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
8656         BUG_ON(root_eb == NULL);
8657
8658         if (!test_bit(BTRFS_FS_QUOTA_ENABLED, &root->fs_info->flags))
8659                 return 0;
8660
8661         if (!extent_buffer_uptodate(root_eb)) {
8662                 ret = btrfs_read_buffer(root_eb, root_gen);
8663                 if (ret)
8664                         goto out;
8665         }
8666
8667         if (root_level == 0) {
8668                 ret = account_leaf_items(trans, root, root_eb);
8669                 goto out;
8670         }
8671
8672         path = btrfs_alloc_path();
8673         if (!path)
8674                 return -ENOMEM;
8675
8676         /*
8677          * Walk down the tree.  Missing extent blocks are filled in as
8678          * we go. Metadata is accounted every time we read a new
8679          * extent block.
8680          *
8681          * When we reach a leaf, we account for file extent items in it,
8682          * walk back up the tree (adjusting slot pointers as we go)
8683          * and restart the search process.
8684          */
8685         extent_buffer_get(root_eb); /* For path */
8686         path->nodes[root_level] = root_eb;
8687         path->slots[root_level] = 0;
8688         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
8689 walk_down:
8690         level = root_level;
8691         while (level >= 0) {
8692                 if (path->nodes[level] == NULL) {
8693                         int parent_slot;
8694                         u64 child_gen;
8695                         u64 child_bytenr;
8696
8697                         /* We need to get child blockptr/gen from
8698                          * parent before we can read it. */
8699                         eb = path->nodes[level + 1];
8700                         parent_slot = path->slots[level + 1];
8701                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8702                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8703
8704                         eb = read_tree_block(root, child_bytenr, child_gen);
8705                         if (IS_ERR(eb)) {
8706                                 ret = PTR_ERR(eb);
8707                                 goto out;
8708                         } else if (!extent_buffer_uptodate(eb)) {
8709                                 free_extent_buffer(eb);
8710                                 ret = -EIO;
8711                                 goto out;
8712                         }
8713
8714                         path->nodes[level] = eb;
8715                         path->slots[level] = 0;
8716
8717                         btrfs_tree_read_lock(eb);
8718                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8719                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8720
8721                         ret = btrfs_qgroup_insert_dirty_extent(trans,
8722                                         root->fs_info, child_bytenr,
8723                                         root->nodesize, GFP_NOFS);
8724                         if (ret)
8725                                 goto out;
8726                 }
8727
8728                 if (level == 0) {
8729                         ret = account_leaf_items(trans, root, path->nodes[level]);
8730                         if (ret)
8731                                 goto out;
8732
8733                         /* Nonzero return here means we completed our search */
8734                         ret = adjust_slots_upwards(root, path, root_level);
8735                         if (ret)
8736                                 break;
8737
8738                         /* Restart search with new slots */
8739                         goto walk_down;
8740                 }
8741
8742                 level--;
8743         }
8744
8745         ret = 0;
8746 out:
8747         btrfs_free_path(path);
8748
8749         return ret;
8750 }
8751
8752 /*
8753  * helper to process tree block while walking down the tree.
8754  *
8755  * when wc->stage == UPDATE_BACKREF, this function updates
8756  * back refs for pointers in the block.
8757  *
8758  * NOTE: return value 1 means we should stop walking down.
8759  */
8760 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8761                                    struct btrfs_root *root,
8762                                    struct btrfs_path *path,
8763                                    struct walk_control *wc, int lookup_info)
8764 {
8765         int level = wc->level;
8766         struct extent_buffer *eb = path->nodes[level];
8767         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8768         int ret;
8769
8770         if (wc->stage == UPDATE_BACKREF &&
8771             btrfs_header_owner(eb) != root->root_key.objectid)
8772                 return 1;
8773
8774         /*
8775          * when reference count of tree block is 1, it won't increase
8776          * again. once full backref flag is set, we never clear it.
8777          */
8778         if (lookup_info &&
8779             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8780              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8781                 BUG_ON(!path->locks[level]);
8782                 ret = btrfs_lookup_extent_info(trans, root,
8783                                                eb->start, level, 1,
8784                                                &wc->refs[level],
8785                                                &wc->flags[level]);
8786                 BUG_ON(ret == -ENOMEM);
8787                 if (ret)
8788                         return ret;
8789                 BUG_ON(wc->refs[level] == 0);
8790         }
8791
8792         if (wc->stage == DROP_REFERENCE) {
8793                 if (wc->refs[level] > 1)
8794                         return 1;
8795
8796                 if (path->locks[level] && !wc->keep_locks) {
8797                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8798                         path->locks[level] = 0;
8799                 }
8800                 return 0;
8801         }
8802
8803         /* wc->stage == UPDATE_BACKREF */
8804         if (!(wc->flags[level] & flag)) {
8805                 BUG_ON(!path->locks[level]);
8806                 ret = btrfs_inc_ref(trans, root, eb, 1);
8807                 BUG_ON(ret); /* -ENOMEM */
8808                 ret = btrfs_dec_ref(trans, root, eb, 0);
8809                 BUG_ON(ret); /* -ENOMEM */
8810                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8811                                                   eb->len, flag,
8812                                                   btrfs_header_level(eb), 0);
8813                 BUG_ON(ret); /* -ENOMEM */
8814                 wc->flags[level] |= flag;
8815         }
8816
8817         /*
8818          * the block is shared by multiple trees, so it's not good to
8819          * keep the tree lock
8820          */
8821         if (path->locks[level] && level > 0) {
8822                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8823                 path->locks[level] = 0;
8824         }
8825         return 0;
8826 }
8827
8828 /*
8829  * helper to process tree block pointer.
8830  *
8831  * when wc->stage == DROP_REFERENCE, this function checks
8832  * reference count of the block pointed to. if the block
8833  * is shared and we need update back refs for the subtree
8834  * rooted at the block, this function changes wc->stage to
8835  * UPDATE_BACKREF. if the block is shared and there is no
8836  * need to update back, this function drops the reference
8837  * to the block.
8838  *
8839  * NOTE: return value 1 means we should stop walking down.
8840  */
8841 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8842                                  struct btrfs_root *root,
8843                                  struct btrfs_path *path,
8844                                  struct walk_control *wc, int *lookup_info)
8845 {
8846         u64 bytenr;
8847         u64 generation;
8848         u64 parent;
8849         u32 blocksize;
8850         struct btrfs_key key;
8851         struct extent_buffer *next;
8852         int level = wc->level;
8853         int reada = 0;
8854         int ret = 0;
8855         bool need_account = false;
8856
8857         generation = btrfs_node_ptr_generation(path->nodes[level],
8858                                                path->slots[level]);
8859         /*
8860          * if the lower level block was created before the snapshot
8861          * was created, we know there is no need to update back refs
8862          * for the subtree
8863          */
8864         if (wc->stage == UPDATE_BACKREF &&
8865             generation <= root->root_key.offset) {
8866                 *lookup_info = 1;
8867                 return 1;
8868         }
8869
8870         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8871         blocksize = root->nodesize;
8872
8873         next = btrfs_find_tree_block(root->fs_info, bytenr);
8874         if (!next) {
8875                 next = btrfs_find_create_tree_block(root, bytenr);
8876                 if (IS_ERR(next))
8877                         return PTR_ERR(next);
8878
8879                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8880                                                level - 1);
8881                 reada = 1;
8882         }
8883         btrfs_tree_lock(next);
8884         btrfs_set_lock_blocking(next);
8885
8886         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8887                                        &wc->refs[level - 1],
8888                                        &wc->flags[level - 1]);
8889         if (ret < 0)
8890                 goto out_unlock;
8891
8892         if (unlikely(wc->refs[level - 1] == 0)) {
8893                 btrfs_err(root->fs_info, "Missing references.");
8894                 ret = -EIO;
8895                 goto out_unlock;
8896         }
8897         *lookup_info = 0;
8898
8899         if (wc->stage == DROP_REFERENCE) {
8900                 if (wc->refs[level - 1] > 1) {
8901                         need_account = true;
8902                         if (level == 1 &&
8903                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8904                                 goto skip;
8905
8906                         if (!wc->update_ref ||
8907                             generation <= root->root_key.offset)
8908                                 goto skip;
8909
8910                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8911                                               path->slots[level]);
8912                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8913                         if (ret < 0)
8914                                 goto skip;
8915
8916                         wc->stage = UPDATE_BACKREF;
8917                         wc->shared_level = level - 1;
8918                 }
8919         } else {
8920                 if (level == 1 &&
8921                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8922                         goto skip;
8923         }
8924
8925         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8926                 btrfs_tree_unlock(next);
8927                 free_extent_buffer(next);
8928                 next = NULL;
8929                 *lookup_info = 1;
8930         }
8931
8932         if (!next) {
8933                 if (reada && level == 1)
8934                         reada_walk_down(trans, root, wc, path);
8935                 next = read_tree_block(root, bytenr, generation);
8936                 if (IS_ERR(next)) {
8937                         return PTR_ERR(next);
8938                 } else if (!extent_buffer_uptodate(next)) {
8939                         free_extent_buffer(next);
8940                         return -EIO;
8941                 }
8942                 btrfs_tree_lock(next);
8943                 btrfs_set_lock_blocking(next);
8944         }
8945
8946         level--;
8947         ASSERT(level == btrfs_header_level(next));
8948         if (level != btrfs_header_level(next)) {
8949                 btrfs_err(root->fs_info, "mismatched level");
8950                 ret = -EIO;
8951                 goto out_unlock;
8952         }
8953         path->nodes[level] = next;
8954         path->slots[level] = 0;
8955         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8956         wc->level = level;
8957         if (wc->level == 1)
8958                 wc->reada_slot = 0;
8959         return 0;
8960 skip:
8961         wc->refs[level - 1] = 0;
8962         wc->flags[level - 1] = 0;
8963         if (wc->stage == DROP_REFERENCE) {
8964                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8965                         parent = path->nodes[level]->start;
8966                 } else {
8967                         ASSERT(root->root_key.objectid ==
8968                                btrfs_header_owner(path->nodes[level]));
8969                         if (root->root_key.objectid !=
8970                             btrfs_header_owner(path->nodes[level])) {
8971                                 btrfs_err(root->fs_info,
8972                                                 "mismatched block owner");
8973                                 ret = -EIO;
8974                                 goto out_unlock;
8975                         }
8976                         parent = 0;
8977                 }
8978
8979                 if (need_account) {
8980                         ret = account_shared_subtree(trans, root, next,
8981                                                      generation, level - 1);
8982                         if (ret) {
8983                                 btrfs_err_rl(root->fs_info,
8984                                              "Error %d accounting shared subtree. Quota is out of sync, rescan required.",
8985                                              ret);
8986                         }
8987                 }
8988                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8989                                 root->root_key.objectid, level - 1, 0);
8990                 if (ret)
8991                         goto out_unlock;
8992         }
8993
8994         *lookup_info = 1;
8995         ret = 1;
8996
8997 out_unlock:
8998         btrfs_tree_unlock(next);
8999         free_extent_buffer(next);
9000
9001         return ret;
9002 }
9003
9004 /*
9005  * helper to process tree block while walking up the tree.
9006  *
9007  * when wc->stage == DROP_REFERENCE, this function drops
9008  * reference count on the block.
9009  *
9010  * when wc->stage == UPDATE_BACKREF, this function changes
9011  * wc->stage back to DROP_REFERENCE if we changed wc->stage
9012  * to UPDATE_BACKREF previously while processing the block.
9013  *
9014  * NOTE: return value 1 means we should stop walking up.
9015  */
9016 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9017                                  struct btrfs_root *root,
9018                                  struct btrfs_path *path,
9019                                  struct walk_control *wc)
9020 {
9021         int ret;
9022         int level = wc->level;
9023         struct extent_buffer *eb = path->nodes[level];
9024         u64 parent = 0;
9025
9026         if (wc->stage == UPDATE_BACKREF) {
9027                 BUG_ON(wc->shared_level < level);
9028                 if (level < wc->shared_level)
9029                         goto out;
9030
9031                 ret = find_next_key(path, level + 1, &wc->update_progress);
9032                 if (ret > 0)
9033                         wc->update_ref = 0;
9034
9035                 wc->stage = DROP_REFERENCE;
9036                 wc->shared_level = -1;
9037                 path->slots[level] = 0;
9038
9039                 /*
9040                  * check reference count again if the block isn't locked.
9041                  * we should start walking down the tree again if reference
9042                  * count is one.
9043                  */
9044                 if (!path->locks[level]) {
9045                         BUG_ON(level == 0);
9046                         btrfs_tree_lock(eb);
9047                         btrfs_set_lock_blocking(eb);
9048                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9049
9050                         ret = btrfs_lookup_extent_info(trans, root,
9051                                                        eb->start, level, 1,
9052                                                        &wc->refs[level],
9053                                                        &wc->flags[level]);
9054                         if (ret < 0) {
9055                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9056                                 path->locks[level] = 0;
9057                                 return ret;
9058                         }
9059                         BUG_ON(wc->refs[level] == 0);
9060                         if (wc->refs[level] == 1) {
9061                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9062                                 path->locks[level] = 0;
9063                                 return 1;
9064                         }
9065                 }
9066         }
9067
9068         /* wc->stage == DROP_REFERENCE */
9069         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9070
9071         if (wc->refs[level] == 1) {
9072                 if (level == 0) {
9073                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9074                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9075                         else
9076                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9077                         BUG_ON(ret); /* -ENOMEM */
9078                         ret = account_leaf_items(trans, root, eb);
9079                         if (ret) {
9080                                 btrfs_err_rl(root->fs_info,
9081                                              "error %d accounting leaf items. Quota is out of sync, rescan required.",
9082                                              ret);
9083                         }
9084                 }
9085                 /* make block locked assertion in clean_tree_block happy */
9086                 if (!path->locks[level] &&
9087                     btrfs_header_generation(eb) == trans->transid) {
9088                         btrfs_tree_lock(eb);
9089                         btrfs_set_lock_blocking(eb);
9090                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9091                 }
9092                 clean_tree_block(trans, root->fs_info, eb);
9093         }
9094
9095         if (eb == root->node) {
9096                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9097                         parent = eb->start;
9098                 else
9099                         BUG_ON(root->root_key.objectid !=
9100                                btrfs_header_owner(eb));
9101         } else {
9102                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9103                         parent = path->nodes[level + 1]->start;
9104                 else
9105                         BUG_ON(root->root_key.objectid !=
9106                                btrfs_header_owner(path->nodes[level + 1]));
9107         }
9108
9109         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
9110 out:
9111         wc->refs[level] = 0;
9112         wc->flags[level] = 0;
9113         return 0;
9114 }
9115
9116 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
9117                                    struct btrfs_root *root,
9118                                    struct btrfs_path *path,
9119                                    struct walk_control *wc)
9120 {
9121         int level = wc->level;
9122         int lookup_info = 1;
9123         int ret;
9124
9125         while (level >= 0) {
9126                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
9127                 if (ret > 0)
9128                         break;
9129
9130                 if (level == 0)
9131                         break;
9132
9133                 if (path->slots[level] >=
9134                     btrfs_header_nritems(path->nodes[level]))
9135                         break;
9136
9137                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
9138                 if (ret > 0) {
9139                         path->slots[level]++;
9140                         continue;
9141                 } else if (ret < 0)
9142                         return ret;
9143                 level = wc->level;
9144         }
9145         return 0;
9146 }
9147
9148 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
9149                                  struct btrfs_root *root,
9150                                  struct btrfs_path *path,
9151                                  struct walk_control *wc, int max_level)
9152 {
9153         int level = wc->level;
9154         int ret;
9155
9156         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
9157         while (level < max_level && path->nodes[level]) {
9158                 wc->level = level;
9159                 if (path->slots[level] + 1 <
9160                     btrfs_header_nritems(path->nodes[level])) {
9161                         path->slots[level]++;
9162                         return 0;
9163                 } else {
9164                         ret = walk_up_proc(trans, root, path, wc);
9165                         if (ret > 0)
9166                                 return 0;
9167
9168                         if (path->locks[level]) {
9169                                 btrfs_tree_unlock_rw(path->nodes[level],
9170                                                      path->locks[level]);
9171                                 path->locks[level] = 0;
9172                         }
9173                         free_extent_buffer(path->nodes[level]);
9174                         path->nodes[level] = NULL;
9175                         level++;
9176                 }
9177         }
9178         return 1;
9179 }
9180
9181 /*
9182  * drop a subvolume tree.
9183  *
9184  * this function traverses the tree freeing any blocks that only
9185  * referenced by the tree.
9186  *
9187  * when a shared tree block is found. this function decreases its
9188  * reference count by one. if update_ref is true, this function
9189  * also make sure backrefs for the shared block and all lower level
9190  * blocks are properly updated.
9191  *
9192  * If called with for_reloc == 0, may exit early with -EAGAIN
9193  */
9194 int btrfs_drop_snapshot(struct btrfs_root *root,
9195                          struct btrfs_block_rsv *block_rsv, int update_ref,
9196                          int for_reloc)
9197 {
9198         struct btrfs_fs_info *fs_info = root->fs_info;
9199         struct btrfs_path *path;
9200         struct btrfs_trans_handle *trans;
9201         struct btrfs_root *tree_root = fs_info->tree_root;
9202         struct btrfs_root_item *root_item = &root->root_item;
9203         struct walk_control *wc;
9204         struct btrfs_key key;
9205         int err = 0;
9206         int ret;
9207         int level;
9208         bool root_dropped = false;
9209
9210         btrfs_debug(fs_info, "Drop subvolume %llu", root->objectid);
9211
9212         path = btrfs_alloc_path();
9213         if (!path) {
9214                 err = -ENOMEM;
9215                 goto out;
9216         }
9217
9218         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9219         if (!wc) {
9220                 btrfs_free_path(path);
9221                 err = -ENOMEM;
9222                 goto out;
9223         }
9224
9225         trans = btrfs_start_transaction(tree_root, 0);
9226         if (IS_ERR(trans)) {
9227                 err = PTR_ERR(trans);
9228                 goto out_free;
9229         }
9230
9231         if (block_rsv)
9232                 trans->block_rsv = block_rsv;
9233
9234         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9235                 level = btrfs_header_level(root->node);
9236                 path->nodes[level] = btrfs_lock_root_node(root);
9237                 btrfs_set_lock_blocking(path->nodes[level]);
9238                 path->slots[level] = 0;
9239                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9240                 memset(&wc->update_progress, 0,
9241                        sizeof(wc->update_progress));
9242         } else {
9243                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9244                 memcpy(&wc->update_progress, &key,
9245                        sizeof(wc->update_progress));
9246
9247                 level = root_item->drop_level;
9248                 BUG_ON(level == 0);
9249                 path->lowest_level = level;
9250                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9251                 path->lowest_level = 0;
9252                 if (ret < 0) {
9253                         err = ret;
9254                         goto out_end_trans;
9255                 }
9256                 WARN_ON(ret > 0);
9257
9258                 /*
9259                  * unlock our path, this is safe because only this
9260                  * function is allowed to delete this snapshot
9261                  */
9262                 btrfs_unlock_up_safe(path, 0);
9263
9264                 level = btrfs_header_level(root->node);
9265                 while (1) {
9266                         btrfs_tree_lock(path->nodes[level]);
9267                         btrfs_set_lock_blocking(path->nodes[level]);
9268                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9269
9270                         ret = btrfs_lookup_extent_info(trans, root,
9271                                                 path->nodes[level]->start,
9272                                                 level, 1, &wc->refs[level],
9273                                                 &wc->flags[level]);
9274                         if (ret < 0) {
9275                                 err = ret;
9276                                 goto out_end_trans;
9277                         }
9278                         BUG_ON(wc->refs[level] == 0);
9279
9280                         if (level == root_item->drop_level)
9281                                 break;
9282
9283                         btrfs_tree_unlock(path->nodes[level]);
9284                         path->locks[level] = 0;
9285                         WARN_ON(wc->refs[level] != 1);
9286                         level--;
9287                 }
9288         }
9289
9290         wc->level = level;
9291         wc->shared_level = -1;
9292         wc->stage = DROP_REFERENCE;
9293         wc->update_ref = update_ref;
9294         wc->keep_locks = 0;
9295         wc->for_reloc = for_reloc;
9296         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9297
9298         while (1) {
9299
9300                 ret = walk_down_tree(trans, root, path, wc);
9301                 if (ret < 0) {
9302                         err = ret;
9303                         break;
9304                 }
9305
9306                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9307                 if (ret < 0) {
9308                         err = ret;
9309                         break;
9310                 }
9311
9312                 if (ret > 0) {
9313                         BUG_ON(wc->stage != DROP_REFERENCE);
9314                         break;
9315                 }
9316
9317                 if (wc->stage == DROP_REFERENCE) {
9318                         level = wc->level;
9319                         btrfs_node_key(path->nodes[level],
9320                                        &root_item->drop_progress,
9321                                        path->slots[level]);
9322                         root_item->drop_level = level;
9323                 }
9324
9325                 BUG_ON(wc->level == 0);
9326                 if (btrfs_should_end_transaction(trans, tree_root) ||
9327                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
9328                         ret = btrfs_update_root(trans, tree_root,
9329                                                 &root->root_key,
9330                                                 root_item);
9331                         if (ret) {
9332                                 btrfs_abort_transaction(trans, ret);
9333                                 err = ret;
9334                                 goto out_end_trans;
9335                         }
9336
9337                         btrfs_end_transaction_throttle(trans, tree_root);
9338                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
9339                                 btrfs_debug(fs_info,
9340                                             "drop snapshot early exit");
9341                                 err = -EAGAIN;
9342                                 goto out_free;
9343                         }
9344
9345                         trans = btrfs_start_transaction(tree_root, 0);
9346                         if (IS_ERR(trans)) {
9347                                 err = PTR_ERR(trans);
9348                                 goto out_free;
9349                         }
9350                         if (block_rsv)
9351                                 trans->block_rsv = block_rsv;
9352                 }
9353         }
9354         btrfs_release_path(path);
9355         if (err)
9356                 goto out_end_trans;
9357
9358         ret = btrfs_del_root(trans, tree_root, &root->root_key);
9359         if (ret) {
9360                 btrfs_abort_transaction(trans, ret);
9361                 goto out_end_trans;
9362         }
9363
9364         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9365                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9366                                       NULL, NULL);
9367                 if (ret < 0) {
9368                         btrfs_abort_transaction(trans, ret);
9369                         err = ret;
9370                         goto out_end_trans;
9371                 } else if (ret > 0) {
9372                         /* if we fail to delete the orphan item this time
9373                          * around, it'll get picked up the next time.
9374                          *
9375                          * The most common failure here is just -ENOENT.
9376                          */
9377                         btrfs_del_orphan_item(trans, tree_root,
9378                                               root->root_key.objectid);
9379                 }
9380         }
9381
9382         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9383                 btrfs_add_dropped_root(trans, root);
9384         } else {
9385                 free_extent_buffer(root->node);
9386                 free_extent_buffer(root->commit_root);
9387                 btrfs_put_fs_root(root);
9388         }
9389         root_dropped = true;
9390 out_end_trans:
9391         btrfs_end_transaction_throttle(trans, tree_root);
9392 out_free:
9393         kfree(wc);
9394         btrfs_free_path(path);
9395 out:
9396         /*
9397          * So if we need to stop dropping the snapshot for whatever reason we
9398          * need to make sure to add it back to the dead root list so that we
9399          * keep trying to do the work later.  This also cleans up roots if we
9400          * don't have it in the radix (like when we recover after a power fail
9401          * or unmount) so we don't leak memory.
9402          */
9403         if (!for_reloc && root_dropped == false)
9404                 btrfs_add_dead_root(root);
9405         if (err && err != -EAGAIN)
9406                 btrfs_handle_fs_error(fs_info, err, NULL);
9407         return err;
9408 }
9409
9410 /*
9411  * drop subtree rooted at tree block 'node'.
9412  *
9413  * NOTE: this function will unlock and release tree block 'node'
9414  * only used by relocation code
9415  */
9416 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9417                         struct btrfs_root *root,
9418                         struct extent_buffer *node,
9419                         struct extent_buffer *parent)
9420 {
9421         struct btrfs_path *path;
9422         struct walk_control *wc;
9423         int level;
9424         int parent_level;
9425         int ret = 0;
9426         int wret;
9427
9428         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9429
9430         path = btrfs_alloc_path();
9431         if (!path)
9432                 return -ENOMEM;
9433
9434         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9435         if (!wc) {
9436                 btrfs_free_path(path);
9437                 return -ENOMEM;
9438         }
9439
9440         btrfs_assert_tree_locked(parent);
9441         parent_level = btrfs_header_level(parent);
9442         extent_buffer_get(parent);
9443         path->nodes[parent_level] = parent;
9444         path->slots[parent_level] = btrfs_header_nritems(parent);
9445
9446         btrfs_assert_tree_locked(node);
9447         level = btrfs_header_level(node);
9448         path->nodes[level] = node;
9449         path->slots[level] = 0;
9450         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9451
9452         wc->refs[parent_level] = 1;
9453         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9454         wc->level = level;
9455         wc->shared_level = -1;
9456         wc->stage = DROP_REFERENCE;
9457         wc->update_ref = 0;
9458         wc->keep_locks = 1;
9459         wc->for_reloc = 1;
9460         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9461
9462         while (1) {
9463                 wret = walk_down_tree(trans, root, path, wc);
9464                 if (wret < 0) {
9465                         ret = wret;
9466                         break;
9467                 }
9468
9469                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9470                 if (wret < 0)
9471                         ret = wret;
9472                 if (wret != 0)
9473                         break;
9474         }
9475
9476         kfree(wc);
9477         btrfs_free_path(path);
9478         return ret;
9479 }
9480
9481 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
9482 {
9483         u64 num_devices;
9484         u64 stripped;
9485
9486         /*
9487          * if restripe for this chunk_type is on pick target profile and
9488          * return, otherwise do the usual balance
9489          */
9490         stripped = get_restripe_target(root->fs_info, flags);
9491         if (stripped)
9492                 return extended_to_chunk(stripped);
9493
9494         num_devices = root->fs_info->fs_devices->rw_devices;
9495
9496         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9497                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9498                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9499
9500         if (num_devices == 1) {
9501                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9502                 stripped = flags & ~stripped;
9503
9504                 /* turn raid0 into single device chunks */
9505                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9506                         return stripped;
9507
9508                 /* turn mirroring into duplication */
9509                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9510                              BTRFS_BLOCK_GROUP_RAID10))
9511                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9512         } else {
9513                 /* they already had raid on here, just return */
9514                 if (flags & stripped)
9515                         return flags;
9516
9517                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9518                 stripped = flags & ~stripped;
9519
9520                 /* switch duplicated blocks with raid1 */
9521                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9522                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9523
9524                 /* this is drive concat, leave it alone */
9525         }
9526
9527         return flags;
9528 }
9529
9530 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9531 {
9532         struct btrfs_space_info *sinfo = cache->space_info;
9533         u64 num_bytes;
9534         u64 min_allocable_bytes;
9535         int ret = -ENOSPC;
9536
9537         /*
9538          * We need some metadata space and system metadata space for
9539          * allocating chunks in some corner cases until we force to set
9540          * it to be readonly.
9541          */
9542         if ((sinfo->flags &
9543              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9544             !force)
9545                 min_allocable_bytes = SZ_1M;
9546         else
9547                 min_allocable_bytes = 0;
9548
9549         spin_lock(&sinfo->lock);
9550         spin_lock(&cache->lock);
9551
9552         if (cache->ro) {
9553                 cache->ro++;
9554                 ret = 0;
9555                 goto out;
9556         }
9557
9558         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9559                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9560
9561         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
9562             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
9563             min_allocable_bytes <= sinfo->total_bytes) {
9564                 sinfo->bytes_readonly += num_bytes;
9565                 cache->ro++;
9566                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9567                 ret = 0;
9568         }
9569 out:
9570         spin_unlock(&cache->lock);
9571         spin_unlock(&sinfo->lock);
9572         return ret;
9573 }
9574
9575 int btrfs_inc_block_group_ro(struct btrfs_root *root,
9576                              struct btrfs_block_group_cache *cache)
9577
9578 {
9579         struct btrfs_trans_handle *trans;
9580         u64 alloc_flags;
9581         int ret;
9582
9583 again:
9584         trans = btrfs_join_transaction(root);
9585         if (IS_ERR(trans))
9586                 return PTR_ERR(trans);
9587
9588         /*
9589          * we're not allowed to set block groups readonly after the dirty
9590          * block groups cache has started writing.  If it already started,
9591          * back off and let this transaction commit
9592          */
9593         mutex_lock(&root->fs_info->ro_block_group_mutex);
9594         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9595                 u64 transid = trans->transid;
9596
9597                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
9598                 btrfs_end_transaction(trans, root);
9599
9600                 ret = btrfs_wait_for_commit(root, transid);
9601                 if (ret)
9602                         return ret;
9603                 goto again;
9604         }
9605
9606         /*
9607          * if we are changing raid levels, try to allocate a corresponding
9608          * block group with the new raid level.
9609          */
9610         alloc_flags = update_block_group_flags(root, cache->flags);
9611         if (alloc_flags != cache->flags) {
9612                 ret = do_chunk_alloc(trans, root, alloc_flags,
9613                                      CHUNK_ALLOC_FORCE);
9614                 /*
9615                  * ENOSPC is allowed here, we may have enough space
9616                  * already allocated at the new raid level to
9617                  * carry on
9618                  */
9619                 if (ret == -ENOSPC)
9620                         ret = 0;
9621                 if (ret < 0)
9622                         goto out;
9623         }
9624
9625         ret = inc_block_group_ro(cache, 0);
9626         if (!ret)
9627                 goto out;
9628         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
9629         ret = do_chunk_alloc(trans, root, alloc_flags,
9630                              CHUNK_ALLOC_FORCE);
9631         if (ret < 0)
9632                 goto out;
9633         ret = inc_block_group_ro(cache, 0);
9634 out:
9635         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9636                 alloc_flags = update_block_group_flags(root, cache->flags);
9637                 lock_chunks(root->fs_info->chunk_root);
9638                 check_system_chunk(trans, root, alloc_flags);
9639                 unlock_chunks(root->fs_info->chunk_root);
9640         }
9641         mutex_unlock(&root->fs_info->ro_block_group_mutex);
9642
9643         btrfs_end_transaction(trans, root);
9644         return ret;
9645 }
9646
9647 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9648                             struct btrfs_root *root, u64 type)
9649 {
9650         u64 alloc_flags = get_alloc_profile(root, type);
9651         return do_chunk_alloc(trans, root, alloc_flags,
9652                               CHUNK_ALLOC_FORCE);
9653 }
9654
9655 /*
9656  * helper to account the unused space of all the readonly block group in the
9657  * space_info. takes mirrors into account.
9658  */
9659 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9660 {
9661         struct btrfs_block_group_cache *block_group;
9662         u64 free_bytes = 0;
9663         int factor;
9664
9665         /* It's df, we don't care if it's racy */
9666         if (list_empty(&sinfo->ro_bgs))
9667                 return 0;
9668
9669         spin_lock(&sinfo->lock);
9670         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9671                 spin_lock(&block_group->lock);
9672
9673                 if (!block_group->ro) {
9674                         spin_unlock(&block_group->lock);
9675                         continue;
9676                 }
9677
9678                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9679                                           BTRFS_BLOCK_GROUP_RAID10 |
9680                                           BTRFS_BLOCK_GROUP_DUP))
9681                         factor = 2;
9682                 else
9683                         factor = 1;
9684
9685                 free_bytes += (block_group->key.offset -
9686                                btrfs_block_group_used(&block_group->item)) *
9687                                factor;
9688
9689                 spin_unlock(&block_group->lock);
9690         }
9691         spin_unlock(&sinfo->lock);
9692
9693         return free_bytes;
9694 }
9695
9696 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9697                               struct btrfs_block_group_cache *cache)
9698 {
9699         struct btrfs_space_info *sinfo = cache->space_info;
9700         u64 num_bytes;
9701
9702         BUG_ON(!cache->ro);
9703
9704         spin_lock(&sinfo->lock);
9705         spin_lock(&cache->lock);
9706         if (!--cache->ro) {
9707                 num_bytes = cache->key.offset - cache->reserved -
9708                             cache->pinned - cache->bytes_super -
9709                             btrfs_block_group_used(&cache->item);
9710                 sinfo->bytes_readonly -= num_bytes;
9711                 list_del_init(&cache->ro_list);
9712         }
9713         spin_unlock(&cache->lock);
9714         spin_unlock(&sinfo->lock);
9715 }
9716
9717 /*
9718  * checks to see if its even possible to relocate this block group.
9719  *
9720  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9721  * ok to go ahead and try.
9722  */
9723 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9724 {
9725         struct btrfs_block_group_cache *block_group;
9726         struct btrfs_space_info *space_info;
9727         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9728         struct btrfs_device *device;
9729         struct btrfs_trans_handle *trans;
9730         u64 min_free;
9731         u64 dev_min = 1;
9732         u64 dev_nr = 0;
9733         u64 target;
9734         int debug;
9735         int index;
9736         int full = 0;
9737         int ret = 0;
9738
9739         debug = btrfs_test_opt(root->fs_info, ENOSPC_DEBUG);
9740
9741         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9742
9743         /* odd, couldn't find the block group, leave it alone */
9744         if (!block_group) {
9745                 if (debug)
9746                         btrfs_warn(root->fs_info,
9747                                    "can't find block group for bytenr %llu",
9748                                    bytenr);
9749                 return -1;
9750         }
9751
9752         min_free = btrfs_block_group_used(&block_group->item);
9753
9754         /* no bytes used, we're good */
9755         if (!min_free)
9756                 goto out;
9757
9758         space_info = block_group->space_info;
9759         spin_lock(&space_info->lock);
9760
9761         full = space_info->full;
9762
9763         /*
9764          * if this is the last block group we have in this space, we can't
9765          * relocate it unless we're able to allocate a new chunk below.
9766          *
9767          * Otherwise, we need to make sure we have room in the space to handle
9768          * all of the extents from this block group.  If we can, we're good
9769          */
9770         if ((space_info->total_bytes != block_group->key.offset) &&
9771             (space_info->bytes_used + space_info->bytes_reserved +
9772              space_info->bytes_pinned + space_info->bytes_readonly +
9773              min_free < space_info->total_bytes)) {
9774                 spin_unlock(&space_info->lock);
9775                 goto out;
9776         }
9777         spin_unlock(&space_info->lock);
9778
9779         /*
9780          * ok we don't have enough space, but maybe we have free space on our
9781          * devices to allocate new chunks for relocation, so loop through our
9782          * alloc devices and guess if we have enough space.  if this block
9783          * group is going to be restriped, run checks against the target
9784          * profile instead of the current one.
9785          */
9786         ret = -1;
9787
9788         /*
9789          * index:
9790          *      0: raid10
9791          *      1: raid1
9792          *      2: dup
9793          *      3: raid0
9794          *      4: single
9795          */
9796         target = get_restripe_target(root->fs_info, block_group->flags);
9797         if (target) {
9798                 index = __get_raid_index(extended_to_chunk(target));
9799         } else {
9800                 /*
9801                  * this is just a balance, so if we were marked as full
9802                  * we know there is no space for a new chunk
9803                  */
9804                 if (full) {
9805                         if (debug)
9806                                 btrfs_warn(root->fs_info,
9807                                         "no space to alloc new chunk for block group %llu",
9808                                         block_group->key.objectid);
9809                         goto out;
9810                 }
9811
9812                 index = get_block_group_index(block_group);
9813         }
9814
9815         if (index == BTRFS_RAID_RAID10) {
9816                 dev_min = 4;
9817                 /* Divide by 2 */
9818                 min_free >>= 1;
9819         } else if (index == BTRFS_RAID_RAID1) {
9820                 dev_min = 2;
9821         } else if (index == BTRFS_RAID_DUP) {
9822                 /* Multiply by 2 */
9823                 min_free <<= 1;
9824         } else if (index == BTRFS_RAID_RAID0) {
9825                 dev_min = fs_devices->rw_devices;
9826                 min_free = div64_u64(min_free, dev_min);
9827         }
9828
9829         /* We need to do this so that we can look at pending chunks */
9830         trans = btrfs_join_transaction(root);
9831         if (IS_ERR(trans)) {
9832                 ret = PTR_ERR(trans);
9833                 goto out;
9834         }
9835
9836         mutex_lock(&root->fs_info->chunk_mutex);
9837         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9838                 u64 dev_offset;
9839
9840                 /*
9841                  * check to make sure we can actually find a chunk with enough
9842                  * space to fit our block group in.
9843                  */
9844                 if (device->total_bytes > device->bytes_used + min_free &&
9845                     !device->is_tgtdev_for_dev_replace) {
9846                         ret = find_free_dev_extent(trans, device, min_free,
9847                                                    &dev_offset, NULL);
9848                         if (!ret)
9849                                 dev_nr++;
9850
9851                         if (dev_nr >= dev_min)
9852                                 break;
9853
9854                         ret = -1;
9855                 }
9856         }
9857         if (debug && ret == -1)
9858                 btrfs_warn(root->fs_info,
9859                         "no space to allocate a new chunk for block group %llu",
9860                         block_group->key.objectid);
9861         mutex_unlock(&root->fs_info->chunk_mutex);
9862         btrfs_end_transaction(trans, root);
9863 out:
9864         btrfs_put_block_group(block_group);
9865         return ret;
9866 }
9867
9868 static int find_first_block_group(struct btrfs_root *root,
9869                 struct btrfs_path *path, struct btrfs_key *key)
9870 {
9871         int ret = 0;
9872         struct btrfs_key found_key;
9873         struct extent_buffer *leaf;
9874         int slot;
9875
9876         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9877         if (ret < 0)
9878                 goto out;
9879
9880         while (1) {
9881                 slot = path->slots[0];
9882                 leaf = path->nodes[0];
9883                 if (slot >= btrfs_header_nritems(leaf)) {
9884                         ret = btrfs_next_leaf(root, path);
9885                         if (ret == 0)
9886                                 continue;
9887                         if (ret < 0)
9888                                 goto out;
9889                         break;
9890                 }
9891                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9892
9893                 if (found_key.objectid >= key->objectid &&
9894                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9895                         struct extent_map_tree *em_tree;
9896                         struct extent_map *em;
9897
9898                         em_tree = &root->fs_info->mapping_tree.map_tree;
9899                         read_lock(&em_tree->lock);
9900                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9901                                                    found_key.offset);
9902                         read_unlock(&em_tree->lock);
9903                         if (!em) {
9904                                 btrfs_err(root->fs_info,
9905                         "logical %llu len %llu found bg but no related chunk",
9906                                           found_key.objectid, found_key.offset);
9907                                 ret = -ENOENT;
9908                         } else {
9909                                 ret = 0;
9910                         }
9911                         free_extent_map(em);
9912                         goto out;
9913                 }
9914                 path->slots[0]++;
9915         }
9916 out:
9917         return ret;
9918 }
9919
9920 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9921 {
9922         struct btrfs_block_group_cache *block_group;
9923         u64 last = 0;
9924
9925         while (1) {
9926                 struct inode *inode;
9927
9928                 block_group = btrfs_lookup_first_block_group(info, last);
9929                 while (block_group) {
9930                         spin_lock(&block_group->lock);
9931                         if (block_group->iref)
9932                                 break;
9933                         spin_unlock(&block_group->lock);
9934                         block_group = next_block_group(info->tree_root,
9935                                                        block_group);
9936                 }
9937                 if (!block_group) {
9938                         if (last == 0)
9939                                 break;
9940                         last = 0;
9941                         continue;
9942                 }
9943
9944                 inode = block_group->inode;
9945                 block_group->iref = 0;
9946                 block_group->inode = NULL;
9947                 spin_unlock(&block_group->lock);
9948                 ASSERT(block_group->io_ctl.inode == NULL);
9949                 iput(inode);
9950                 last = block_group->key.objectid + block_group->key.offset;
9951                 btrfs_put_block_group(block_group);
9952         }
9953 }
9954
9955 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9956 {
9957         struct btrfs_block_group_cache *block_group;
9958         struct btrfs_space_info *space_info;
9959         struct btrfs_caching_control *caching_ctl;
9960         struct rb_node *n;
9961
9962         down_write(&info->commit_root_sem);
9963         while (!list_empty(&info->caching_block_groups)) {
9964                 caching_ctl = list_entry(info->caching_block_groups.next,
9965                                          struct btrfs_caching_control, list);
9966                 list_del(&caching_ctl->list);
9967                 put_caching_control(caching_ctl);
9968         }
9969         up_write(&info->commit_root_sem);
9970
9971         spin_lock(&info->unused_bgs_lock);
9972         while (!list_empty(&info->unused_bgs)) {
9973                 block_group = list_first_entry(&info->unused_bgs,
9974                                                struct btrfs_block_group_cache,
9975                                                bg_list);
9976                 list_del_init(&block_group->bg_list);
9977                 btrfs_put_block_group(block_group);
9978         }
9979         spin_unlock(&info->unused_bgs_lock);
9980
9981         spin_lock(&info->block_group_cache_lock);
9982         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9983                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9984                                        cache_node);
9985                 rb_erase(&block_group->cache_node,
9986                          &info->block_group_cache_tree);
9987                 RB_CLEAR_NODE(&block_group->cache_node);
9988                 spin_unlock(&info->block_group_cache_lock);
9989
9990                 down_write(&block_group->space_info->groups_sem);
9991                 list_del(&block_group->list);
9992                 up_write(&block_group->space_info->groups_sem);
9993
9994                 if (block_group->cached == BTRFS_CACHE_STARTED)
9995                         wait_block_group_cache_done(block_group);
9996
9997                 /*
9998                  * We haven't cached this block group, which means we could
9999                  * possibly have excluded extents on this block group.
10000                  */
10001                 if (block_group->cached == BTRFS_CACHE_NO ||
10002                     block_group->cached == BTRFS_CACHE_ERROR)
10003                         free_excluded_extents(info->extent_root, block_group);
10004
10005                 btrfs_remove_free_space_cache(block_group);
10006                 ASSERT(list_empty(&block_group->dirty_list));
10007                 ASSERT(list_empty(&block_group->io_list));
10008                 ASSERT(list_empty(&block_group->bg_list));
10009                 ASSERT(atomic_read(&block_group->count) == 1);
10010                 btrfs_put_block_group(block_group);
10011
10012                 spin_lock(&info->block_group_cache_lock);
10013         }
10014         spin_unlock(&info->block_group_cache_lock);
10015
10016         /* now that all the block groups are freed, go through and
10017          * free all the space_info structs.  This is only called during
10018          * the final stages of unmount, and so we know nobody is
10019          * using them.  We call synchronize_rcu() once before we start,
10020          * just to be on the safe side.
10021          */
10022         synchronize_rcu();
10023
10024         release_global_block_rsv(info);
10025
10026         while (!list_empty(&info->space_info)) {
10027                 int i;
10028
10029                 space_info = list_entry(info->space_info.next,
10030                                         struct btrfs_space_info,
10031                                         list);
10032
10033                 /*
10034                  * Do not hide this behind enospc_debug, this is actually
10035                  * important and indicates a real bug if this happens.
10036                  */
10037                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10038                             space_info->bytes_reserved > 0 ||
10039                             space_info->bytes_may_use > 0))
10040                         dump_space_info(info, space_info, 0, 0);
10041                 list_del(&space_info->list);
10042                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10043                         struct kobject *kobj;
10044                         kobj = space_info->block_group_kobjs[i];
10045                         space_info->block_group_kobjs[i] = NULL;
10046                         if (kobj) {
10047                                 kobject_del(kobj);
10048                                 kobject_put(kobj);
10049                         }
10050                 }
10051                 kobject_del(&space_info->kobj);
10052                 kobject_put(&space_info->kobj);
10053         }
10054         return 0;
10055 }
10056
10057 static void __link_block_group(struct btrfs_space_info *space_info,
10058                                struct btrfs_block_group_cache *cache)
10059 {
10060         int index = get_block_group_index(cache);
10061         bool first = false;
10062
10063         down_write(&space_info->groups_sem);
10064         if (list_empty(&space_info->block_groups[index]))
10065                 first = true;
10066         list_add_tail(&cache->list, &space_info->block_groups[index]);
10067         up_write(&space_info->groups_sem);
10068
10069         if (first) {
10070                 struct raid_kobject *rkobj;
10071                 int ret;
10072
10073                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10074                 if (!rkobj)
10075                         goto out_err;
10076                 rkobj->raid_type = index;
10077                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10078                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10079                                   "%s", get_raid_name(index));
10080                 if (ret) {
10081                         kobject_put(&rkobj->kobj);
10082                         goto out_err;
10083                 }
10084                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10085         }
10086
10087         return;
10088 out_err:
10089         btrfs_warn(cache->fs_info,
10090                    "failed to add kobject for block cache, ignoring");
10091 }
10092
10093 static struct btrfs_block_group_cache *
10094 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
10095 {
10096         struct btrfs_block_group_cache *cache;
10097
10098         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10099         if (!cache)
10100                 return NULL;
10101
10102         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10103                                         GFP_NOFS);
10104         if (!cache->free_space_ctl) {
10105                 kfree(cache);
10106                 return NULL;
10107         }
10108
10109         cache->key.objectid = start;
10110         cache->key.offset = size;
10111         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10112
10113         cache->sectorsize = root->sectorsize;
10114         cache->fs_info = root->fs_info;
10115         cache->full_stripe_len = btrfs_full_stripe_len(root,
10116                                                &root->fs_info->mapping_tree,
10117                                                start);
10118         set_free_space_tree_thresholds(cache);
10119
10120         atomic_set(&cache->count, 1);
10121         spin_lock_init(&cache->lock);
10122         init_rwsem(&cache->data_rwsem);
10123         INIT_LIST_HEAD(&cache->list);
10124         INIT_LIST_HEAD(&cache->cluster_list);
10125         INIT_LIST_HEAD(&cache->bg_list);
10126         INIT_LIST_HEAD(&cache->ro_list);
10127         INIT_LIST_HEAD(&cache->dirty_list);
10128         INIT_LIST_HEAD(&cache->io_list);
10129         btrfs_init_free_space_ctl(cache);
10130         atomic_set(&cache->trimming, 0);
10131         mutex_init(&cache->free_space_lock);
10132
10133         return cache;
10134 }
10135
10136 int btrfs_read_block_groups(struct btrfs_root *root)
10137 {
10138         struct btrfs_path *path;
10139         int ret;
10140         struct btrfs_block_group_cache *cache;
10141         struct btrfs_fs_info *info = root->fs_info;
10142         struct btrfs_space_info *space_info;
10143         struct btrfs_key key;
10144         struct btrfs_key found_key;
10145         struct extent_buffer *leaf;
10146         int need_clear = 0;
10147         u64 cache_gen;
10148         u64 feature;
10149         int mixed;
10150
10151         feature = btrfs_super_incompat_flags(info->super_copy);
10152         mixed = !!(feature & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS);
10153
10154         root = info->extent_root;
10155         key.objectid = 0;
10156         key.offset = 0;
10157         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10158         path = btrfs_alloc_path();
10159         if (!path)
10160                 return -ENOMEM;
10161         path->reada = READA_FORWARD;
10162
10163         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
10164         if (btrfs_test_opt(root->fs_info, SPACE_CACHE) &&
10165             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
10166                 need_clear = 1;
10167         if (btrfs_test_opt(root->fs_info, CLEAR_CACHE))
10168                 need_clear = 1;
10169
10170         while (1) {
10171                 ret = find_first_block_group(root, path, &key);
10172                 if (ret > 0)
10173                         break;
10174                 if (ret != 0)
10175                         goto error;
10176
10177                 leaf = path->nodes[0];
10178                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10179
10180                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
10181                                                        found_key.offset);
10182                 if (!cache) {
10183                         ret = -ENOMEM;
10184                         goto error;
10185                 }
10186
10187                 if (need_clear) {
10188                         /*
10189                          * When we mount with old space cache, we need to
10190                          * set BTRFS_DC_CLEAR and set dirty flag.
10191                          *
10192                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10193                          *    truncate the old free space cache inode and
10194                          *    setup a new one.
10195                          * b) Setting 'dirty flag' makes sure that we flush
10196                          *    the new space cache info onto disk.
10197                          */
10198                         if (btrfs_test_opt(root->fs_info, SPACE_CACHE))
10199                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10200                 }
10201
10202                 read_extent_buffer(leaf, &cache->item,
10203                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10204                                    sizeof(cache->item));
10205                 cache->flags = btrfs_block_group_flags(&cache->item);
10206                 if (!mixed &&
10207                     ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
10208                     (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
10209                         btrfs_err(info,
10210 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
10211                                   cache->key.objectid);
10212                         ret = -EINVAL;
10213                         goto error;
10214                 }
10215
10216                 key.objectid = found_key.objectid + found_key.offset;
10217                 btrfs_release_path(path);
10218
10219                 /*
10220                  * We need to exclude the super stripes now so that the space
10221                  * info has super bytes accounted for, otherwise we'll think
10222                  * we have more space than we actually do.
10223                  */
10224                 ret = exclude_super_stripes(root, cache);
10225                 if (ret) {
10226                         /*
10227                          * We may have excluded something, so call this just in
10228                          * case.
10229                          */
10230                         free_excluded_extents(root, cache);
10231                         btrfs_put_block_group(cache);
10232                         goto error;
10233                 }
10234
10235                 /*
10236                  * check for two cases, either we are full, and therefore
10237                  * don't need to bother with the caching work since we won't
10238                  * find any space, or we are empty, and we can just add all
10239                  * the space in and be done with it.  This saves us _alot_ of
10240                  * time, particularly in the full case.
10241                  */
10242                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10243                         cache->last_byte_to_unpin = (u64)-1;
10244                         cache->cached = BTRFS_CACHE_FINISHED;
10245                         free_excluded_extents(root, cache);
10246                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10247                         cache->last_byte_to_unpin = (u64)-1;
10248                         cache->cached = BTRFS_CACHE_FINISHED;
10249                         add_new_free_space(cache, root->fs_info,
10250                                            found_key.objectid,
10251                                            found_key.objectid +
10252                                            found_key.offset);
10253                         free_excluded_extents(root, cache);
10254                 }
10255
10256                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
10257                 if (ret) {
10258                         btrfs_remove_free_space_cache(cache);
10259                         btrfs_put_block_group(cache);
10260                         goto error;
10261                 }
10262
10263                 trace_btrfs_add_block_group(root->fs_info, cache, 0);
10264                 ret = update_space_info(info, cache->flags, found_key.offset,
10265                                         btrfs_block_group_used(&cache->item),
10266                                         cache->bytes_super, &space_info);
10267                 if (ret) {
10268                         btrfs_remove_free_space_cache(cache);
10269                         spin_lock(&info->block_group_cache_lock);
10270                         rb_erase(&cache->cache_node,
10271                                  &info->block_group_cache_tree);
10272                         RB_CLEAR_NODE(&cache->cache_node);
10273                         spin_unlock(&info->block_group_cache_lock);
10274                         btrfs_put_block_group(cache);
10275                         goto error;
10276                 }
10277
10278                 cache->space_info = space_info;
10279
10280                 __link_block_group(space_info, cache);
10281
10282                 set_avail_alloc_bits(root->fs_info, cache->flags);
10283                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
10284                         inc_block_group_ro(cache, 1);
10285                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10286                         spin_lock(&info->unused_bgs_lock);
10287                         /* Should always be true but just in case. */
10288                         if (list_empty(&cache->bg_list)) {
10289                                 btrfs_get_block_group(cache);
10290                                 list_add_tail(&cache->bg_list,
10291                                               &info->unused_bgs);
10292                         }
10293                         spin_unlock(&info->unused_bgs_lock);
10294                 }
10295         }
10296
10297         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
10298                 if (!(get_alloc_profile(root, space_info->flags) &
10299                       (BTRFS_BLOCK_GROUP_RAID10 |
10300                        BTRFS_BLOCK_GROUP_RAID1 |
10301                        BTRFS_BLOCK_GROUP_RAID5 |
10302                        BTRFS_BLOCK_GROUP_RAID6 |
10303                        BTRFS_BLOCK_GROUP_DUP)))
10304                         continue;
10305                 /*
10306                  * avoid allocating from un-mirrored block group if there are
10307                  * mirrored block groups.
10308                  */
10309                 list_for_each_entry(cache,
10310                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10311                                 list)
10312                         inc_block_group_ro(cache, 1);
10313                 list_for_each_entry(cache,
10314                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10315                                 list)
10316                         inc_block_group_ro(cache, 1);
10317         }
10318
10319         init_global_block_rsv(info);
10320         ret = 0;
10321 error:
10322         btrfs_free_path(path);
10323         return ret;
10324 }
10325
10326 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
10327                                        struct btrfs_root *root)
10328 {
10329         struct btrfs_block_group_cache *block_group, *tmp;
10330         struct btrfs_root *extent_root = root->fs_info->extent_root;
10331         struct btrfs_block_group_item item;
10332         struct btrfs_key key;
10333         int ret = 0;
10334         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10335
10336         trans->can_flush_pending_bgs = false;
10337         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10338                 if (ret)
10339                         goto next;
10340
10341                 spin_lock(&block_group->lock);
10342                 memcpy(&item, &block_group->item, sizeof(item));
10343                 memcpy(&key, &block_group->key, sizeof(key));
10344                 spin_unlock(&block_group->lock);
10345
10346                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10347                                         sizeof(item));
10348                 if (ret)
10349                         btrfs_abort_transaction(trans, ret);
10350                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
10351                                                key.objectid, key.offset);
10352                 if (ret)
10353                         btrfs_abort_transaction(trans, ret);
10354                 add_block_group_free_space(trans, root->fs_info, block_group);
10355                 /* already aborted the transaction if it failed. */
10356 next:
10357                 list_del_init(&block_group->bg_list);
10358         }
10359         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10360 }
10361
10362 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10363                            struct btrfs_root *root, u64 bytes_used,
10364                            u64 type, u64 chunk_objectid, u64 chunk_offset,
10365                            u64 size)
10366 {
10367         int ret;
10368         struct btrfs_root *extent_root;
10369         struct btrfs_block_group_cache *cache;
10370         extent_root = root->fs_info->extent_root;
10371
10372         btrfs_set_log_full_commit(root->fs_info, trans);
10373
10374         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
10375         if (!cache)
10376                 return -ENOMEM;
10377
10378         btrfs_set_block_group_used(&cache->item, bytes_used);
10379         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
10380         btrfs_set_block_group_flags(&cache->item, type);
10381
10382         cache->flags = type;
10383         cache->last_byte_to_unpin = (u64)-1;
10384         cache->cached = BTRFS_CACHE_FINISHED;
10385         cache->needs_free_space = 1;
10386         ret = exclude_super_stripes(root, cache);
10387         if (ret) {
10388                 /*
10389                  * We may have excluded something, so call this just in
10390                  * case.
10391                  */
10392                 free_excluded_extents(root, cache);
10393                 btrfs_put_block_group(cache);
10394                 return ret;
10395         }
10396
10397         add_new_free_space(cache, root->fs_info, chunk_offset,
10398                            chunk_offset + size);
10399
10400         free_excluded_extents(root, cache);
10401
10402 #ifdef CONFIG_BTRFS_DEBUG
10403         if (btrfs_should_fragment_free_space(root, cache)) {
10404                 u64 new_bytes_used = size - bytes_used;
10405
10406                 bytes_used += new_bytes_used >> 1;
10407                 fragment_free_space(root, cache);
10408         }
10409 #endif
10410         /*
10411          * Call to ensure the corresponding space_info object is created and
10412          * assigned to our block group, but don't update its counters just yet.
10413          * We want our bg to be added to the rbtree with its ->space_info set.
10414          */
10415         ret = update_space_info(root->fs_info, cache->flags, 0, 0, 0,
10416                                 &cache->space_info);
10417         if (ret) {
10418                 btrfs_remove_free_space_cache(cache);
10419                 btrfs_put_block_group(cache);
10420                 return ret;
10421         }
10422
10423         ret = btrfs_add_block_group_cache(root->fs_info, cache);
10424         if (ret) {
10425                 btrfs_remove_free_space_cache(cache);
10426                 btrfs_put_block_group(cache);
10427                 return ret;
10428         }
10429
10430         /*
10431          * Now that our block group has its ->space_info set and is inserted in
10432          * the rbtree, update the space info's counters.
10433          */
10434         trace_btrfs_add_block_group(root->fs_info, cache, 1);
10435         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
10436                                 cache->bytes_super, &cache->space_info);
10437         if (ret) {
10438                 btrfs_remove_free_space_cache(cache);
10439                 spin_lock(&root->fs_info->block_group_cache_lock);
10440                 rb_erase(&cache->cache_node,
10441                          &root->fs_info->block_group_cache_tree);
10442                 RB_CLEAR_NODE(&cache->cache_node);
10443                 spin_unlock(&root->fs_info->block_group_cache_lock);
10444                 btrfs_put_block_group(cache);
10445                 return ret;
10446         }
10447         update_global_block_rsv(root->fs_info);
10448
10449         __link_block_group(cache->space_info, cache);
10450
10451         list_add_tail(&cache->bg_list, &trans->new_bgs);
10452
10453         set_avail_alloc_bits(extent_root->fs_info, type);
10454         return 0;
10455 }
10456
10457 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10458 {
10459         u64 extra_flags = chunk_to_extended(flags) &
10460                                 BTRFS_EXTENDED_PROFILE_MASK;
10461
10462         write_seqlock(&fs_info->profiles_lock);
10463         if (flags & BTRFS_BLOCK_GROUP_DATA)
10464                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10465         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10466                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10467         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10468                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10469         write_sequnlock(&fs_info->profiles_lock);
10470 }
10471
10472 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10473                              struct btrfs_root *root, u64 group_start,
10474                              struct extent_map *em)
10475 {
10476         struct btrfs_path *path;
10477         struct btrfs_block_group_cache *block_group;
10478         struct btrfs_free_cluster *cluster;
10479         struct btrfs_root *tree_root = root->fs_info->tree_root;
10480         struct btrfs_key key;
10481         struct inode *inode;
10482         struct kobject *kobj = NULL;
10483         int ret;
10484         int index;
10485         int factor;
10486         struct btrfs_caching_control *caching_ctl = NULL;
10487         bool remove_em;
10488
10489         root = root->fs_info->extent_root;
10490
10491         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
10492         BUG_ON(!block_group);
10493         BUG_ON(!block_group->ro);
10494
10495         /*
10496          * Free the reserved super bytes from this block group before
10497          * remove it.
10498          */
10499         free_excluded_extents(root, block_group);
10500
10501         memcpy(&key, &block_group->key, sizeof(key));
10502         index = get_block_group_index(block_group);
10503         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10504                                   BTRFS_BLOCK_GROUP_RAID1 |
10505                                   BTRFS_BLOCK_GROUP_RAID10))
10506                 factor = 2;
10507         else
10508                 factor = 1;
10509
10510         /* make sure this block group isn't part of an allocation cluster */
10511         cluster = &root->fs_info->data_alloc_cluster;
10512         spin_lock(&cluster->refill_lock);
10513         btrfs_return_cluster_to_free_space(block_group, cluster);
10514         spin_unlock(&cluster->refill_lock);
10515
10516         /*
10517          * make sure this block group isn't part of a metadata
10518          * allocation cluster
10519          */
10520         cluster = &root->fs_info->meta_alloc_cluster;
10521         spin_lock(&cluster->refill_lock);
10522         btrfs_return_cluster_to_free_space(block_group, cluster);
10523         spin_unlock(&cluster->refill_lock);
10524
10525         path = btrfs_alloc_path();
10526         if (!path) {
10527                 ret = -ENOMEM;
10528                 goto out;
10529         }
10530
10531         /*
10532          * get the inode first so any iput calls done for the io_list
10533          * aren't the final iput (no unlinks allowed now)
10534          */
10535         inode = lookup_free_space_inode(tree_root, block_group, path);
10536
10537         mutex_lock(&trans->transaction->cache_write_mutex);
10538         /*
10539          * make sure our free spache cache IO is done before remove the
10540          * free space inode
10541          */
10542         spin_lock(&trans->transaction->dirty_bgs_lock);
10543         if (!list_empty(&block_group->io_list)) {
10544                 list_del_init(&block_group->io_list);
10545
10546                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10547
10548                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10549                 btrfs_wait_cache_io(root, trans, block_group,
10550                                     &block_group->io_ctl, path,
10551                                     block_group->key.objectid);
10552                 btrfs_put_block_group(block_group);
10553                 spin_lock(&trans->transaction->dirty_bgs_lock);
10554         }
10555
10556         if (!list_empty(&block_group->dirty_list)) {
10557                 list_del_init(&block_group->dirty_list);
10558                 btrfs_put_block_group(block_group);
10559         }
10560         spin_unlock(&trans->transaction->dirty_bgs_lock);
10561         mutex_unlock(&trans->transaction->cache_write_mutex);
10562
10563         if (!IS_ERR(inode)) {
10564                 ret = btrfs_orphan_add(trans, inode);
10565                 if (ret) {
10566                         btrfs_add_delayed_iput(inode);
10567                         goto out;
10568                 }
10569                 clear_nlink(inode);
10570                 /* One for the block groups ref */
10571                 spin_lock(&block_group->lock);
10572                 if (block_group->iref) {
10573                         block_group->iref = 0;
10574                         block_group->inode = NULL;
10575                         spin_unlock(&block_group->lock);
10576                         iput(inode);
10577                 } else {
10578                         spin_unlock(&block_group->lock);
10579                 }
10580                 /* One for our lookup ref */
10581                 btrfs_add_delayed_iput(inode);
10582         }
10583
10584         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10585         key.offset = block_group->key.objectid;
10586         key.type = 0;
10587
10588         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10589         if (ret < 0)
10590                 goto out;
10591         if (ret > 0)
10592                 btrfs_release_path(path);
10593         if (ret == 0) {
10594                 ret = btrfs_del_item(trans, tree_root, path);
10595                 if (ret)
10596                         goto out;
10597                 btrfs_release_path(path);
10598         }
10599
10600         spin_lock(&root->fs_info->block_group_cache_lock);
10601         rb_erase(&block_group->cache_node,
10602                  &root->fs_info->block_group_cache_tree);
10603         RB_CLEAR_NODE(&block_group->cache_node);
10604
10605         if (root->fs_info->first_logical_byte == block_group->key.objectid)
10606                 root->fs_info->first_logical_byte = (u64)-1;
10607         spin_unlock(&root->fs_info->block_group_cache_lock);
10608
10609         down_write(&block_group->space_info->groups_sem);
10610         /*
10611          * we must use list_del_init so people can check to see if they
10612          * are still on the list after taking the semaphore
10613          */
10614         list_del_init(&block_group->list);
10615         if (list_empty(&block_group->space_info->block_groups[index])) {
10616                 kobj = block_group->space_info->block_group_kobjs[index];
10617                 block_group->space_info->block_group_kobjs[index] = NULL;
10618                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
10619         }
10620         up_write(&block_group->space_info->groups_sem);
10621         if (kobj) {
10622                 kobject_del(kobj);
10623                 kobject_put(kobj);
10624         }
10625
10626         if (block_group->has_caching_ctl)
10627                 caching_ctl = get_caching_control(block_group);
10628         if (block_group->cached == BTRFS_CACHE_STARTED)
10629                 wait_block_group_cache_done(block_group);
10630         if (block_group->has_caching_ctl) {
10631                 down_write(&root->fs_info->commit_root_sem);
10632                 if (!caching_ctl) {
10633                         struct btrfs_caching_control *ctl;
10634
10635                         list_for_each_entry(ctl,
10636                                     &root->fs_info->caching_block_groups, list)
10637                                 if (ctl->block_group == block_group) {
10638                                         caching_ctl = ctl;
10639                                         atomic_inc(&caching_ctl->count);
10640                                         break;
10641                                 }
10642                 }
10643                 if (caching_ctl)
10644                         list_del_init(&caching_ctl->list);
10645                 up_write(&root->fs_info->commit_root_sem);
10646                 if (caching_ctl) {
10647                         /* Once for the caching bgs list and once for us. */
10648                         put_caching_control(caching_ctl);
10649                         put_caching_control(caching_ctl);
10650                 }
10651         }
10652
10653         spin_lock(&trans->transaction->dirty_bgs_lock);
10654         if (!list_empty(&block_group->dirty_list)) {
10655                 WARN_ON(1);
10656         }
10657         if (!list_empty(&block_group->io_list)) {
10658                 WARN_ON(1);
10659         }
10660         spin_unlock(&trans->transaction->dirty_bgs_lock);
10661         btrfs_remove_free_space_cache(block_group);
10662
10663         spin_lock(&block_group->space_info->lock);
10664         list_del_init(&block_group->ro_list);
10665
10666         if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
10667                 WARN_ON(block_group->space_info->total_bytes
10668                         < block_group->key.offset);
10669                 WARN_ON(block_group->space_info->bytes_readonly
10670                         < block_group->key.offset);
10671                 WARN_ON(block_group->space_info->disk_total
10672                         < block_group->key.offset * factor);
10673         }
10674         block_group->space_info->total_bytes -= block_group->key.offset;
10675         block_group->space_info->bytes_readonly -= block_group->key.offset;
10676         block_group->space_info->disk_total -= block_group->key.offset * factor;
10677
10678         spin_unlock(&block_group->space_info->lock);
10679
10680         memcpy(&key, &block_group->key, sizeof(key));
10681
10682         lock_chunks(root);
10683         if (!list_empty(&em->list)) {
10684                 /* We're in the transaction->pending_chunks list. */
10685                 free_extent_map(em);
10686         }
10687         spin_lock(&block_group->lock);
10688         block_group->removed = 1;
10689         /*
10690          * At this point trimming can't start on this block group, because we
10691          * removed the block group from the tree fs_info->block_group_cache_tree
10692          * so no one can't find it anymore and even if someone already got this
10693          * block group before we removed it from the rbtree, they have already
10694          * incremented block_group->trimming - if they didn't, they won't find
10695          * any free space entries because we already removed them all when we
10696          * called btrfs_remove_free_space_cache().
10697          *
10698          * And we must not remove the extent map from the fs_info->mapping_tree
10699          * to prevent the same logical address range and physical device space
10700          * ranges from being reused for a new block group. This is because our
10701          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10702          * completely transactionless, so while it is trimming a range the
10703          * currently running transaction might finish and a new one start,
10704          * allowing for new block groups to be created that can reuse the same
10705          * physical device locations unless we take this special care.
10706          *
10707          * There may also be an implicit trim operation if the file system
10708          * is mounted with -odiscard. The same protections must remain
10709          * in place until the extents have been discarded completely when
10710          * the transaction commit has completed.
10711          */
10712         remove_em = (atomic_read(&block_group->trimming) == 0);
10713         /*
10714          * Make sure a trimmer task always sees the em in the pinned_chunks list
10715          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10716          * before checking block_group->removed).
10717          */
10718         if (!remove_em) {
10719                 /*
10720                  * Our em might be in trans->transaction->pending_chunks which
10721                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10722                  * and so is the fs_info->pinned_chunks list.
10723                  *
10724                  * So at this point we must be holding the chunk_mutex to avoid
10725                  * any races with chunk allocation (more specifically at
10726                  * volumes.c:contains_pending_extent()), to ensure it always
10727                  * sees the em, either in the pending_chunks list or in the
10728                  * pinned_chunks list.
10729                  */
10730                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10731         }
10732         spin_unlock(&block_group->lock);
10733
10734         if (remove_em) {
10735                 struct extent_map_tree *em_tree;
10736
10737                 em_tree = &root->fs_info->mapping_tree.map_tree;
10738                 write_lock(&em_tree->lock);
10739                 /*
10740                  * The em might be in the pending_chunks list, so make sure the
10741                  * chunk mutex is locked, since remove_extent_mapping() will
10742                  * delete us from that list.
10743                  */
10744                 remove_extent_mapping(em_tree, em);
10745                 write_unlock(&em_tree->lock);
10746                 /* once for the tree */
10747                 free_extent_map(em);
10748         }
10749
10750         unlock_chunks(root);
10751
10752         ret = remove_block_group_free_space(trans, root->fs_info, block_group);
10753         if (ret)
10754                 goto out;
10755
10756         btrfs_put_block_group(block_group);
10757         btrfs_put_block_group(block_group);
10758
10759         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10760         if (ret > 0)
10761                 ret = -EIO;
10762         if (ret < 0)
10763                 goto out;
10764
10765         ret = btrfs_del_item(trans, root, path);
10766 out:
10767         btrfs_free_path(path);
10768         return ret;
10769 }
10770
10771 struct btrfs_trans_handle *
10772 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10773                                      const u64 chunk_offset)
10774 {
10775         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10776         struct extent_map *em;
10777         struct map_lookup *map;
10778         unsigned int num_items;
10779
10780         read_lock(&em_tree->lock);
10781         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10782         read_unlock(&em_tree->lock);
10783         ASSERT(em && em->start == chunk_offset);
10784
10785         /*
10786          * We need to reserve 3 + N units from the metadata space info in order
10787          * to remove a block group (done at btrfs_remove_chunk() and at
10788          * btrfs_remove_block_group()), which are used for:
10789          *
10790          * 1 unit for adding the free space inode's orphan (located in the tree
10791          * of tree roots).
10792          * 1 unit for deleting the block group item (located in the extent
10793          * tree).
10794          * 1 unit for deleting the free space item (located in tree of tree
10795          * roots).
10796          * N units for deleting N device extent items corresponding to each
10797          * stripe (located in the device tree).
10798          *
10799          * In order to remove a block group we also need to reserve units in the
10800          * system space info in order to update the chunk tree (update one or
10801          * more device items and remove one chunk item), but this is done at
10802          * btrfs_remove_chunk() through a call to check_system_chunk().
10803          */
10804         map = em->map_lookup;
10805         num_items = 3 + map->num_stripes;
10806         free_extent_map(em);
10807
10808         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10809                                                            num_items, 1);
10810 }
10811
10812 /*
10813  * Process the unused_bgs list and remove any that don't have any allocated
10814  * space inside of them.
10815  */
10816 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10817 {
10818         struct btrfs_block_group_cache *block_group;
10819         struct btrfs_space_info *space_info;
10820         struct btrfs_root *root = fs_info->extent_root;
10821         struct btrfs_trans_handle *trans;
10822         int ret = 0;
10823
10824         if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
10825                 return;
10826
10827         spin_lock(&fs_info->unused_bgs_lock);
10828         while (!list_empty(&fs_info->unused_bgs)) {
10829                 u64 start, end;
10830                 int trimming;
10831
10832                 block_group = list_first_entry(&fs_info->unused_bgs,
10833                                                struct btrfs_block_group_cache,
10834                                                bg_list);
10835                 list_del_init(&block_group->bg_list);
10836
10837                 space_info = block_group->space_info;
10838
10839                 if (ret || btrfs_mixed_space_info(space_info)) {
10840                         btrfs_put_block_group(block_group);
10841                         continue;
10842                 }
10843                 spin_unlock(&fs_info->unused_bgs_lock);
10844
10845                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10846
10847                 /* Don't want to race with allocators so take the groups_sem */
10848                 down_write(&space_info->groups_sem);
10849                 spin_lock(&block_group->lock);
10850                 if (block_group->reserved ||
10851                     btrfs_block_group_used(&block_group->item) ||
10852                     block_group->ro ||
10853                     list_is_singular(&block_group->list)) {
10854                         /*
10855                          * We want to bail if we made new allocations or have
10856                          * outstanding allocations in this block group.  We do
10857                          * the ro check in case balance is currently acting on
10858                          * this block group.
10859                          */
10860                         spin_unlock(&block_group->lock);
10861                         up_write(&space_info->groups_sem);
10862                         goto next;
10863                 }
10864                 spin_unlock(&block_group->lock);
10865
10866                 /* We don't want to force the issue, only flip if it's ok. */
10867                 ret = inc_block_group_ro(block_group, 0);
10868                 up_write(&space_info->groups_sem);
10869                 if (ret < 0) {
10870                         ret = 0;
10871                         goto next;
10872                 }
10873
10874                 /*
10875                  * Want to do this before we do anything else so we can recover
10876                  * properly if we fail to join the transaction.
10877                  */
10878                 trans = btrfs_start_trans_remove_block_group(fs_info,
10879                                                      block_group->key.objectid);
10880                 if (IS_ERR(trans)) {
10881                         btrfs_dec_block_group_ro(root, block_group);
10882                         ret = PTR_ERR(trans);
10883                         goto next;
10884                 }
10885
10886                 /*
10887                  * We could have pending pinned extents for this block group,
10888                  * just delete them, we don't care about them anymore.
10889                  */
10890                 start = block_group->key.objectid;
10891                 end = start + block_group->key.offset - 1;
10892                 /*
10893                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10894                  * btrfs_finish_extent_commit(). If we are at transaction N,
10895                  * another task might be running finish_extent_commit() for the
10896                  * previous transaction N - 1, and have seen a range belonging
10897                  * to the block group in freed_extents[] before we were able to
10898                  * clear the whole block group range from freed_extents[]. This
10899                  * means that task can lookup for the block group after we
10900                  * unpinned it from freed_extents[] and removed it, leading to
10901                  * a BUG_ON() at btrfs_unpin_extent_range().
10902                  */
10903                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10904                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10905                                   EXTENT_DIRTY);
10906                 if (ret) {
10907                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10908                         btrfs_dec_block_group_ro(root, block_group);
10909                         goto end_trans;
10910                 }
10911                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10912                                   EXTENT_DIRTY);
10913                 if (ret) {
10914                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10915                         btrfs_dec_block_group_ro(root, block_group);
10916                         goto end_trans;
10917                 }
10918                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10919
10920                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10921                 spin_lock(&space_info->lock);
10922                 spin_lock(&block_group->lock);
10923
10924                 space_info->bytes_pinned -= block_group->pinned;
10925                 space_info->bytes_readonly += block_group->pinned;
10926                 percpu_counter_add(&space_info->total_bytes_pinned,
10927                                    -block_group->pinned);
10928                 block_group->pinned = 0;
10929
10930                 spin_unlock(&block_group->lock);
10931                 spin_unlock(&space_info->lock);
10932
10933                 /* DISCARD can flip during remount */
10934                 trimming = btrfs_test_opt(root->fs_info, DISCARD);
10935
10936                 /* Implicit trim during transaction commit. */
10937                 if (trimming)
10938                         btrfs_get_block_group_trimming(block_group);
10939
10940                 /*
10941                  * Btrfs_remove_chunk will abort the transaction if things go
10942                  * horribly wrong.
10943                  */
10944                 ret = btrfs_remove_chunk(trans, root,
10945                                          block_group->key.objectid);
10946
10947                 if (ret) {
10948                         if (trimming)
10949                                 btrfs_put_block_group_trimming(block_group);
10950                         goto end_trans;
10951                 }
10952
10953                 /*
10954                  * If we're not mounted with -odiscard, we can just forget
10955                  * about this block group. Otherwise we'll need to wait
10956                  * until transaction commit to do the actual discard.
10957                  */
10958                 if (trimming) {
10959                         spin_lock(&fs_info->unused_bgs_lock);
10960                         /*
10961                          * A concurrent scrub might have added us to the list
10962                          * fs_info->unused_bgs, so use a list_move operation
10963                          * to add the block group to the deleted_bgs list.
10964                          */
10965                         list_move(&block_group->bg_list,
10966                                   &trans->transaction->deleted_bgs);
10967                         spin_unlock(&fs_info->unused_bgs_lock);
10968                         btrfs_get_block_group(block_group);
10969                 }
10970 end_trans:
10971                 btrfs_end_transaction(trans, root);
10972 next:
10973                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10974                 btrfs_put_block_group(block_group);
10975                 spin_lock(&fs_info->unused_bgs_lock);
10976         }
10977         spin_unlock(&fs_info->unused_bgs_lock);
10978 }
10979
10980 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10981 {
10982         struct btrfs_space_info *space_info;
10983         struct btrfs_super_block *disk_super;
10984         u64 features;
10985         u64 flags;
10986         int mixed = 0;
10987         int ret;
10988
10989         disk_super = fs_info->super_copy;
10990         if (!btrfs_super_root(disk_super))
10991                 return -EINVAL;
10992
10993         features = btrfs_super_incompat_flags(disk_super);
10994         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10995                 mixed = 1;
10996
10997         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10998         ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10999         if (ret)
11000                 goto out;
11001
11002         if (mixed) {
11003                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
11004                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
11005         } else {
11006                 flags = BTRFS_BLOCK_GROUP_METADATA;
11007                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
11008                 if (ret)
11009                         goto out;
11010
11011                 flags = BTRFS_BLOCK_GROUP_DATA;
11012                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
11013         }
11014 out:
11015         return ret;
11016 }
11017
11018 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
11019 {
11020         return unpin_extent_range(root, start, end, false);
11021 }
11022
11023 /*
11024  * It used to be that old block groups would be left around forever.
11025  * Iterating over them would be enough to trim unused space.  Since we
11026  * now automatically remove them, we also need to iterate over unallocated
11027  * space.
11028  *
11029  * We don't want a transaction for this since the discard may take a
11030  * substantial amount of time.  We don't require that a transaction be
11031  * running, but we do need to take a running transaction into account
11032  * to ensure that we're not discarding chunks that were released in
11033  * the current transaction.
11034  *
11035  * Holding the chunks lock will prevent other threads from allocating
11036  * or releasing chunks, but it won't prevent a running transaction
11037  * from committing and releasing the memory that the pending chunks
11038  * list head uses.  For that, we need to take a reference to the
11039  * transaction.
11040  */
11041 static int btrfs_trim_free_extents(struct btrfs_device *device,
11042                                    u64 minlen, u64 *trimmed)
11043 {
11044         u64 start = 0, len = 0;
11045         int ret;
11046
11047         *trimmed = 0;
11048
11049         /* Not writeable = nothing to do. */
11050         if (!device->writeable)
11051                 return 0;
11052
11053         /* No free space = nothing to do. */
11054         if (device->total_bytes <= device->bytes_used)
11055                 return 0;
11056
11057         ret = 0;
11058
11059         while (1) {
11060                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
11061                 struct btrfs_transaction *trans;
11062                 u64 bytes;
11063
11064                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11065                 if (ret)
11066                         return ret;
11067
11068                 down_read(&fs_info->commit_root_sem);
11069
11070                 spin_lock(&fs_info->trans_lock);
11071                 trans = fs_info->running_transaction;
11072                 if (trans)
11073                         atomic_inc(&trans->use_count);
11074                 spin_unlock(&fs_info->trans_lock);
11075
11076                 ret = find_free_dev_extent_start(trans, device, minlen, start,
11077                                                  &start, &len);
11078                 if (trans)
11079                         btrfs_put_transaction(trans);
11080
11081                 if (ret) {
11082                         up_read(&fs_info->commit_root_sem);
11083                         mutex_unlock(&fs_info->chunk_mutex);
11084                         if (ret == -ENOSPC)
11085                                 ret = 0;
11086                         break;
11087                 }
11088
11089                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
11090                 up_read(&fs_info->commit_root_sem);
11091                 mutex_unlock(&fs_info->chunk_mutex);
11092
11093                 if (ret)
11094                         break;
11095
11096                 start += len;
11097                 *trimmed += bytes;
11098
11099                 if (fatal_signal_pending(current)) {
11100                         ret = -ERESTARTSYS;
11101                         break;
11102                 }
11103
11104                 cond_resched();
11105         }
11106
11107         return ret;
11108 }
11109
11110 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
11111 {
11112         struct btrfs_fs_info *fs_info = root->fs_info;
11113         struct btrfs_block_group_cache *cache = NULL;
11114         struct btrfs_device *device;
11115         struct list_head *devices;
11116         u64 group_trimmed;
11117         u64 start;
11118         u64 end;
11119         u64 trimmed = 0;
11120         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
11121         int ret = 0;
11122
11123         /*
11124          * try to trim all FS space, our block group may start from non-zero.
11125          */
11126         if (range->len == total_bytes)
11127                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
11128         else
11129                 cache = btrfs_lookup_block_group(fs_info, range->start);
11130
11131         while (cache) {
11132                 if (cache->key.objectid >= (range->start + range->len)) {
11133                         btrfs_put_block_group(cache);
11134                         break;
11135                 }
11136
11137                 start = max(range->start, cache->key.objectid);
11138                 end = min(range->start + range->len,
11139                                 cache->key.objectid + cache->key.offset);
11140
11141                 if (end - start >= range->minlen) {
11142                         if (!block_group_cache_done(cache)) {
11143                                 ret = cache_block_group(cache, 0);
11144                                 if (ret) {
11145                                         btrfs_put_block_group(cache);
11146                                         break;
11147                                 }
11148                                 ret = wait_block_group_cache_done(cache);
11149                                 if (ret) {
11150                                         btrfs_put_block_group(cache);
11151                                         break;
11152                                 }
11153                         }
11154                         ret = btrfs_trim_block_group(cache,
11155                                                      &group_trimmed,
11156                                                      start,
11157                                                      end,
11158                                                      range->minlen);
11159
11160                         trimmed += group_trimmed;
11161                         if (ret) {
11162                                 btrfs_put_block_group(cache);
11163                                 break;
11164                         }
11165                 }
11166
11167                 cache = next_block_group(fs_info->tree_root, cache);
11168         }
11169
11170         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
11171         devices = &root->fs_info->fs_devices->alloc_list;
11172         list_for_each_entry(device, devices, dev_alloc_list) {
11173                 ret = btrfs_trim_free_extents(device, range->minlen,
11174                                               &group_trimmed);
11175                 if (ret)
11176                         break;
11177
11178                 trimmed += group_trimmed;
11179         }
11180         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
11181
11182         range->len = trimmed;
11183         return ret;
11184 }
11185
11186 /*
11187  * btrfs_{start,end}_write_no_snapshoting() are similar to
11188  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11189  * data into the page cache through nocow before the subvolume is snapshoted,
11190  * but flush the data into disk after the snapshot creation, or to prevent
11191  * operations while snapshoting is ongoing and that cause the snapshot to be
11192  * inconsistent (writes followed by expanding truncates for example).
11193  */
11194 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
11195 {
11196         percpu_counter_dec(&root->subv_writers->counter);
11197         /*
11198          * Make sure counter is updated before we wake up waiters.
11199          */
11200         smp_mb();
11201         if (waitqueue_active(&root->subv_writers->wait))
11202                 wake_up(&root->subv_writers->wait);
11203 }
11204
11205 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
11206 {
11207         if (atomic_read(&root->will_be_snapshoted))
11208                 return 0;
11209
11210         percpu_counter_inc(&root->subv_writers->counter);
11211         /*
11212          * Make sure counter is updated before we check for snapshot creation.
11213          */
11214         smp_mb();
11215         if (atomic_read(&root->will_be_snapshoted)) {
11216                 btrfs_end_write_no_snapshoting(root);
11217                 return 0;
11218         }
11219         return 1;
11220 }
11221
11222 static int wait_snapshoting_atomic_t(atomic_t *a)
11223 {
11224         schedule();
11225         return 0;
11226 }
11227
11228 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11229 {
11230         while (true) {
11231                 int ret;
11232
11233                 ret = btrfs_start_write_no_snapshoting(root);
11234                 if (ret)
11235                         break;
11236                 wait_on_atomic_t(&root->will_be_snapshoted,
11237                                  wait_snapshoting_atomic_t,
11238                                  TASK_UNINTERRUPTIBLE);
11239         }
11240 }