fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include <linux/percpu_counter.h>
  28 #include "hash.h"
  29 #include "tree-log.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "volumes.h"
  33 #include "raid56.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36 #include "free-space-tree.h"
  37 #include "math.h"
  38 #include "sysfs.h"
  39 #include "qgroup.h"
  40
  41 #undef SCRAMBLE_DELAYED_REFS
  42
  43 /*
  44  * control flags for do_chunk_alloc's force field
  45  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  46  * if we really need one.
  47  *
  48  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  49  * if we have very few chunks already allocated.  This is
  50  * used as part of the clustering code to help make sure
  51  * we have a good pool of storage to cluster in, without
  52  * filling the FS with empty chunks
  53  *
  54  * CHUNK_ALLOC_FORCE means it must try to allocate one
  55  *
  56  */
  57 enum {
  58         CHUNK_ALLOC_NO_FORCE = 0,
  59         CHUNK_ALLOC_LIMITED = 1,
  60         CHUNK_ALLOC_FORCE = 2,
  61 };
  62
  63 /*
  64  * Control how reservations are dealt with.
  65  *
  66  * RESERVE_FREE - freeing a reservation.
  67  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  68  *   ENOSPC accounting
  69  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  70  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  71  */
  72 enum {
  73         RESERVE_FREE = 0,
  74         RESERVE_ALLOC = 1,
  75         RESERVE_ALLOC_NO_ACCOUNT = 2,
  76 };
  77
  78 static int update_block_group(struct btrfs_trans_handle *trans,
  79                               struct btrfs_root *root, u64 bytenr,
  80                               u64 num_bytes, int alloc);
  81 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  82                                 struct btrfs_root *root,
  83                                 struct btrfs_delayed_ref_node *node, u64 parent,
  84                                 u64 root_objectid, u64 owner_objectid,
  85                                 u64 owner_offset, int refs_to_drop,
  86                                 struct btrfs_delayed_extent_op *extra_op);
  87 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  88                                     struct extent_buffer *leaf,
  89                                     struct btrfs_extent_item *ei);
  90 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  91                                       struct btrfs_root *root,
  92                                       u64 parent, u64 root_objectid,
  93                                       u64 flags, u64 owner, u64 offset,
  94                                       struct btrfs_key *ins, int ref_mod);
  95 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  96                                      struct btrfs_root *root,
  97                                      u64 parent, u64 root_objectid,
  98                                      u64 flags, struct btrfs_disk_key *key,
  99                                      int level, struct btrfs_key *ins);
 100 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
 101                           struct btrfs_root *extent_root, u64 flags,
 102                           int force);
 103 static int find_next_key(struct btrfs_path *path, int level,
 104                          struct btrfs_key *key);
 105 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 106                             int dump_block_groups);
 107 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 108                                        u64 num_bytes, int reserve,
 109                                        int delalloc);
 110 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 111                                u64 num_bytes);
 112 int btrfs_pin_extent(struct btrfs_root *root,
 113                      u64 bytenr, u64 num_bytes, int reserved);
 114 static int __reserve_metadata_bytes(struct btrfs_root *root,
 115                                     struct btrfs_space_info *space_info,
 116                                     u64 orig_bytes,
 117                                     enum btrfs_reserve_flush_enum flush);
 118 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
 119                                      struct btrfs_space_info *space_info,
 120                                      u64 num_bytes);
 121 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
 122                                      struct btrfs_space_info *space_info,
 123                                      u64 num_bytes);
 124
 125 static noinline int
 126 block_group_cache_done(struct btrfs_block_group_cache *cache)
 127 {
 128         smp_mb();
 129         return cache->cached == BTRFS_CACHE_FINISHED ||
 130                 cache->cached == BTRFS_CACHE_ERROR;
 131 }
 132
 133 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 134 {
 135         return (cache->flags & bits) == bits;
 136 }
 137
 138 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 139 {
 140         atomic_inc(&cache->count);
 141 }
 142
 143 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 144 {
 145         if (atomic_dec_and_test(&cache->count)) {
 146                 WARN_ON(cache->pinned > 0);
 147                 WARN_ON(cache->reserved > 0);
 148                 kfree(cache->free_space_ctl);
 149                 kfree(cache);
 150         }
 151 }
 152
 153 /*
 154  * this adds the block group to the fs_info rb tree for the block group
 155  * cache
 156  */
 157 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 158                                 struct btrfs_block_group_cache *block_group)
 159 {
 160         struct rb_node **p;
 161         struct rb_node *parent = NULL;
 162         struct btrfs_block_group_cache *cache;
 163
 164         spin_lock(&info->block_group_cache_lock);
 165         p = &info->block_group_cache_tree.rb_node;
 166
 167         while (*p) {
 168                 parent = *p;
 169                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 170                                  cache_node);
 171                 if (block_group->key.objectid < cache->key.objectid) {
 172                         p = &(*p)->rb_left;
 173                 } else if (block_group->key.objectid > cache->key.objectid) {
 174                         p = &(*p)->rb_right;
 175                 } else {
 176                         spin_unlock(&info->block_group_cache_lock);
 177                         return -EEXIST;
 178                 }
 179         }
 180
 181         rb_link_node(&block_group->cache_node, parent, p);
 182         rb_insert_color(&block_group->cache_node,
 183                         &info->block_group_cache_tree);
 184
 185         if (info->first_logical_byte > block_group->key.objectid)
 186                 info->first_logical_byte = block_group->key.objectid;
 187
 188         spin_unlock(&info->block_group_cache_lock);
 189
 190         return 0;
 191 }
 192
 193 /*
 194  * This will return the block group at or after bytenr if contains is 0, else
 195  * it will return the block group that contains the bytenr
 196  */
 197 static struct btrfs_block_group_cache *
 198 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 199                               int contains)
 200 {
 201         struct btrfs_block_group_cache *cache, *ret = NULL;
 202         struct rb_node *n;
 203         u64 end, start;
 204
 205         spin_lock(&info->block_group_cache_lock);
 206         n = info->block_group_cache_tree.rb_node;
 207
 208         while (n) {
 209                 cache = rb_entry(n, struct btrfs_block_group_cache,
 210                                  cache_node);
 211                 end = cache->key.objectid + cache->key.offset - 1;
 212                 start = cache->key.objectid;
 213
 214                 if (bytenr < start) {
 215                         if (!contains && (!ret || start < ret->key.objectid))
 216                                 ret = cache;
 217                         n = n->rb_left;
 218                 } else if (bytenr > start) {
 219                         if (contains && bytenr <= end) {
 220                                 ret = cache;
 221                                 break;
 222                         }
 223                         n = n->rb_right;
 224                 } else {
 225                         ret = cache;
 226                         break;
 227                 }
 228         }
 229         if (ret) {
 230                 btrfs_get_block_group(ret);
 231                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 232                         info->first_logical_byte = ret->key.objectid;
 233         }
 234         spin_unlock(&info->block_group_cache_lock);
 235
 236         return ret;
 237 }
 238
 239 static int add_excluded_extent(struct btrfs_root *root,
 240                                u64 start, u64 num_bytes)
 241 {
 242         u64 end = start + num_bytes - 1;
 243         set_extent_bits(&root->fs_info->freed_extents[0],
 244                         start, end, EXTENT_UPTODATE);
 245         set_extent_bits(&root->fs_info->freed_extents[1],
 246                         start, end, EXTENT_UPTODATE);
 247         return 0;
 248 }
 249
 250 static void free_excluded_extents(struct btrfs_root *root,
 251                                   struct btrfs_block_group_cache *cache)
 252 {
 253         u64 start, end;
 254
 255         start = cache->key.objectid;
 256         end = start + cache->key.offset - 1;
 257
 258         clear_extent_bits(&root->fs_info->freed_extents[0],
 259                           start, end, EXTENT_UPTODATE);
 260         clear_extent_bits(&root->fs_info->freed_extents[1],
 261                           start, end, EXTENT_UPTODATE);
 262 }
 263
 264 static int exclude_super_stripes(struct btrfs_root *root,
 265                                  struct btrfs_block_group_cache *cache)
 266 {
 267         u64 bytenr;
 268         u64 *logical;
 269         int stripe_len;
 270         int i, nr, ret;
 271
 272         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 273                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 274                 cache->bytes_super += stripe_len;
 275                 ret = add_excluded_extent(root, cache->key.objectid,
 276                                           stripe_len);
 277                 if (ret)
 278                         return ret;
 279         }
 280
 281         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 282                 bytenr = btrfs_sb_offset(i);
 283                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 284                                        cache->key.objectid, bytenr,
 285                                        0, &logical, &nr, &stripe_len);
 286                 if (ret)
 287                         return ret;
 288
 289                 while (nr--) {
 290                         u64 start, len;
 291
 292                         if (logical[nr] > cache->key.objectid +
 293                             cache->key.offset)
 294                                 continue;
 295
 296                         if (logical[nr] + stripe_len <= cache->key.objectid)
 297                                 continue;
 298
 299                         start = logical[nr];
 300                         if (start < cache->key.objectid) {
 301                                 start = cache->key.objectid;
 302                                 len = (logical[nr] + stripe_len) - start;
 303                         } else {
 304                                 len = min_t(u64, stripe_len,
 305                                             cache->key.objectid +
 306                                             cache->key.offset - start);
 307                         }
 308
 309                         cache->bytes_super += len;
 310                         ret = add_excluded_extent(root, start, len);
 311                         if (ret) {
 312                                 kfree(logical);
 313                                 return ret;
 314                         }
 315                 }
 316
 317                 kfree(logical);
 318         }
 319         return 0;
 320 }
 321
 322 static struct btrfs_caching_control *
 323 get_caching_control(struct btrfs_block_group_cache *cache)
 324 {
 325         struct btrfs_caching_control *ctl;
 326
 327         spin_lock(&cache->lock);
 328         if (!cache->caching_ctl) {
 329                 spin_unlock(&cache->lock);
 330                 return NULL;
 331         }
 332
 333         ctl = cache->caching_ctl;
 334         atomic_inc(&ctl->count);
 335         spin_unlock(&cache->lock);
 336         return ctl;
 337 }
 338
 339 static void put_caching_control(struct btrfs_caching_control *ctl)
 340 {
 341         if (atomic_dec_and_test(&ctl->count))
 342                 kfree(ctl);
 343 }
 344
 345 #ifdef CONFIG_BTRFS_DEBUG
 346 static void fragment_free_space(struct btrfs_root *root,
 347                                 struct btrfs_block_group_cache *block_group)
 348 {
 349         u64 start = block_group->key.objectid;
 350         u64 len = block_group->key.offset;
 351         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 352                 root->nodesize : root->sectorsize;
 353         u64 step = chunk << 1;
 354
 355         while (len > chunk) {
 356                 btrfs_remove_free_space(block_group, start, chunk);
 357                 start += step;
 358                 if (len < step)
 359                         len = 0;
 360                 else
 361                         len -= step;
 362         }
 363 }
 364 #endif
 365
 366 /*
 367  * this is only called by cache_block_group, since we could have freed extents
 368  * we need to check the pinned_extents for any extents that can't be used yet
 369  * since their free space will be released as soon as the transaction commits.
 370  */
 371 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 372                        struct btrfs_fs_info *info, u64 start, u64 end)
 373 {
 374         u64 extent_start, extent_end, size, total_added = 0;
 375         int ret;
 376
 377         while (start < end) {
 378                 ret = find_first_extent_bit(info->pinned_extents, start,
 379                                             &extent_start, &extent_end,
 380                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 381                                             NULL);
 382                 if (ret)
 383                         break;
 384
 385                 if (extent_start <= start) {
 386                         start = extent_end + 1;
 387                 } else if (extent_start > start && extent_start < end) {
 388                         size = extent_start - start;
 389                         total_added += size;
 390                         ret = btrfs_add_free_space(block_group, start,
 391                                                    size);
 392                         BUG_ON(ret); /* -ENOMEM or logic error */
 393                         start = extent_end + 1;
 394                 } else {
 395                         break;
 396                 }
 397         }
 398
 399         if (start < end) {
 400                 size = end - start;
 401                 total_added += size;
 402                 ret = btrfs_add_free_space(block_group, start, size);
 403                 BUG_ON(ret); /* -ENOMEM or logic error */
 404         }
 405
 406         return total_added;
 407 }
 408
 409 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 410 {
 411         struct btrfs_block_group_cache *block_group;
 412         struct btrfs_fs_info *fs_info;
 413         struct btrfs_root *extent_root;
 414         struct btrfs_path *path;
 415         struct extent_buffer *leaf;
 416         struct btrfs_key key;
 417         u64 total_found = 0;
 418         u64 last = 0;
 419         u32 nritems;
 420         int ret;
 421         bool wakeup = true;
 422
 423         block_group = caching_ctl->block_group;
 424         fs_info = block_group->fs_info;
 425         extent_root = fs_info->extent_root;
 426
 427         path = btrfs_alloc_path();
 428         if (!path)
 429                 return -ENOMEM;
 430
 431         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 432
 433 #ifdef CONFIG_BTRFS_DEBUG
 434         /*
 435          * If we're fragmenting we don't want to make anybody think we can
 436          * allocate from this block group until we've had a chance to fragment
 437          * the free space.
 438          */
 439         if (btrfs_should_fragment_free_space(extent_root, block_group))
 440                 wakeup = false;
 441 #endif
 442         /*
 443          * We don't want to deadlock with somebody trying to allocate a new
 444          * extent for the extent root while also trying to search the extent
 445          * root to add free space.  So we skip locking and search the commit
 446          * root, since its read-only
 447          */
 448         path->skip_locking = 1;
 449         path->search_commit_root = 1;
 450         path->reada = READA_FORWARD;
 451
 452         key.objectid = last;
 453         key.offset = 0;
 454         key.type = BTRFS_EXTENT_ITEM_KEY;
 455
 456 next:
 457         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 458         if (ret < 0)
 459                 goto out;
 460
 461         leaf = path->nodes[0];
 462         nritems = btrfs_header_nritems(leaf);
 463
 464         while (1) {
 465                 if (btrfs_fs_closing(fs_info) > 1) {
 466                         last = (u64)-1;
 467                         break;
 468                 }
 469
 470                 if (path->slots[0] < nritems) {
 471                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 472                 } else {
 473                         ret = find_next_key(path, 0, &key);
 474                         if (ret)
 475                                 break;
 476
 477                         if (need_resched() ||
 478                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 479                                 if (wakeup)
 480                                         caching_ctl->progress = last;
 481                                 btrfs_release_path(path);
 482                                 up_read(&fs_info->commit_root_sem);
 483                                 mutex_unlock(&caching_ctl->mutex);
 484                                 cond_resched();
 485                                 mutex_lock(&caching_ctl->mutex);
 486                                 down_read(&fs_info->commit_root_sem);
 487                                 goto next;
 488                         }
 489
 490                         ret = btrfs_next_leaf(extent_root, path);
 491                         if (ret < 0)
 492                                 goto out;
 493                         if (ret)
 494                                 break;
 495                         leaf = path->nodes[0];
 496                         nritems = btrfs_header_nritems(leaf);
 497                         continue;
 498                 }
 499
 500                 if (key.objectid < last) {
 501                         key.objectid = last;
 502                         key.offset = 0;
 503                         key.type = BTRFS_EXTENT_ITEM_KEY;
 504
 505                         if (wakeup)
 506                                 caching_ctl->progress = last;
 507                         btrfs_release_path(path);
 508                         goto next;
 509                 }
 510
 511                 if (key.objectid < block_group->key.objectid) {
 512                         path->slots[0]++;
 513                         continue;
 514                 }
 515
 516                 if (key.objectid >= block_group->key.objectid +
 517                     block_group->key.offset)
 518                         break;
 519
 520                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 521                     key.type == BTRFS_METADATA_ITEM_KEY) {
 522                         total_found += add_new_free_space(block_group,
 523                                                           fs_info, last,
 524                                                           key.objectid);
 525                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 526                                 last = key.objectid +
 527                                         fs_info->tree_root->nodesize;
 528                         else
 529                                 last = key.objectid + key.offset;
 530
 531                         if (total_found > CACHING_CTL_WAKE_UP) {
 532                                 total_found = 0;
 533                                 if (wakeup)
 534                                         wake_up(&caching_ctl->wait);
 535                         }
 536                 }
 537                 path->slots[0]++;
 538         }
 539         ret = 0;
 540
 541         total_found += add_new_free_space(block_group, fs_info, last,
 542                                           block_group->key.objectid +
 543                                           block_group->key.offset);
 544         caching_ctl->progress = (u64)-1;
 545
 546 out:
 547         btrfs_free_path(path);
 548         return ret;
 549 }
 550
 551 static noinline void caching_thread(struct btrfs_work *work)
 552 {
 553         struct btrfs_block_group_cache *block_group;
 554         struct btrfs_fs_info *fs_info;
 555         struct btrfs_caching_control *caching_ctl;
 556         struct btrfs_root *extent_root;
 557         int ret;
 558
 559         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 560         block_group = caching_ctl->block_group;
 561         fs_info = block_group->fs_info;
 562         extent_root = fs_info->extent_root;
 563
 564         mutex_lock(&caching_ctl->mutex);
 565         down_read(&fs_info->commit_root_sem);
 566
 567         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 568                 ret = load_free_space_tree(caching_ctl);
 569         else
 570                 ret = load_extent_tree_free(caching_ctl);
 571
 572         spin_lock(&block_group->lock);
 573         block_group->caching_ctl = NULL;
 574         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 575         spin_unlock(&block_group->lock);
 576
 577 #ifdef CONFIG_BTRFS_DEBUG
 578         if (btrfs_should_fragment_free_space(extent_root, block_group)) {
 579                 u64 bytes_used;
 580
 581                 spin_lock(&block_group->space_info->lock);
 582                 spin_lock(&block_group->lock);
 583                 bytes_used = block_group->key.offset -
 584                         btrfs_block_group_used(&block_group->item);
 585                 block_group->space_info->bytes_used += bytes_used >> 1;
 586                 spin_unlock(&block_group->lock);
 587                 spin_unlock(&block_group->space_info->lock);
 588                 fragment_free_space(extent_root, block_group);
 589         }
 590 #endif
 591
 592         caching_ctl->progress = (u64)-1;
 593
 594         up_read(&fs_info->commit_root_sem);
 595         free_excluded_extents(fs_info->extent_root, block_group);
 596         mutex_unlock(&caching_ctl->mutex);
 597
 598         wake_up(&caching_ctl->wait);
 599
 600         put_caching_control(caching_ctl);
 601         btrfs_put_block_group(block_group);
 602 }
 603
 604 static int cache_block_group(struct btrfs_block_group_cache *cache,
 605                              int load_cache_only)
 606 {
 607         DEFINE_WAIT(wait);
 608         struct btrfs_fs_info *fs_info = cache->fs_info;
 609         struct btrfs_caching_control *caching_ctl;
 610         int ret = 0;
 611
 612         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 613         if (!caching_ctl)
 614                 return -ENOMEM;
 615
 616         INIT_LIST_HEAD(&caching_ctl->list);
 617         mutex_init(&caching_ctl->mutex);
 618         init_waitqueue_head(&caching_ctl->wait);
 619         caching_ctl->block_group = cache;
 620         caching_ctl->progress = cache->key.objectid;
 621         atomic_set(&caching_ctl->count, 1);
 622         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 623                         caching_thread, NULL, NULL);
 624
 625         spin_lock(&cache->lock);
 626         /*
 627          * This should be a rare occasion, but this could happen I think in the
 628          * case where one thread starts to load the space cache info, and then
 629          * some other thread starts a transaction commit which tries to do an
 630          * allocation while the other thread is still loading the space cache
 631          * info.  The previous loop should have kept us from choosing this block
 632          * group, but if we've moved to the state where we will wait on caching
 633          * block groups we need to first check if we're doing a fast load here,
 634          * so we can wait for it to finish, otherwise we could end up allocating
 635          * from a block group who's cache gets evicted for one reason or
 636          * another.
 637          */
 638         while (cache->cached == BTRFS_CACHE_FAST) {
 639                 struct btrfs_caching_control *ctl;
 640
 641                 ctl = cache->caching_ctl;
 642                 atomic_inc(&ctl->count);
 643                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 644                 spin_unlock(&cache->lock);
 645
 646                 schedule();
 647
 648                 finish_wait(&ctl->wait, &wait);
 649                 put_caching_control(ctl);
 650                 spin_lock(&cache->lock);
 651         }
 652
 653         if (cache->cached != BTRFS_CACHE_NO) {
 654                 spin_unlock(&cache->lock);
 655                 kfree(caching_ctl);
 656                 return 0;
 657         }
 658         WARN_ON(cache->caching_ctl);
 659         cache->caching_ctl = caching_ctl;
 660         cache->cached = BTRFS_CACHE_FAST;
 661         spin_unlock(&cache->lock);
 662
 663         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 664                 mutex_lock(&caching_ctl->mutex);
 665                 ret = load_free_space_cache(fs_info, cache);
 666
 667                 spin_lock(&cache->lock);
 668                 if (ret == 1) {
 669                         cache->caching_ctl = NULL;
 670                         cache->cached = BTRFS_CACHE_FINISHED;
 671                         cache->last_byte_to_unpin = (u64)-1;
 672                         caching_ctl->progress = (u64)-1;
 673                 } else {
 674                         if (load_cache_only) {
 675                                 cache->caching_ctl = NULL;
 676                                 cache->cached = BTRFS_CACHE_NO;
 677                         } else {
 678                                 cache->cached = BTRFS_CACHE_STARTED;
 679                                 cache->has_caching_ctl = 1;
 680                         }
 681                 }
 682                 spin_unlock(&cache->lock);
 683 #ifdef CONFIG_BTRFS_DEBUG
 684                 if (ret == 1 &&
 685                     btrfs_should_fragment_free_space(fs_info->extent_root,
 686                                                      cache)) {
 687                         u64 bytes_used;
 688
 689                         spin_lock(&cache->space_info->lock);
 690                         spin_lock(&cache->lock);
 691                         bytes_used = cache->key.offset -
 692                                 btrfs_block_group_used(&cache->item);
 693                         cache->space_info->bytes_used += bytes_used >> 1;
 694                         spin_unlock(&cache->lock);
 695                         spin_unlock(&cache->space_info->lock);
 696                         fragment_free_space(fs_info->extent_root, cache);
 697                 }
 698 #endif
 699                 mutex_unlock(&caching_ctl->mutex);
 700
 701                 wake_up(&caching_ctl->wait);
 702                 if (ret == 1) {
 703                         put_caching_control(caching_ctl);
 704                         free_excluded_extents(fs_info->extent_root, cache);
 705                         return 0;
 706                 }
 707         } else {
 708                 /*
 709                  * We're either using the free space tree or no caching at all.
 710                  * Set cached to the appropriate value and wakeup any waiters.
 711                  */
 712                 spin_lock(&cache->lock);
 713                 if (load_cache_only) {
 714                         cache->caching_ctl = NULL;
 715                         cache->cached = BTRFS_CACHE_NO;
 716                 } else {
 717                         cache->cached = BTRFS_CACHE_STARTED;
 718                         cache->has_caching_ctl = 1;
 719                 }
 720                 spin_unlock(&cache->lock);
 721                 wake_up(&caching_ctl->wait);
 722         }
 723
 724         if (load_cache_only) {
 725                 put_caching_control(caching_ctl);
 726                 return 0;
 727         }
 728
 729         down_write(&fs_info->commit_root_sem);
 730         atomic_inc(&caching_ctl->count);
 731         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 732         up_write(&fs_info->commit_root_sem);
 733
 734         btrfs_get_block_group(cache);
 735
 736         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 737
 738         return ret;
 739 }
 740
 741 /*
 742  * return the block group that starts at or after bytenr
 743  */
 744 static struct btrfs_block_group_cache *
 745 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 746 {
 747         struct btrfs_block_group_cache *cache;
 748
 749         cache = block_group_cache_tree_search(info, bytenr, 0);
 750
 751         return cache;
 752 }
 753
 754 /*
 755  * return the block group that contains the given bytenr
 756  */
 757 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 758                                                  struct btrfs_fs_info *info,
 759                                                  u64 bytenr)
 760 {
 761         struct btrfs_block_group_cache *cache;
 762
 763         cache = block_group_cache_tree_search(info, bytenr, 1);
 764
 765         return cache;
 766 }
 767
 768 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 769                                                   u64 flags)
 770 {
 771         struct list_head *head = &info->space_info;
 772         struct btrfs_space_info *found;
 773
 774         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 775
 776         rcu_read_lock();
 777         list_for_each_entry_rcu(found, head, list) {
 778                 if (found->flags & flags) {
 779                         rcu_read_unlock();
 780                         return found;
 781                 }
 782         }
 783         rcu_read_unlock();
 784         return NULL;
 785 }
 786
 787 /*
 788  * after adding space to the filesystem, we need to clear the full flags
 789  * on all the space infos.
 790  */
 791 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 792 {
 793         struct list_head *head = &info->space_info;
 794         struct btrfs_space_info *found;
 795
 796         rcu_read_lock();
 797         list_for_each_entry_rcu(found, head, list)
 798                 found->full = 0;
 799         rcu_read_unlock();
 800 }
 801
 802 /* simple helper to search for an existing data extent at a given offset */
 803 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 804 {
 805         int ret;
 806         struct btrfs_key key;
 807         struct btrfs_path *path;
 808
 809         path = btrfs_alloc_path();
 810         if (!path)
 811                 return -ENOMEM;
 812
 813         key.objectid = start;
 814         key.offset = len;
 815         key.type = BTRFS_EXTENT_ITEM_KEY;
 816         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 817                                 0, 0);
 818         btrfs_free_path(path);
 819         return ret;
 820 }
 821
 822 /*
 823  * helper function to lookup reference count and flags of a tree block.
 824  *
 825  * the head node for delayed ref is used to store the sum of all the
 826  * reference count modifications queued up in the rbtree. the head
 827  * node may also store the extent flags to set. This way you can check
 828  * to see what the reference count and extent flags would be if all of
 829  * the delayed refs are not processed.
 830  */
 831 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 832                              struct btrfs_root *root, u64 bytenr,
 833                              u64 offset, int metadata, u64 *refs, u64 *flags)
 834 {
 835         struct btrfs_delayed_ref_head *head;
 836         struct btrfs_delayed_ref_root *delayed_refs;
 837         struct btrfs_path *path;
 838         struct btrfs_extent_item *ei;
 839         struct extent_buffer *leaf;
 840         struct btrfs_key key;
 841         u32 item_size;
 842         u64 num_refs;
 843         u64 extent_flags;
 844         int ret;
 845
 846         /*
 847          * If we don't have skinny metadata, don't bother doing anything
 848          * different
 849          */
 850         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 851                 offset = root->nodesize;
 852                 metadata = 0;
 853         }
 854
 855         path = btrfs_alloc_path();
 856         if (!path)
 857                 return -ENOMEM;
 858
 859         if (!trans) {
 860                 path->skip_locking = 1;
 861                 path->search_commit_root = 1;
 862         }
 863
 864 search_again:
 865         key.objectid = bytenr;
 866         key.offset = offset;
 867         if (metadata)
 868                 key.type = BTRFS_METADATA_ITEM_KEY;
 869         else
 870                 key.type = BTRFS_EXTENT_ITEM_KEY;
 871
 872         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 873                                 &key, path, 0, 0);
 874         if (ret < 0)
 875                 goto out_free;
 876
 877         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 878                 if (path->slots[0]) {
 879                         path->slots[0]--;
 880                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 881                                               path->slots[0]);
 882                         if (key.objectid == bytenr &&
 883                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 884                             key.offset == root->nodesize)
 885                                 ret = 0;
 886                 }
 887         }
 888
 889         if (ret == 0) {
 890                 leaf = path->nodes[0];
 891                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 892                 if (item_size >= sizeof(*ei)) {
 893                         ei = btrfs_item_ptr(leaf, path->slots[0],
 894                                             struct btrfs_extent_item);
 895                         num_refs = btrfs_extent_refs(leaf, ei);
 896                         extent_flags = btrfs_extent_flags(leaf, ei);
 897                 } else {
 898 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 899                         struct btrfs_extent_item_v0 *ei0;
 900                         BUG_ON(item_size != sizeof(*ei0));
 901                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 902                                              struct btrfs_extent_item_v0);
 903                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 904                         /* FIXME: this isn't correct for data */
 905                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 906 #else
 907                         BUG();
 908 #endif
 909                 }
 910                 BUG_ON(num_refs == 0);
 911         } else {
 912                 num_refs = 0;
 913                 extent_flags = 0;
 914                 ret = 0;
 915         }
 916
 917         if (!trans)
 918                 goto out;
 919
 920         delayed_refs = &trans->transaction->delayed_refs;
 921         spin_lock(&delayed_refs->lock);
 922         head = btrfs_find_delayed_ref_head(trans, bytenr);
 923         if (head) {
 924                 if (!mutex_trylock(&head->mutex)) {
 925                         atomic_inc(&head->node.refs);
 926                         spin_unlock(&delayed_refs->lock);
 927
 928                         btrfs_release_path(path);
 929
 930                         /*
 931                          * Mutex was contended, block until it's released and try
 932                          * again
 933                          */
 934                         mutex_lock(&head->mutex);
 935                         mutex_unlock(&head->mutex);
 936                         btrfs_put_delayed_ref(&head->node);
 937                         goto search_again;
 938                 }
 939                 spin_lock(&head->lock);
 940                 if (head->extent_op && head->extent_op->update_flags)
 941                         extent_flags |= head->extent_op->flags_to_set;
 942                 else
 943                         BUG_ON(num_refs == 0);
 944
 945                 num_refs += head->node.ref_mod;
 946                 spin_unlock(&head->lock);
 947                 mutex_unlock(&head->mutex);
 948         }
 949         spin_unlock(&delayed_refs->lock);
 950 out:
 951         WARN_ON(num_refs == 0);
 952         if (refs)
 953                 *refs = num_refs;
 954         if (flags)
 955                 *flags = extent_flags;
 956 out_free:
 957         btrfs_free_path(path);
 958         return ret;
 959 }
 960
 961 /*
 962  * Back reference rules.  Back refs have three main goals:
 963  *
 964  * 1) differentiate between all holders of references to an extent so that
 965  *    when a reference is dropped we can make sure it was a valid reference
 966  *    before freeing the extent.
 967  *
 968  * 2) Provide enough information to quickly find the holders of an extent
 969  *    if we notice a given block is corrupted or bad.
 970  *
 971  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 972  *    maintenance.  This is actually the same as #2, but with a slightly
 973  *    different use case.
 974  *
 975  * There are two kinds of back refs. The implicit back refs is optimized
 976  * for pointers in non-shared tree blocks. For a given pointer in a block,
 977  * back refs of this kind provide information about the block's owner tree
 978  * and the pointer's key. These information allow us to find the block by
 979  * b-tree searching. The full back refs is for pointers in tree blocks not
 980  * referenced by their owner trees. The location of tree block is recorded
 981  * in the back refs. Actually the full back refs is generic, and can be
 982  * used in all cases the implicit back refs is used. The major shortcoming
 983  * of the full back refs is its overhead. Every time a tree block gets
 984  * COWed, we have to update back refs entry for all pointers in it.
 985  *
 986  * For a newly allocated tree block, we use implicit back refs for
 987  * pointers in it. This means most tree related operations only involve
 988  * implicit back refs. For a tree block created in old transaction, the
 989  * only way to drop a reference to it is COW it. So we can detect the
 990  * event that tree block loses its owner tree's reference and do the
 991  * back refs conversion.
 992  *
 993  * When a tree block is COWed through a tree, there are four cases:
 994  *
 995  * The reference count of the block is one and the tree is the block's
 996  * owner tree. Nothing to do in this case.
 997  *
 998  * The reference count of the block is one and the tree is not the
 999  * block's owner tree. In this case, full back refs is used for pointers
1000  * in the block. Remove these full back refs, add implicit back refs for
1001  * every pointers in the new block.
1002  *
1003  * The reference count of the block is greater than one and the tree is
1004  * the block's owner tree. In this case, implicit back refs is used for
1005  * pointers in the block. Add full back refs for every pointers in the
1006  * block, increase lower level extents' reference counts. The original
1007  * implicit back refs are entailed to the new block.
1008  *
1009  * The reference count of the block is greater than one and the tree is
1010  * not the block's owner tree. Add implicit back refs for every pointer in
1011  * the new block, increase lower level extents' reference count.
1012  *
1013  * Back Reference Key composing:
1014  *
1015  * The key objectid corresponds to the first byte in the extent,
1016  * The key type is used to differentiate between types of back refs.
1017  * There are different meanings of the key offset for different types
1018  * of back refs.
1019  *
1020  * File extents can be referenced by:
1021  *
1022  * - multiple snapshots, subvolumes, or different generations in one subvol
1023  * - different files inside a single subvolume
1024  * - different offsets inside a file (bookend extents in file.c)
1025  *
1026  * The extent ref structure for the implicit back refs has fields for:
1027  *
1028  * - Objectid of the subvolume root
1029  * - objectid of the file holding the reference
1030  * - original offset in the file
1031  * - how many bookend extents
1032  *
1033  * The key offset for the implicit back refs is hash of the first
1034  * three fields.
1035  *
1036  * The extent ref structure for the full back refs has field for:
1037  *
1038  * - number of pointers in the tree leaf
1039  *
1040  * The key offset for the implicit back refs is the first byte of
1041  * the tree leaf
1042  *
1043  * When a file extent is allocated, The implicit back refs is used.
1044  * the fields are filled in:
1045  *
1046  *     (root_key.objectid, inode objectid, offset in file, 1)
1047  *
1048  * When a file extent is removed file truncation, we find the
1049  * corresponding implicit back refs and check the following fields:
1050  *
1051  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1052  *
1053  * Btree extents can be referenced by:
1054  *
1055  * - Different subvolumes
1056  *
1057  * Both the implicit back refs and the full back refs for tree blocks
1058  * only consist of key. The key offset for the implicit back refs is
1059  * objectid of block's owner tree. The key offset for the full back refs
1060  * is the first byte of parent block.
1061  *
1062  * When implicit back refs is used, information about the lowest key and
1063  * level of the tree block are required. These information are stored in
1064  * tree block info structure.
1065  */
1066
1067 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1068 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1069                                   struct btrfs_root *root,
1070                                   struct btrfs_path *path,
1071                                   u64 owner, u32 extra_size)
1072 {
1073         struct btrfs_extent_item *item;
1074         struct btrfs_extent_item_v0 *ei0;
1075         struct btrfs_extent_ref_v0 *ref0;
1076         struct btrfs_tree_block_info *bi;
1077         struct extent_buffer *leaf;
1078         struct btrfs_key key;
1079         struct btrfs_key found_key;
1080         u32 new_size = sizeof(*item);
1081         u64 refs;
1082         int ret;
1083
1084         leaf = path->nodes[0];
1085         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1086
1087         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1088         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1089                              struct btrfs_extent_item_v0);
1090         refs = btrfs_extent_refs_v0(leaf, ei0);
1091
1092         if (owner == (u64)-1) {
1093                 while (1) {
1094                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1095                                 ret = btrfs_next_leaf(root, path);
1096                                 if (ret < 0)
1097                                         return ret;
1098                                 BUG_ON(ret > 0); /* Corruption */
1099                                 leaf = path->nodes[0];
1100                         }
1101                         btrfs_item_key_to_cpu(leaf, &found_key,
1102                                               path->slots[0]);
1103                         BUG_ON(key.objectid != found_key.objectid);
1104                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1105                                 path->slots[0]++;
1106                                 continue;
1107                         }
1108                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1109                                               struct btrfs_extent_ref_v0);
1110                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1111                         break;
1112                 }
1113         }
1114         btrfs_release_path(path);
1115
1116         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1117                 new_size += sizeof(*bi);
1118
1119         new_size -= sizeof(*ei0);
1120         ret = btrfs_search_slot(trans, root, &key, path,
1121                                 new_size + extra_size, 1);
1122         if (ret < 0)
1123                 return ret;
1124         BUG_ON(ret); /* Corruption */
1125
1126         btrfs_extend_item(root, path, new_size);
1127
1128         leaf = path->nodes[0];
1129         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1130         btrfs_set_extent_refs(leaf, item, refs);
1131         /* FIXME: get real generation */
1132         btrfs_set_extent_generation(leaf, item, 0);
1133         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1134                 btrfs_set_extent_flags(leaf, item,
1135                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1136                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1137                 bi = (struct btrfs_tree_block_info *)(item + 1);
1138                 /* FIXME: get first key of the block */
1139                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1140                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1141         } else {
1142                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1143         }
1144         btrfs_mark_buffer_dirty(leaf);
1145         return 0;
1146 }
1147 #endif
1148
1149 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1150 {
1151         u32 high_crc = ~(u32)0;
1152         u32 low_crc = ~(u32)0;
1153         __le64 lenum;
1154
1155         lenum = cpu_to_le64(root_objectid);
1156         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1157         lenum = cpu_to_le64(owner);
1158         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1159         lenum = cpu_to_le64(offset);
1160         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1161
1162         return ((u64)high_crc << 31) ^ (u64)low_crc;
1163 }
1164
1165 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1166                                      struct btrfs_extent_data_ref *ref)
1167 {
1168         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1169                                     btrfs_extent_data_ref_objectid(leaf, ref),
1170                                     btrfs_extent_data_ref_offset(leaf, ref));
1171 }
1172
1173 static int match_extent_data_ref(struct extent_buffer *leaf,
1174                                  struct btrfs_extent_data_ref *ref,
1175                                  u64 root_objectid, u64 owner, u64 offset)
1176 {
1177         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1178             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1179             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1180                 return 0;
1181         return 1;
1182 }
1183
1184 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1185                                            struct btrfs_root *root,
1186                                            struct btrfs_path *path,
1187                                            u64 bytenr, u64 parent,
1188                                            u64 root_objectid,
1189                                            u64 owner, u64 offset)
1190 {
1191         struct btrfs_key key;
1192         struct btrfs_extent_data_ref *ref;
1193         struct extent_buffer *leaf;
1194         u32 nritems;
1195         int ret;
1196         int recow;
1197         int err = -ENOENT;
1198
1199         key.objectid = bytenr;
1200         if (parent) {
1201                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1202                 key.offset = parent;
1203         } else {
1204                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1205                 key.offset = hash_extent_data_ref(root_objectid,
1206                                                   owner, offset);
1207         }
1208 again:
1209         recow = 0;
1210         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1211         if (ret < 0) {
1212                 err = ret;
1213                 goto fail;
1214         }
1215
1216         if (parent) {
1217                 if (!ret)
1218                         return 0;
1219 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1220                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1221                 btrfs_release_path(path);
1222                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1223                 if (ret < 0) {
1224                         err = ret;
1225                         goto fail;
1226                 }
1227                 if (!ret)
1228                         return 0;
1229 #endif
1230                 goto fail;
1231         }
1232
1233         leaf = path->nodes[0];
1234         nritems = btrfs_header_nritems(leaf);
1235         while (1) {
1236                 if (path->slots[0] >= nritems) {
1237                         ret = btrfs_next_leaf(root, path);
1238                         if (ret < 0)
1239                                 err = ret;
1240                         if (ret)
1241                                 goto fail;
1242
1243                         leaf = path->nodes[0];
1244                         nritems = btrfs_header_nritems(leaf);
1245                         recow = 1;
1246                 }
1247
1248                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1249                 if (key.objectid != bytenr ||
1250                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1251                         goto fail;
1252
1253                 ref = btrfs_item_ptr(leaf, path->slots[0],
1254                                      struct btrfs_extent_data_ref);
1255
1256                 if (match_extent_data_ref(leaf, ref, root_objectid,
1257                                           owner, offset)) {
1258                         if (recow) {
1259                                 btrfs_release_path(path);
1260                                 goto again;
1261                         }
1262                         err = 0;
1263                         break;
1264                 }
1265                 path->slots[0]++;
1266         }
1267 fail:
1268         return err;
1269 }
1270
1271 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1272                                            struct btrfs_root *root,
1273                                            struct btrfs_path *path,
1274                                            u64 bytenr, u64 parent,
1275                                            u64 root_objectid, u64 owner,
1276                                            u64 offset, int refs_to_add)
1277 {
1278         struct btrfs_key key;
1279         struct extent_buffer *leaf;
1280         u32 size;
1281         u32 num_refs;
1282         int ret;
1283
1284         key.objectid = bytenr;
1285         if (parent) {
1286                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1287                 key.offset = parent;
1288                 size = sizeof(struct btrfs_shared_data_ref);
1289         } else {
1290                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1291                 key.offset = hash_extent_data_ref(root_objectid,
1292                                                   owner, offset);
1293                 size = sizeof(struct btrfs_extent_data_ref);
1294         }
1295
1296         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1297         if (ret && ret != -EEXIST)
1298                 goto fail;
1299
1300         leaf = path->nodes[0];
1301         if (parent) {
1302                 struct btrfs_shared_data_ref *ref;
1303                 ref = btrfs_item_ptr(leaf, path->slots[0],
1304                                      struct btrfs_shared_data_ref);
1305                 if (ret == 0) {
1306                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1307                 } else {
1308                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1309                         num_refs += refs_to_add;
1310                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1311                 }
1312         } else {
1313                 struct btrfs_extent_data_ref *ref;
1314                 while (ret == -EEXIST) {
1315                         ref = btrfs_item_ptr(leaf, path->slots[0],
1316                                              struct btrfs_extent_data_ref);
1317                         if (match_extent_data_ref(leaf, ref, root_objectid,
1318                                                   owner, offset))
1319                                 break;
1320                         btrfs_release_path(path);
1321                         key.offset++;
1322                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1323                                                       size);
1324                         if (ret && ret != -EEXIST)
1325                                 goto fail;
1326
1327                         leaf = path->nodes[0];
1328                 }
1329                 ref = btrfs_item_ptr(leaf, path->slots[0],
1330                                      struct btrfs_extent_data_ref);
1331                 if (ret == 0) {
1332                         btrfs_set_extent_data_ref_root(leaf, ref,
1333                                                        root_objectid);
1334                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1335                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1336                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1337                 } else {
1338                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1339                         num_refs += refs_to_add;
1340                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1341                 }
1342         }
1343         btrfs_mark_buffer_dirty(leaf);
1344         ret = 0;
1345 fail:
1346         btrfs_release_path(path);
1347         return ret;
1348 }
1349
1350 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1351                                            struct btrfs_root *root,
1352                                            struct btrfs_path *path,
1353                                            int refs_to_drop, int *last_ref)
1354 {
1355         struct btrfs_key key;
1356         struct btrfs_extent_data_ref *ref1 = NULL;
1357         struct btrfs_shared_data_ref *ref2 = NULL;
1358         struct extent_buffer *leaf;
1359         u32 num_refs = 0;
1360         int ret = 0;
1361
1362         leaf = path->nodes[0];
1363         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1364
1365         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1366                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1367                                       struct btrfs_extent_data_ref);
1368                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1369         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1370                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1371                                       struct btrfs_shared_data_ref);
1372                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1373 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1374         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1375                 struct btrfs_extent_ref_v0 *ref0;
1376                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1377                                       struct btrfs_extent_ref_v0);
1378                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1379 #endif
1380         } else {
1381                 BUG();
1382         }
1383
1384         BUG_ON(num_refs < refs_to_drop);
1385         num_refs -= refs_to_drop;
1386
1387         if (num_refs == 0) {
1388                 ret = btrfs_del_item(trans, root, path);
1389                 *last_ref = 1;
1390         } else {
1391                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1392                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1393                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1394                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1395 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1396                 else {
1397                         struct btrfs_extent_ref_v0 *ref0;
1398                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1399                                         struct btrfs_extent_ref_v0);
1400                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1401                 }
1402 #endif
1403                 btrfs_mark_buffer_dirty(leaf);
1404         }
1405         return ret;
1406 }
1407
1408 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1409                                           struct btrfs_extent_inline_ref *iref)
1410 {
1411         struct btrfs_key key;
1412         struct extent_buffer *leaf;
1413         struct btrfs_extent_data_ref *ref1;
1414         struct btrfs_shared_data_ref *ref2;
1415         u32 num_refs = 0;
1416
1417         leaf = path->nodes[0];
1418         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1419         if (iref) {
1420                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1421                     BTRFS_EXTENT_DATA_REF_KEY) {
1422                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1423                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1424                 } else {
1425                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1426                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1427                 }
1428         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1429                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1430                                       struct btrfs_extent_data_ref);
1431                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1432         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1433                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1434                                       struct btrfs_shared_data_ref);
1435                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1436 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1437         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1438                 struct btrfs_extent_ref_v0 *ref0;
1439                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1440                                       struct btrfs_extent_ref_v0);
1441                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1442 #endif
1443         } else {
1444                 WARN_ON(1);
1445         }
1446         return num_refs;
1447 }
1448
1449 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1450                                           struct btrfs_root *root,
1451                                           struct btrfs_path *path,
1452                                           u64 bytenr, u64 parent,
1453                                           u64 root_objectid)
1454 {
1455         struct btrfs_key key;
1456         int ret;
1457
1458         key.objectid = bytenr;
1459         if (parent) {
1460                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1461                 key.offset = parent;
1462         } else {
1463                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1464                 key.offset = root_objectid;
1465         }
1466
1467         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1468         if (ret > 0)
1469                 ret = -ENOENT;
1470 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1471         if (ret == -ENOENT && parent) {
1472                 btrfs_release_path(path);
1473                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1474                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1475                 if (ret > 0)
1476                         ret = -ENOENT;
1477         }
1478 #endif
1479         return ret;
1480 }
1481
1482 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1483                                           struct btrfs_root *root,
1484                                           struct btrfs_path *path,
1485                                           u64 bytenr, u64 parent,
1486                                           u64 root_objectid)
1487 {
1488         struct btrfs_key key;
1489         int ret;
1490
1491         key.objectid = bytenr;
1492         if (parent) {
1493                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1494                 key.offset = parent;
1495         } else {
1496                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1497                 key.offset = root_objectid;
1498         }
1499
1500         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1501         btrfs_release_path(path);
1502         return ret;
1503 }
1504
1505 static inline int extent_ref_type(u64 parent, u64 owner)
1506 {
1507         int type;
1508         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1509                 if (parent > 0)
1510                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1511                 else
1512                         type = BTRFS_TREE_BLOCK_REF_KEY;
1513         } else {
1514                 if (parent > 0)
1515                         type = BTRFS_SHARED_DATA_REF_KEY;
1516                 else
1517                         type = BTRFS_EXTENT_DATA_REF_KEY;
1518         }
1519         return type;
1520 }
1521
1522 static int find_next_key(struct btrfs_path *path, int level,
1523                          struct btrfs_key *key)
1524
1525 {
1526         for (; level < BTRFS_MAX_LEVEL; level++) {
1527                 if (!path->nodes[level])
1528                         break;
1529                 if (path->slots[level] + 1 >=
1530                     btrfs_header_nritems(path->nodes[level]))
1531                         continue;
1532                 if (level == 0)
1533                         btrfs_item_key_to_cpu(path->nodes[level], key,
1534                                               path->slots[level] + 1);
1535                 else
1536                         btrfs_node_key_to_cpu(path->nodes[level], key,
1537                                               path->slots[level] + 1);
1538                 return 0;
1539         }
1540         return 1;
1541 }
1542
1543 /*
1544  * look for inline back ref. if back ref is found, *ref_ret is set
1545  * to the address of inline back ref, and 0 is returned.
1546  *
1547  * if back ref isn't found, *ref_ret is set to the address where it
1548  * should be inserted, and -ENOENT is returned.
1549  *
1550  * if insert is true and there are too many inline back refs, the path
1551  * points to the extent item, and -EAGAIN is returned.
1552  *
1553  * NOTE: inline back refs are ordered in the same way that back ref
1554  *       items in the tree are ordered.
1555  */
1556 static noinline_for_stack
1557 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1558                                  struct btrfs_root *root,
1559                                  struct btrfs_path *path,
1560                                  struct btrfs_extent_inline_ref **ref_ret,
1561                                  u64 bytenr, u64 num_bytes,
1562                                  u64 parent, u64 root_objectid,
1563                                  u64 owner, u64 offset, int insert)
1564 {
1565         struct btrfs_key key;
1566         struct extent_buffer *leaf;
1567         struct btrfs_extent_item *ei;
1568         struct btrfs_extent_inline_ref *iref;
1569         u64 flags;
1570         u64 item_size;
1571         unsigned long ptr;
1572         unsigned long end;
1573         int extra_size;
1574         int type;
1575         int want;
1576         int ret;
1577         int err = 0;
1578         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1579                                                  SKINNY_METADATA);
1580
1581         key.objectid = bytenr;
1582         key.type = BTRFS_EXTENT_ITEM_KEY;
1583         key.offset = num_bytes;
1584
1585         want = extent_ref_type(parent, owner);
1586         if (insert) {
1587                 extra_size = btrfs_extent_inline_ref_size(want);
1588                 path->keep_locks = 1;
1589         } else
1590                 extra_size = -1;
1591
1592         /*
1593          * Owner is our parent level, so we can just add one to get the level
1594          * for the block we are interested in.
1595          */
1596         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1597                 key.type = BTRFS_METADATA_ITEM_KEY;
1598                 key.offset = owner;
1599         }
1600
1601 again:
1602         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1603         if (ret < 0) {
1604                 err = ret;
1605                 goto out;
1606         }
1607
1608         /*
1609          * We may be a newly converted file system which still has the old fat
1610          * extent entries for metadata, so try and see if we have one of those.
1611          */
1612         if (ret > 0 && skinny_metadata) {
1613                 skinny_metadata = false;
1614                 if (path->slots[0]) {
1615                         path->slots[0]--;
1616                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1617                                               path->slots[0]);
1618                         if (key.objectid == bytenr &&
1619                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1620                             key.offset == num_bytes)
1621                                 ret = 0;
1622                 }
1623                 if (ret) {
1624                         key.objectid = bytenr;
1625                         key.type = BTRFS_EXTENT_ITEM_KEY;
1626                         key.offset = num_bytes;
1627                         btrfs_release_path(path);
1628                         goto again;
1629                 }
1630         }
1631
1632         if (ret && !insert) {
1633                 err = -ENOENT;
1634                 goto out;
1635         } else if (WARN_ON(ret)) {
1636                 err = -EIO;
1637                 goto out;
1638         }
1639
1640         leaf = path->nodes[0];
1641         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1642 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1643         if (item_size < sizeof(*ei)) {
1644                 if (!insert) {
1645                         err = -ENOENT;
1646                         goto out;
1647                 }
1648                 ret = convert_extent_item_v0(trans, root, path, owner,
1649                                              extra_size);
1650                 if (ret < 0) {
1651                         err = ret;
1652                         goto out;
1653                 }
1654                 leaf = path->nodes[0];
1655                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1656         }
1657 #endif
1658         BUG_ON(item_size < sizeof(*ei));
1659
1660         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1661         flags = btrfs_extent_flags(leaf, ei);
1662
1663         ptr = (unsigned long)(ei + 1);
1664         end = (unsigned long)ei + item_size;
1665
1666         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1667                 ptr += sizeof(struct btrfs_tree_block_info);
1668                 BUG_ON(ptr > end);
1669         }
1670
1671         err = -ENOENT;
1672         while (1) {
1673                 if (ptr >= end) {
1674                         WARN_ON(ptr > end);
1675                         break;
1676                 }
1677                 iref = (struct btrfs_extent_inline_ref *)ptr;
1678                 type = btrfs_extent_inline_ref_type(leaf, iref);
1679                 if (want < type)
1680                         break;
1681                 if (want > type) {
1682                         ptr += btrfs_extent_inline_ref_size(type);
1683                         continue;
1684                 }
1685
1686                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1687                         struct btrfs_extent_data_ref *dref;
1688                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1689                         if (match_extent_data_ref(leaf, dref, root_objectid,
1690                                                   owner, offset)) {
1691                                 err = 0;
1692                                 break;
1693                         }
1694                         if (hash_extent_data_ref_item(leaf, dref) <
1695                             hash_extent_data_ref(root_objectid, owner, offset))
1696                                 break;
1697                 } else {
1698                         u64 ref_offset;
1699                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1700                         if (parent > 0) {
1701                                 if (parent == ref_offset) {
1702                                         err = 0;
1703                                         break;
1704                                 }
1705                                 if (ref_offset < parent)
1706                                         break;
1707                         } else {
1708                                 if (root_objectid == ref_offset) {
1709                                         err = 0;
1710                                         break;
1711                                 }
1712                                 if (ref_offset < root_objectid)
1713                                         break;
1714                         }
1715                 }
1716                 ptr += btrfs_extent_inline_ref_size(type);
1717         }
1718         if (err == -ENOENT && insert) {
1719                 if (item_size + extra_size >=
1720                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1721                         err = -EAGAIN;
1722                         goto out;
1723                 }
1724                 /*
1725                  * To add new inline back ref, we have to make sure
1726                  * there is no corresponding back ref item.
1727                  * For simplicity, we just do not add new inline back
1728                  * ref if there is any kind of item for this block
1729                  */
1730                 if (find_next_key(path, 0, &key) == 0 &&
1731                     key.objectid == bytenr &&
1732                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1733                         err = -EAGAIN;
1734                         goto out;
1735                 }
1736         }
1737         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1738 out:
1739         if (insert) {
1740                 path->keep_locks = 0;
1741                 btrfs_unlock_up_safe(path, 1);
1742         }
1743         return err;
1744 }
1745
1746 /*
1747  * helper to add new inline back ref
1748  */
1749 static noinline_for_stack
1750 void setup_inline_extent_backref(struct btrfs_root *root,
1751                                  struct btrfs_path *path,
1752                                  struct btrfs_extent_inline_ref *iref,
1753                                  u64 parent, u64 root_objectid,
1754                                  u64 owner, u64 offset, int refs_to_add,
1755                                  struct btrfs_delayed_extent_op *extent_op)
1756 {
1757         struct extent_buffer *leaf;
1758         struct btrfs_extent_item *ei;
1759         unsigned long ptr;
1760         unsigned long end;
1761         unsigned long item_offset;
1762         u64 refs;
1763         int size;
1764         int type;
1765
1766         leaf = path->nodes[0];
1767         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1768         item_offset = (unsigned long)iref - (unsigned long)ei;
1769
1770         type = extent_ref_type(parent, owner);
1771         size = btrfs_extent_inline_ref_size(type);
1772
1773         btrfs_extend_item(root, path, size);
1774
1775         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1776         refs = btrfs_extent_refs(leaf, ei);
1777         refs += refs_to_add;
1778         btrfs_set_extent_refs(leaf, ei, refs);
1779         if (extent_op)
1780                 __run_delayed_extent_op(extent_op, leaf, ei);
1781
1782         ptr = (unsigned long)ei + item_offset;
1783         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1784         if (ptr < end - size)
1785                 memmove_extent_buffer(leaf, ptr + size, ptr,
1786                                       end - size - ptr);
1787
1788         iref = (struct btrfs_extent_inline_ref *)ptr;
1789         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1790         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1791                 struct btrfs_extent_data_ref *dref;
1792                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1793                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1794                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1795                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1796                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1797         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1798                 struct btrfs_shared_data_ref *sref;
1799                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1800                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1801                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1802         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1803                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1804         } else {
1805                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1806         }
1807         btrfs_mark_buffer_dirty(leaf);
1808 }
1809
1810 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1811                                  struct btrfs_root *root,
1812                                  struct btrfs_path *path,
1813                                  struct btrfs_extent_inline_ref **ref_ret,
1814                                  u64 bytenr, u64 num_bytes, u64 parent,
1815                                  u64 root_objectid, u64 owner, u64 offset)
1816 {
1817         int ret;
1818
1819         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1820                                            bytenr, num_bytes, parent,
1821                                            root_objectid, owner, offset, 0);
1822         if (ret != -ENOENT)
1823                 return ret;
1824
1825         btrfs_release_path(path);
1826         *ref_ret = NULL;
1827
1828         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1829                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1830                                             root_objectid);
1831         } else {
1832                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1833                                              root_objectid, owner, offset);
1834         }
1835         return ret;
1836 }
1837
1838 /*
1839  * helper to update/remove inline back ref
1840  */
1841 static noinline_for_stack
1842 void update_inline_extent_backref(struct btrfs_root *root,
1843                                   struct btrfs_path *path,
1844                                   struct btrfs_extent_inline_ref *iref,
1845                                   int refs_to_mod,
1846                                   struct btrfs_delayed_extent_op *extent_op,
1847                                   int *last_ref)
1848 {
1849         struct extent_buffer *leaf;
1850         struct btrfs_extent_item *ei;
1851         struct btrfs_extent_data_ref *dref = NULL;
1852         struct btrfs_shared_data_ref *sref = NULL;
1853         unsigned long ptr;
1854         unsigned long end;
1855         u32 item_size;
1856         int size;
1857         int type;
1858         u64 refs;
1859
1860         leaf = path->nodes[0];
1861         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1862         refs = btrfs_extent_refs(leaf, ei);
1863         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1864         refs += refs_to_mod;
1865         btrfs_set_extent_refs(leaf, ei, refs);
1866         if (extent_op)
1867                 __run_delayed_extent_op(extent_op, leaf, ei);
1868
1869         type = btrfs_extent_inline_ref_type(leaf, iref);
1870
1871         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1872                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1873                 refs = btrfs_extent_data_ref_count(leaf, dref);
1874         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1875                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1876                 refs = btrfs_shared_data_ref_count(leaf, sref);
1877         } else {
1878                 refs = 1;
1879                 BUG_ON(refs_to_mod != -1);
1880         }
1881
1882         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1883         refs += refs_to_mod;
1884
1885         if (refs > 0) {
1886                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1887                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1888                 else
1889                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1890         } else {
1891                 *last_ref = 1;
1892                 size =  btrfs_extent_inline_ref_size(type);
1893                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1894                 ptr = (unsigned long)iref;
1895                 end = (unsigned long)ei + item_size;
1896                 if (ptr + size < end)
1897                         memmove_extent_buffer(leaf, ptr, ptr + size,
1898                                               end - ptr - size);
1899                 item_size -= size;
1900                 btrfs_truncate_item(root, path, item_size, 1);
1901         }
1902         btrfs_mark_buffer_dirty(leaf);
1903 }
1904
1905 static noinline_for_stack
1906 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1907                                  struct btrfs_root *root,
1908                                  struct btrfs_path *path,
1909                                  u64 bytenr, u64 num_bytes, u64 parent,
1910                                  u64 root_objectid, u64 owner,
1911                                  u64 offset, int refs_to_add,
1912                                  struct btrfs_delayed_extent_op *extent_op)
1913 {
1914         struct btrfs_extent_inline_ref *iref;
1915         int ret;
1916
1917         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1918                                            bytenr, num_bytes, parent,
1919                                            root_objectid, owner, offset, 1);
1920         if (ret == 0) {
1921                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1922                 update_inline_extent_backref(root, path, iref,
1923                                              refs_to_add, extent_op, NULL);
1924         } else if (ret == -ENOENT) {
1925                 setup_inline_extent_backref(root, path, iref, parent,
1926                                             root_objectid, owner, offset,
1927                                             refs_to_add, extent_op);
1928                 ret = 0;
1929         }
1930         return ret;
1931 }
1932
1933 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1934                                  struct btrfs_root *root,
1935                                  struct btrfs_path *path,
1936                                  u64 bytenr, u64 parent, u64 root_objectid,
1937                                  u64 owner, u64 offset, int refs_to_add)
1938 {
1939         int ret;
1940         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1941                 BUG_ON(refs_to_add != 1);
1942                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1943                                             parent, root_objectid);
1944         } else {
1945                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1946                                              parent, root_objectid,
1947                                              owner, offset, refs_to_add);
1948         }
1949         return ret;
1950 }
1951
1952 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1953                                  struct btrfs_root *root,
1954                                  struct btrfs_path *path,
1955                                  struct btrfs_extent_inline_ref *iref,
1956                                  int refs_to_drop, int is_data, int *last_ref)
1957 {
1958         int ret = 0;
1959
1960         BUG_ON(!is_data && refs_to_drop != 1);
1961         if (iref) {
1962                 update_inline_extent_backref(root, path, iref,
1963                                              -refs_to_drop, NULL, last_ref);
1964         } else if (is_data) {
1965                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1966                                              last_ref);
1967         } else {
1968                 *last_ref = 1;
1969                 ret = btrfs_del_item(trans, root, path);
1970         }
1971         return ret;
1972 }
1973
1974 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1975 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1976                                u64 *discarded_bytes)
1977 {
1978         int j, ret = 0;
1979         u64 bytes_left, end;
1980         u64 aligned_start = ALIGN(start, 1 << 9);
1981
1982         if (WARN_ON(start != aligned_start)) {
1983                 len -= aligned_start - start;
1984                 len = round_down(len, 1 << 9);
1985                 start = aligned_start;
1986         }
1987
1988         *discarded_bytes = 0;
1989
1990         if (!len)
1991                 return 0;
1992
1993         end = start + len;
1994         bytes_left = len;
1995
1996         /* Skip any superblocks on this device. */
1997         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1998                 u64 sb_start = btrfs_sb_offset(j);
1999                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
2000                 u64 size = sb_start - start;
2001
2002                 if (!in_range(sb_start, start, bytes_left) &&
2003                     !in_range(sb_end, start, bytes_left) &&
2004                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
2005                         continue;
2006
2007                 /*
2008                  * Superblock spans beginning of range.  Adjust start and
2009                  * try again.
2010                  */
2011                 if (sb_start <= start) {
2012                         start += sb_end - start;
2013                         if (start > end) {
2014                                 bytes_left = 0;
2015                                 break;
2016                         }
2017                         bytes_left = end - start;
2018                         continue;
2019                 }
2020
2021                 if (size) {
2022                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2023                                                    GFP_NOFS, 0);
2024                         if (!ret)
2025                                 *discarded_bytes += size;
2026                         else if (ret != -EOPNOTSUPP)
2027                                 return ret;
2028                 }
2029
2030                 start = sb_end;
2031                 if (start > end) {
2032                         bytes_left = 0;
2033                         break;
2034                 }
2035                 bytes_left = end - start;
2036         }
2037
2038         if (bytes_left) {
2039                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2040                                            GFP_NOFS, 0);
2041                 if (!ret)
2042                         *discarded_bytes += bytes_left;
2043         }
2044         return ret;
2045 }
2046
2047 int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
2048                          u64 num_bytes, u64 *actual_bytes)
2049 {
2050         int ret;
2051         u64 discarded_bytes = 0;
2052         struct btrfs_bio *bbio = NULL;
2053
2054
2055         /*
2056          * Avoid races with device replace and make sure our bbio has devices
2057          * associated to its stripes that don't go away while we are discarding.
2058          */
2059         btrfs_bio_counter_inc_blocked(root->fs_info);
2060         /* Tell the block device(s) that the sectors can be discarded */
2061         ret = btrfs_map_block(root->fs_info, REQ_OP_DISCARD,
2062                               bytenr, &num_bytes, &bbio, 0);
2063         /* Error condition is -ENOMEM */
2064         if (!ret) {
2065                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2066                 int i;
2067
2068
2069                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2070                         u64 bytes;
2071                         if (!stripe->dev->can_discard)
2072                                 continue;
2073
2074                         ret = btrfs_issue_discard(stripe->dev->bdev,
2075                                                   stripe->physical,
2076                                                   stripe->length,
2077                                                   &bytes);
2078                         if (!ret)
2079                                 discarded_bytes += bytes;
2080                         else if (ret != -EOPNOTSUPP)
2081                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2082
2083                         /*
2084                          * Just in case we get back EOPNOTSUPP for some reason,
2085                          * just ignore the return value so we don't screw up
2086                          * people calling discard_extent.
2087                          */
2088                         ret = 0;
2089                 }
2090                 btrfs_put_bbio(bbio);
2091         }
2092         btrfs_bio_counter_dec(root->fs_info);
2093
2094         if (actual_bytes)
2095                 *actual_bytes = discarded_bytes;
2096
2097
2098         if (ret == -EOPNOTSUPP)
2099                 ret = 0;
2100         return ret;
2101 }
2102
2103 /* Can return -ENOMEM */
2104 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2105                          struct btrfs_root *root,
2106                          u64 bytenr, u64 num_bytes, u64 parent,
2107                          u64 root_objectid, u64 owner, u64 offset)
2108 {
2109         int ret;
2110         struct btrfs_fs_info *fs_info = root->fs_info;
2111
2112         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2113                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2114
2115         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2116                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2117                                         num_bytes,
2118                                         parent, root_objectid, (int)owner,
2119                                         BTRFS_ADD_DELAYED_REF, NULL);
2120         } else {
2121                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2122                                         num_bytes, parent, root_objectid,
2123                                         owner, offset, 0,
2124                                         BTRFS_ADD_DELAYED_REF, NULL);
2125         }
2126         return ret;
2127 }
2128
2129 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2130                                   struct btrfs_root *root,
2131                                   struct btrfs_delayed_ref_node *node,
2132                                   u64 parent, u64 root_objectid,
2133                                   u64 owner, u64 offset, int refs_to_add,
2134                                   struct btrfs_delayed_extent_op *extent_op)
2135 {
2136         struct btrfs_fs_info *fs_info = root->fs_info;
2137         struct btrfs_path *path;
2138         struct extent_buffer *leaf;
2139         struct btrfs_extent_item *item;
2140         struct btrfs_key key;
2141         u64 bytenr = node->bytenr;
2142         u64 num_bytes = node->num_bytes;
2143         u64 refs;
2144         int ret;
2145
2146         path = btrfs_alloc_path();
2147         if (!path)
2148                 return -ENOMEM;
2149
2150         path->reada = READA_FORWARD;
2151         path->leave_spinning = 1;
2152         /* this will setup the path even if it fails to insert the back ref */
2153         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2154                                            bytenr, num_bytes, parent,
2155                                            root_objectid, owner, offset,
2156                                            refs_to_add, extent_op);
2157         if ((ret < 0 && ret != -EAGAIN) || !ret)
2158                 goto out;
2159
2160         /*
2161          * Ok we had -EAGAIN which means we didn't have space to insert and
2162          * inline extent ref, so just update the reference count and add a
2163          * normal backref.
2164          */
2165         leaf = path->nodes[0];
2166         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2167         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2168         refs = btrfs_extent_refs(leaf, item);
2169         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2170         if (extent_op)
2171                 __run_delayed_extent_op(extent_op, leaf, item);
2172
2173         btrfs_mark_buffer_dirty(leaf);
2174         btrfs_release_path(path);
2175
2176         path->reada = READA_FORWARD;
2177         path->leave_spinning = 1;
2178         /* now insert the actual backref */
2179         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2180                                     path, bytenr, parent, root_objectid,
2181                                     owner, offset, refs_to_add);
2182         if (ret)
2183                 btrfs_abort_transaction(trans, ret);
2184 out:
2185         btrfs_free_path(path);
2186         return ret;
2187 }
2188
2189 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2190                                 struct btrfs_root *root,
2191                                 struct btrfs_delayed_ref_node *node,
2192                                 struct btrfs_delayed_extent_op *extent_op,
2193                                 int insert_reserved)
2194 {
2195         int ret = 0;
2196         struct btrfs_delayed_data_ref *ref;
2197         struct btrfs_key ins;
2198         u64 parent = 0;
2199         u64 ref_root = 0;
2200         u64 flags = 0;
2201
2202         ins.objectid = node->bytenr;
2203         ins.offset = node->num_bytes;
2204         ins.type = BTRFS_EXTENT_ITEM_KEY;
2205
2206         ref = btrfs_delayed_node_to_data_ref(node);
2207         trace_run_delayed_data_ref(root->fs_info, node, ref, node->action);
2208
2209         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2210                 parent = ref->parent;
2211         ref_root = ref->root;
2212
2213         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2214                 if (extent_op)
2215                         flags |= extent_op->flags_to_set;
2216                 ret = alloc_reserved_file_extent(trans, root,
2217                                                  parent, ref_root, flags,
2218                                                  ref->objectid, ref->offset,
2219                                                  &ins, node->ref_mod);
2220         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2221                 ret = __btrfs_inc_extent_ref(trans, root, node, parent,
2222                                              ref_root, ref->objectid,
2223                                              ref->offset, node->ref_mod,
2224                                              extent_op);
2225         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2226                 ret = __btrfs_free_extent(trans, root, node, parent,
2227                                           ref_root, ref->objectid,
2228                                           ref->offset, node->ref_mod,
2229                                           extent_op);
2230         } else {
2231                 BUG();
2232         }
2233         return ret;
2234 }
2235
2236 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2237                                     struct extent_buffer *leaf,
2238                                     struct btrfs_extent_item *ei)
2239 {
2240         u64 flags = btrfs_extent_flags(leaf, ei);
2241         if (extent_op->update_flags) {
2242                 flags |= extent_op->flags_to_set;
2243                 btrfs_set_extent_flags(leaf, ei, flags);
2244         }
2245
2246         if (extent_op->update_key) {
2247                 struct btrfs_tree_block_info *bi;
2248                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2249                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2250                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2251         }
2252 }
2253
2254 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2255                                  struct btrfs_root *root,
2256                                  struct btrfs_delayed_ref_node *node,
2257                                  struct btrfs_delayed_extent_op *extent_op)
2258 {
2259         struct btrfs_key key;
2260         struct btrfs_path *path;
2261         struct btrfs_extent_item *ei;
2262         struct extent_buffer *leaf;
2263         u32 item_size;
2264         int ret;
2265         int err = 0;
2266         int metadata = !extent_op->is_data;
2267
2268         if (trans->aborted)
2269                 return 0;
2270
2271         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2272                 metadata = 0;
2273
2274         path = btrfs_alloc_path();
2275         if (!path)
2276                 return -ENOMEM;
2277
2278         key.objectid = node->bytenr;
2279
2280         if (metadata) {
2281                 key.type = BTRFS_METADATA_ITEM_KEY;
2282                 key.offset = extent_op->level;
2283         } else {
2284                 key.type = BTRFS_EXTENT_ITEM_KEY;
2285                 key.offset = node->num_bytes;
2286         }
2287
2288 again:
2289         path->reada = READA_FORWARD;
2290         path->leave_spinning = 1;
2291         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2292                                 path, 0, 1);
2293         if (ret < 0) {
2294                 err = ret;
2295                 goto out;
2296         }
2297         if (ret > 0) {
2298                 if (metadata) {
2299                         if (path->slots[0] > 0) {
2300                                 path->slots[0]--;
2301                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2302                                                       path->slots[0]);
2303                                 if (key.objectid == node->bytenr &&
2304                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2305                                     key.offset == node->num_bytes)
2306                                         ret = 0;
2307                         }
2308                         if (ret > 0) {
2309                                 btrfs_release_path(path);
2310                                 metadata = 0;
2311
2312                                 key.objectid = node->bytenr;
2313                                 key.offset = node->num_bytes;
2314                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2315                                 goto again;
2316                         }
2317                 } else {
2318                         err = -EIO;
2319                         goto out;
2320                 }
2321         }
2322
2323         leaf = path->nodes[0];
2324         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2325 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2326         if (item_size < sizeof(*ei)) {
2327                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2328                                              path, (u64)-1, 0);
2329                 if (ret < 0) {
2330                         err = ret;
2331                         goto out;
2332                 }
2333                 leaf = path->nodes[0];
2334                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2335         }
2336 #endif
2337         BUG_ON(item_size < sizeof(*ei));
2338         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2339         __run_delayed_extent_op(extent_op, leaf, ei);
2340
2341         btrfs_mark_buffer_dirty(leaf);
2342 out:
2343         btrfs_free_path(path);
2344         return err;
2345 }
2346
2347 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2348                                 struct btrfs_root *root,
2349                                 struct btrfs_delayed_ref_node *node,
2350                                 struct btrfs_delayed_extent_op *extent_op,
2351                                 int insert_reserved)
2352 {
2353         int ret = 0;
2354         struct btrfs_delayed_tree_ref *ref;
2355         struct btrfs_key ins;
2356         u64 parent = 0;
2357         u64 ref_root = 0;
2358         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2359                                                  SKINNY_METADATA);
2360
2361         ref = btrfs_delayed_node_to_tree_ref(node);
2362         trace_run_delayed_tree_ref(root->fs_info, node, ref, node->action);
2363
2364         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2365                 parent = ref->parent;
2366         ref_root = ref->root;
2367
2368         ins.objectid = node->bytenr;
2369         if (skinny_metadata) {
2370                 ins.offset = ref->level;
2371                 ins.type = BTRFS_METADATA_ITEM_KEY;
2372         } else {
2373                 ins.offset = node->num_bytes;
2374                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2375         }
2376
2377         BUG_ON(node->ref_mod != 1);
2378         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2379                 BUG_ON(!extent_op || !extent_op->update_flags);
2380                 ret = alloc_reserved_tree_block(trans, root,
2381                                                 parent, ref_root,
2382                                                 extent_op->flags_to_set,
2383                                                 &extent_op->key,
2384                                                 ref->level, &ins);
2385         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2386                 ret = __btrfs_inc_extent_ref(trans, root, node,
2387                                              parent, ref_root,
2388                                              ref->level, 0, 1,
2389                                              extent_op);
2390         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2391                 ret = __btrfs_free_extent(trans, root, node,
2392                                           parent, ref_root,
2393                                           ref->level, 0, 1, extent_op);
2394         } else {
2395                 BUG();
2396         }
2397         return ret;
2398 }
2399
2400 /* helper function to actually process a single delayed ref entry */
2401 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2402                                struct btrfs_root *root,
2403                                struct btrfs_delayed_ref_node *node,
2404                                struct btrfs_delayed_extent_op *extent_op,
2405                                int insert_reserved)
2406 {
2407         int ret = 0;
2408
2409         if (trans->aborted) {
2410                 if (insert_reserved)
2411                         btrfs_pin_extent(root, node->bytenr,
2412                                          node->num_bytes, 1);
2413                 return 0;
2414         }
2415
2416         if (btrfs_delayed_ref_is_head(node)) {
2417                 struct btrfs_delayed_ref_head *head;
2418                 /*
2419                  * we've hit the end of the chain and we were supposed
2420                  * to insert this extent into the tree.  But, it got
2421                  * deleted before we ever needed to insert it, so all
2422                  * we have to do is clean up the accounting
2423                  */
2424                 BUG_ON(extent_op);
2425                 head = btrfs_delayed_node_to_head(node);
2426                 trace_run_delayed_ref_head(root->fs_info, node, head,
2427                                            node->action);
2428
2429                 if (insert_reserved) {
2430                         btrfs_pin_extent(root, node->bytenr,
2431                                          node->num_bytes, 1);
2432                         if (head->is_data) {
2433                                 ret = btrfs_del_csums(trans, root,
2434                                                       node->bytenr,
2435                                                       node->num_bytes);
2436                         }
2437                 }
2438
2439                 /* Also free its reserved qgroup space */
2440                 btrfs_qgroup_free_delayed_ref(root->fs_info,
2441                                               head->qgroup_ref_root,
2442                                               head->qgroup_reserved);
2443                 return ret;
2444         }
2445
2446         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2447             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2448                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2449                                            insert_reserved);
2450         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2451                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2452                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2453                                            insert_reserved);
2454         else
2455                 BUG();
2456         return ret;
2457 }
2458
2459 static inline struct btrfs_delayed_ref_node *
2460 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2461 {
2462         struct btrfs_delayed_ref_node *ref;
2463
2464         if (list_empty(&head->ref_list))
2465                 return NULL;
2466
2467         /*
2468          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2469          * This is to prevent a ref count from going down to zero, which deletes
2470          * the extent item from the extent tree, when there still are references
2471          * to add, which would fail because they would not find the extent item.
2472          */
2473         list_for_each_entry(ref, &head->ref_list, list) {
2474                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2475                         return ref;
2476         }
2477
2478         return list_entry(head->ref_list.next, struct btrfs_delayed_ref_node,
2479                           list);
2480 }
2481
2482 /*
2483  * Returns 0 on success or if called with an already aborted transaction.
2484  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2485  */
2486 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2487                                              struct btrfs_root *root,
2488                                              unsigned long nr)
2489 {
2490         struct btrfs_delayed_ref_root *delayed_refs;
2491         struct btrfs_delayed_ref_node *ref;
2492         struct btrfs_delayed_ref_head *locked_ref = NULL;
2493         struct btrfs_delayed_extent_op *extent_op;
2494         struct btrfs_fs_info *fs_info = root->fs_info;
2495         ktime_t start = ktime_get();
2496         int ret;
2497         unsigned long count = 0;
2498         unsigned long actual_count = 0;
2499         int must_insert_reserved = 0;
2500
2501         delayed_refs = &trans->transaction->delayed_refs;
2502         while (1) {
2503                 if (!locked_ref) {
2504                         if (count >= nr)
2505                                 break;
2506
2507                         spin_lock(&delayed_refs->lock);
2508                         locked_ref = btrfs_select_ref_head(trans);
2509                         if (!locked_ref) {
2510                                 spin_unlock(&delayed_refs->lock);
2511                                 break;
2512                         }
2513
2514                         /* grab the lock that says we are going to process
2515                          * all the refs for this head */
2516                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2517                         spin_unlock(&delayed_refs->lock);
2518                         /*
2519                          * we may have dropped the spin lock to get the head
2520                          * mutex lock, and that might have given someone else
2521                          * time to free the head.  If that's true, it has been
2522                          * removed from our list and we can move on.
2523                          */
2524                         if (ret == -EAGAIN) {
2525                                 locked_ref = NULL;
2526                                 count++;
2527                                 continue;
2528                         }
2529                 }
2530
2531                 /*
2532                  * We need to try and merge add/drops of the same ref since we
2533                  * can run into issues with relocate dropping the implicit ref
2534                  * and then it being added back again before the drop can
2535                  * finish.  If we merged anything we need to re-loop so we can
2536                  * get a good ref.
2537                  * Or we can get node references of the same type that weren't
2538                  * merged when created due to bumps in the tree mod seq, and
2539                  * we need to merge them to prevent adding an inline extent
2540                  * backref before dropping it (triggering a BUG_ON at
2541                  * insert_inline_extent_backref()).
2542                  */
2543                 spin_lock(&locked_ref->lock);
2544                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2545                                          locked_ref);
2546
2547                 /*
2548                  * locked_ref is the head node, so we have to go one
2549                  * node back for any delayed ref updates
2550                  */
2551                 ref = select_delayed_ref(locked_ref);
2552
2553                 if (ref && ref->seq &&
2554                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2555                         spin_unlock(&locked_ref->lock);
2556                         btrfs_delayed_ref_unlock(locked_ref);
2557                         spin_lock(&delayed_refs->lock);
2558                         locked_ref->processing = 0;
2559                         delayed_refs->num_heads_ready++;
2560                         spin_unlock(&delayed_refs->lock);
2561                         locked_ref = NULL;
2562                         cond_resched();
2563                         count++;
2564                         continue;
2565                 }
2566
2567                 /*
2568                  * record the must insert reserved flag before we
2569                  * drop the spin lock.
2570                  */
2571                 must_insert_reserved = locked_ref->must_insert_reserved;
2572                 locked_ref->must_insert_reserved = 0;
2573
2574                 extent_op = locked_ref->extent_op;
2575                 locked_ref->extent_op = NULL;
2576
2577                 if (!ref) {
2578
2579
2580                         /* All delayed refs have been processed, Go ahead
2581                          * and send the head node to run_one_delayed_ref,
2582                          * so that any accounting fixes can happen
2583                          */
2584                         ref = &locked_ref->node;
2585
2586                         if (extent_op && must_insert_reserved) {
2587                                 btrfs_free_delayed_extent_op(extent_op);
2588                                 extent_op = NULL;
2589                         }
2590
2591                         if (extent_op) {
2592                                 spin_unlock(&locked_ref->lock);
2593                                 ret = run_delayed_extent_op(trans, root,
2594                                                             ref, extent_op);
2595                                 btrfs_free_delayed_extent_op(extent_op);
2596
2597                                 if (ret) {
2598                                         /*
2599                                          * Need to reset must_insert_reserved if
2600                                          * there was an error so the abort stuff
2601                                          * can cleanup the reserved space
2602                                          * properly.
2603                                          */
2604                                         if (must_insert_reserved)
2605                                                 locked_ref->must_insert_reserved = 1;
2606                                         locked_ref->processing = 0;
2607                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2608                                         btrfs_delayed_ref_unlock(locked_ref);
2609                                         return ret;
2610                                 }
2611                                 continue;
2612                         }
2613
2614                         /*
2615                          * Need to drop our head ref lock and re-acquire the
2616                          * delayed ref lock and then re-check to make sure
2617                          * nobody got added.
2618                          */
2619                         spin_unlock(&locked_ref->lock);
2620                         spin_lock(&delayed_refs->lock);
2621                         spin_lock(&locked_ref->lock);
2622                         if (!list_empty(&locked_ref->ref_list) ||
2623                             locked_ref->extent_op) {
2624                                 spin_unlock(&locked_ref->lock);
2625                                 spin_unlock(&delayed_refs->lock);
2626                                 continue;
2627                         }
2628                         ref->in_tree = 0;
2629                         delayed_refs->num_heads--;
2630                         rb_erase(&locked_ref->href_node,
2631                                  &delayed_refs->href_root);
2632                         spin_unlock(&delayed_refs->lock);
2633                 } else {
2634                         actual_count++;
2635                         ref->in_tree = 0;
2636                         list_del(&ref->list);
2637                 }
2638                 atomic_dec(&delayed_refs->num_entries);
2639
2640                 if (!btrfs_delayed_ref_is_head(ref)) {
2641                         /*
2642                          * when we play the delayed ref, also correct the
2643                          * ref_mod on head
2644                          */
2645                         switch (ref->action) {
2646                         case BTRFS_ADD_DELAYED_REF:
2647                         case BTRFS_ADD_DELAYED_EXTENT:
2648                                 locked_ref->node.ref_mod -= ref->ref_mod;
2649                                 break;
2650                         case BTRFS_DROP_DELAYED_REF:
2651                                 locked_ref->node.ref_mod += ref->ref_mod;
2652                                 break;
2653                         default:
2654                                 WARN_ON(1);
2655                         }
2656                 }
2657                 spin_unlock(&locked_ref->lock);
2658
2659                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2660                                           must_insert_reserved);
2661
2662                 btrfs_free_delayed_extent_op(extent_op);
2663                 if (ret) {
2664                         locked_ref->processing = 0;
2665                         btrfs_delayed_ref_unlock(locked_ref);
2666                         btrfs_put_delayed_ref(ref);
2667                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2668                         return ret;
2669                 }
2670
2671                 /*
2672                  * If this node is a head, that means all the refs in this head
2673                  * have been dealt with, and we will pick the next head to deal
2674                  * with, so we must unlock the head and drop it from the cluster
2675                  * list before we release it.
2676                  */
2677                 if (btrfs_delayed_ref_is_head(ref)) {
2678                         if (locked_ref->is_data &&
2679                             locked_ref->total_ref_mod < 0) {
2680                                 spin_lock(&delayed_refs->lock);
2681                                 delayed_refs->pending_csums -= ref->num_bytes;
2682                                 spin_unlock(&delayed_refs->lock);
2683                         }
2684                         btrfs_delayed_ref_unlock(locked_ref);
2685                         locked_ref = NULL;
2686                 }
2687                 btrfs_put_delayed_ref(ref);
2688                 count++;
2689                 cond_resched();
2690         }
2691
2692         /*
2693          * We don't want to include ref heads since we can have empty ref heads
2694          * and those will drastically skew our runtime down since we just do
2695          * accounting, no actual extent tree updates.
2696          */
2697         if (actual_count > 0) {
2698                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2699                 u64 avg;
2700
2701                 /*
2702                  * We weigh the current average higher than our current runtime
2703                  * to avoid large swings in the average.
2704                  */
2705                 spin_lock(&delayed_refs->lock);
2706                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2707                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2708                 spin_unlock(&delayed_refs->lock);
2709         }
2710         return 0;
2711 }
2712
2713 #ifdef SCRAMBLE_DELAYED_REFS
2714 /*
2715  * Normally delayed refs get processed in ascending bytenr order. This
2716  * correlates in most cases to the order added. To expose dependencies on this
2717  * order, we start to process the tree in the middle instead of the beginning
2718  */
2719 static u64 find_middle(struct rb_root *root)
2720 {
2721         struct rb_node *n = root->rb_node;
2722         struct btrfs_delayed_ref_node *entry;
2723         int alt = 1;
2724         u64 middle;
2725         u64 first = 0, last = 0;
2726
2727         n = rb_first(root);
2728         if (n) {
2729                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2730                 first = entry->bytenr;
2731         }
2732         n = rb_last(root);
2733         if (n) {
2734                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2735                 last = entry->bytenr;
2736         }
2737         n = root->rb_node;
2738
2739         while (n) {
2740                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2741                 WARN_ON(!entry->in_tree);
2742
2743                 middle = entry->bytenr;
2744
2745                 if (alt)
2746                         n = n->rb_left;
2747                 else
2748                         n = n->rb_right;
2749
2750                 alt = 1 - alt;
2751         }
2752         return middle;
2753 }
2754 #endif
2755
2756 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2757 {
2758         u64 num_bytes;
2759
2760         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2761                              sizeof(struct btrfs_extent_inline_ref));
2762         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2763                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2764
2765         /*
2766          * We don't ever fill up leaves all the way so multiply by 2 just to be
2767          * closer to what we're really going to want to use.
2768          */
2769         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2770 }
2771
2772 /*
2773  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2774  * would require to store the csums for that many bytes.
2775  */
2776 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2777 {
2778         u64 csum_size;
2779         u64 num_csums_per_leaf;
2780         u64 num_csums;
2781
2782         csum_size = BTRFS_MAX_ITEM_SIZE(root);
2783         num_csums_per_leaf = div64_u64(csum_size,
2784                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2785         num_csums = div64_u64(csum_bytes, root->sectorsize);
2786         num_csums += num_csums_per_leaf - 1;
2787         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2788         return num_csums;
2789 }
2790
2791 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2792                                        struct btrfs_root *root)
2793 {
2794         struct btrfs_block_rsv *global_rsv;
2795         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2796         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2797         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2798         u64 num_bytes, num_dirty_bgs_bytes;
2799         int ret = 0;
2800
2801         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2802         num_heads = heads_to_leaves(root, num_heads);
2803         if (num_heads > 1)
2804                 num_bytes += (num_heads - 1) * root->nodesize;
2805         num_bytes <<= 1;
2806         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2807         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2808                                                              num_dirty_bgs);
2809         global_rsv = &root->fs_info->global_block_rsv;
2810
2811         /*
2812          * If we can't allocate any more chunks lets make sure we have _lots_ of
2813          * wiggle room since running delayed refs can create more delayed refs.
2814          */
2815         if (global_rsv->space_info->full) {
2816                 num_dirty_bgs_bytes <<= 1;
2817                 num_bytes <<= 1;
2818         }
2819
2820         spin_lock(&global_rsv->lock);
2821         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2822                 ret = 1;
2823         spin_unlock(&global_rsv->lock);
2824         return ret;
2825 }
2826
2827 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2828                                        struct btrfs_root *root)
2829 {
2830         struct btrfs_fs_info *fs_info = root->fs_info;
2831         u64 num_entries =
2832                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2833         u64 avg_runtime;
2834         u64 val;
2835
2836         smp_mb();
2837         avg_runtime = fs_info->avg_delayed_ref_runtime;
2838         val = num_entries * avg_runtime;
2839         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2840                 return 1;
2841         if (val >= NSEC_PER_SEC / 2)
2842                 return 2;
2843
2844         return btrfs_check_space_for_delayed_refs(trans, root);
2845 }
2846
2847 struct async_delayed_refs {
2848         struct btrfs_root *root;
2849         u64 transid;
2850         int count;
2851         int error;
2852         int sync;
2853         struct completion wait;
2854         struct btrfs_work work;
2855 };
2856
2857 static void delayed_ref_async_start(struct btrfs_work *work)
2858 {
2859         struct async_delayed_refs *async;
2860         struct btrfs_trans_handle *trans;
2861         int ret;
2862
2863         async = container_of(work, struct async_delayed_refs, work);
2864
2865         /* if the commit is already started, we don't need to wait here */
2866         if (btrfs_transaction_blocked(async->root->fs_info))
2867                 goto done;
2868
2869         trans = btrfs_join_transaction(async->root);
2870         if (IS_ERR(trans)) {
2871                 async->error = PTR_ERR(trans);
2872                 goto done;
2873         }
2874
2875         /*
2876          * trans->sync means that when we call end_transaction, we won't
2877          * wait on delayed refs
2878          */
2879         trans->sync = true;
2880
2881         /* Don't bother flushing if we got into a different transaction */
2882         if (trans->transid > async->transid)
2883                 goto end;
2884
2885         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2886         if (ret)
2887                 async->error = ret;
2888 end:
2889         ret = btrfs_end_transaction(trans, async->root);
2890         if (ret && !async->error)
2891                 async->error = ret;
2892 done:
2893         if (async->sync)
2894                 complete(&async->wait);
2895         else
2896                 kfree(async);
2897 }
2898
2899 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2900                                  unsigned long count, u64 transid, int wait)
2901 {
2902         struct async_delayed_refs *async;
2903         int ret;
2904
2905         async = kmalloc(sizeof(*async), GFP_NOFS);
2906         if (!async)
2907                 return -ENOMEM;
2908
2909         async->root = root->fs_info->tree_root;
2910         async->count = count;
2911         async->error = 0;
2912         async->transid = transid;
2913         if (wait)
2914                 async->sync = 1;
2915         else
2916                 async->sync = 0;
2917         init_completion(&async->wait);
2918
2919         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2920                         delayed_ref_async_start, NULL, NULL);
2921
2922         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2923
2924         if (wait) {
2925                 wait_for_completion(&async->wait);
2926                 ret = async->error;
2927                 kfree(async);
2928                 return ret;
2929         }
2930         return 0;
2931 }
2932
2933 /*
2934  * this starts processing the delayed reference count updates and
2935  * extent insertions we have queued up so far.  count can be
2936  * 0, which means to process everything in the tree at the start
2937  * of the run (but not newly added entries), or it can be some target
2938  * number you'd like to process.
2939  *
2940  * Returns 0 on success or if called with an aborted transaction
2941  * Returns <0 on error and aborts the transaction
2942  */
2943 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2944                            struct btrfs_root *root, unsigned long count)
2945 {
2946         struct rb_node *node;
2947         struct btrfs_delayed_ref_root *delayed_refs;
2948         struct btrfs_delayed_ref_head *head;
2949         int ret;
2950         int run_all = count == (unsigned long)-1;
2951         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2952
2953         /* We'll clean this up in btrfs_cleanup_transaction */
2954         if (trans->aborted)
2955                 return 0;
2956
2957         if (root->fs_info->creating_free_space_tree)
2958                 return 0;
2959
2960         if (root == root->fs_info->extent_root)
2961                 root = root->fs_info->tree_root;
2962
2963         delayed_refs = &trans->transaction->delayed_refs;
2964         if (count == 0)
2965                 count = atomic_read(&delayed_refs->num_entries) * 2;
2966
2967 again:
2968 #ifdef SCRAMBLE_DELAYED_REFS
2969         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2970 #endif
2971         trans->can_flush_pending_bgs = false;
2972         ret = __btrfs_run_delayed_refs(trans, root, count);
2973         if (ret < 0) {
2974                 btrfs_abort_transaction(trans, ret);
2975                 return ret;
2976         }
2977
2978         if (run_all) {
2979                 if (!list_empty(&trans->new_bgs))
2980                         btrfs_create_pending_block_groups(trans, root);
2981
2982                 spin_lock(&delayed_refs->lock);
2983                 node = rb_first(&delayed_refs->href_root);
2984                 if (!node) {
2985                         spin_unlock(&delayed_refs->lock);
2986                         goto out;
2987                 }
2988                 count = (unsigned long)-1;
2989
2990                 while (node) {
2991                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2992                                         href_node);
2993                         if (btrfs_delayed_ref_is_head(&head->node)) {
2994                                 struct btrfs_delayed_ref_node *ref;
2995
2996                                 ref = &head->node;
2997                                 atomic_inc(&ref->refs);
2998
2999                                 spin_unlock(&delayed_refs->lock);
3000                                 /*
3001                                  * Mutex was contended, block until it's
3002                                  * released and try again
3003                                  */
3004                                 mutex_lock(&head->mutex);
3005                                 mutex_unlock(&head->mutex);
3006
3007                                 btrfs_put_delayed_ref(ref);
3008                                 cond_resched();
3009                                 goto again;
3010                         } else {
3011                                 WARN_ON(1);
3012                         }
3013                         node = rb_next(node);
3014                 }
3015                 spin_unlock(&delayed_refs->lock);
3016                 cond_resched();
3017                 goto again;
3018         }
3019 out:
3020         assert_qgroups_uptodate(trans);
3021         trans->can_flush_pending_bgs = can_flush_pending_bgs;
3022         return 0;
3023 }
3024
3025 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
3026                                 struct btrfs_root *root,
3027                                 u64 bytenr, u64 num_bytes, u64 flags,
3028                                 int level, int is_data)
3029 {
3030         struct btrfs_delayed_extent_op *extent_op;
3031         int ret;
3032
3033         extent_op = btrfs_alloc_delayed_extent_op();
3034         if (!extent_op)
3035                 return -ENOMEM;
3036
3037         extent_op->flags_to_set = flags;
3038         extent_op->update_flags = true;
3039         extent_op->update_key = false;
3040         extent_op->is_data = is_data ? true : false;
3041         extent_op->level = level;
3042
3043         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
3044                                           num_bytes, extent_op);
3045         if (ret)
3046                 btrfs_free_delayed_extent_op(extent_op);
3047         return ret;
3048 }
3049
3050 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
3051                                       struct btrfs_root *root,
3052                                       struct btrfs_path *path,
3053                                       u64 objectid, u64 offset, u64 bytenr)
3054 {
3055         struct btrfs_delayed_ref_head *head;
3056         struct btrfs_delayed_ref_node *ref;
3057         struct btrfs_delayed_data_ref *data_ref;
3058         struct btrfs_delayed_ref_root *delayed_refs;
3059         int ret = 0;
3060
3061         delayed_refs = &trans->transaction->delayed_refs;
3062         spin_lock(&delayed_refs->lock);
3063         head = btrfs_find_delayed_ref_head(trans, bytenr);
3064         if (!head) {
3065                 spin_unlock(&delayed_refs->lock);
3066                 return 0;
3067         }
3068
3069         if (!mutex_trylock(&head->mutex)) {
3070                 atomic_inc(&head->node.refs);
3071                 spin_unlock(&delayed_refs->lock);
3072
3073                 btrfs_release_path(path);
3074
3075                 /*
3076                  * Mutex was contended, block until it's released and let
3077                  * caller try again
3078                  */
3079                 mutex_lock(&head->mutex);
3080                 mutex_unlock(&head->mutex);
3081                 btrfs_put_delayed_ref(&head->node);
3082                 return -EAGAIN;
3083         }
3084         spin_unlock(&delayed_refs->lock);
3085
3086         spin_lock(&head->lock);
3087         list_for_each_entry(ref, &head->ref_list, list) {
3088                 /* If it's a shared ref we know a cross reference exists */
3089                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3090                         ret = 1;
3091                         break;
3092                 }
3093
3094                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3095
3096                 /*
3097                  * If our ref doesn't match the one we're currently looking at
3098                  * then we have a cross reference.
3099                  */
3100                 if (data_ref->root != root->root_key.objectid ||
3101                     data_ref->objectid != objectid ||
3102                     data_ref->offset != offset) {
3103                         ret = 1;
3104                         break;
3105                 }
3106         }
3107         spin_unlock(&head->lock);
3108         mutex_unlock(&head->mutex);
3109         return ret;
3110 }
3111
3112 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
3113                                         struct btrfs_root *root,
3114                                         struct btrfs_path *path,
3115                                         u64 objectid, u64 offset, u64 bytenr)
3116 {
3117         struct btrfs_root *extent_root = root->fs_info->extent_root;
3118         struct extent_buffer *leaf;
3119         struct btrfs_extent_data_ref *ref;
3120         struct btrfs_extent_inline_ref *iref;
3121         struct btrfs_extent_item *ei;
3122         struct btrfs_key key;
3123         u32 item_size;
3124         int ret;
3125
3126         key.objectid = bytenr;
3127         key.offset = (u64)-1;
3128         key.type = BTRFS_EXTENT_ITEM_KEY;
3129
3130         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3131         if (ret < 0)
3132                 goto out;
3133         BUG_ON(ret == 0); /* Corruption */
3134
3135         ret = -ENOENT;
3136         if (path->slots[0] == 0)
3137                 goto out;
3138
3139         path->slots[0]--;
3140         leaf = path->nodes[0];
3141         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3142
3143         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3144                 goto out;
3145
3146         ret = 1;
3147         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3148 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3149         if (item_size < sizeof(*ei)) {
3150                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3151                 goto out;
3152         }
3153 #endif
3154         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3155
3156         if (item_size != sizeof(*ei) +
3157             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3158                 goto out;
3159
3160         if (btrfs_extent_generation(leaf, ei) <=
3161             btrfs_root_last_snapshot(&root->root_item))
3162                 goto out;
3163
3164         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3165         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3166             BTRFS_EXTENT_DATA_REF_KEY)
3167                 goto out;
3168
3169         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3170         if (btrfs_extent_refs(leaf, ei) !=
3171             btrfs_extent_data_ref_count(leaf, ref) ||
3172             btrfs_extent_data_ref_root(leaf, ref) !=
3173             root->root_key.objectid ||
3174             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3175             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3176                 goto out;
3177
3178         ret = 0;
3179 out:
3180         return ret;
3181 }
3182
3183 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3184                           struct btrfs_root *root,
3185                           u64 objectid, u64 offset, u64 bytenr)
3186 {
3187         struct btrfs_path *path;
3188         int ret;
3189         int ret2;
3190
3191         path = btrfs_alloc_path();
3192         if (!path)
3193                 return -ENOENT;
3194
3195         do {
3196                 ret = check_committed_ref(trans, root, path, objectid,
3197                                           offset, bytenr);
3198                 if (ret && ret != -ENOENT)
3199                         goto out;
3200
3201                 ret2 = check_delayed_ref(trans, root, path, objectid,
3202                                          offset, bytenr);
3203         } while (ret2 == -EAGAIN);
3204
3205         if (ret2 && ret2 != -ENOENT) {
3206                 ret = ret2;
3207                 goto out;
3208         }
3209
3210         if (ret != -ENOENT || ret2 != -ENOENT)
3211                 ret = 0;
3212 out:
3213         btrfs_free_path(path);
3214         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3215                 WARN_ON(ret > 0);
3216         return ret;
3217 }
3218
3219 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3220                            struct btrfs_root *root,
3221                            struct extent_buffer *buf,
3222                            int full_backref, int inc)
3223 {
3224         u64 bytenr;
3225         u64 num_bytes;
3226         u64 parent;
3227         u64 ref_root;
3228         u32 nritems;
3229         struct btrfs_key key;
3230         struct btrfs_file_extent_item *fi;
3231         int i;
3232         int level;
3233         int ret = 0;
3234         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3235                             u64, u64, u64, u64, u64, u64);
3236
3237
3238         if (btrfs_is_testing(root->fs_info))
3239                 return 0;
3240
3241         ref_root = btrfs_header_owner(buf);
3242         nritems = btrfs_header_nritems(buf);
3243         level = btrfs_header_level(buf);
3244
3245         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3246                 return 0;
3247
3248         if (inc)
3249                 process_func = btrfs_inc_extent_ref;
3250         else
3251                 process_func = btrfs_free_extent;
3252
3253         if (full_backref)
3254                 parent = buf->start;
3255         else
3256                 parent = 0;
3257
3258         for (i = 0; i < nritems; i++) {
3259                 if (level == 0) {
3260                         btrfs_item_key_to_cpu(buf, &key, i);
3261                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3262                                 continue;
3263                         fi = btrfs_item_ptr(buf, i,
3264                                             struct btrfs_file_extent_item);
3265                         if (btrfs_file_extent_type(buf, fi) ==
3266                             BTRFS_FILE_EXTENT_INLINE)
3267                                 continue;
3268                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3269                         if (bytenr == 0)
3270                                 continue;
3271
3272                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3273                         key.offset -= btrfs_file_extent_offset(buf, fi);
3274                         ret = process_func(trans, root, bytenr, num_bytes,
3275                                            parent, ref_root, key.objectid,
3276                                            key.offset);
3277                         if (ret)
3278                                 goto fail;
3279                 } else {
3280                         bytenr = btrfs_node_blockptr(buf, i);
3281                         num_bytes = root->nodesize;
3282                         ret = process_func(trans, root, bytenr, num_bytes,
3283                                            parent, ref_root, level - 1, 0);
3284                         if (ret)
3285                                 goto fail;
3286                 }
3287         }
3288         return 0;
3289 fail:
3290         return ret;
3291 }
3292
3293 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3294                   struct extent_buffer *buf, int full_backref)
3295 {
3296         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3297 }
3298
3299 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3300                   struct extent_buffer *buf, int full_backref)
3301 {
3302         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3303 }
3304
3305 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3306                                  struct btrfs_root *root,
3307                                  struct btrfs_path *path,
3308                                  struct btrfs_block_group_cache *cache)
3309 {
3310         int ret;
3311         struct btrfs_root *extent_root = root->fs_info->extent_root;
3312         unsigned long bi;
3313         struct extent_buffer *leaf;
3314
3315         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3316         if (ret) {
3317                 if (ret > 0)
3318                         ret = -ENOENT;
3319                 goto fail;
3320         }
3321
3322         leaf = path->nodes[0];
3323         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3324         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3325         btrfs_mark_buffer_dirty(leaf);
3326 fail:
3327         btrfs_release_path(path);
3328         return ret;
3329
3330 }
3331
3332 static struct btrfs_block_group_cache *
3333 next_block_group(struct btrfs_root *root,
3334                  struct btrfs_block_group_cache *cache)
3335 {
3336         struct rb_node *node;
3337
3338         spin_lock(&root->fs_info->block_group_cache_lock);
3339
3340         /* If our block group was removed, we need a full search. */
3341         if (RB_EMPTY_NODE(&cache->cache_node)) {
3342                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3343
3344                 spin_unlock(&root->fs_info->block_group_cache_lock);
3345                 btrfs_put_block_group(cache);
3346                 cache = btrfs_lookup_first_block_group(root->fs_info,
3347                                                        next_bytenr);
3348                 return cache;
3349         }
3350         node = rb_next(&cache->cache_node);
3351         btrfs_put_block_group(cache);
3352         if (node) {
3353                 cache = rb_entry(node, struct btrfs_block_group_cache,
3354                                  cache_node);
3355                 btrfs_get_block_group(cache);
3356         } else
3357                 cache = NULL;
3358         spin_unlock(&root->fs_info->block_group_cache_lock);
3359         return cache;
3360 }
3361
3362 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3363                             struct btrfs_trans_handle *trans,
3364                             struct btrfs_path *path)
3365 {
3366         struct btrfs_root *root = block_group->fs_info->tree_root;
3367         struct inode *inode = NULL;
3368         u64 alloc_hint = 0;
3369         int dcs = BTRFS_DC_ERROR;
3370         u64 num_pages = 0;
3371         int retries = 0;
3372         int ret = 0;
3373
3374         /*
3375          * If this block group is smaller than 100 megs don't bother caching the
3376          * block group.
3377          */
3378         if (block_group->key.offset < (100 * SZ_1M)) {
3379                 spin_lock(&block_group->lock);
3380                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3381                 spin_unlock(&block_group->lock);
3382                 return 0;
3383         }
3384
3385         if (trans->aborted)
3386                 return 0;
3387 again:
3388         inode = lookup_free_space_inode(root, block_group, path);
3389         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3390                 ret = PTR_ERR(inode);
3391                 btrfs_release_path(path);
3392                 goto out;
3393         }
3394
3395         if (IS_ERR(inode)) {
3396                 BUG_ON(retries);
3397                 retries++;
3398
3399                 if (block_group->ro)
3400                         goto out_free;
3401
3402                 ret = create_free_space_inode(root, trans, block_group, path);
3403                 if (ret)
3404                         goto out_free;
3405                 goto again;
3406         }
3407
3408         /* We've already setup this transaction, go ahead and exit */
3409         if (block_group->cache_generation == trans->transid &&
3410             i_size_read(inode)) {
3411                 dcs = BTRFS_DC_SETUP;
3412                 goto out_put;
3413         }
3414
3415         /*
3416          * We want to set the generation to 0, that way if anything goes wrong
3417          * from here on out we know not to trust this cache when we load up next
3418          * time.
3419          */
3420         BTRFS_I(inode)->generation = 0;
3421         ret = btrfs_update_inode(trans, root, inode);
3422         if (ret) {
3423                 /*
3424                  * So theoretically we could recover from this, simply set the
3425                  * super cache generation to 0 so we know to invalidate the
3426                  * cache, but then we'd have to keep track of the block groups
3427                  * that fail this way so we know we _have_ to reset this cache
3428                  * before the next commit or risk reading stale cache.  So to
3429                  * limit our exposure to horrible edge cases lets just abort the
3430                  * transaction, this only happens in really bad situations
3431                  * anyway.
3432                  */
3433                 btrfs_abort_transaction(trans, ret);
3434                 goto out_put;
3435         }
3436         WARN_ON(ret);
3437
3438         if (i_size_read(inode) > 0) {
3439                 ret = btrfs_check_trunc_cache_free_space(root,
3440                                         &root->fs_info->global_block_rsv);
3441                 if (ret)
3442                         goto out_put;
3443
3444                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3445                 if (ret)
3446                         goto out_put;
3447         }
3448
3449         spin_lock(&block_group->lock);
3450         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3451             !btrfs_test_opt(root->fs_info, SPACE_CACHE)) {
3452                 /*
3453                  * don't bother trying to write stuff out _if_
3454                  * a) we're not cached,
3455                  * b) we're with nospace_cache mount option.
3456                  */
3457                 dcs = BTRFS_DC_WRITTEN;
3458                 spin_unlock(&block_group->lock);
3459                 goto out_put;
3460         }
3461         spin_unlock(&block_group->lock);
3462
3463         /*
3464          * We hit an ENOSPC when setting up the cache in this transaction, just
3465          * skip doing the setup, we've already cleared the cache so we're safe.
3466          */
3467         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3468                 ret = -ENOSPC;
3469                 goto out_put;
3470         }
3471
3472         /*
3473          * Try to preallocate enough space based on how big the block group is.
3474          * Keep in mind this has to include any pinned space which could end up
3475          * taking up quite a bit since it's not folded into the other space
3476          * cache.
3477          */
3478         num_pages = div_u64(block_group->key.offset, SZ_256M);
3479         if (!num_pages)
3480                 num_pages = 1;
3481
3482         num_pages *= 16;
3483         num_pages *= PAGE_SIZE;
3484
3485         ret = btrfs_check_data_free_space(inode, 0, num_pages);
3486         if (ret)
3487                 goto out_put;
3488
3489         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3490                                               num_pages, num_pages,
3491                                               &alloc_hint);
3492         /*
3493          * Our cache requires contiguous chunks so that we don't modify a bunch
3494          * of metadata or split extents when writing the cache out, which means
3495          * we can enospc if we are heavily fragmented in addition to just normal
3496          * out of space conditions.  So if we hit this just skip setting up any
3497          * other block groups for this transaction, maybe we'll unpin enough
3498          * space the next time around.
3499          */
3500         if (!ret)
3501                 dcs = BTRFS_DC_SETUP;
3502         else if (ret == -ENOSPC)
3503                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3504         btrfs_free_reserved_data_space(inode, 0, num_pages);
3505
3506 out_put:
3507         iput(inode);
3508 out_free:
3509         btrfs_release_path(path);
3510 out:
3511         spin_lock(&block_group->lock);
3512         if (!ret && dcs == BTRFS_DC_SETUP)
3513                 block_group->cache_generation = trans->transid;
3514         block_group->disk_cache_state = dcs;
3515         spin_unlock(&block_group->lock);
3516
3517         return ret;
3518 }
3519
3520 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3521                             struct btrfs_root *root)
3522 {
3523         struct btrfs_block_group_cache *cache, *tmp;
3524         struct btrfs_transaction *cur_trans = trans->transaction;
3525         struct btrfs_path *path;
3526
3527         if (list_empty(&cur_trans->dirty_bgs) ||
3528             !btrfs_test_opt(root->fs_info, SPACE_CACHE))
3529                 return 0;
3530
3531         path = btrfs_alloc_path();
3532         if (!path)
3533                 return -ENOMEM;
3534
3535         /* Could add new block groups, use _safe just in case */
3536         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3537                                  dirty_list) {
3538                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3539                         cache_save_setup(cache, trans, path);
3540         }
3541
3542         btrfs_free_path(path);
3543         return 0;
3544 }
3545
3546 /*
3547  * transaction commit does final block group cache writeback during a
3548  * critical section where nothing is allowed to change the FS.  This is
3549  * required in order for the cache to actually match the block group,
3550  * but can introduce a lot of latency into the commit.
3551  *
3552  * So, btrfs_start_dirty_block_groups is here to kick off block group
3553  * cache IO.  There's a chance we'll have to redo some of it if the
3554  * block group changes again during the commit, but it greatly reduces
3555  * the commit latency by getting rid of the easy block groups while
3556  * we're still allowing others to join the commit.
3557  */
3558 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3559                                    struct btrfs_root *root)
3560 {
3561         struct btrfs_block_group_cache *cache;
3562         struct btrfs_transaction *cur_trans = trans->transaction;
3563         int ret = 0;
3564         int should_put;
3565         struct btrfs_path *path = NULL;
3566         LIST_HEAD(dirty);
3567         struct list_head *io = &cur_trans->io_bgs;
3568         int num_started = 0;
3569         int loops = 0;
3570
3571         spin_lock(&cur_trans->dirty_bgs_lock);
3572         if (list_empty(&cur_trans->dirty_bgs)) {
3573                 spin_unlock(&cur_trans->dirty_bgs_lock);
3574                 return 0;
3575         }
3576         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3577         spin_unlock(&cur_trans->dirty_bgs_lock);
3578
3579 again:
3580         /*
3581          * make sure all the block groups on our dirty list actually
3582          * exist
3583          */
3584         btrfs_create_pending_block_groups(trans, root);
3585
3586         if (!path) {
3587                 path = btrfs_alloc_path();
3588                 if (!path)
3589                         return -ENOMEM;
3590         }
3591
3592         /*
3593          * cache_write_mutex is here only to save us from balance or automatic
3594          * removal of empty block groups deleting this block group while we are
3595          * writing out the cache
3596          */
3597         mutex_lock(&trans->transaction->cache_write_mutex);
3598         while (!list_empty(&dirty)) {
3599                 cache = list_first_entry(&dirty,
3600                                          struct btrfs_block_group_cache,
3601                                          dirty_list);
3602                 /*
3603                  * this can happen if something re-dirties a block
3604                  * group that is already under IO.  Just wait for it to
3605                  * finish and then do it all again
3606                  */
3607                 if (!list_empty(&cache->io_list)) {
3608                         list_del_init(&cache->io_list);
3609                         btrfs_wait_cache_io(root, trans, cache,
3610                                             &cache->io_ctl, path,
3611                                             cache->key.objectid);
3612                         btrfs_put_block_group(cache);
3613                 }
3614
3615
3616                 /*
3617                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3618                  * if it should update the cache_state.  Don't delete
3619                  * until after we wait.
3620                  *
3621                  * Since we're not running in the commit critical section
3622                  * we need the dirty_bgs_lock to protect from update_block_group
3623                  */
3624                 spin_lock(&cur_trans->dirty_bgs_lock);
3625                 list_del_init(&cache->dirty_list);
3626                 spin_unlock(&cur_trans->dirty_bgs_lock);
3627
3628                 should_put = 1;
3629
3630                 cache_save_setup(cache, trans, path);
3631
3632                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3633                         cache->io_ctl.inode = NULL;
3634                         ret = btrfs_write_out_cache(root, trans, cache, path);
3635                         if (ret == 0 && cache->io_ctl.inode) {
3636                                 num_started++;
3637                                 should_put = 0;
3638
3639                                 /*
3640                                  * the cache_write_mutex is protecting
3641                                  * the io_list
3642                                  */
3643                                 list_add_tail(&cache->io_list, io);
3644                         } else {
3645                                 /*
3646                                  * if we failed to write the cache, the
3647                                  * generation will be bad and life goes on
3648                                  */
3649                                 ret = 0;
3650                         }
3651                 }
3652                 if (!ret) {
3653                         ret = write_one_cache_group(trans, root, path, cache);
3654                         /*
3655                          * Our block group might still be attached to the list
3656                          * of new block groups in the transaction handle of some
3657                          * other task (struct btrfs_trans_handle->new_bgs). This
3658                          * means its block group item isn't yet in the extent
3659                          * tree. If this happens ignore the error, as we will
3660                          * try again later in the critical section of the
3661                          * transaction commit.
3662                          */
3663                         if (ret == -ENOENT) {
3664                                 ret = 0;
3665                                 spin_lock(&cur_trans->dirty_bgs_lock);
3666                                 if (list_empty(&cache->dirty_list)) {
3667                                         list_add_tail(&cache->dirty_list,
3668                                                       &cur_trans->dirty_bgs);
3669                                         btrfs_get_block_group(cache);
3670                                 }
3671                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3672                         } else if (ret) {
3673                                 btrfs_abort_transaction(trans, ret);
3674                         }
3675                 }
3676
3677                 /* if its not on the io list, we need to put the block group */
3678                 if (should_put)
3679                         btrfs_put_block_group(cache);
3680
3681                 if (ret)
3682                         break;
3683
3684                 /*
3685                  * Avoid blocking other tasks for too long. It might even save
3686                  * us from writing caches for block groups that are going to be
3687                  * removed.
3688                  */
3689                 mutex_unlock(&trans->transaction->cache_write_mutex);
3690                 mutex_lock(&trans->transaction->cache_write_mutex);
3691         }
3692         mutex_unlock(&trans->transaction->cache_write_mutex);
3693
3694         /*
3695          * go through delayed refs for all the stuff we've just kicked off
3696          * and then loop back (just once)
3697          */
3698         ret = btrfs_run_delayed_refs(trans, root, 0);
3699         if (!ret && loops == 0) {
3700                 loops++;
3701                 spin_lock(&cur_trans->dirty_bgs_lock);
3702                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3703                 /*
3704                  * dirty_bgs_lock protects us from concurrent block group
3705                  * deletes too (not just cache_write_mutex).
3706                  */
3707                 if (!list_empty(&dirty)) {
3708                         spin_unlock(&cur_trans->dirty_bgs_lock);
3709                         goto again;
3710                 }
3711                 spin_unlock(&cur_trans->dirty_bgs_lock);
3712         }
3713
3714         btrfs_free_path(path);
3715         return ret;
3716 }
3717
3718 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3719                                    struct btrfs_root *root)
3720 {
3721         struct btrfs_block_group_cache *cache;
3722         struct btrfs_transaction *cur_trans = trans->transaction;
3723         int ret = 0;
3724         int should_put;
3725         struct btrfs_path *path;
3726         struct list_head *io = &cur_trans->io_bgs;
3727         int num_started = 0;
3728
3729         path = btrfs_alloc_path();
3730         if (!path)
3731                 return -ENOMEM;
3732
3733         /*
3734          * Even though we are in the critical section of the transaction commit,
3735          * we can still have concurrent tasks adding elements to this
3736          * transaction's list of dirty block groups. These tasks correspond to
3737          * endio free space workers started when writeback finishes for a
3738          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3739          * allocate new block groups as a result of COWing nodes of the root
3740          * tree when updating the free space inode. The writeback for the space
3741          * caches is triggered by an earlier call to
3742          * btrfs_start_dirty_block_groups() and iterations of the following
3743          * loop.
3744          * Also we want to do the cache_save_setup first and then run the
3745          * delayed refs to make sure we have the best chance at doing this all
3746          * in one shot.
3747          */
3748         spin_lock(&cur_trans->dirty_bgs_lock);
3749         while (!list_empty(&cur_trans->dirty_bgs)) {
3750                 cache = list_first_entry(&cur_trans->dirty_bgs,
3751                                          struct btrfs_block_group_cache,
3752                                          dirty_list);
3753
3754                 /*
3755                  * this can happen if cache_save_setup re-dirties a block
3756                  * group that is already under IO.  Just wait for it to
3757                  * finish and then do it all again
3758                  */
3759                 if (!list_empty(&cache->io_list)) {
3760                         spin_unlock(&cur_trans->dirty_bgs_lock);
3761                         list_del_init(&cache->io_list);
3762                         btrfs_wait_cache_io(root, trans, cache,
3763                                             &cache->io_ctl, path,
3764                                             cache->key.objectid);
3765                         btrfs_put_block_group(cache);
3766                         spin_lock(&cur_trans->dirty_bgs_lock);
3767                 }
3768
3769                 /*
3770                  * don't remove from the dirty list until after we've waited
3771                  * on any pending IO
3772                  */
3773                 list_del_init(&cache->dirty_list);
3774                 spin_unlock(&cur_trans->dirty_bgs_lock);
3775                 should_put = 1;
3776
3777                 cache_save_setup(cache, trans, path);
3778
3779                 if (!ret)
3780                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3781
3782                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3783                         cache->io_ctl.inode = NULL;
3784                         ret = btrfs_write_out_cache(root, trans, cache, path);
3785                         if (ret == 0 && cache->io_ctl.inode) {
3786                                 num_started++;
3787                                 should_put = 0;
3788                                 list_add_tail(&cache->io_list, io);
3789                         } else {
3790                                 /*
3791                                  * if we failed to write the cache, the
3792                                  * generation will be bad and life goes on
3793                                  */
3794                                 ret = 0;
3795                         }
3796                 }
3797                 if (!ret) {
3798                         ret = write_one_cache_group(trans, root, path, cache);
3799                         /*
3800                          * One of the free space endio workers might have
3801                          * created a new block group while updating a free space
3802                          * cache's inode (at inode.c:btrfs_finish_ordered_io())
3803                          * and hasn't released its transaction handle yet, in
3804                          * which case the new block group is still attached to
3805                          * its transaction handle and its creation has not
3806                          * finished yet (no block group item in the extent tree
3807                          * yet, etc). If this is the case, wait for all free
3808                          * space endio workers to finish and retry. This is a
3809                          * a very rare case so no need for a more efficient and
3810                          * complex approach.
3811                          */
3812                         if (ret == -ENOENT) {
3813                                 wait_event(cur_trans->writer_wait,
3814                                    atomic_read(&cur_trans->num_writers) == 1);
3815                                 ret = write_one_cache_group(trans, root, path,
3816                                                             cache);
3817                         }
3818                         if (ret)
3819                                 btrfs_abort_transaction(trans, ret);
3820                 }
3821
3822                 /* if its not on the io list, we need to put the block group */
3823                 if (should_put)
3824                         btrfs_put_block_group(cache);
3825                 spin_lock(&cur_trans->dirty_bgs_lock);
3826         }
3827         spin_unlock(&cur_trans->dirty_bgs_lock);
3828
3829         while (!list_empty(io)) {
3830                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3831                                          io_list);
3832                 list_del_init(&cache->io_list);
3833                 btrfs_wait_cache_io(root, trans, cache,
3834                                     &cache->io_ctl, path, cache->key.objectid);
3835                 btrfs_put_block_group(cache);
3836         }
3837
3838         btrfs_free_path(path);
3839         return ret;
3840 }
3841
3842 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3843 {
3844         struct btrfs_block_group_cache *block_group;
3845         int readonly = 0;
3846
3847         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3848         if (!block_group || block_group->ro)
3849                 readonly = 1;
3850         if (block_group)
3851                 btrfs_put_block_group(block_group);
3852         return readonly;
3853 }
3854
3855 bool btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3856 {
3857         struct btrfs_block_group_cache *bg;
3858         bool ret = true;
3859
3860         bg = btrfs_lookup_block_group(fs_info, bytenr);
3861         if (!bg)
3862                 return false;
3863
3864         spin_lock(&bg->lock);
3865         if (bg->ro)
3866                 ret = false;
3867         else
3868                 atomic_inc(&bg->nocow_writers);
3869         spin_unlock(&bg->lock);
3870
3871         /* no put on block group, done by btrfs_dec_nocow_writers */
3872         if (!ret)
3873                 btrfs_put_block_group(bg);
3874
3875         return ret;
3876
3877 }
3878
3879 void btrfs_dec_nocow_writers(struct btrfs_fs_info *fs_info, u64 bytenr)
3880 {
3881         struct btrfs_block_group_cache *bg;
3882
3883         bg = btrfs_lookup_block_group(fs_info, bytenr);
3884         ASSERT(bg);
3885         if (atomic_dec_and_test(&bg->nocow_writers))
3886                 wake_up_atomic_t(&bg->nocow_writers);
3887         /*
3888          * Once for our lookup and once for the lookup done by a previous call
3889          * to btrfs_inc_nocow_writers()
3890          */
3891         btrfs_put_block_group(bg);
3892         btrfs_put_block_group(bg);
3893 }
3894
3895 static int btrfs_wait_nocow_writers_atomic_t(atomic_t *a)
3896 {
3897         schedule();
3898         return 0;
3899 }
3900
3901 void btrfs_wait_nocow_writers(struct btrfs_block_group_cache *bg)
3902 {
3903         wait_on_atomic_t(&bg->nocow_writers,
3904                          btrfs_wait_nocow_writers_atomic_t,
3905                          TASK_UNINTERRUPTIBLE);
3906 }
3907
3908 static const char *alloc_name(u64 flags)
3909 {
3910         switch (flags) {
3911         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3912                 return "mixed";
3913         case BTRFS_BLOCK_GROUP_METADATA:
3914                 return "metadata";
3915         case BTRFS_BLOCK_GROUP_DATA:
3916                 return "data";
3917         case BTRFS_BLOCK_GROUP_SYSTEM:
3918                 return "system";
3919         default:
3920                 WARN_ON(1);
3921                 return "invalid-combination";
3922         };
3923 }
3924
3925 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3926                              u64 total_bytes, u64 bytes_used,
3927                              u64 bytes_readonly,
3928                              struct btrfs_space_info **space_info)
3929 {
3930         struct btrfs_space_info *found;
3931         int i;
3932         int factor;
3933         int ret;
3934
3935         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3936                      BTRFS_BLOCK_GROUP_RAID10))
3937                 factor = 2;
3938         else
3939                 factor = 1;
3940
3941         found = __find_space_info(info, flags);
3942         if (found) {
3943                 spin_lock(&found->lock);
3944                 found->total_bytes += total_bytes;
3945                 found->disk_total += total_bytes * factor;
3946                 found->bytes_used += bytes_used;
3947                 found->disk_used += bytes_used * factor;
3948                 found->bytes_readonly += bytes_readonly;
3949                 if (total_bytes > 0)
3950                         found->full = 0;
3951                 space_info_add_new_bytes(info, found, total_bytes -
3952                                          bytes_used - bytes_readonly);
3953                 spin_unlock(&found->lock);
3954                 *space_info = found;
3955                 return 0;
3956         }
3957         found = kzalloc(sizeof(*found), GFP_NOFS);
3958         if (!found)
3959                 return -ENOMEM;
3960
3961         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3962         if (ret) {
3963                 kfree(found);
3964                 return ret;
3965         }
3966
3967         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3968                 INIT_LIST_HEAD(&found->block_groups[i]);
3969         init_rwsem(&found->groups_sem);
3970         spin_lock_init(&found->lock);
3971         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3972         found->total_bytes = total_bytes;
3973         found->disk_total = total_bytes * factor;
3974         found->bytes_used = bytes_used;
3975         found->disk_used = bytes_used * factor;
3976         found->bytes_pinned = 0;
3977         found->bytes_reserved = 0;
3978         found->bytes_readonly = bytes_readonly;
3979         found->bytes_may_use = 0;
3980         found->full = 0;
3981         found->max_extent_size = 0;
3982         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3983         found->chunk_alloc = 0;
3984         found->flush = 0;
3985         init_waitqueue_head(&found->wait);
3986         INIT_LIST_HEAD(&found->ro_bgs);
3987         INIT_LIST_HEAD(&found->tickets);
3988         INIT_LIST_HEAD(&found->priority_tickets);
3989
3990         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3991                                     info->space_info_kobj, "%s",
3992                                     alloc_name(found->flags));
3993         if (ret) {
3994                 kfree(found);
3995                 return ret;
3996         }
3997
3998         *space_info = found;
3999         list_add_rcu(&found->list, &info->space_info);
4000         if (flags & BTRFS_BLOCK_GROUP_DATA)
4001                 info->data_sinfo = found;
4002
4003         return ret;
4004 }
4005
4006 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
4007 {
4008         u64 extra_flags = chunk_to_extended(flags) &
4009                                 BTRFS_EXTENDED_PROFILE_MASK;
4010
4011         write_seqlock(&fs_info->profiles_lock);
4012         if (flags & BTRFS_BLOCK_GROUP_DATA)
4013                 fs_info->avail_data_alloc_bits |= extra_flags;
4014         if (flags & BTRFS_BLOCK_GROUP_METADATA)
4015                 fs_info->avail_metadata_alloc_bits |= extra_flags;
4016         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4017                 fs_info->avail_system_alloc_bits |= extra_flags;
4018         write_sequnlock(&fs_info->profiles_lock);
4019 }
4020
4021 /*
4022  * returns target flags in extended format or 0 if restripe for this
4023  * chunk_type is not in progress
4024  *
4025  * should be called with either volume_mutex or balance_lock held
4026  */
4027 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
4028 {
4029         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
4030         u64 target = 0;
4031
4032         if (!bctl)
4033                 return 0;
4034
4035         if (flags & BTRFS_BLOCK_GROUP_DATA &&
4036             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4037                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
4038         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
4039                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4040                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
4041         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
4042                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
4043                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
4044         }
4045
4046         return target;
4047 }
4048
4049 /*
4050  * @flags: available profiles in extended format (see ctree.h)
4051  *
4052  * Returns reduced profile in chunk format.  If profile changing is in
4053  * progress (either running or paused) picks the target profile (if it's
4054  * already available), otherwise falls back to plain reducing.
4055  */
4056 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
4057 {
4058         u64 num_devices = root->fs_info->fs_devices->rw_devices;
4059         u64 target;
4060         u64 raid_type;
4061         u64 allowed = 0;
4062
4063         /*
4064          * see if restripe for this chunk_type is in progress, if so
4065          * try to reduce to the target profile
4066          */
4067         spin_lock(&root->fs_info->balance_lock);
4068         target = get_restripe_target(root->fs_info, flags);
4069         if (target) {
4070                 /* pick target profile only if it's already available */
4071                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
4072                         spin_unlock(&root->fs_info->balance_lock);
4073                         return extended_to_chunk(target);
4074                 }
4075         }
4076         spin_unlock(&root->fs_info->balance_lock);
4077
4078         /* First, mask out the RAID levels which aren't possible */
4079         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
4080                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
4081                         allowed |= btrfs_raid_group[raid_type];
4082         }
4083         allowed &= flags;
4084
4085         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
4086                 allowed = BTRFS_BLOCK_GROUP_RAID6;
4087         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
4088                 allowed = BTRFS_BLOCK_GROUP_RAID5;
4089         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
4090                 allowed = BTRFS_BLOCK_GROUP_RAID10;
4091         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
4092                 allowed = BTRFS_BLOCK_GROUP_RAID1;
4093         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
4094                 allowed = BTRFS_BLOCK_GROUP_RAID0;
4095
4096         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
4097
4098         return extended_to_chunk(flags | allowed);
4099 }
4100
4101 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
4102 {
4103         unsigned seq;
4104         u64 flags;
4105
4106         do {
4107                 flags = orig_flags;
4108                 seq = read_seqbegin(&root->fs_info->profiles_lock);
4109
4110                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4111                         flags |= root->fs_info->avail_data_alloc_bits;
4112                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4113                         flags |= root->fs_info->avail_system_alloc_bits;
4114                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4115                         flags |= root->fs_info->avail_metadata_alloc_bits;
4116         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
4117
4118         return btrfs_reduce_alloc_profile(root, flags);
4119 }
4120
4121 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
4122 {
4123         u64 flags;
4124         u64 ret;
4125
4126         if (data)
4127                 flags = BTRFS_BLOCK_GROUP_DATA;
4128         else if (root == root->fs_info->chunk_root)
4129                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4130         else
4131                 flags = BTRFS_BLOCK_GROUP_METADATA;
4132
4133         ret = get_alloc_profile(root, flags);
4134         return ret;
4135 }
4136
4137 int btrfs_alloc_data_chunk_ondemand(struct inode *inode, u64 bytes)
4138 {
4139         struct btrfs_space_info *data_sinfo;
4140         struct btrfs_root *root = BTRFS_I(inode)->root;
4141         struct btrfs_fs_info *fs_info = root->fs_info;
4142         u64 used;
4143         int ret = 0;
4144         int need_commit = 2;
4145         int have_pinned_space;
4146
4147         /* make sure bytes are sectorsize aligned */
4148         bytes = ALIGN(bytes, root->sectorsize);
4149
4150         if (btrfs_is_free_space_inode(inode)) {
4151                 need_commit = 0;
4152                 ASSERT(current->journal_info);
4153         }
4154
4155         data_sinfo = fs_info->data_sinfo;
4156         if (!data_sinfo)
4157                 goto alloc;
4158
4159 again:
4160         /* make sure we have enough space to handle the data first */
4161         spin_lock(&data_sinfo->lock);
4162         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
4163                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
4164                 data_sinfo->bytes_may_use;
4165
4166         if (used + bytes > data_sinfo->total_bytes) {
4167                 struct btrfs_trans_handle *trans;
4168
4169                 /*
4170                  * if we don't have enough free bytes in this space then we need
4171                  * to alloc a new chunk.
4172                  */
4173                 if (!data_sinfo->full) {
4174                         u64 alloc_target;
4175
4176                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4177                         spin_unlock(&data_sinfo->lock);
4178 alloc:
4179                         alloc_target = btrfs_get_alloc_profile(root, 1);
4180                         /*
4181                          * It is ugly that we don't call nolock join
4182                          * transaction for the free space inode case here.
4183                          * But it is safe because we only do the data space
4184                          * reservation for the free space cache in the
4185                          * transaction context, the common join transaction
4186                          * just increase the counter of the current transaction
4187                          * handler, doesn't try to acquire the trans_lock of
4188                          * the fs.
4189                          */
4190                         trans = btrfs_join_transaction(root);
4191                         if (IS_ERR(trans))
4192                                 return PTR_ERR(trans);
4193
4194                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4195                                              alloc_target,
4196                                              CHUNK_ALLOC_NO_FORCE);
4197                         btrfs_end_transaction(trans, root);
4198                         if (ret < 0) {
4199                                 if (ret != -ENOSPC)
4200                                         return ret;
4201                                 else {
4202                                         have_pinned_space = 1;
4203                                         goto commit_trans;
4204                                 }
4205                         }
4206
4207                         if (!data_sinfo)
4208                                 data_sinfo = fs_info->data_sinfo;
4209
4210                         goto again;
4211                 }
4212
4213                 /*
4214                  * If we don't have enough pinned space to deal with this
4215                  * allocation, and no removed chunk in current transaction,
4216                  * don't bother committing the transaction.
4217                  */
4218                 have_pinned_space = percpu_counter_compare(
4219                         &data_sinfo->total_bytes_pinned,
4220                         used + bytes - data_sinfo->total_bytes);
4221                 spin_unlock(&data_sinfo->lock);
4222
4223                 /* commit the current transaction and try again */
4224 commit_trans:
4225                 if (need_commit &&
4226                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
4227                         need_commit--;
4228
4229                         if (need_commit > 0) {
4230                                 btrfs_start_delalloc_roots(fs_info, 0, -1);
4231                                 btrfs_wait_ordered_roots(fs_info, -1, 0, (u64)-1);
4232                         }
4233
4234                         trans = btrfs_join_transaction(root);
4235                         if (IS_ERR(trans))
4236                                 return PTR_ERR(trans);
4237                         if (have_pinned_space >= 0 ||
4238                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4239                                      &trans->transaction->flags) ||
4240                             need_commit > 0) {
4241                                 ret = btrfs_commit_transaction(trans, root);
4242                                 if (ret)
4243                                         return ret;
4244                                 /*
4245                                  * The cleaner kthread might still be doing iput
4246                                  * operations. Wait for it to finish so that
4247                                  * more space is released.
4248                                  */
4249                                 mutex_lock(&root->fs_info->cleaner_delayed_iput_mutex);
4250                                 mutex_unlock(&root->fs_info->cleaner_delayed_iput_mutex);
4251                                 goto again;
4252                         } else {
4253                                 btrfs_end_transaction(trans, root);
4254                         }
4255                 }
4256
4257                 trace_btrfs_space_reservation(root->fs_info,
4258                                               "space_info:enospc",
4259                                               data_sinfo->flags, bytes, 1);
4260                 return -ENOSPC;
4261         }
4262         data_sinfo->bytes_may_use += bytes;
4263         trace_btrfs_space_reservation(root->fs_info, "space_info",
4264                                       data_sinfo->flags, bytes, 1);
4265         spin_unlock(&data_sinfo->lock);
4266
4267         return ret;
4268 }
4269
4270 /*
4271  * New check_data_free_space() with ability for precious data reservation
4272  * Will replace old btrfs_check_data_free_space(), but for patch split,
4273  * add a new function first and then replace it.
4274  */
4275 int btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4276 {
4277         struct btrfs_root *root = BTRFS_I(inode)->root;
4278         int ret;
4279
4280         /* align the range */
4281         len = round_up(start + len, root->sectorsize) -
4282               round_down(start, root->sectorsize);
4283         start = round_down(start, root->sectorsize);
4284
4285         ret = btrfs_alloc_data_chunk_ondemand(inode, len);
4286         if (ret < 0)
4287                 return ret;
4288
4289         /*
4290          * Use new btrfs_qgroup_reserve_data to reserve precious data space
4291          *
4292          * TODO: Find a good method to avoid reserve data space for NOCOW
4293          * range, but don't impact performance on quota disable case.
4294          */
4295         ret = btrfs_qgroup_reserve_data(inode, start, len);
4296         return ret;
4297 }
4298
4299 /*
4300  * Called if we need to clear a data reservation for this inode
4301  * Normally in a error case.
4302  *
4303  * This one will *NOT* use accurate qgroup reserved space API, just for case
4304  * which we can't sleep and is sure it won't affect qgroup reserved space.
4305  * Like clear_bit_hook().
4306  */
4307 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4308                                             u64 len)
4309 {
4310         struct btrfs_root *root = BTRFS_I(inode)->root;
4311         struct btrfs_space_info *data_sinfo;
4312
4313         /* Make sure the range is aligned to sectorsize */
4314         len = round_up(start + len, root->sectorsize) -
4315               round_down(start, root->sectorsize);
4316         start = round_down(start, root->sectorsize);
4317
4318         data_sinfo = root->fs_info->data_sinfo;
4319         spin_lock(&data_sinfo->lock);
4320         if (WARN_ON(data_sinfo->bytes_may_use < len))
4321                 data_sinfo->bytes_may_use = 0;
4322         else
4323                 data_sinfo->bytes_may_use -= len;
4324         trace_btrfs_space_reservation(root->fs_info, "space_info",
4325                                       data_sinfo->flags, len, 0);
4326         spin_unlock(&data_sinfo->lock);
4327 }
4328
4329 /*
4330  * Called if we need to clear a data reservation for this inode
4331  * Normally in a error case.
4332  *
4333  * This one will handle the per-inode data rsv map for accurate reserved
4334  * space framework.
4335  */
4336 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4337 {
4338         btrfs_free_reserved_data_space_noquota(inode, start, len);
4339         btrfs_qgroup_free_data(inode, start, len);
4340 }
4341
4342 static void force_metadata_allocation(struct btrfs_fs_info *info)
4343 {
4344         struct list_head *head = &info->space_info;
4345         struct btrfs_space_info *found;
4346
4347         rcu_read_lock();
4348         list_for_each_entry_rcu(found, head, list) {
4349                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4350                         found->force_alloc = CHUNK_ALLOC_FORCE;
4351         }
4352         rcu_read_unlock();
4353 }
4354
4355 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4356 {
4357         return (global->size << 1);
4358 }
4359
4360 static int should_alloc_chunk(struct btrfs_root *root,
4361                               struct btrfs_space_info *sinfo, int force)
4362 {
4363         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4364         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4365         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4366         u64 thresh;
4367
4368         if (force == CHUNK_ALLOC_FORCE)
4369                 return 1;
4370
4371         /*
4372          * We need to take into account the global rsv because for all intents
4373          * and purposes it's used space.  Don't worry about locking the
4374          * global_rsv, it doesn't change except when the transaction commits.
4375          */
4376         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4377                 num_allocated += calc_global_rsv_need_space(global_rsv);
4378
4379         /*
4380          * in limited mode, we want to have some free space up to
4381          * about 1% of the FS size.
4382          */
4383         if (force == CHUNK_ALLOC_LIMITED) {
4384                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4385                 thresh = max_t(u64, SZ_64M, div_factor_fine(thresh, 1));
4386
4387                 if (num_bytes - num_allocated < thresh)
4388                         return 1;
4389         }
4390
4391         if (num_allocated + SZ_2M < div_factor(num_bytes, 8))
4392                 return 0;
4393         return 1;
4394 }
4395
4396 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4397 {
4398         u64 num_dev;
4399
4400         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4401                     BTRFS_BLOCK_GROUP_RAID0 |
4402                     BTRFS_BLOCK_GROUP_RAID5 |
4403                     BTRFS_BLOCK_GROUP_RAID6))
4404                 num_dev = root->fs_info->fs_devices->rw_devices;
4405         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4406                 num_dev = 2;
4407         else
4408                 num_dev = 1;    /* DUP or single */
4409
4410         return num_dev;
4411 }
4412
4413 /*
4414  * If @is_allocation is true, reserve space in the system space info necessary
4415  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4416  * removing a chunk.
4417  */
4418 void check_system_chunk(struct btrfs_trans_handle *trans,
4419                         struct btrfs_root *root,
4420                         u64 type)
4421 {
4422         struct btrfs_space_info *info;
4423         u64 left;
4424         u64 thresh;
4425         int ret = 0;
4426         u64 num_devs;
4427
4428         /*
4429          * Needed because we can end up allocating a system chunk and for an
4430          * atomic and race free space reservation in the chunk block reserve.
4431          */
4432         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4433
4434         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4435         spin_lock(&info->lock);
4436         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4437                 info->bytes_reserved - info->bytes_readonly -
4438                 info->bytes_may_use;
4439         spin_unlock(&info->lock);
4440
4441         num_devs = get_profile_num_devs(root, type);
4442
4443         /* num_devs device items to update and 1 chunk item to add or remove */
4444         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4445                 btrfs_calc_trans_metadata_size(root, 1);
4446
4447         if (left < thresh && btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
4448                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4449                         left, thresh, type);
4450                 dump_space_info(info, 0, 0);
4451         }
4452
4453         if (left < thresh) {
4454                 u64 flags;
4455
4456                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4457                 /*
4458                  * Ignore failure to create system chunk. We might end up not
4459                  * needing it, as we might not need to COW all nodes/leafs from
4460                  * the paths we visit in the chunk tree (they were already COWed
4461                  * or created in the current transaction for example).
4462                  */
4463                 ret = btrfs_alloc_chunk(trans, root, flags);
4464         }
4465
4466         if (!ret) {
4467                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4468                                           &root->fs_info->chunk_block_rsv,
4469                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4470                 if (!ret)
4471                         trans->chunk_bytes_reserved += thresh;
4472         }
4473 }
4474
4475 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4476                           struct btrfs_root *extent_root, u64 flags, int force)
4477 {
4478         struct btrfs_space_info *space_info;
4479         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4480         int wait_for_alloc = 0;
4481         int ret = 0;
4482
4483         /* Don't re-enter if we're already allocating a chunk */
4484         if (trans->allocating_chunk)
4485                 return -ENOSPC;
4486
4487         space_info = __find_space_info(extent_root->fs_info, flags);
4488         if (!space_info) {
4489                 ret = update_space_info(extent_root->fs_info, flags,
4490                                         0, 0, 0, &space_info);
4491                 BUG_ON(ret); /* -ENOMEM */
4492         }
4493         BUG_ON(!space_info); /* Logic error */
4494
4495 again:
4496         spin_lock(&space_info->lock);
4497         if (force < space_info->force_alloc)
4498                 force = space_info->force_alloc;
4499         if (space_info->full) {
4500                 if (should_alloc_chunk(extent_root, space_info, force))
4501                         ret = -ENOSPC;
4502                 else
4503                         ret = 0;
4504                 spin_unlock(&space_info->lock);
4505                 return ret;
4506         }
4507
4508         if (!should_alloc_chunk(extent_root, space_info, force)) {
4509                 spin_unlock(&space_info->lock);
4510                 return 0;
4511         } else if (space_info->chunk_alloc) {
4512                 wait_for_alloc = 1;
4513         } else {
4514                 space_info->chunk_alloc = 1;
4515         }
4516
4517         spin_unlock(&space_info->lock);
4518
4519         mutex_lock(&fs_info->chunk_mutex);
4520
4521         /*
4522          * The chunk_mutex is held throughout the entirety of a chunk
4523          * allocation, so once we've acquired the chunk_mutex we know that the
4524          * other guy is done and we need to recheck and see if we should
4525          * allocate.
4526          */
4527         if (wait_for_alloc) {
4528                 mutex_unlock(&fs_info->chunk_mutex);
4529                 wait_for_alloc = 0;
4530                 goto again;
4531         }
4532
4533         trans->allocating_chunk = true;
4534
4535         /*
4536          * If we have mixed data/metadata chunks we want to make sure we keep
4537          * allocating mixed chunks instead of individual chunks.
4538          */
4539         if (btrfs_mixed_space_info(space_info))
4540                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4541
4542         /*
4543          * if we're doing a data chunk, go ahead and make sure that
4544          * we keep a reasonable number of metadata chunks allocated in the
4545          * FS as well.
4546          */
4547         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4548                 fs_info->data_chunk_allocations++;
4549                 if (!(fs_info->data_chunk_allocations %
4550                       fs_info->metadata_ratio))
4551                         force_metadata_allocation(fs_info);
4552         }
4553
4554         /*
4555          * Check if we have enough space in SYSTEM chunk because we may need
4556          * to update devices.
4557          */
4558         check_system_chunk(trans, extent_root, flags);
4559
4560         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4561         trans->allocating_chunk = false;
4562
4563         spin_lock(&space_info->lock);
4564         if (ret < 0 && ret != -ENOSPC)
4565                 goto out;
4566         if (ret)
4567                 space_info->full = 1;
4568         else
4569                 ret = 1;
4570
4571         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4572 out:
4573         space_info->chunk_alloc = 0;
4574         spin_unlock(&space_info->lock);
4575         mutex_unlock(&fs_info->chunk_mutex);
4576         /*
4577          * When we allocate a new chunk we reserve space in the chunk block
4578          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4579          * add new nodes/leafs to it if we end up needing to do it when
4580          * inserting the chunk item and updating device items as part of the
4581          * second phase of chunk allocation, performed by
4582          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4583          * large number of new block groups to create in our transaction
4584          * handle's new_bgs list to avoid exhausting the chunk block reserve
4585          * in extreme cases - like having a single transaction create many new
4586          * block groups when starting to write out the free space caches of all
4587          * the block groups that were made dirty during the lifetime of the
4588          * transaction.
4589          */
4590         if (trans->can_flush_pending_bgs &&
4591             trans->chunk_bytes_reserved >= (u64)SZ_2M) {
4592                 btrfs_create_pending_block_groups(trans, extent_root);
4593                 btrfs_trans_release_chunk_metadata(trans);
4594         }
4595         return ret;
4596 }
4597
4598 static int can_overcommit(struct btrfs_root *root,
4599                           struct btrfs_space_info *space_info, u64 bytes,
4600                           enum btrfs_reserve_flush_enum flush)
4601 {
4602         struct btrfs_block_rsv *global_rsv;
4603         u64 profile;
4604         u64 space_size;
4605         u64 avail;
4606         u64 used;
4607
4608         /* Don't overcommit when in mixed mode. */
4609         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
4610                 return 0;
4611
4612         BUG_ON(root->fs_info == NULL);
4613         global_rsv = &root->fs_info->global_block_rsv;
4614         profile = btrfs_get_alloc_profile(root, 0);
4615         used = space_info->bytes_used + space_info->bytes_reserved +
4616                 space_info->bytes_pinned + space_info->bytes_readonly;
4617
4618         /*
4619          * We only want to allow over committing if we have lots of actual space
4620          * free, but if we don't have enough space to handle the global reserve
4621          * space then we could end up having a real enospc problem when trying
4622          * to allocate a chunk or some other such important allocation.
4623          */
4624         spin_lock(&global_rsv->lock);
4625         space_size = calc_global_rsv_need_space(global_rsv);
4626         spin_unlock(&global_rsv->lock);
4627         if (used + space_size >= space_info->total_bytes)
4628                 return 0;
4629
4630         used += space_info->bytes_may_use;
4631
4632         spin_lock(&root->fs_info->free_chunk_lock);
4633         avail = root->fs_info->free_chunk_space;
4634         spin_unlock(&root->fs_info->free_chunk_lock);
4635
4636         /*
4637          * If we have dup, raid1 or raid10 then only half of the free
4638          * space is actually useable.  For raid56, the space info used
4639          * doesn't include the parity drive, so we don't have to
4640          * change the math
4641          */
4642         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4643                        BTRFS_BLOCK_GROUP_RAID1 |
4644                        BTRFS_BLOCK_GROUP_RAID10))
4645                 avail >>= 1;
4646
4647         /*
4648          * If we aren't flushing all things, let us overcommit up to
4649          * 1/2th of the space. If we can flush, don't let us overcommit
4650          * too much, let it overcommit up to 1/8 of the space.
4651          */
4652         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4653                 avail >>= 3;
4654         else
4655                 avail >>= 1;
4656
4657         if (used + bytes < space_info->total_bytes + avail)
4658                 return 1;
4659         return 0;
4660 }
4661
4662 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4663                                          unsigned long nr_pages, int nr_items)
4664 {
4665         struct super_block *sb = root->fs_info->sb;
4666
4667         if (down_read_trylock(&sb->s_umount)) {
4668                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4669                 up_read(&sb->s_umount);
4670         } else {
4671                 /*
4672                  * We needn't worry the filesystem going from r/w to r/o though
4673                  * we don't acquire ->s_umount mutex, because the filesystem
4674                  * should guarantee the delalloc inodes list be empty after
4675                  * the filesystem is readonly(all dirty pages are written to
4676                  * the disk).
4677                  */
4678                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4679                 if (!current->journal_info)
4680                         btrfs_wait_ordered_roots(root->fs_info, nr_items,
4681                                                  0, (u64)-1);
4682         }
4683 }
4684
4685 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4686 {
4687         u64 bytes;
4688         int nr;
4689
4690         bytes = btrfs_calc_trans_metadata_size(root, 1);
4691         nr = (int)div64_u64(to_reclaim, bytes);
4692         if (!nr)
4693                 nr = 1;
4694         return nr;
4695 }
4696
4697 #define EXTENT_SIZE_PER_ITEM    SZ_256K
4698
4699 /*
4700  * shrink metadata reservation for delalloc
4701  */
4702 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4703                             bool wait_ordered)
4704 {
4705         struct btrfs_block_rsv *block_rsv;
4706         struct btrfs_space_info *space_info;
4707         struct btrfs_trans_handle *trans;
4708         u64 delalloc_bytes;
4709         u64 max_reclaim;
4710         long time_left;
4711         unsigned long nr_pages;
4712         int loops;
4713         int items;
4714         enum btrfs_reserve_flush_enum flush;
4715
4716         /* Calc the number of the pages we need flush for space reservation */
4717         items = calc_reclaim_items_nr(root, to_reclaim);
4718         to_reclaim = (u64)items * EXTENT_SIZE_PER_ITEM;
4719
4720         trans = (struct btrfs_trans_handle *)current->journal_info;
4721         block_rsv = &root->fs_info->delalloc_block_rsv;
4722         space_info = block_rsv->space_info;
4723
4724         delalloc_bytes = percpu_counter_sum_positive(
4725                                                 &root->fs_info->delalloc_bytes);
4726         if (delalloc_bytes == 0) {
4727                 if (trans)
4728                         return;
4729                 if (wait_ordered)
4730                         btrfs_wait_ordered_roots(root->fs_info, items,
4731                                                  0, (u64)-1);
4732                 return;
4733         }
4734
4735         loops = 0;
4736         while (delalloc_bytes && loops < 3) {
4737                 max_reclaim = min(delalloc_bytes, to_reclaim);
4738                 nr_pages = max_reclaim >> PAGE_SHIFT;
4739                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4740                 /*
4741                  * We need to wait for the async pages to actually start before
4742                  * we do anything.
4743                  */
4744                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4745                 if (!max_reclaim)
4746                         goto skip_async;
4747
4748                 if (max_reclaim <= nr_pages)
4749                         max_reclaim = 0;
4750                 else
4751                         max_reclaim -= nr_pages;
4752
4753                 wait_event(root->fs_info->async_submit_wait,
4754                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4755                            (int)max_reclaim);
4756 skip_async:
4757                 if (!trans)
4758                         flush = BTRFS_RESERVE_FLUSH_ALL;
4759                 else
4760                         flush = BTRFS_RESERVE_NO_FLUSH;
4761                 spin_lock(&space_info->lock);
4762                 if (can_overcommit(root, space_info, orig, flush)) {
4763                         spin_unlock(&space_info->lock);
4764                         break;
4765                 }
4766                 if (list_empty(&space_info->tickets) &&
4767                     list_empty(&space_info->priority_tickets)) {
4768                         spin_unlock(&space_info->lock);
4769                         break;
4770                 }
4771                 spin_unlock(&space_info->lock);
4772
4773                 loops++;
4774                 if (wait_ordered && !trans) {
4775                         btrfs_wait_ordered_roots(root->fs_info, items,
4776                                                  0, (u64)-1);
4777                 } else {
4778                         time_left = schedule_timeout_killable(1);
4779                         if (time_left)
4780                                 break;
4781                 }
4782                 delalloc_bytes = percpu_counter_sum_positive(
4783                                                 &root->fs_info->delalloc_bytes);
4784         }
4785 }
4786
4787 /**
4788  * maybe_commit_transaction - possibly commit the transaction if its ok to
4789  * @root - the root we're allocating for
4790  * @bytes - the number of bytes we want to reserve
4791  * @force - force the commit
4792  *
4793  * This will check to make sure that committing the transaction will actually
4794  * get us somewhere and then commit the transaction if it does.  Otherwise it
4795  * will return -ENOSPC.
4796  */
4797 static int may_commit_transaction(struct btrfs_root *root,
4798                                   struct btrfs_space_info *space_info,
4799                                   u64 bytes, int force)
4800 {
4801         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4802         struct btrfs_trans_handle *trans;
4803
4804         trans = (struct btrfs_trans_handle *)current->journal_info;
4805         if (trans)
4806                 return -EAGAIN;
4807
4808         if (force)
4809                 goto commit;
4810
4811         /* See if there is enough pinned space to make this reservation */
4812         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4813                                    bytes) >= 0)
4814                 goto commit;
4815
4816         /*
4817          * See if there is some space in the delayed insertion reservation for
4818          * this reservation.
4819          */
4820         if (space_info != delayed_rsv->space_info)
4821                 return -ENOSPC;
4822
4823         spin_lock(&delayed_rsv->lock);
4824         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4825                                    bytes - delayed_rsv->size) >= 0) {
4826                 spin_unlock(&delayed_rsv->lock);
4827                 return -ENOSPC;
4828         }
4829         spin_unlock(&delayed_rsv->lock);
4830
4831 commit:
4832         trans = btrfs_join_transaction(root);
4833         if (IS_ERR(trans))
4834                 return -ENOSPC;
4835
4836         return btrfs_commit_transaction(trans, root);
4837 }
4838
4839 struct reserve_ticket {
4840         u64 bytes;
4841         int error;
4842         struct list_head list;
4843         wait_queue_head_t wait;
4844 };
4845
4846 static int flush_space(struct btrfs_root *root,
4847                        struct btrfs_space_info *space_info, u64 num_bytes,
4848                        u64 orig_bytes, int state)
4849 {
4850         struct btrfs_trans_handle *trans;
4851         int nr;
4852         int ret = 0;
4853
4854         switch (state) {
4855         case FLUSH_DELAYED_ITEMS_NR:
4856         case FLUSH_DELAYED_ITEMS:
4857                 if (state == FLUSH_DELAYED_ITEMS_NR)
4858                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4859                 else
4860                         nr = -1;
4861
4862                 trans = btrfs_join_transaction(root);
4863                 if (IS_ERR(trans)) {
4864                         ret = PTR_ERR(trans);
4865                         break;
4866                 }
4867                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4868                 btrfs_end_transaction(trans, root);
4869                 break;
4870         case FLUSH_DELALLOC:
4871         case FLUSH_DELALLOC_WAIT:
4872                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4873                                 state == FLUSH_DELALLOC_WAIT);
4874                 break;
4875         case ALLOC_CHUNK:
4876                 trans = btrfs_join_transaction(root);
4877                 if (IS_ERR(trans)) {
4878                         ret = PTR_ERR(trans);
4879                         break;
4880                 }
4881                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4882                                      btrfs_get_alloc_profile(root, 0),
4883                                      CHUNK_ALLOC_NO_FORCE);
4884                 btrfs_end_transaction(trans, root);
4885                 if (ret == -ENOSPC)
4886                         ret = 0;
4887                 break;
4888         case COMMIT_TRANS:
4889                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4890                 break;
4891         default:
4892                 ret = -ENOSPC;
4893                 break;
4894         }
4895
4896         trace_btrfs_flush_space(root->fs_info, space_info->flags, num_bytes,
4897                                 orig_bytes, state, ret);
4898         return ret;
4899 }
4900
4901 static inline u64
4902 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4903                                  struct btrfs_space_info *space_info)
4904 {
4905         struct reserve_ticket *ticket;
4906         u64 used;
4907         u64 expected;
4908         u64 to_reclaim = 0;
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         list_for_each_entry(ticket, &space_info->tickets, list)
4916                 to_reclaim += ticket->bytes;
4917         list_for_each_entry(ticket, &space_info->priority_tickets, list)
4918                 to_reclaim += ticket->bytes;
4919         if (to_reclaim)
4920                 return to_reclaim;
4921
4922         used = space_info->bytes_used + space_info->bytes_reserved +
4923                space_info->bytes_pinned + space_info->bytes_readonly +
4924                space_info->bytes_may_use;
4925         if (can_overcommit(root, space_info, SZ_1M, BTRFS_RESERVE_FLUSH_ALL))
4926                 expected = div_factor_fine(space_info->total_bytes, 95);
4927         else
4928                 expected = div_factor_fine(space_info->total_bytes, 90);
4929
4930         if (used > expected)
4931                 to_reclaim = used - expected;
4932         else
4933                 to_reclaim = 0;
4934         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4935                                      space_info->bytes_reserved);
4936         return to_reclaim;
4937 }
4938
4939 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4940                                         struct btrfs_root *root, u64 used)
4941 {
4942         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4943
4944         /* If we're just plain full then async reclaim just slows us down. */
4945         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
4946                 return 0;
4947
4948         if (!btrfs_calc_reclaim_metadata_size(root, space_info))
4949                 return 0;
4950
4951         return (used >= thresh && !btrfs_fs_closing(root->fs_info) &&
4952                 !test_bit(BTRFS_FS_STATE_REMOUNTING,
4953                           &root->fs_info->fs_state));
4954 }
4955
4956 static void wake_all_tickets(struct list_head *head)
4957 {
4958         struct reserve_ticket *ticket;
4959
4960         while (!list_empty(head)) {
4961                 ticket = list_first_entry(head, struct reserve_ticket, list);
4962                 list_del_init(&ticket->list);
4963                 ticket->error = -ENOSPC;
4964                 wake_up(&ticket->wait);
4965         }
4966 }
4967
4968 /*
4969  * This is for normal flushers, we can wait all goddamned day if we want to.  We
4970  * will loop and continuously try to flush as long as we are making progress.
4971  * We count progress as clearing off tickets each time we have to loop.
4972  */
4973 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4974 {
4975         struct reserve_ticket *last_ticket = NULL;
4976         struct btrfs_fs_info *fs_info;
4977         struct btrfs_space_info *space_info;
4978         u64 to_reclaim;
4979         int flush_state;
4980         int commit_cycles = 0;
4981
4982         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4983         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4984
4985         spin_lock(&space_info->lock);
4986         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4987                                                       space_info);
4988         if (!to_reclaim) {
4989                 space_info->flush = 0;
4990                 spin_unlock(&space_info->lock);
4991                 return;
4992         }
4993         last_ticket = list_first_entry(&space_info->tickets,
4994                                        struct reserve_ticket, list);
4995         spin_unlock(&space_info->lock);
4996
4997         flush_state = FLUSH_DELAYED_ITEMS_NR;
4998         do {
4999                 struct reserve_ticket *ticket;
5000                 int ret;
5001
5002                 ret = flush_space(fs_info->fs_root, space_info, to_reclaim,
5003                             to_reclaim, flush_state);
5004                 spin_lock(&space_info->lock);
5005                 if (list_empty(&space_info->tickets)) {
5006                         space_info->flush = 0;
5007                         spin_unlock(&space_info->lock);
5008                         return;
5009                 }
5010                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5011                                                               space_info);
5012                 ticket = list_first_entry(&space_info->tickets,
5013                                           struct reserve_ticket, list);
5014                 if (last_ticket == ticket) {
5015                         flush_state++;
5016                 } else {
5017                         last_ticket = ticket;
5018                         flush_state = FLUSH_DELAYED_ITEMS_NR;
5019                         if (commit_cycles)
5020                                 commit_cycles--;
5021                 }
5022
5023                 if (flush_state > COMMIT_TRANS) {
5024                         commit_cycles++;
5025                         if (commit_cycles > 2) {
5026                                 wake_all_tickets(&space_info->tickets);
5027                                 space_info->flush = 0;
5028                         } else {
5029                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
5030                         }
5031                 }
5032                 spin_unlock(&space_info->lock);
5033         } while (flush_state <= COMMIT_TRANS);
5034 }
5035
5036 void btrfs_init_async_reclaim_work(struct work_struct *work)
5037 {
5038         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
5039 }
5040
5041 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
5042                                             struct btrfs_space_info *space_info,
5043                                             struct reserve_ticket *ticket)
5044 {
5045         u64 to_reclaim;
5046         int flush_state = FLUSH_DELAYED_ITEMS_NR;
5047
5048         spin_lock(&space_info->lock);
5049         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
5050                                                       space_info);
5051         if (!to_reclaim) {
5052                 spin_unlock(&space_info->lock);
5053                 return;
5054         }
5055         spin_unlock(&space_info->lock);
5056
5057         do {
5058                 flush_space(fs_info->fs_root, space_info, to_reclaim,
5059                             to_reclaim, flush_state);
5060                 flush_state++;
5061                 spin_lock(&space_info->lock);
5062                 if (ticket->bytes == 0) {
5063                         spin_unlock(&space_info->lock);
5064                         return;
5065                 }
5066                 spin_unlock(&space_info->lock);
5067
5068                 /*
5069                  * Priority flushers can't wait on delalloc without
5070                  * deadlocking.
5071                  */
5072                 if (flush_state == FLUSH_DELALLOC ||
5073                     flush_state == FLUSH_DELALLOC_WAIT)
5074                         flush_state = ALLOC_CHUNK;
5075         } while (flush_state < COMMIT_TRANS);
5076 }
5077
5078 static int wait_reserve_ticket(struct btrfs_fs_info *fs_info,
5079                                struct btrfs_space_info *space_info,
5080                                struct reserve_ticket *ticket, u64 orig_bytes)
5081
5082 {
5083         DEFINE_WAIT(wait);
5084         int ret = 0;
5085
5086         spin_lock(&space_info->lock);
5087         while (ticket->bytes > 0 && ticket->error == 0) {
5088                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
5089                 if (ret) {
5090                         ret = -EINTR;
5091                         break;
5092                 }
5093                 spin_unlock(&space_info->lock);
5094
5095                 schedule();
5096
5097                 finish_wait(&ticket->wait, &wait);
5098                 spin_lock(&space_info->lock);
5099         }
5100         if (!ret)
5101                 ret = ticket->error;
5102         if (!list_empty(&ticket->list))
5103                 list_del_init(&ticket->list);
5104         if (ticket->bytes && ticket->bytes < orig_bytes) {
5105                 u64 num_bytes = orig_bytes - ticket->bytes;
5106                 space_info->bytes_may_use -= num_bytes;
5107                 trace_btrfs_space_reservation(fs_info, "space_info",
5108                                               space_info->flags, num_bytes, 0);
5109         }
5110         spin_unlock(&space_info->lock);
5111
5112         return ret;
5113 }
5114
5115 /**
5116  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5117  * @root - the root we're allocating for
5118  * @space_info - the space info we want to allocate from
5119  * @orig_bytes - the number of bytes we want
5120  * @flush - whether or not we can flush to make our reservation
5121  *
5122  * This will reserve orig_bytes number of bytes from the space info associated
5123  * with the block_rsv.  If there is not enough space it will make an attempt to
5124  * flush out space to make room.  It will do this by flushing delalloc if
5125  * possible or committing the transaction.  If flush is 0 then no attempts to
5126  * regain reservations will be made and this will fail if there is not enough
5127  * space already.
5128  */
5129 static int __reserve_metadata_bytes(struct btrfs_root *root,
5130                                     struct btrfs_space_info *space_info,
5131                                     u64 orig_bytes,
5132                                     enum btrfs_reserve_flush_enum flush)
5133 {
5134         struct reserve_ticket ticket;
5135         u64 used;
5136         int ret = 0;
5137
5138         ASSERT(orig_bytes);
5139         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
5140
5141         spin_lock(&space_info->lock);
5142         ret = -ENOSPC;
5143         used = space_info->bytes_used + space_info->bytes_reserved +
5144                 space_info->bytes_pinned + space_info->bytes_readonly +
5145                 space_info->bytes_may_use;
5146
5147         /*
5148          * If we have enough space then hooray, make our reservation and carry
5149          * on.  If not see if we can overcommit, and if we can, hooray carry on.
5150          * If not things get more complicated.
5151          */
5152         if (used + orig_bytes <= space_info->total_bytes) {
5153                 space_info->bytes_may_use += orig_bytes;
5154                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5155                                               space_info->flags, orig_bytes,
5156                                               1);
5157                 ret = 0;
5158         } else if (can_overcommit(root, space_info, orig_bytes, flush)) {
5159                 space_info->bytes_may_use += orig_bytes;
5160                 trace_btrfs_space_reservation(root->fs_info, "space_info",
5161                                               space_info->flags, orig_bytes,
5162                                               1);
5163                 ret = 0;
5164         }
5165
5166         /*
5167          * If we couldn't make a reservation then setup our reservation ticket
5168          * and kick the async worker if it's not already running.
5169          *
5170          * If we are a priority flusher then we just need to add our ticket to
5171          * the list and we will do our own flushing further down.
5172          */
5173         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
5174                 ticket.bytes = orig_bytes;
5175                 ticket.error = 0;
5176                 init_waitqueue_head(&ticket.wait);
5177                 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
5178                         list_add_tail(&ticket.list, &space_info->tickets);
5179                         if (!space_info->flush) {
5180                                 space_info->flush = 1;
5181                                 trace_btrfs_trigger_flush(root->fs_info,
5182                                                           space_info->flags,
5183                                                           orig_bytes, flush,
5184                                                           "enospc");
5185                                 queue_work(system_unbound_wq,
5186                                            &root->fs_info->async_reclaim_work);
5187                         }
5188                 } else {
5189                         list_add_tail(&ticket.list,
5190                                       &space_info->priority_tickets);
5191                 }
5192         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
5193                 used += orig_bytes;
5194                 /*
5195                  * We will do the space reservation dance during log replay,
5196                  * which means we won't have fs_info->fs_root set, so don't do
5197                  * the async reclaim as we will panic.
5198                  */
5199                 if (!root->fs_info->log_root_recovering &&
5200                     need_do_async_reclaim(space_info, root, used) &&
5201                     !work_busy(&root->fs_info->async_reclaim_work)) {
5202                         trace_btrfs_trigger_flush(root->fs_info,
5203                                                   space_info->flags,
5204                                                   orig_bytes, flush,
5205                                                   "preempt");
5206                         queue_work(system_unbound_wq,
5207                                    &root->fs_info->async_reclaim_work);
5208                 }
5209         }
5210         spin_unlock(&space_info->lock);
5211         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5212                 return ret;
5213
5214         if (flush == BTRFS_RESERVE_FLUSH_ALL)
5215                 return wait_reserve_ticket(root->fs_info, space_info, &ticket,
5216                                            orig_bytes);
5217
5218         ret = 0;
5219         priority_reclaim_metadata_space(root->fs_info, space_info, &ticket);
5220         spin_lock(&space_info->lock);
5221         if (ticket.bytes) {
5222                 if (ticket.bytes < orig_bytes) {
5223                         u64 num_bytes = orig_bytes - ticket.bytes;
5224                         space_info->bytes_may_use -= num_bytes;
5225                         trace_btrfs_space_reservation(root->fs_info,
5226                                         "space_info", space_info->flags,
5227                                         num_bytes, 0);
5228
5229                 }
5230                 list_del_init(&ticket.list);
5231                 ret = -ENOSPC;
5232         }
5233         spin_unlock(&space_info->lock);
5234         ASSERT(list_empty(&ticket.list));
5235         return ret;
5236 }
5237
5238 /**
5239  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
5240  * @root - the root we're allocating for
5241  * @block_rsv - the block_rsv we're allocating for
5242  * @orig_bytes - the number of bytes we want
5243  * @flush - whether or not we can flush to make our reservation
5244  *
5245  * This will reserve orgi_bytes number of bytes from the space info associated
5246  * with the block_rsv.  If there is not enough space it will make an attempt to
5247  * flush out space to make room.  It will do this by flushing delalloc if
5248  * possible or committing the transaction.  If flush is 0 then no attempts to
5249  * regain reservations will be made and this will fail if there is not enough
5250  * space already.
5251  */
5252 static int reserve_metadata_bytes(struct btrfs_root *root,
5253                                   struct btrfs_block_rsv *block_rsv,
5254                                   u64 orig_bytes,
5255                                   enum btrfs_reserve_flush_enum flush)
5256 {
5257         int ret;
5258
5259         ret = __reserve_metadata_bytes(root, block_rsv->space_info, orig_bytes,
5260                                        flush);
5261         if (ret == -ENOSPC &&
5262             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5263                 struct btrfs_block_rsv *global_rsv =
5264                         &root->fs_info->global_block_rsv;
5265
5266                 if (block_rsv != global_rsv &&
5267                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5268                         ret = 0;
5269         }
5270         if (ret == -ENOSPC)
5271                 trace_btrfs_space_reservation(root->fs_info,
5272                                               "space_info:enospc",
5273                                               block_rsv->space_info->flags,
5274                                               orig_bytes, 1);
5275         return ret;
5276 }
5277
5278 static struct btrfs_block_rsv *get_block_rsv(
5279                                         const struct btrfs_trans_handle *trans,
5280                                         const struct btrfs_root *root)
5281 {
5282         struct btrfs_block_rsv *block_rsv = NULL;
5283
5284         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5285             (root == root->fs_info->csum_root && trans->adding_csums) ||
5286              (root == root->fs_info->uuid_root))
5287                 block_rsv = trans->block_rsv;
5288
5289         if (!block_rsv)
5290                 block_rsv = root->block_rsv;
5291
5292         if (!block_rsv)
5293                 block_rsv = &root->fs_info->empty_block_rsv;
5294
5295         return block_rsv;
5296 }
5297
5298 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5299                                u64 num_bytes)
5300 {
5301         int ret = -ENOSPC;
5302         spin_lock(&block_rsv->lock);
5303         if (block_rsv->reserved >= num_bytes) {
5304                 block_rsv->reserved -= num_bytes;
5305                 if (block_rsv->reserved < block_rsv->size)
5306                         block_rsv->full = 0;
5307                 ret = 0;
5308         }
5309         spin_unlock(&block_rsv->lock);
5310         return ret;
5311 }
5312
5313 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5314                                 u64 num_bytes, int update_size)
5315 {
5316         spin_lock(&block_rsv->lock);
5317         block_rsv->reserved += num_bytes;
5318         if (update_size)
5319                 block_rsv->size += num_bytes;
5320         else if (block_rsv->reserved >= block_rsv->size)
5321                 block_rsv->full = 1;
5322         spin_unlock(&block_rsv->lock);
5323 }
5324
5325 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5326                              struct btrfs_block_rsv *dest, u64 num_bytes,
5327                              int min_factor)
5328 {
5329         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5330         u64 min_bytes;
5331
5332         if (global_rsv->space_info != dest->space_info)
5333                 return -ENOSPC;
5334
5335         spin_lock(&global_rsv->lock);
5336         min_bytes = div_factor(global_rsv->size, min_factor);
5337         if (global_rsv->reserved < min_bytes + num_bytes) {
5338                 spin_unlock(&global_rsv->lock);
5339                 return -ENOSPC;
5340         }
5341         global_rsv->reserved -= num_bytes;
5342         if (global_rsv->reserved < global_rsv->size)
5343                 global_rsv->full = 0;
5344         spin_unlock(&global_rsv->lock);
5345
5346         block_rsv_add_bytes(dest, num_bytes, 1);
5347         return 0;
5348 }
5349
5350 /*
5351  * This is for space we already have accounted in space_info->bytes_may_use, so
5352  * basically when we're returning space from block_rsv's.
5353  */
5354 static void space_info_add_old_bytes(struct btrfs_fs_info *fs_info,
5355                                      struct btrfs_space_info *space_info,
5356                                      u64 num_bytes)
5357 {
5358         struct reserve_ticket *ticket;
5359         struct list_head *head;
5360         u64 used;
5361         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
5362         bool check_overcommit = false;
5363
5364         spin_lock(&space_info->lock);
5365         head = &space_info->priority_tickets;
5366
5367         /*
5368          * If we are over our limit then we need to check and see if we can
5369          * overcommit, and if we can't then we just need to free up our space
5370          * and not satisfy any requests.
5371          */
5372         used = space_info->bytes_used + space_info->bytes_reserved +
5373                 space_info->bytes_pinned + space_info->bytes_readonly +
5374                 space_info->bytes_may_use;
5375         if (used - num_bytes >= space_info->total_bytes)
5376                 check_overcommit = true;
5377 again:
5378         while (!list_empty(head) && num_bytes) {
5379                 ticket = list_first_entry(head, struct reserve_ticket,
5380                                           list);
5381                 /*
5382                  * We use 0 bytes because this space is already reserved, so
5383                  * adding the ticket space would be a double count.
5384                  */
5385                 if (check_overcommit &&
5386                     !can_overcommit(fs_info->extent_root, space_info, 0,
5387                                     flush))
5388                         break;
5389                 if (num_bytes >= ticket->bytes) {
5390                         list_del_init(&ticket->list);
5391                         num_bytes -= ticket->bytes;
5392                         ticket->bytes = 0;
5393                         wake_up(&ticket->wait);
5394                 } else {
5395                         ticket->bytes -= num_bytes;
5396                         num_bytes = 0;
5397                 }
5398         }
5399
5400         if (num_bytes && head == &space_info->priority_tickets) {
5401                 head = &space_info->tickets;
5402                 flush = BTRFS_RESERVE_FLUSH_ALL;
5403                 goto again;
5404         }
5405         space_info->bytes_may_use -= num_bytes;
5406         trace_btrfs_space_reservation(fs_info, "space_info",
5407                                       space_info->flags, num_bytes, 0);
5408         spin_unlock(&space_info->lock);
5409 }
5410
5411 /*
5412  * This is for newly allocated space that isn't accounted in
5413  * space_info->bytes_may_use yet.  So if we allocate a chunk or unpin an extent
5414  * we use this helper.
5415  */
5416 static void space_info_add_new_bytes(struct btrfs_fs_info *fs_info,
5417                                      struct btrfs_space_info *space_info,
5418                                      u64 num_bytes)
5419 {
5420         struct reserve_ticket *ticket;
5421         struct list_head *head = &space_info->priority_tickets;
5422
5423 again:
5424         while (!list_empty(head) && num_bytes) {
5425                 ticket = list_first_entry(head, struct reserve_ticket,
5426                                           list);
5427                 if (num_bytes >= ticket->bytes) {
5428                         trace_btrfs_space_reservation(fs_info, "space_info",
5429                                                       space_info->flags,
5430                                                       ticket->bytes, 1);
5431                         list_del_init(&ticket->list);
5432                         num_bytes -= ticket->bytes;
5433                         space_info->bytes_may_use += ticket->bytes;
5434                         ticket->bytes = 0;
5435                         wake_up(&ticket->wait);
5436                 } else {
5437                         trace_btrfs_space_reservation(fs_info, "space_info",
5438                                                       space_info->flags,
5439                                                       num_bytes, 1);
5440                         space_info->bytes_may_use += num_bytes;
5441                         ticket->bytes -= num_bytes;
5442                         num_bytes = 0;
5443                 }
5444         }
5445
5446         if (num_bytes && head == &space_info->priority_tickets) {
5447                 head = &space_info->tickets;
5448                 goto again;
5449         }
5450 }
5451
5452 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5453                                     struct btrfs_block_rsv *block_rsv,
5454                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5455 {
5456         struct btrfs_space_info *space_info = block_rsv->space_info;
5457
5458         spin_lock(&block_rsv->lock);
5459         if (num_bytes == (u64)-1)
5460                 num_bytes = block_rsv->size;
5461         block_rsv->size -= num_bytes;
5462         if (block_rsv->reserved >= block_rsv->size) {
5463                 num_bytes = block_rsv->reserved - block_rsv->size;
5464                 block_rsv->reserved = block_rsv->size;
5465                 block_rsv->full = 1;
5466         } else {
5467                 num_bytes = 0;
5468         }
5469         spin_unlock(&block_rsv->lock);
5470
5471         if (num_bytes > 0) {
5472                 if (dest) {
5473                         spin_lock(&dest->lock);
5474                         if (!dest->full) {
5475                                 u64 bytes_to_add;
5476
5477                                 bytes_to_add = dest->size - dest->reserved;
5478                                 bytes_to_add = min(num_bytes, bytes_to_add);
5479                                 dest->reserved += bytes_to_add;
5480                                 if (dest->reserved >= dest->size)
5481                                         dest->full = 1;
5482                                 num_bytes -= bytes_to_add;
5483                         }
5484                         spin_unlock(&dest->lock);
5485                 }
5486                 if (num_bytes)
5487                         space_info_add_old_bytes(fs_info, space_info,
5488                                                  num_bytes);
5489         }
5490 }
5491
5492 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src,
5493                             struct btrfs_block_rsv *dst, u64 num_bytes,
5494                             int update_size)
5495 {
5496         int ret;
5497
5498         ret = block_rsv_use_bytes(src, num_bytes);
5499         if (ret)
5500                 return ret;
5501
5502         block_rsv_add_bytes(dst, num_bytes, update_size);
5503         return 0;
5504 }
5505
5506 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5507 {
5508         memset(rsv, 0, sizeof(*rsv));
5509         spin_lock_init(&rsv->lock);
5510         rsv->type = type;
5511 }
5512
5513 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5514                                               unsigned short type)
5515 {
5516         struct btrfs_block_rsv *block_rsv;
5517         struct btrfs_fs_info *fs_info = root->fs_info;
5518
5519         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5520         if (!block_rsv)
5521                 return NULL;
5522
5523         btrfs_init_block_rsv(block_rsv, type);
5524         block_rsv->space_info = __find_space_info(fs_info,
5525                                                   BTRFS_BLOCK_GROUP_METADATA);
5526         return block_rsv;
5527 }
5528
5529 void btrfs_free_block_rsv(struct btrfs_root *root,
5530                           struct btrfs_block_rsv *rsv)
5531 {
5532         if (!rsv)
5533                 return;
5534         btrfs_block_rsv_release(root, rsv, (u64)-1);
5535         kfree(rsv);
5536 }
5537
5538 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5539 {
5540         kfree(rsv);
5541 }
5542
5543 int btrfs_block_rsv_add(struct btrfs_root *root,
5544                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5545                         enum btrfs_reserve_flush_enum flush)
5546 {
5547         int ret;
5548
5549         if (num_bytes == 0)
5550                 return 0;
5551
5552         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5553         if (!ret) {
5554                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5555                 return 0;
5556         }
5557
5558         return ret;
5559 }
5560
5561 int btrfs_block_rsv_check(struct btrfs_root *root,
5562                           struct btrfs_block_rsv *block_rsv, int min_factor)
5563 {
5564         u64 num_bytes = 0;
5565         int ret = -ENOSPC;
5566
5567         if (!block_rsv)
5568                 return 0;
5569
5570         spin_lock(&block_rsv->lock);
5571         num_bytes = div_factor(block_rsv->size, min_factor);
5572         if (block_rsv->reserved >= num_bytes)
5573                 ret = 0;
5574         spin_unlock(&block_rsv->lock);
5575
5576         return ret;
5577 }
5578
5579 int btrfs_block_rsv_refill(struct btrfs_root *root,
5580                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5581                            enum btrfs_reserve_flush_enum flush)
5582 {
5583         u64 num_bytes = 0;
5584         int ret = -ENOSPC;
5585
5586         if (!block_rsv)
5587                 return 0;
5588
5589         spin_lock(&block_rsv->lock);
5590         num_bytes = min_reserved;
5591         if (block_rsv->reserved >= num_bytes)
5592                 ret = 0;
5593         else
5594                 num_bytes -= block_rsv->reserved;
5595         spin_unlock(&block_rsv->lock);
5596
5597         if (!ret)
5598                 return 0;
5599
5600         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5601         if (!ret) {
5602                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5603                 return 0;
5604         }
5605
5606         return ret;
5607 }
5608
5609 void btrfs_block_rsv_release(struct btrfs_root *root,
5610                              struct btrfs_block_rsv *block_rsv,
5611                              u64 num_bytes)
5612 {
5613         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5614         if (global_rsv == block_rsv ||
5615             block_rsv->space_info != global_rsv->space_info)
5616                 global_rsv = NULL;
5617         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5618                                 num_bytes);
5619 }
5620
5621 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5622 {
5623         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5624         struct btrfs_space_info *sinfo = block_rsv->space_info;
5625         u64 num_bytes;
5626
5627         /*
5628          * The global block rsv is based on the size of the extent tree, the
5629          * checksum tree and the root tree.  If the fs is empty we want to set
5630          * it to a minimal amount for safety.
5631          */
5632         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item) +
5633                 btrfs_root_used(&fs_info->csum_root->root_item) +
5634                 btrfs_root_used(&fs_info->tree_root->root_item);
5635         num_bytes = max_t(u64, num_bytes, SZ_16M);
5636
5637         spin_lock(&sinfo->lock);
5638         spin_lock(&block_rsv->lock);
5639
5640         block_rsv->size = min_t(u64, num_bytes, SZ_512M);
5641
5642         if (block_rsv->reserved < block_rsv->size) {
5643                 num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5644                         sinfo->bytes_reserved + sinfo->bytes_readonly +
5645                         sinfo->bytes_may_use;
5646                 if (sinfo->total_bytes > num_bytes) {
5647                         num_bytes = sinfo->total_bytes - num_bytes;
5648                         num_bytes = min(num_bytes,
5649                                         block_rsv->size - block_rsv->reserved);
5650                         block_rsv->reserved += num_bytes;
5651                         sinfo->bytes_may_use += num_bytes;
5652                         trace_btrfs_space_reservation(fs_info, "space_info",
5653                                                       sinfo->flags, num_bytes,
5654                                                       1);
5655                 }
5656         } else if (block_rsv->reserved > block_rsv->size) {
5657                 num_bytes = block_rsv->reserved - block_rsv->size;
5658                 sinfo->bytes_may_use -= num_bytes;
5659                 trace_btrfs_space_reservation(fs_info, "space_info",
5660                                       sinfo->flags, num_bytes, 0);
5661                 block_rsv->reserved = block_rsv->size;
5662         }
5663
5664         if (block_rsv->reserved == block_rsv->size)
5665                 block_rsv->full = 1;
5666         else
5667                 block_rsv->full = 0;
5668
5669         spin_unlock(&block_rsv->lock);
5670         spin_unlock(&sinfo->lock);
5671 }
5672
5673 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5674 {
5675         struct btrfs_space_info *space_info;
5676
5677         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5678         fs_info->chunk_block_rsv.space_info = space_info;
5679
5680         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5681         fs_info->global_block_rsv.space_info = space_info;
5682         fs_info->delalloc_block_rsv.space_info = space_info;
5683         fs_info->trans_block_rsv.space_info = space_info;
5684         fs_info->empty_block_rsv.space_info = space_info;
5685         fs_info->delayed_block_rsv.space_info = space_info;
5686
5687         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5688         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5689         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5690         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5691         if (fs_info->quota_root)
5692                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5693         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5694
5695         update_global_block_rsv(fs_info);
5696 }
5697
5698 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5699 {
5700         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5701                                 (u64)-1);
5702         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5703         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5704         WARN_ON(fs_info->trans_block_rsv.size > 0);
5705         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5706         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5707         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5708         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5709         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5710 }
5711
5712 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5713                                   struct btrfs_root *root)
5714 {
5715         if (!trans->block_rsv)
5716                 return;
5717
5718         if (!trans->bytes_reserved)
5719                 return;
5720
5721         trace_btrfs_space_reservation(root->fs_info, "transaction",
5722                                       trans->transid, trans->bytes_reserved, 0);
5723         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5724         trans->bytes_reserved = 0;
5725 }
5726
5727 /*
5728  * To be called after all the new block groups attached to the transaction
5729  * handle have been created (btrfs_create_pending_block_groups()).
5730  */
5731 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5732 {
5733         struct btrfs_fs_info *fs_info = trans->fs_info;
5734
5735         if (!trans->chunk_bytes_reserved)
5736                 return;
5737
5738         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5739
5740         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5741                                 trans->chunk_bytes_reserved);
5742         trans->chunk_bytes_reserved = 0;
5743 }
5744
5745 /* Can only return 0 or -ENOSPC */
5746 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5747                                   struct inode *inode)
5748 {
5749         struct btrfs_root *root = BTRFS_I(inode)->root;
5750         /*
5751          * We always use trans->block_rsv here as we will have reserved space
5752          * for our orphan when starting the transaction, using get_block_rsv()
5753          * here will sometimes make us choose the wrong block rsv as we could be
5754          * doing a reloc inode for a non refcounted root.
5755          */
5756         struct btrfs_block_rsv *src_rsv = trans->block_rsv;
5757         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5758
5759         /*
5760          * We need to hold space in order to delete our orphan item once we've
5761          * added it, so this takes the reservation so we can release it later
5762          * when we are truly done with the orphan item.
5763          */
5764         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5765         trace_btrfs_space_reservation(root->fs_info, "orphan",
5766                                       btrfs_ino(inode), num_bytes, 1);
5767         return btrfs_block_rsv_migrate(src_rsv, dst_rsv, num_bytes, 1);
5768 }
5769
5770 void btrfs_orphan_release_metadata(struct inode *inode)
5771 {
5772         struct btrfs_root *root = BTRFS_I(inode)->root;
5773         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5774         trace_btrfs_space_reservation(root->fs_info, "orphan",
5775                                       btrfs_ino(inode), num_bytes, 0);
5776         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5777 }
5778
5779 /*
5780  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5781  * root: the root of the parent directory
5782  * rsv: block reservation
5783  * items: the number of items that we need do reservation
5784  * qgroup_reserved: used to return the reserved size in qgroup
5785  *
5786  * This function is used to reserve the space for snapshot/subvolume
5787  * creation and deletion. Those operations are different with the
5788  * common file/directory operations, they change two fs/file trees
5789  * and root tree, the number of items that the qgroup reserves is
5790  * different with the free space reservation. So we can not use
5791  * the space reservation mechanism in start_transaction().
5792  */
5793 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5794                                      struct btrfs_block_rsv *rsv,
5795                                      int items,
5796                                      u64 *qgroup_reserved,
5797                                      bool use_global_rsv)
5798 {
5799         u64 num_bytes;
5800         int ret;
5801         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5802
5803         if (root->fs_info->quota_enabled) {
5804                 /* One for parent inode, two for dir entries */
5805                 num_bytes = 3 * root->nodesize;
5806                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5807                 if (ret)
5808                         return ret;
5809         } else {
5810                 num_bytes = 0;
5811         }
5812
5813         *qgroup_reserved = num_bytes;
5814
5815         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5816         rsv->space_info = __find_space_info(root->fs_info,
5817                                             BTRFS_BLOCK_GROUP_METADATA);
5818         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5819                                   BTRFS_RESERVE_FLUSH_ALL);
5820
5821         if (ret == -ENOSPC && use_global_rsv)
5822                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes, 1);
5823
5824         if (ret && *qgroup_reserved)
5825                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5826
5827         return ret;
5828 }
5829
5830 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5831                                       struct btrfs_block_rsv *rsv,
5832                                       u64 qgroup_reserved)
5833 {
5834         btrfs_block_rsv_release(root, rsv, (u64)-1);
5835 }
5836
5837 /**
5838  * drop_outstanding_extent - drop an outstanding extent
5839  * @inode: the inode we're dropping the extent for
5840  * @num_bytes: the number of bytes we're releasing.
5841  *
5842  * This is called when we are freeing up an outstanding extent, either called
5843  * after an error or after an extent is written.  This will return the number of
5844  * reserved extents that need to be freed.  This must be called with
5845  * BTRFS_I(inode)->lock held.
5846  */
5847 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5848 {
5849         unsigned drop_inode_space = 0;
5850         unsigned dropped_extents = 0;
5851         unsigned num_extents = 0;
5852
5853         num_extents = (unsigned)div64_u64(num_bytes +
5854                                           BTRFS_MAX_EXTENT_SIZE - 1,
5855                                           BTRFS_MAX_EXTENT_SIZE);
5856         ASSERT(num_extents);
5857         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5858         BTRFS_I(inode)->outstanding_extents -= num_extents;
5859
5860         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5861             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5862                                &BTRFS_I(inode)->runtime_flags))
5863                 drop_inode_space = 1;
5864
5865         /*
5866          * If we have more or the same amount of outstanding extents than we have
5867          * reserved then we need to leave the reserved extents count alone.
5868          */
5869         if (BTRFS_I(inode)->outstanding_extents >=
5870             BTRFS_I(inode)->reserved_extents)
5871                 return drop_inode_space;
5872
5873         dropped_extents = BTRFS_I(inode)->reserved_extents -
5874                 BTRFS_I(inode)->outstanding_extents;
5875         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5876         return dropped_extents + drop_inode_space;
5877 }
5878
5879 /**
5880  * calc_csum_metadata_size - return the amount of metadata space that must be
5881  *      reserved/freed for the given bytes.
5882  * @inode: the inode we're manipulating
5883  * @num_bytes: the number of bytes in question
5884  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5885  *
5886  * This adjusts the number of csum_bytes in the inode and then returns the
5887  * correct amount of metadata that must either be reserved or freed.  We
5888  * calculate how many checksums we can fit into one leaf and then divide the
5889  * number of bytes that will need to be checksumed by this value to figure out
5890  * how many checksums will be required.  If we are adding bytes then the number
5891  * may go up and we will return the number of additional bytes that must be
5892  * reserved.  If it is going down we will return the number of bytes that must
5893  * be freed.
5894  *
5895  * This must be called with BTRFS_I(inode)->lock held.
5896  */
5897 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5898                                    int reserve)
5899 {
5900         struct btrfs_root *root = BTRFS_I(inode)->root;
5901         u64 old_csums, num_csums;
5902
5903         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5904             BTRFS_I(inode)->csum_bytes == 0)
5905                 return 0;
5906
5907         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5908         if (reserve)
5909                 BTRFS_I(inode)->csum_bytes += num_bytes;
5910         else
5911                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5912         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5913
5914         /* No change, no need to reserve more */
5915         if (old_csums == num_csums)
5916                 return 0;
5917
5918         if (reserve)
5919                 return btrfs_calc_trans_metadata_size(root,
5920                                                       num_csums - old_csums);
5921
5922         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5923 }
5924
5925 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5926 {
5927         struct btrfs_root *root = BTRFS_I(inode)->root;
5928         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5929         u64 to_reserve = 0;
5930         u64 csum_bytes;
5931         unsigned nr_extents = 0;
5932         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5933         int ret = 0;
5934         bool delalloc_lock = true;
5935         u64 to_free = 0;
5936         unsigned dropped;
5937         bool release_extra = false;
5938
5939         /* If we are a free space inode we need to not flush since we will be in
5940          * the middle of a transaction commit.  We also don't need the delalloc
5941          * mutex since we won't race with anybody.  We need this mostly to make
5942          * lockdep shut its filthy mouth.
5943          *
5944          * If we have a transaction open (can happen if we call truncate_block
5945          * from truncate), then we need FLUSH_LIMIT so we don't deadlock.
5946          */
5947         if (btrfs_is_free_space_inode(inode)) {
5948                 flush = BTRFS_RESERVE_NO_FLUSH;
5949                 delalloc_lock = false;
5950         } else if (current->journal_info) {
5951                 flush = BTRFS_RESERVE_FLUSH_LIMIT;
5952         }
5953
5954         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5955             btrfs_transaction_in_commit(root->fs_info))
5956                 schedule_timeout(1);
5957
5958         if (delalloc_lock)
5959                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5960
5961         num_bytes = ALIGN(num_bytes, root->sectorsize);
5962
5963         spin_lock(&BTRFS_I(inode)->lock);
5964         nr_extents = (unsigned)div64_u64(num_bytes +
5965                                          BTRFS_MAX_EXTENT_SIZE - 1,
5966                                          BTRFS_MAX_EXTENT_SIZE);
5967         BTRFS_I(inode)->outstanding_extents += nr_extents;
5968
5969         nr_extents = 0;
5970         if (BTRFS_I(inode)->outstanding_extents >
5971             BTRFS_I(inode)->reserved_extents)
5972                 nr_extents += BTRFS_I(inode)->outstanding_extents -
5973                         BTRFS_I(inode)->reserved_extents;
5974
5975         /* We always want to reserve a slot for updating the inode. */
5976         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents + 1);
5977         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5978         csum_bytes = BTRFS_I(inode)->csum_bytes;
5979         spin_unlock(&BTRFS_I(inode)->lock);
5980
5981         if (root->fs_info->quota_enabled) {
5982                 ret = btrfs_qgroup_reserve_meta(root,
5983                                 nr_extents * root->nodesize);
5984                 if (ret)
5985                         goto out_fail;
5986         }
5987
5988         ret = btrfs_block_rsv_add(root, block_rsv, to_reserve, flush);
5989         if (unlikely(ret)) {
5990                 btrfs_qgroup_free_meta(root, nr_extents * root->nodesize);
5991                 goto out_fail;
5992         }
5993
5994         spin_lock(&BTRFS_I(inode)->lock);
5995         if (test_and_set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5996                              &BTRFS_I(inode)->runtime_flags)) {
5997                 to_reserve -= btrfs_calc_trans_metadata_size(root, 1);
5998                 release_extra = true;
5999         }
6000         BTRFS_I(inode)->reserved_extents += nr_extents;
6001         spin_unlock(&BTRFS_I(inode)->lock);
6002
6003         if (delalloc_lock)
6004                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6005
6006         if (to_reserve)
6007                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
6008                                               btrfs_ino(inode), to_reserve, 1);
6009         if (release_extra)
6010                 btrfs_block_rsv_release(root, block_rsv,
6011                                         btrfs_calc_trans_metadata_size(root,
6012                                                                        1));
6013         return 0;
6014
6015 out_fail:
6016         spin_lock(&BTRFS_I(inode)->lock);
6017         dropped = drop_outstanding_extent(inode, num_bytes);
6018         /*
6019          * If the inodes csum_bytes is the same as the original
6020          * csum_bytes then we know we haven't raced with any free()ers
6021          * so we can just reduce our inodes csum bytes and carry on.
6022          */
6023         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
6024                 calc_csum_metadata_size(inode, num_bytes, 0);
6025         } else {
6026                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
6027                 u64 bytes;
6028
6029                 /*
6030                  * This is tricky, but first we need to figure out how much we
6031                  * freed from any free-ers that occurred during this
6032                  * reservation, so we reset ->csum_bytes to the csum_bytes
6033                  * before we dropped our lock, and then call the free for the
6034                  * number of bytes that were freed while we were trying our
6035                  * reservation.
6036                  */
6037                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
6038                 BTRFS_I(inode)->csum_bytes = csum_bytes;
6039                 to_free = calc_csum_metadata_size(inode, bytes, 0);
6040
6041
6042                 /*
6043                  * Now we need to see how much we would have freed had we not
6044                  * been making this reservation and our ->csum_bytes were not
6045                  * artificially inflated.
6046                  */
6047                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
6048                 bytes = csum_bytes - orig_csum_bytes;
6049                 bytes = calc_csum_metadata_size(inode, bytes, 0);
6050
6051                 /*
6052                  * Now reset ->csum_bytes to what it should be.  If bytes is
6053                  * more than to_free then we would have freed more space had we
6054                  * not had an artificially high ->csum_bytes, so we need to free
6055                  * the remainder.  If bytes is the same or less then we don't
6056                  * need to do anything, the other free-ers did the correct
6057                  * thing.
6058                  */
6059                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
6060                 if (bytes > to_free)
6061                         to_free = bytes - to_free;
6062                 else
6063                         to_free = 0;
6064         }
6065         spin_unlock(&BTRFS_I(inode)->lock);
6066         if (dropped)
6067                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
6068
6069         if (to_free) {
6070                 btrfs_block_rsv_release(root, block_rsv, to_free);
6071                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
6072                                               btrfs_ino(inode), to_free, 0);
6073         }
6074         if (delalloc_lock)
6075                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
6076         return ret;
6077 }
6078
6079 /**
6080  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
6081  * @inode: the inode to release the reservation for
6082  * @num_bytes: the number of bytes we're releasing
6083  *
6084  * This will release the metadata reservation for an inode.  This can be called
6085  * once we complete IO for a given set of bytes to release their metadata
6086  * reservations.
6087  */
6088 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
6089 {
6090         struct btrfs_root *root = BTRFS_I(inode)->root;
6091         u64 to_free = 0;
6092         unsigned dropped;
6093
6094         num_bytes = ALIGN(num_bytes, root->sectorsize);
6095         spin_lock(&BTRFS_I(inode)->lock);
6096         dropped = drop_outstanding_extent(inode, num_bytes);
6097
6098         if (num_bytes)
6099                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
6100         spin_unlock(&BTRFS_I(inode)->lock);
6101         if (dropped > 0)
6102                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
6103
6104         if (btrfs_is_testing(root->fs_info))
6105                 return;
6106
6107         trace_btrfs_space_reservation(root->fs_info, "delalloc",
6108                                       btrfs_ino(inode), to_free, 0);
6109
6110         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
6111                                 to_free);
6112 }
6113
6114 /**
6115  * btrfs_delalloc_reserve_space - reserve data and metadata space for
6116  * delalloc
6117  * @inode: inode we're writing to
6118  * @start: start range we are writing to
6119  * @len: how long the range we are writing to
6120  *
6121  * TODO: This function will finally replace old btrfs_delalloc_reserve_space()
6122  *
6123  * This will do the following things
6124  *
6125  * o reserve space in data space info for num bytes
6126  *   and reserve precious corresponding qgroup space
6127  *   (Done in check_data_free_space)
6128  *
6129  * o reserve space for metadata space, based on the number of outstanding
6130  *   extents and how much csums will be needed
6131  *   also reserve metadata space in a per root over-reserve method.
6132  * o add to the inodes->delalloc_bytes
6133  * o add it to the fs_info's delalloc inodes list.
6134  *   (Above 3 all done in delalloc_reserve_metadata)
6135  *
6136  * Return 0 for success
6137  * Return <0 for error(-ENOSPC or -EQUOT)
6138  */
6139 int btrfs_delalloc_reserve_space(struct inode *inode, u64 start, u64 len)
6140 {
6141         int ret;
6142
6143         ret = btrfs_check_data_free_space(inode, start, len);
6144         if (ret < 0)
6145                 return ret;
6146         ret = btrfs_delalloc_reserve_metadata(inode, len);
6147         if (ret < 0)
6148                 btrfs_free_reserved_data_space(inode, start, len);
6149         return ret;
6150 }
6151
6152 /**
6153  * btrfs_delalloc_release_space - release data and metadata space for delalloc
6154  * @inode: inode we're releasing space for
6155  * @start: start position of the space already reserved
6156  * @len: the len of the space already reserved
6157  *
6158  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
6159  * called in the case that we don't need the metadata AND data reservations
6160  * anymore.  So if there is an error or we insert an inline extent.
6161  *
6162  * This function will release the metadata space that was not used and will
6163  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
6164  * list if there are no delalloc bytes left.
6165  * Also it will handle the qgroup reserved space.
6166  */
6167 void btrfs_delalloc_release_space(struct inode *inode, u64 start, u64 len)
6168 {
6169         btrfs_delalloc_release_metadata(inode, len);
6170         btrfs_free_reserved_data_space(inode, start, len);
6171 }
6172
6173 static int update_block_group(struct btrfs_trans_handle *trans,
6174                               struct btrfs_root *root, u64 bytenr,
6175                               u64 num_bytes, int alloc)
6176 {
6177         struct btrfs_block_group_cache *cache = NULL;
6178         struct btrfs_fs_info *info = root->fs_info;
6179         u64 total = num_bytes;
6180         u64 old_val;
6181         u64 byte_in_group;
6182         int factor;
6183
6184         /* block accounting for super block */
6185         spin_lock(&info->delalloc_root_lock);
6186         old_val = btrfs_super_bytes_used(info->super_copy);
6187         if (alloc)
6188                 old_val += num_bytes;
6189         else
6190                 old_val -= num_bytes;
6191         btrfs_set_super_bytes_used(info->super_copy, old_val);
6192         spin_unlock(&info->delalloc_root_lock);
6193
6194         while (total) {
6195                 cache = btrfs_lookup_block_group(info, bytenr);
6196                 if (!cache)
6197                         return -ENOENT;
6198                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
6199                                     BTRFS_BLOCK_GROUP_RAID1 |
6200                                     BTRFS_BLOCK_GROUP_RAID10))
6201                         factor = 2;
6202                 else
6203                         factor = 1;
6204                 /*
6205                  * If this block group has free space cache written out, we
6206                  * need to make sure to load it if we are removing space.  This
6207                  * is because we need the unpinning stage to actually add the
6208                  * space back to the block group, otherwise we will leak space.
6209                  */
6210                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
6211                         cache_block_group(cache, 1);
6212
6213                 byte_in_group = bytenr - cache->key.objectid;
6214                 WARN_ON(byte_in_group > cache->key.offset);
6215
6216                 spin_lock(&cache->space_info->lock);
6217                 spin_lock(&cache->lock);
6218
6219                 if (btrfs_test_opt(root->fs_info, SPACE_CACHE) &&
6220                     cache->disk_cache_state < BTRFS_DC_CLEAR)
6221                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6222
6223                 old_val = btrfs_block_group_used(&cache->item);
6224                 num_bytes = min(total, cache->key.offset - byte_in_group);
6225                 if (alloc) {
6226                         old_val += num_bytes;
6227                         btrfs_set_block_group_used(&cache->item, old_val);
6228                         cache->reserved -= num_bytes;
6229                         cache->space_info->bytes_reserved -= num_bytes;
6230                         cache->space_info->bytes_used += num_bytes;
6231                         cache->space_info->disk_used += num_bytes * factor;
6232                         spin_unlock(&cache->lock);
6233                         spin_unlock(&cache->space_info->lock);
6234                 } else {
6235                         old_val -= num_bytes;
6236                         btrfs_set_block_group_used(&cache->item, old_val);
6237                         cache->pinned += num_bytes;
6238                         cache->space_info->bytes_pinned += num_bytes;
6239                         cache->space_info->bytes_used -= num_bytes;
6240                         cache->space_info->disk_used -= num_bytes * factor;
6241                         spin_unlock(&cache->lock);
6242                         spin_unlock(&cache->space_info->lock);
6243
6244                         trace_btrfs_space_reservation(root->fs_info, "pinned",
6245                                                       cache->space_info->flags,
6246                                                       num_bytes, 1);
6247                         set_extent_dirty(info->pinned_extents,
6248                                          bytenr, bytenr + num_bytes - 1,
6249                                          GFP_NOFS | __GFP_NOFAIL);
6250                 }
6251
6252                 spin_lock(&trans->transaction->dirty_bgs_lock);
6253                 if (list_empty(&cache->dirty_list)) {
6254                         list_add_tail(&cache->dirty_list,
6255                                       &trans->transaction->dirty_bgs);
6256                                 trans->transaction->num_dirty_bgs++;
6257                         btrfs_get_block_group(cache);
6258                 }
6259                 spin_unlock(&trans->transaction->dirty_bgs_lock);
6260
6261                 /*
6262                  * No longer have used bytes in this block group, queue it for
6263                  * deletion. We do this after adding the block group to the
6264                  * dirty list to avoid races between cleaner kthread and space
6265                  * cache writeout.
6266                  */
6267                 if (!alloc && old_val == 0) {
6268                         spin_lock(&info->unused_bgs_lock);
6269                         if (list_empty(&cache->bg_list)) {
6270                                 btrfs_get_block_group(cache);
6271                                 list_add_tail(&cache->bg_list,
6272                                               &info->unused_bgs);
6273                         }
6274                         spin_unlock(&info->unused_bgs_lock);
6275                 }
6276
6277                 btrfs_put_block_group(cache);
6278                 total -= num_bytes;
6279                 bytenr += num_bytes;
6280         }
6281         return 0;
6282 }
6283
6284 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
6285 {
6286         struct btrfs_block_group_cache *cache;
6287         u64 bytenr;
6288
6289         spin_lock(&root->fs_info->block_group_cache_lock);
6290         bytenr = root->fs_info->first_logical_byte;
6291         spin_unlock(&root->fs_info->block_group_cache_lock);
6292
6293         if (bytenr < (u64)-1)
6294                 return bytenr;
6295
6296         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
6297         if (!cache)
6298                 return 0;
6299
6300         bytenr = cache->key.objectid;
6301         btrfs_put_block_group(cache);
6302
6303         return bytenr;
6304 }
6305
6306 static int pin_down_extent(struct btrfs_root *root,
6307                            struct btrfs_block_group_cache *cache,
6308                            u64 bytenr, u64 num_bytes, int reserved)
6309 {
6310         spin_lock(&cache->space_info->lock);
6311         spin_lock(&cache->lock);
6312         cache->pinned += num_bytes;
6313         cache->space_info->bytes_pinned += num_bytes;
6314         if (reserved) {
6315                 cache->reserved -= num_bytes;
6316                 cache->space_info->bytes_reserved -= num_bytes;
6317         }
6318         spin_unlock(&cache->lock);
6319         spin_unlock(&cache->space_info->lock);
6320
6321         trace_btrfs_space_reservation(root->fs_info, "pinned",
6322                                       cache->space_info->flags, num_bytes, 1);
6323         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
6324                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6325         return 0;
6326 }
6327
6328 /*
6329  * this function must be called within transaction
6330  */
6331 int btrfs_pin_extent(struct btrfs_root *root,
6332                      u64 bytenr, u64 num_bytes, int reserved)
6333 {
6334         struct btrfs_block_group_cache *cache;
6335
6336         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6337         BUG_ON(!cache); /* Logic error */
6338
6339         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
6340
6341         btrfs_put_block_group(cache);
6342         return 0;
6343 }
6344
6345 /*
6346  * this function must be called within transaction
6347  */
6348 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
6349                                     u64 bytenr, u64 num_bytes)
6350 {
6351         struct btrfs_block_group_cache *cache;
6352         int ret;
6353
6354         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6355         if (!cache)
6356                 return -EINVAL;
6357
6358         /*
6359          * pull in the free space cache (if any) so that our pin
6360          * removes the free space from the cache.  We have load_only set
6361          * to one because the slow code to read in the free extents does check
6362          * the pinned extents.
6363          */
6364         cache_block_group(cache, 1);
6365
6366         pin_down_extent(root, cache, bytenr, num_bytes, 0);
6367
6368         /* remove us from the free space cache (if we're there at all) */
6369         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6370         btrfs_put_block_group(cache);
6371         return ret;
6372 }
6373
6374 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
6375 {
6376         int ret;
6377         struct btrfs_block_group_cache *block_group;
6378         struct btrfs_caching_control *caching_ctl;
6379
6380         block_group = btrfs_lookup_block_group(root->fs_info, start);
6381         if (!block_group)
6382                 return -EINVAL;
6383
6384         cache_block_group(block_group, 0);
6385         caching_ctl = get_caching_control(block_group);
6386
6387         if (!caching_ctl) {
6388                 /* Logic error */
6389                 BUG_ON(!block_group_cache_done(block_group));
6390                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6391         } else {
6392                 mutex_lock(&caching_ctl->mutex);
6393
6394                 if (start >= caching_ctl->progress) {
6395                         ret = add_excluded_extent(root, start, num_bytes);
6396                 } else if (start + num_bytes <= caching_ctl->progress) {
6397                         ret = btrfs_remove_free_space(block_group,
6398                                                       start, num_bytes);
6399                 } else {
6400                         num_bytes = caching_ctl->progress - start;
6401                         ret = btrfs_remove_free_space(block_group,
6402                                                       start, num_bytes);
6403                         if (ret)
6404                                 goto out_lock;
6405
6406                         num_bytes = (start + num_bytes) -
6407                                 caching_ctl->progress;
6408                         start = caching_ctl->progress;
6409                         ret = add_excluded_extent(root, start, num_bytes);
6410                 }
6411 out_lock:
6412                 mutex_unlock(&caching_ctl->mutex);
6413                 put_caching_control(caching_ctl);
6414         }
6415         btrfs_put_block_group(block_group);
6416         return ret;
6417 }
6418
6419 int btrfs_exclude_logged_extents(struct btrfs_root *log,
6420                                  struct extent_buffer *eb)
6421 {
6422         struct btrfs_file_extent_item *item;
6423         struct btrfs_key key;
6424         int found_type;
6425         int i;
6426
6427         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
6428                 return 0;
6429
6430         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6431                 btrfs_item_key_to_cpu(eb, &key, i);
6432                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6433                         continue;
6434                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6435                 found_type = btrfs_file_extent_type(eb, item);
6436                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6437                         continue;
6438                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6439                         continue;
6440                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6441                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6442                 __exclude_logged_extent(log, key.objectid, key.offset);
6443         }
6444
6445         return 0;
6446 }
6447
6448 static void
6449 btrfs_inc_block_group_reservations(struct btrfs_block_group_cache *bg)
6450 {
6451         atomic_inc(&bg->reservations);
6452 }
6453
6454 void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
6455                                         const u64 start)
6456 {
6457         struct btrfs_block_group_cache *bg;
6458
6459         bg = btrfs_lookup_block_group(fs_info, start);
6460         ASSERT(bg);
6461         if (atomic_dec_and_test(&bg->reservations))
6462                 wake_up_atomic_t(&bg->reservations);
6463         btrfs_put_block_group(bg);
6464 }
6465
6466 static int btrfs_wait_bg_reservations_atomic_t(atomic_t *a)
6467 {
6468         schedule();
6469         return 0;
6470 }
6471
6472 void btrfs_wait_block_group_reservations(struct btrfs_block_group_cache *bg)
6473 {
6474         struct btrfs_space_info *space_info = bg->space_info;
6475
6476         ASSERT(bg->ro);
6477
6478         if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
6479                 return;
6480
6481         /*
6482          * Our block group is read only but before we set it to read only,
6483          * some task might have had allocated an extent from it already, but it
6484          * has not yet created a respective ordered extent (and added it to a
6485          * root's list of ordered extents).
6486          * Therefore wait for any task currently allocating extents, since the
6487          * block group's reservations counter is incremented while a read lock
6488          * on the groups' semaphore is held and decremented after releasing
6489          * the read access on that semaphore and creating the ordered extent.
6490          */
6491         down_write(&space_info->groups_sem);
6492         up_write(&space_info->groups_sem);
6493
6494         wait_on_atomic_t(&bg->reservations,
6495                          btrfs_wait_bg_reservations_atomic_t,
6496                          TASK_UNINTERRUPTIBLE);
6497 }
6498
6499 /**
6500  * btrfs_update_reserved_bytes - update the block_group and space info counters
6501  * @cache:      The cache we are manipulating
6502  * @num_bytes:  The number of bytes in question
6503  * @reserve:    One of the reservation enums
6504  * @delalloc:   The blocks are allocated for the delalloc write
6505  *
6506  * This is called by the allocator when it reserves space, or by somebody who is
6507  * freeing space that was never actually used on disk.  For example if you
6508  * reserve some space for a new leaf in transaction A and before transaction A
6509  * commits you free that leaf, you call this with reserve set to 0 in order to
6510  * clear the reservation.
6511  *
6512  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
6513  * ENOSPC accounting.  For data we handle the reservation through clearing the
6514  * delalloc bits in the io_tree.  We have to do this since we could end up
6515  * allocating less disk space for the amount of data we have reserved in the
6516  * case of compression.
6517  *
6518  * If this is a reservation and the block group has become read only we cannot
6519  * make the reservation and return -EAGAIN, otherwise this function always
6520  * succeeds.
6521  */
6522 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
6523                                        u64 num_bytes, int reserve, int delalloc)
6524 {
6525         struct btrfs_space_info *space_info = cache->space_info;
6526         int ret = 0;
6527
6528         spin_lock(&space_info->lock);
6529         spin_lock(&cache->lock);
6530         if (reserve != RESERVE_FREE) {
6531                 if (cache->ro) {
6532                         ret = -EAGAIN;
6533                 } else {
6534                         cache->reserved += num_bytes;
6535                         space_info->bytes_reserved += num_bytes;
6536                         if (reserve == RESERVE_ALLOC) {
6537                                 trace_btrfs_space_reservation(cache->fs_info,
6538                                                 "space_info", space_info->flags,
6539                                                 num_bytes, 0);
6540                                 space_info->bytes_may_use -= num_bytes;
6541                         }
6542
6543                         if (delalloc)
6544                                 cache->delalloc_bytes += num_bytes;
6545                 }
6546         } else {
6547                 if (cache->ro)
6548                         space_info->bytes_readonly += num_bytes;
6549                 cache->reserved -= num_bytes;
6550                 space_info->bytes_reserved -= num_bytes;
6551
6552                 if (delalloc)
6553                         cache->delalloc_bytes -= num_bytes;
6554         }
6555         spin_unlock(&cache->lock);
6556         spin_unlock(&space_info->lock);
6557         return ret;
6558 }
6559
6560 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6561                                 struct btrfs_root *root)
6562 {
6563         struct btrfs_fs_info *fs_info = root->fs_info;
6564         struct btrfs_caching_control *next;
6565         struct btrfs_caching_control *caching_ctl;
6566         struct btrfs_block_group_cache *cache;
6567
6568         down_write(&fs_info->commit_root_sem);
6569
6570         list_for_each_entry_safe(caching_ctl, next,
6571                                  &fs_info->caching_block_groups, list) {
6572                 cache = caching_ctl->block_group;
6573                 if (block_group_cache_done(cache)) {
6574                         cache->last_byte_to_unpin = (u64)-1;
6575                         list_del_init(&caching_ctl->list);
6576                         put_caching_control(caching_ctl);
6577                 } else {
6578                         cache->last_byte_to_unpin = caching_ctl->progress;
6579                 }
6580         }
6581
6582         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6583                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6584         else
6585                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6586
6587         up_write(&fs_info->commit_root_sem);
6588
6589         update_global_block_rsv(fs_info);
6590 }
6591
6592 /*
6593  * Returns the free cluster for the given space info and sets empty_cluster to
6594  * what it should be based on the mount options.
6595  */
6596 static struct btrfs_free_cluster *
6597 fetch_cluster_info(struct btrfs_root *root, struct btrfs_space_info *space_info,
6598                    u64 *empty_cluster)
6599 {
6600         struct btrfs_free_cluster *ret = NULL;
6601         bool ssd = btrfs_test_opt(root->fs_info, SSD);
6602
6603         *empty_cluster = 0;
6604         if (btrfs_mixed_space_info(space_info))
6605                 return ret;
6606
6607         if (ssd)
6608                 *empty_cluster = SZ_2M;
6609         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6610                 ret = &root->fs_info->meta_alloc_cluster;
6611                 if (!ssd)
6612                         *empty_cluster = SZ_64K;
6613         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) && ssd) {
6614                 ret = &root->fs_info->data_alloc_cluster;
6615         }
6616
6617         return ret;
6618 }
6619
6620 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
6621                               const bool return_free_space)
6622 {
6623         struct btrfs_fs_info *fs_info = root->fs_info;
6624         struct btrfs_block_group_cache *cache = NULL;
6625         struct btrfs_space_info *space_info;
6626         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6627         struct btrfs_free_cluster *cluster = NULL;
6628         u64 len;
6629         u64 total_unpinned = 0;
6630         u64 empty_cluster = 0;
6631         bool readonly;
6632
6633         while (start <= end) {
6634                 readonly = false;
6635                 if (!cache ||
6636                     start >= cache->key.objectid + cache->key.offset) {
6637                         if (cache)
6638                                 btrfs_put_block_group(cache);
6639                         total_unpinned = 0;
6640                         cache = btrfs_lookup_block_group(fs_info, start);
6641                         BUG_ON(!cache); /* Logic error */
6642
6643                         cluster = fetch_cluster_info(root,
6644                                                      cache->space_info,
6645                                                      &empty_cluster);
6646                         empty_cluster <<= 1;
6647                 }
6648
6649                 len = cache->key.objectid + cache->key.offset - start;
6650                 len = min(len, end + 1 - start);
6651
6652                 if (start < cache->last_byte_to_unpin) {
6653                         len = min(len, cache->last_byte_to_unpin - start);
6654                         if (return_free_space)
6655                                 btrfs_add_free_space(cache, start, len);
6656                 }
6657
6658                 start += len;
6659                 total_unpinned += len;
6660                 space_info = cache->space_info;
6661
6662                 /*
6663                  * If this space cluster has been marked as fragmented and we've
6664                  * unpinned enough in this block group to potentially allow a
6665                  * cluster to be created inside of it go ahead and clear the
6666                  * fragmented check.
6667                  */
6668                 if (cluster && cluster->fragmented &&
6669                     total_unpinned > empty_cluster) {
6670                         spin_lock(&cluster->lock);
6671                         cluster->fragmented = 0;
6672                         spin_unlock(&cluster->lock);
6673                 }
6674
6675                 spin_lock(&space_info->lock);
6676                 spin_lock(&cache->lock);
6677                 cache->pinned -= len;
6678                 space_info->bytes_pinned -= len;
6679
6680                 trace_btrfs_space_reservation(fs_info, "pinned",
6681                                               space_info->flags, len, 0);
6682                 space_info->max_extent_size = 0;
6683                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6684                 if (cache->ro) {
6685                         space_info->bytes_readonly += len;
6686                         readonly = true;
6687                 }
6688                 spin_unlock(&cache->lock);
6689                 if (!readonly && return_free_space &&
6690                     global_rsv->space_info == space_info) {
6691                         u64 to_add = len;
6692                         WARN_ON(!return_free_space);
6693                         spin_lock(&global_rsv->lock);
6694                         if (!global_rsv->full) {
6695                                 to_add = min(len, global_rsv->size -
6696                                              global_rsv->reserved);
6697                                 global_rsv->reserved += to_add;
6698                                 space_info->bytes_may_use += to_add;
6699                                 if (global_rsv->reserved >= global_rsv->size)
6700                                         global_rsv->full = 1;
6701                                 trace_btrfs_space_reservation(fs_info,
6702                                                               "space_info",
6703                                                               space_info->flags,
6704                                                               to_add, 1);
6705                                 len -= to_add;
6706                         }
6707                         spin_unlock(&global_rsv->lock);
6708                         /* Add to any tickets we may have */
6709                         if (len)
6710                                 space_info_add_new_bytes(fs_info, space_info,
6711                                                          len);
6712                 }
6713                 spin_unlock(&space_info->lock);
6714         }
6715
6716         if (cache)
6717                 btrfs_put_block_group(cache);
6718         return 0;
6719 }
6720
6721 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6722                                struct btrfs_root *root)
6723 {
6724         struct btrfs_fs_info *fs_info = root->fs_info;
6725         struct btrfs_block_group_cache *block_group, *tmp;
6726         struct list_head *deleted_bgs;
6727         struct extent_io_tree *unpin;
6728         u64 start;
6729         u64 end;
6730         int ret;
6731
6732         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6733                 unpin = &fs_info->freed_extents[1];
6734         else
6735                 unpin = &fs_info->freed_extents[0];
6736
6737         while (!trans->aborted) {
6738                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6739                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6740                                             EXTENT_DIRTY, NULL);
6741                 if (ret) {
6742                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6743                         break;
6744                 }
6745
6746                 if (btrfs_test_opt(root->fs_info, DISCARD))
6747                         ret = btrfs_discard_extent(root, start,
6748                                                    end + 1 - start, NULL);
6749
6750                 clear_extent_dirty(unpin, start, end);
6751                 unpin_extent_range(root, start, end, true);
6752                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6753                 cond_resched();
6754         }
6755
6756         /*
6757          * Transaction is finished.  We don't need the lock anymore.  We
6758          * do need to clean up the block groups in case of a transaction
6759          * abort.
6760          */
6761         deleted_bgs = &trans->transaction->deleted_bgs;
6762         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6763                 u64 trimmed = 0;
6764
6765                 ret = -EROFS;
6766                 if (!trans->aborted)
6767                         ret = btrfs_discard_extent(root,
6768                                                    block_group->key.objectid,
6769                                                    block_group->key.offset,
6770                                                    &trimmed);
6771
6772                 list_del_init(&block_group->bg_list);
6773                 btrfs_put_block_group_trimming(block_group);
6774                 btrfs_put_block_group(block_group);
6775
6776                 if (ret) {
6777                         const char *errstr = btrfs_decode_error(ret);
6778                         btrfs_warn(fs_info,
6779                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6780                                    ret, errstr);
6781                 }
6782         }
6783
6784         return 0;
6785 }
6786
6787 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6788                              u64 owner, u64 root_objectid)
6789 {
6790         struct btrfs_space_info *space_info;
6791         u64 flags;
6792
6793         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6794                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6795                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6796                 else
6797                         flags = BTRFS_BLOCK_GROUP_METADATA;
6798         } else {
6799                 flags = BTRFS_BLOCK_GROUP_DATA;
6800         }
6801
6802         space_info = __find_space_info(fs_info, flags);
6803         BUG_ON(!space_info); /* Logic bug */
6804         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6805 }
6806
6807
6808 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6809                                 struct btrfs_root *root,
6810                                 struct btrfs_delayed_ref_node *node, u64 parent,
6811                                 u64 root_objectid, u64 owner_objectid,
6812                                 u64 owner_offset, int refs_to_drop,
6813                                 struct btrfs_delayed_extent_op *extent_op)
6814 {
6815         struct btrfs_key key;
6816         struct btrfs_path *path;
6817         struct btrfs_fs_info *info = root->fs_info;
6818         struct btrfs_root *extent_root = info->extent_root;
6819         struct extent_buffer *leaf;
6820         struct btrfs_extent_item *ei;
6821         struct btrfs_extent_inline_ref *iref;
6822         int ret;
6823         int is_data;
6824         int extent_slot = 0;
6825         int found_extent = 0;
6826         int num_to_del = 1;
6827         u32 item_size;
6828         u64 refs;
6829         u64 bytenr = node->bytenr;
6830         u64 num_bytes = node->num_bytes;
6831         int last_ref = 0;
6832         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6833                                                  SKINNY_METADATA);
6834
6835         path = btrfs_alloc_path();
6836         if (!path)
6837                 return -ENOMEM;
6838
6839         path->reada = READA_FORWARD;
6840         path->leave_spinning = 1;
6841
6842         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6843         BUG_ON(!is_data && refs_to_drop != 1);
6844
6845         if (is_data)
6846                 skinny_metadata = 0;
6847
6848         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6849                                     bytenr, num_bytes, parent,
6850                                     root_objectid, owner_objectid,
6851                                     owner_offset);
6852         if (ret == 0) {
6853                 extent_slot = path->slots[0];
6854                 while (extent_slot >= 0) {
6855                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6856                                               extent_slot);
6857                         if (key.objectid != bytenr)
6858                                 break;
6859                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6860                             key.offset == num_bytes) {
6861                                 found_extent = 1;
6862                                 break;
6863                         }
6864                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6865                             key.offset == owner_objectid) {
6866                                 found_extent = 1;
6867                                 break;
6868                         }
6869                         if (path->slots[0] - extent_slot > 5)
6870                                 break;
6871                         extent_slot--;
6872                 }
6873 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6874                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6875                 if (found_extent && item_size < sizeof(*ei))
6876                         found_extent = 0;
6877 #endif
6878                 if (!found_extent) {
6879                         BUG_ON(iref);
6880                         ret = remove_extent_backref(trans, extent_root, path,
6881                                                     NULL, refs_to_drop,
6882                                                     is_data, &last_ref);
6883                         if (ret) {
6884                                 btrfs_abort_transaction(trans, ret);
6885                                 goto out;
6886                         }
6887                         btrfs_release_path(path);
6888                         path->leave_spinning = 1;
6889
6890                         key.objectid = bytenr;
6891                         key.type = BTRFS_EXTENT_ITEM_KEY;
6892                         key.offset = num_bytes;
6893
6894                         if (!is_data && skinny_metadata) {
6895                                 key.type = BTRFS_METADATA_ITEM_KEY;
6896                                 key.offset = owner_objectid;
6897                         }
6898
6899                         ret = btrfs_search_slot(trans, extent_root,
6900                                                 &key, path, -1, 1);
6901                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6902                                 /*
6903                                  * Couldn't find our skinny metadata item,
6904                                  * see if we have ye olde extent item.
6905                                  */
6906                                 path->slots[0]--;
6907                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6908                                                       path->slots[0]);
6909                                 if (key.objectid == bytenr &&
6910                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6911                                     key.offset == num_bytes)
6912                                         ret = 0;
6913                         }
6914
6915                         if (ret > 0 && skinny_metadata) {
6916                                 skinny_metadata = false;
6917                                 key.objectid = bytenr;
6918                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6919                                 key.offset = num_bytes;
6920                                 btrfs_release_path(path);
6921                                 ret = btrfs_search_slot(trans, extent_root,
6922                                                         &key, path, -1, 1);
6923                         }
6924
6925                         if (ret) {
6926                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6927                                         ret, bytenr);
6928                                 if (ret > 0)
6929                                         btrfs_print_leaf(extent_root,
6930                                                          path->nodes[0]);
6931                         }
6932                         if (ret < 0) {
6933                                 btrfs_abort_transaction(trans, ret);
6934                                 goto out;
6935                         }
6936                         extent_slot = path->slots[0];
6937                 }
6938         } else if (WARN_ON(ret == -ENOENT)) {
6939                 btrfs_print_leaf(extent_root, path->nodes[0]);
6940                 btrfs_err(info,
6941                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6942                         bytenr, parent, root_objectid, owner_objectid,
6943                         owner_offset);
6944                 btrfs_abort_transaction(trans, ret);
6945                 goto out;
6946         } else {
6947                 btrfs_abort_transaction(trans, ret);
6948                 goto out;
6949         }
6950
6951         leaf = path->nodes[0];
6952         item_size = btrfs_item_size_nr(leaf, extent_slot);
6953 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6954         if (item_size < sizeof(*ei)) {
6955                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6956                 ret = convert_extent_item_v0(trans, extent_root, path,
6957                                              owner_objectid, 0);
6958                 if (ret < 0) {
6959                         btrfs_abort_transaction(trans, ret);
6960                         goto out;
6961                 }
6962
6963                 btrfs_release_path(path);
6964                 path->leave_spinning = 1;
6965
6966                 key.objectid = bytenr;
6967                 key.type = BTRFS_EXTENT_ITEM_KEY;
6968                 key.offset = num_bytes;
6969
6970                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6971                                         -1, 1);
6972                 if (ret) {
6973                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6974                                 ret, bytenr);
6975                         btrfs_print_leaf(extent_root, path->nodes[0]);
6976                 }
6977                 if (ret < 0) {
6978                         btrfs_abort_transaction(trans, ret);
6979                         goto out;
6980                 }
6981
6982                 extent_slot = path->slots[0];
6983                 leaf = path->nodes[0];
6984                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6985         }
6986 #endif
6987         BUG_ON(item_size < sizeof(*ei));
6988         ei = btrfs_item_ptr(leaf, extent_slot,
6989                             struct btrfs_extent_item);
6990         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6991             key.type == BTRFS_EXTENT_ITEM_KEY) {
6992                 struct btrfs_tree_block_info *bi;
6993                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6994                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6995                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6996         }
6997
6998         refs = btrfs_extent_refs(leaf, ei);
6999         if (refs < refs_to_drop) {
7000                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
7001                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
7002                 ret = -EINVAL;
7003                 btrfs_abort_transaction(trans, ret);
7004                 goto out;
7005         }
7006         refs -= refs_to_drop;
7007
7008         if (refs > 0) {
7009                 if (extent_op)
7010                         __run_delayed_extent_op(extent_op, leaf, ei);
7011                 /*
7012                  * In the case of inline back ref, reference count will
7013                  * be updated by remove_extent_backref
7014                  */
7015                 if (iref) {
7016                         BUG_ON(!found_extent);
7017                 } else {
7018                         btrfs_set_extent_refs(leaf, ei, refs);
7019                         btrfs_mark_buffer_dirty(leaf);
7020                 }
7021                 if (found_extent) {
7022                         ret = remove_extent_backref(trans, extent_root, path,
7023                                                     iref, refs_to_drop,
7024                                                     is_data, &last_ref);
7025                         if (ret) {
7026                                 btrfs_abort_transaction(trans, ret);
7027                                 goto out;
7028                         }
7029                 }
7030                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
7031                                  root_objectid);
7032         } else {
7033                 if (found_extent) {
7034                         BUG_ON(is_data && refs_to_drop !=
7035                                extent_data_ref_count(path, iref));
7036                         if (iref) {
7037                                 BUG_ON(path->slots[0] != extent_slot);
7038                         } else {
7039                                 BUG_ON(path->slots[0] != extent_slot + 1);
7040                                 path->slots[0] = extent_slot;
7041                                 num_to_del = 2;
7042                         }
7043                 }
7044
7045                 last_ref = 1;
7046                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
7047                                       num_to_del);
7048                 if (ret) {
7049                         btrfs_abort_transaction(trans, ret);
7050                         goto out;
7051                 }
7052                 btrfs_release_path(path);
7053
7054                 if (is_data) {
7055                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
7056                         if (ret) {
7057                                 btrfs_abort_transaction(trans, ret);
7058                                 goto out;
7059                         }
7060                 }
7061
7062                 ret = add_to_free_space_tree(trans, root->fs_info, bytenr,
7063                                              num_bytes);
7064                 if (ret) {
7065                         btrfs_abort_transaction(trans, ret);
7066                         goto out;
7067                 }
7068
7069                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
7070                 if (ret) {
7071                         btrfs_abort_transaction(trans, ret);
7072                         goto out;
7073                 }
7074         }
7075         btrfs_release_path(path);
7076
7077 out:
7078         btrfs_free_path(path);
7079         return ret;
7080 }
7081
7082 /*
7083  * when we free an block, it is possible (and likely) that we free the last
7084  * delayed ref for that extent as well.  This searches the delayed ref tree for
7085  * a given extent, and if there are no other delayed refs to be processed, it
7086  * removes it from the tree.
7087  */
7088 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
7089                                       struct btrfs_root *root, u64 bytenr)
7090 {
7091         struct btrfs_delayed_ref_head *head;
7092         struct btrfs_delayed_ref_root *delayed_refs;
7093         int ret = 0;
7094
7095         delayed_refs = &trans->transaction->delayed_refs;
7096         spin_lock(&delayed_refs->lock);
7097         head = btrfs_find_delayed_ref_head(trans, bytenr);
7098         if (!head)
7099                 goto out_delayed_unlock;
7100
7101         spin_lock(&head->lock);
7102         if (!list_empty(&head->ref_list))
7103                 goto out;
7104
7105         if (head->extent_op) {
7106                 if (!head->must_insert_reserved)
7107                         goto out;
7108                 btrfs_free_delayed_extent_op(head->extent_op);
7109                 head->extent_op = NULL;
7110         }
7111
7112         /*
7113          * waiting for the lock here would deadlock.  If someone else has it
7114          * locked they are already in the process of dropping it anyway
7115          */
7116         if (!mutex_trylock(&head->mutex))
7117                 goto out;
7118
7119         /*
7120          * at this point we have a head with no other entries.  Go
7121          * ahead and process it.
7122          */
7123         head->node.in_tree = 0;
7124         rb_erase(&head->href_node, &delayed_refs->href_root);
7125
7126         atomic_dec(&delayed_refs->num_entries);
7127
7128         /*
7129          * we don't take a ref on the node because we're removing it from the
7130          * tree, so we just steal the ref the tree was holding.
7131          */
7132         delayed_refs->num_heads--;
7133         if (head->processing == 0)
7134                 delayed_refs->num_heads_ready--;
7135         head->processing = 0;
7136         spin_unlock(&head->lock);
7137         spin_unlock(&delayed_refs->lock);
7138
7139         BUG_ON(head->extent_op);
7140         if (head->must_insert_reserved)
7141                 ret = 1;
7142
7143         mutex_unlock(&head->mutex);
7144         btrfs_put_delayed_ref(&head->node);
7145         return ret;
7146 out:
7147         spin_unlock(&head->lock);
7148
7149 out_delayed_unlock:
7150         spin_unlock(&delayed_refs->lock);
7151         return 0;
7152 }
7153
7154 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
7155                            struct btrfs_root *root,
7156                            struct extent_buffer *buf,
7157                            u64 parent, int last_ref)
7158 {
7159         int pin = 1;
7160         int ret;
7161
7162         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7163                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7164                                         buf->start, buf->len,
7165                                         parent, root->root_key.objectid,
7166                                         btrfs_header_level(buf),
7167                                         BTRFS_DROP_DELAYED_REF, NULL);
7168                 BUG_ON(ret); /* -ENOMEM */
7169         }
7170
7171         if (!last_ref)
7172                 return;
7173
7174         if (btrfs_header_generation(buf) == trans->transid) {
7175                 struct btrfs_block_group_cache *cache;
7176
7177                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
7178                         ret = check_ref_cleanup(trans, root, buf->start);
7179                         if (!ret)
7180                                 goto out;
7181                 }
7182
7183                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
7184
7185                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
7186                         pin_down_extent(root, cache, buf->start, buf->len, 1);
7187                         btrfs_put_block_group(cache);
7188                         goto out;
7189                 }
7190
7191                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
7192
7193                 btrfs_add_free_space(cache, buf->start, buf->len);
7194                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
7195                 btrfs_put_block_group(cache);
7196                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
7197                 pin = 0;
7198         }
7199 out:
7200         if (pin)
7201                 add_pinned_bytes(root->fs_info, buf->len,
7202                                  btrfs_header_level(buf),
7203                                  root->root_key.objectid);
7204
7205         /*
7206          * Deleting the buffer, clear the corrupt flag since it doesn't matter
7207          * anymore.
7208          */
7209         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
7210 }
7211
7212 /* Can return -ENOMEM */
7213 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7214                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
7215                       u64 owner, u64 offset)
7216 {
7217         int ret;
7218         struct btrfs_fs_info *fs_info = root->fs_info;
7219
7220         if (btrfs_is_testing(fs_info))
7221                 return 0;
7222
7223         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
7224
7225         /*
7226          * tree log blocks never actually go into the extent allocation
7227          * tree, just update pinning info and exit early.
7228          */
7229         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
7230                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
7231                 /* unlocks the pinned mutex */
7232                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
7233                 ret = 0;
7234         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
7235                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
7236                                         num_bytes,
7237                                         parent, root_objectid, (int)owner,
7238                                         BTRFS_DROP_DELAYED_REF, NULL);
7239         } else {
7240                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
7241                                                 num_bytes,
7242                                                 parent, root_objectid, owner,
7243                                                 offset, 0,
7244                                                 BTRFS_DROP_DELAYED_REF, NULL);
7245         }
7246         return ret;
7247 }
7248
7249 /*
7250  * when we wait for progress in the block group caching, its because
7251  * our allocation attempt failed at least once.  So, we must sleep
7252  * and let some progress happen before we try again.
7253  *
7254  * This function will sleep at least once waiting for new free space to
7255  * show up, and then it will check the block group free space numbers
7256  * for our min num_bytes.  Another option is to have it go ahead
7257  * and look in the rbtree for a free extent of a given size, but this
7258  * is a good start.
7259  *
7260  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
7261  * any of the information in this block group.
7262  */
7263 static noinline void
7264 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
7265                                 u64 num_bytes)
7266 {
7267         struct btrfs_caching_control *caching_ctl;
7268
7269         caching_ctl = get_caching_control(cache);
7270         if (!caching_ctl)
7271                 return;
7272
7273         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
7274                    (cache->free_space_ctl->free_space >= num_bytes));
7275
7276         put_caching_control(caching_ctl);
7277 }
7278
7279 static noinline int
7280 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
7281 {
7282         struct btrfs_caching_control *caching_ctl;
7283         int ret = 0;
7284
7285         caching_ctl = get_caching_control(cache);
7286         if (!caching_ctl)
7287                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
7288
7289         wait_event(caching_ctl->wait, block_group_cache_done(cache));
7290         if (cache->cached == BTRFS_CACHE_ERROR)
7291                 ret = -EIO;
7292         put_caching_control(caching_ctl);
7293         return ret;
7294 }
7295
7296 int __get_raid_index(u64 flags)
7297 {
7298         if (flags & BTRFS_BLOCK_GROUP_RAID10)
7299                 return BTRFS_RAID_RAID10;
7300         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
7301                 return BTRFS_RAID_RAID1;
7302         else if (flags & BTRFS_BLOCK_GROUP_DUP)
7303                 return BTRFS_RAID_DUP;
7304         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
7305                 return BTRFS_RAID_RAID0;
7306         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
7307                 return BTRFS_RAID_RAID5;
7308         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
7309                 return BTRFS_RAID_RAID6;
7310
7311         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
7312 }
7313
7314 int get_block_group_index(struct btrfs_block_group_cache *cache)
7315 {
7316         return __get_raid_index(cache->flags);
7317 }
7318
7319 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
7320         [BTRFS_RAID_RAID10]     = "raid10",
7321         [BTRFS_RAID_RAID1]      = "raid1",
7322         [BTRFS_RAID_DUP]        = "dup",
7323         [BTRFS_RAID_RAID0]      = "raid0",
7324         [BTRFS_RAID_SINGLE]     = "single",
7325         [BTRFS_RAID_RAID5]      = "raid5",
7326         [BTRFS_RAID_RAID6]      = "raid6",
7327 };
7328
7329 static const char *get_raid_name(enum btrfs_raid_types type)
7330 {
7331         if (type >= BTRFS_NR_RAID_TYPES)
7332                 return NULL;
7333
7334         return btrfs_raid_type_names[type];
7335 }
7336
7337 enum btrfs_loop_type {
7338         LOOP_CACHING_NOWAIT = 0,
7339         LOOP_CACHING_WAIT = 1,
7340         LOOP_ALLOC_CHUNK = 2,
7341         LOOP_NO_EMPTY_SIZE = 3,
7342 };
7343
7344 static inline void
7345 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
7346                        int delalloc)
7347 {
7348         if (delalloc)
7349                 down_read(&cache->data_rwsem);
7350 }
7351
7352 static inline void
7353 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
7354                        int delalloc)
7355 {
7356         btrfs_get_block_group(cache);
7357         if (delalloc)
7358                 down_read(&cache->data_rwsem);
7359 }
7360
7361 static struct btrfs_block_group_cache *
7362 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
7363                    struct btrfs_free_cluster *cluster,
7364                    int delalloc)
7365 {
7366         struct btrfs_block_group_cache *used_bg = NULL;
7367
7368         spin_lock(&cluster->refill_lock);
7369         while (1) {
7370                 used_bg = cluster->block_group;
7371                 if (!used_bg)
7372                         return NULL;
7373
7374                 if (used_bg == block_group)
7375                         return used_bg;
7376
7377                 btrfs_get_block_group(used_bg);
7378
7379                 if (!delalloc)
7380                         return used_bg;
7381
7382                 if (down_read_trylock(&used_bg->data_rwsem))
7383                         return used_bg;
7384
7385                 spin_unlock(&cluster->refill_lock);
7386
7387                 down_read(&used_bg->data_rwsem);
7388
7389                 spin_lock(&cluster->refill_lock);
7390                 if (used_bg == cluster->block_group)
7391                         return used_bg;
7392
7393                 up_read(&used_bg->data_rwsem);
7394                 btrfs_put_block_group(used_bg);
7395         }
7396 }
7397
7398 static inline void
7399 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7400                          int delalloc)
7401 {
7402         if (delalloc)
7403                 up_read(&cache->data_rwsem);
7404         btrfs_put_block_group(cache);
7405 }
7406
7407 /*
7408  * walks the btree of allocated extents and find a hole of a given size.
7409  * The key ins is changed to record the hole:
7410  * ins->objectid == start position
7411  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7412  * ins->offset == the size of the hole.
7413  * Any available blocks before search_start are skipped.
7414  *
7415  * If there is no suitable free space, we will record the max size of
7416  * the free space extent currently.
7417  */
7418 static noinline int find_free_extent(struct btrfs_root *orig_root,
7419                                      u64 num_bytes, u64 empty_size,
7420                                      u64 hint_byte, struct btrfs_key *ins,
7421                                      u64 flags, int delalloc)
7422 {
7423         int ret = 0;
7424         struct btrfs_root *root = orig_root->fs_info->extent_root;
7425         struct btrfs_free_cluster *last_ptr = NULL;
7426         struct btrfs_block_group_cache *block_group = NULL;
7427         u64 search_start = 0;
7428         u64 max_extent_size = 0;
7429         u64 empty_cluster = 0;
7430         struct btrfs_space_info *space_info;
7431         int loop = 0;
7432         int index = __get_raid_index(flags);
7433         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
7434                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
7435         bool failed_cluster_refill = false;
7436         bool failed_alloc = false;
7437         bool use_cluster = true;
7438         bool have_caching_bg = false;
7439         bool orig_have_caching_bg = false;
7440         bool full_search = false;
7441
7442         WARN_ON(num_bytes < root->sectorsize);
7443         ins->type = BTRFS_EXTENT_ITEM_KEY;
7444         ins->objectid = 0;
7445         ins->offset = 0;
7446
7447         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
7448
7449         space_info = __find_space_info(root->fs_info, flags);
7450         if (!space_info) {
7451                 btrfs_err(root->fs_info, "No space info for %llu", flags);
7452                 return -ENOSPC;
7453         }
7454
7455         /*
7456          * If our free space is heavily fragmented we may not be able to make
7457          * big contiguous allocations, so instead of doing the expensive search
7458          * for free space, simply return ENOSPC with our max_extent_size so we
7459          * can go ahead and search for a more manageable chunk.
7460          *
7461          * If our max_extent_size is large enough for our allocation simply
7462          * disable clustering since we will likely not be able to find enough
7463          * space to create a cluster and induce latency trying.
7464          */
7465         if (unlikely(space_info->max_extent_size)) {
7466                 spin_lock(&space_info->lock);
7467                 if (space_info->max_extent_size &&
7468                     num_bytes > space_info->max_extent_size) {
7469                         ins->offset = space_info->max_extent_size;
7470                         spin_unlock(&space_info->lock);
7471                         return -ENOSPC;
7472                 } else if (space_info->max_extent_size) {
7473                         use_cluster = false;
7474                 }
7475                 spin_unlock(&space_info->lock);
7476         }
7477
7478         last_ptr = fetch_cluster_info(orig_root, space_info, &empty_cluster);
7479         if (last_ptr) {
7480                 spin_lock(&last_ptr->lock);
7481                 if (last_ptr->block_group)
7482                         hint_byte = last_ptr->window_start;
7483                 if (last_ptr->fragmented) {
7484                         /*
7485                          * We still set window_start so we can keep track of the
7486                          * last place we found an allocation to try and save
7487                          * some time.
7488                          */
7489                         hint_byte = last_ptr->window_start;
7490                         use_cluster = false;
7491                 }
7492                 spin_unlock(&last_ptr->lock);
7493         }
7494
7495         search_start = max(search_start, first_logical_byte(root, 0));
7496         search_start = max(search_start, hint_byte);
7497         if (search_start == hint_byte) {
7498                 block_group = btrfs_lookup_block_group(root->fs_info,
7499                                                        search_start);
7500                 /*
7501                  * we don't want to use the block group if it doesn't match our
7502                  * allocation bits, or if its not cached.
7503                  *
7504                  * However if we are re-searching with an ideal block group
7505                  * picked out then we don't care that the block group is cached.
7506                  */
7507                 if (block_group && block_group_bits(block_group, flags) &&
7508                     block_group->cached != BTRFS_CACHE_NO) {
7509                         down_read(&space_info->groups_sem);
7510                         if (list_empty(&block_group->list) ||
7511                             block_group->ro) {
7512                                 /*
7513                                  * someone is removing this block group,
7514                                  * we can't jump into the have_block_group
7515                                  * target because our list pointers are not
7516                                  * valid
7517                                  */
7518                                 btrfs_put_block_group(block_group);
7519                                 up_read(&space_info->groups_sem);
7520                         } else {
7521                                 index = get_block_group_index(block_group);
7522                                 btrfs_lock_block_group(block_group, delalloc);
7523                                 goto have_block_group;
7524                         }
7525                 } else if (block_group) {
7526                         btrfs_put_block_group(block_group);
7527                 }
7528         }
7529 search:
7530         have_caching_bg = false;
7531         if (index == 0 || index == __get_raid_index(flags))
7532                 full_search = true;
7533         down_read(&space_info->groups_sem);
7534         list_for_each_entry(block_group, &space_info->block_groups[index],
7535                             list) {
7536                 u64 offset;
7537                 int cached;
7538
7539                 btrfs_grab_block_group(block_group, delalloc);
7540                 search_start = block_group->key.objectid;
7541
7542                 /*
7543                  * this can happen if we end up cycling through all the
7544                  * raid types, but we want to make sure we only allocate
7545                  * for the proper type.
7546                  */
7547                 if (!block_group_bits(block_group, flags)) {
7548                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7549                                 BTRFS_BLOCK_GROUP_RAID1 |
7550                                 BTRFS_BLOCK_GROUP_RAID5 |
7551                                 BTRFS_BLOCK_GROUP_RAID6 |
7552                                 BTRFS_BLOCK_GROUP_RAID10;
7553
7554                         /*
7555                          * if they asked for extra copies and this block group
7556                          * doesn't provide them, bail.  This does allow us to
7557                          * fill raid0 from raid1.
7558                          */
7559                         if ((flags & extra) && !(block_group->flags & extra))
7560                                 goto loop;
7561                 }
7562
7563 have_block_group:
7564                 cached = block_group_cache_done(block_group);
7565                 if (unlikely(!cached)) {
7566                         have_caching_bg = true;
7567                         ret = cache_block_group(block_group, 0);
7568                         BUG_ON(ret < 0);
7569                         ret = 0;
7570                 }
7571
7572                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7573                         goto loop;
7574                 if (unlikely(block_group->ro))
7575                         goto loop;
7576
7577                 /*
7578                  * Ok we want to try and use the cluster allocator, so
7579                  * lets look there
7580                  */
7581                 if (last_ptr && use_cluster) {
7582                         struct btrfs_block_group_cache *used_block_group;
7583                         unsigned long aligned_cluster;
7584                         /*
7585                          * the refill lock keeps out other
7586                          * people trying to start a new cluster
7587                          */
7588                         used_block_group = btrfs_lock_cluster(block_group,
7589                                                               last_ptr,
7590                                                               delalloc);
7591                         if (!used_block_group)
7592                                 goto refill_cluster;
7593
7594                         if (used_block_group != block_group &&
7595                             (used_block_group->ro ||
7596                              !block_group_bits(used_block_group, flags)))
7597                                 goto release_cluster;
7598
7599                         offset = btrfs_alloc_from_cluster(used_block_group,
7600                                                 last_ptr,
7601                                                 num_bytes,
7602                                                 used_block_group->key.objectid,
7603                                                 &max_extent_size);
7604                         if (offset) {
7605                                 /* we have a block, we're done */
7606                                 spin_unlock(&last_ptr->refill_lock);
7607                                 trace_btrfs_reserve_extent_cluster(root,
7608                                                 used_block_group,
7609                                                 search_start, num_bytes);
7610                                 if (used_block_group != block_group) {
7611                                         btrfs_release_block_group(block_group,
7612                                                                   delalloc);
7613                                         block_group = used_block_group;
7614                                 }
7615                                 goto checks;
7616                         }
7617
7618                         WARN_ON(last_ptr->block_group != used_block_group);
7619 release_cluster:
7620                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7621                          * set up a new clusters, so lets just skip it
7622                          * and let the allocator find whatever block
7623                          * it can find.  If we reach this point, we
7624                          * will have tried the cluster allocator
7625                          * plenty of times and not have found
7626                          * anything, so we are likely way too
7627                          * fragmented for the clustering stuff to find
7628                          * anything.
7629                          *
7630                          * However, if the cluster is taken from the
7631                          * current block group, release the cluster
7632                          * first, so that we stand a better chance of
7633                          * succeeding in the unclustered
7634                          * allocation.  */
7635                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7636                             used_block_group != block_group) {
7637                                 spin_unlock(&last_ptr->refill_lock);
7638                                 btrfs_release_block_group(used_block_group,
7639                                                           delalloc);
7640                                 goto unclustered_alloc;
7641                         }
7642
7643                         /*
7644                          * this cluster didn't work out, free it and
7645                          * start over
7646                          */
7647                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7648
7649                         if (used_block_group != block_group)
7650                                 btrfs_release_block_group(used_block_group,
7651                                                           delalloc);
7652 refill_cluster:
7653                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7654                                 spin_unlock(&last_ptr->refill_lock);
7655                                 goto unclustered_alloc;
7656                         }
7657
7658                         aligned_cluster = max_t(unsigned long,
7659                                                 empty_cluster + empty_size,
7660                                               block_group->full_stripe_len);
7661
7662                         /* allocate a cluster in this block group */
7663                         ret = btrfs_find_space_cluster(root, block_group,
7664                                                        last_ptr, search_start,
7665                                                        num_bytes,
7666                                                        aligned_cluster);
7667                         if (ret == 0) {
7668                                 /*
7669                                  * now pull our allocation out of this
7670                                  * cluster
7671                                  */
7672                                 offset = btrfs_alloc_from_cluster(block_group,
7673                                                         last_ptr,
7674                                                         num_bytes,
7675                                                         search_start,
7676                                                         &max_extent_size);
7677                                 if (offset) {
7678                                         /* we found one, proceed */
7679                                         spin_unlock(&last_ptr->refill_lock);
7680                                         trace_btrfs_reserve_extent_cluster(root,
7681                                                 block_group, search_start,
7682                                                 num_bytes);
7683                                         goto checks;
7684                                 }
7685                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7686                                    && !failed_cluster_refill) {
7687                                 spin_unlock(&last_ptr->refill_lock);
7688
7689                                 failed_cluster_refill = true;
7690                                 wait_block_group_cache_progress(block_group,
7691                                        num_bytes + empty_cluster + empty_size);
7692                                 goto have_block_group;
7693                         }
7694
7695                         /*
7696                          * at this point we either didn't find a cluster
7697                          * or we weren't able to allocate a block from our
7698                          * cluster.  Free the cluster we've been trying
7699                          * to use, and go to the next block group
7700                          */
7701                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7702                         spin_unlock(&last_ptr->refill_lock);
7703                         goto loop;
7704                 }
7705
7706 unclustered_alloc:
7707                 /*
7708                  * We are doing an unclustered alloc, set the fragmented flag so
7709                  * we don't bother trying to setup a cluster again until we get
7710                  * more space.
7711                  */
7712                 if (unlikely(last_ptr)) {
7713                         spin_lock(&last_ptr->lock);
7714                         last_ptr->fragmented = 1;
7715                         spin_unlock(&last_ptr->lock);
7716                 }
7717                 spin_lock(&block_group->free_space_ctl->tree_lock);
7718                 if (cached &&
7719                     block_group->free_space_ctl->free_space <
7720                     num_bytes + empty_cluster + empty_size) {
7721                         if (block_group->free_space_ctl->free_space >
7722                             max_extent_size)
7723                                 max_extent_size =
7724                                         block_group->free_space_ctl->free_space;
7725                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7726                         goto loop;
7727                 }
7728                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7729
7730                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7731                                                     num_bytes, empty_size,
7732                                                     &max_extent_size);
7733                 /*
7734                  * If we didn't find a chunk, and we haven't failed on this
7735                  * block group before, and this block group is in the middle of
7736                  * caching and we are ok with waiting, then go ahead and wait
7737                  * for progress to be made, and set failed_alloc to true.
7738                  *
7739                  * If failed_alloc is true then we've already waited on this
7740                  * block group once and should move on to the next block group.
7741                  */
7742                 if (!offset && !failed_alloc && !cached &&
7743                     loop > LOOP_CACHING_NOWAIT) {
7744                         wait_block_group_cache_progress(block_group,
7745                                                 num_bytes + empty_size);
7746                         failed_alloc = true;
7747                         goto have_block_group;
7748                 } else if (!offset) {
7749                         goto loop;
7750                 }
7751 checks:
7752                 search_start = ALIGN(offset, root->stripesize);
7753
7754                 /* move on to the next group */
7755                 if (search_start + num_bytes >
7756                     block_group->key.objectid + block_group->key.offset) {
7757                         btrfs_add_free_space(block_group, offset, num_bytes);
7758                         goto loop;
7759                 }
7760
7761                 if (offset < search_start)
7762                         btrfs_add_free_space(block_group, offset,
7763                                              search_start - offset);
7764                 BUG_ON(offset > search_start);
7765
7766                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7767                                                   alloc_type, delalloc);
7768                 if (ret == -EAGAIN) {
7769                         btrfs_add_free_space(block_group, offset, num_bytes);
7770                         goto loop;
7771                 }
7772                 btrfs_inc_block_group_reservations(block_group);
7773
7774                 /* we are all good, lets return */
7775                 ins->objectid = search_start;
7776                 ins->offset = num_bytes;
7777
7778                 trace_btrfs_reserve_extent(orig_root, block_group,
7779                                            search_start, num_bytes);
7780                 btrfs_release_block_group(block_group, delalloc);
7781                 break;
7782 loop:
7783                 failed_cluster_refill = false;
7784                 failed_alloc = false;
7785                 BUG_ON(index != get_block_group_index(block_group));
7786                 btrfs_release_block_group(block_group, delalloc);
7787         }
7788         up_read(&space_info->groups_sem);
7789
7790         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7791                 && !orig_have_caching_bg)
7792                 orig_have_caching_bg = true;
7793
7794         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7795                 goto search;
7796
7797         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7798                 goto search;
7799
7800         /*
7801          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7802          *                      caching kthreads as we move along
7803          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7804          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7805          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7806          *                      again
7807          */
7808         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7809                 index = 0;
7810                 if (loop == LOOP_CACHING_NOWAIT) {
7811                         /*
7812                          * We want to skip the LOOP_CACHING_WAIT step if we
7813                          * don't have any uncached bgs and we've already done a
7814                          * full search through.
7815                          */
7816                         if (orig_have_caching_bg || !full_search)
7817                                 loop = LOOP_CACHING_WAIT;
7818                         else
7819                                 loop = LOOP_ALLOC_CHUNK;
7820                 } else {
7821                         loop++;
7822                 }
7823
7824                 if (loop == LOOP_ALLOC_CHUNK) {
7825                         struct btrfs_trans_handle *trans;
7826                         int exist = 0;
7827
7828                         trans = current->journal_info;
7829                         if (trans)
7830                                 exist = 1;
7831                         else
7832                                 trans = btrfs_join_transaction(root);
7833
7834                         if (IS_ERR(trans)) {
7835                                 ret = PTR_ERR(trans);
7836                                 goto out;
7837                         }
7838
7839                         ret = do_chunk_alloc(trans, root, flags,
7840                                              CHUNK_ALLOC_FORCE);
7841
7842                         /*
7843                          * If we can't allocate a new chunk we've already looped
7844                          * through at least once, move on to the NO_EMPTY_SIZE
7845                          * case.
7846                          */
7847                         if (ret == -ENOSPC)
7848                                 loop = LOOP_NO_EMPTY_SIZE;
7849
7850                         /*
7851                          * Do not bail out on ENOSPC since we
7852                          * can do more things.
7853                          */
7854                         if (ret < 0 && ret != -ENOSPC)
7855                                 btrfs_abort_transaction(trans, ret);
7856                         else
7857                                 ret = 0;
7858                         if (!exist)
7859                                 btrfs_end_transaction(trans, root);
7860                         if (ret)
7861                                 goto out;
7862                 }
7863
7864                 if (loop == LOOP_NO_EMPTY_SIZE) {
7865                         /*
7866                          * Don't loop again if we already have no empty_size and
7867                          * no empty_cluster.
7868                          */
7869                         if (empty_size == 0 &&
7870                             empty_cluster == 0) {
7871                                 ret = -ENOSPC;
7872                                 goto out;
7873                         }
7874                         empty_size = 0;
7875                         empty_cluster = 0;
7876                 }
7877
7878                 goto search;
7879         } else if (!ins->objectid) {
7880                 ret = -ENOSPC;
7881         } else if (ins->objectid) {
7882                 if (!use_cluster && last_ptr) {
7883                         spin_lock(&last_ptr->lock);
7884                         last_ptr->window_start = ins->objectid;
7885                         spin_unlock(&last_ptr->lock);
7886                 }
7887                 ret = 0;
7888         }
7889 out:
7890         if (ret == -ENOSPC) {
7891                 spin_lock(&space_info->lock);
7892                 space_info->max_extent_size = max_extent_size;
7893                 spin_unlock(&space_info->lock);
7894                 ins->offset = max_extent_size;
7895         }
7896         return ret;
7897 }
7898
7899 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7900                             int dump_block_groups)
7901 {
7902         struct btrfs_block_group_cache *cache;
7903         int index = 0;
7904
7905         spin_lock(&info->lock);
7906         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7907                info->flags,
7908                info->total_bytes - info->bytes_used - info->bytes_pinned -
7909                info->bytes_reserved - info->bytes_readonly -
7910                info->bytes_may_use, (info->full) ? "" : "not ");
7911         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7912                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7913                info->total_bytes, info->bytes_used, info->bytes_pinned,
7914                info->bytes_reserved, info->bytes_may_use,
7915                info->bytes_readonly);
7916         spin_unlock(&info->lock);
7917
7918         if (!dump_block_groups)
7919                 return;
7920
7921         down_read(&info->groups_sem);
7922 again:
7923         list_for_each_entry(cache, &info->block_groups[index], list) {
7924                 spin_lock(&cache->lock);
7925                 printk(KERN_INFO "BTRFS: "
7926                            "block group %llu has %llu bytes, "
7927                            "%llu used %llu pinned %llu reserved %s\n",
7928                        cache->key.objectid, cache->key.offset,
7929                        btrfs_block_group_used(&cache->item), cache->pinned,
7930                        cache->reserved, cache->ro ? "[readonly]" : "");
7931                 btrfs_dump_free_space(cache, bytes);
7932                 spin_unlock(&cache->lock);
7933         }
7934         if (++index < BTRFS_NR_RAID_TYPES)
7935                 goto again;
7936         up_read(&info->groups_sem);
7937 }
7938
7939 int btrfs_reserve_extent(struct btrfs_root *root,
7940                          u64 num_bytes, u64 min_alloc_size,
7941                          u64 empty_size, u64 hint_byte,
7942                          struct btrfs_key *ins, int is_data, int delalloc)
7943 {
7944         bool final_tried = num_bytes == min_alloc_size;
7945         u64 flags;
7946         int ret;
7947
7948         flags = btrfs_get_alloc_profile(root, is_data);
7949 again:
7950         WARN_ON(num_bytes < root->sectorsize);
7951         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7952                                flags, delalloc);
7953         if (!ret && !is_data) {
7954                 btrfs_dec_block_group_reservations(root->fs_info,
7955                                                    ins->objectid);
7956         } else if (ret == -ENOSPC) {
7957                 if (!final_tried && ins->offset) {
7958                         num_bytes = min(num_bytes >> 1, ins->offset);
7959                         num_bytes = round_down(num_bytes, root->sectorsize);
7960                         num_bytes = max(num_bytes, min_alloc_size);
7961                         if (num_bytes == min_alloc_size)
7962                                 final_tried = true;
7963                         goto again;
7964                 } else if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
7965                         struct btrfs_space_info *sinfo;
7966
7967                         sinfo = __find_space_info(root->fs_info, flags);
7968                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7969                                 flags, num_bytes);
7970                         if (sinfo)
7971                                 dump_space_info(sinfo, num_bytes, 1);
7972                 }
7973         }
7974
7975         return ret;
7976 }
7977
7978 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7979                                         u64 start, u64 len,
7980                                         int pin, int delalloc)
7981 {
7982         struct btrfs_block_group_cache *cache;
7983         int ret = 0;
7984
7985         cache = btrfs_lookup_block_group(root->fs_info, start);
7986         if (!cache) {
7987                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7988                         start);
7989                 return -ENOSPC;
7990         }
7991
7992         if (pin)
7993                 pin_down_extent(root, cache, start, len, 1);
7994         else {
7995                 if (btrfs_test_opt(root->fs_info, DISCARD))
7996                         ret = btrfs_discard_extent(root, start, len, NULL);
7997                 btrfs_add_free_space(cache, start, len);
7998                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7999                 trace_btrfs_reserved_extent_free(root, start, len);
8000         }
8001
8002         btrfs_put_block_group(cache);
8003         return ret;
8004 }
8005
8006 int btrfs_free_reserved_extent(struct btrfs_root *root,
8007                                u64 start, u64 len, int delalloc)
8008 {
8009         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
8010 }
8011
8012 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
8013                                        u64 start, u64 len)
8014 {
8015         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
8016 }
8017
8018 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8019                                       struct btrfs_root *root,
8020                                       u64 parent, u64 root_objectid,
8021                                       u64 flags, u64 owner, u64 offset,
8022                                       struct btrfs_key *ins, int ref_mod)
8023 {
8024         int ret;
8025         struct btrfs_fs_info *fs_info = root->fs_info;
8026         struct btrfs_extent_item *extent_item;
8027         struct btrfs_extent_inline_ref *iref;
8028         struct btrfs_path *path;
8029         struct extent_buffer *leaf;
8030         int type;
8031         u32 size;
8032
8033         if (parent > 0)
8034                 type = BTRFS_SHARED_DATA_REF_KEY;
8035         else
8036                 type = BTRFS_EXTENT_DATA_REF_KEY;
8037
8038         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
8039
8040         path = btrfs_alloc_path();
8041         if (!path)
8042                 return -ENOMEM;
8043
8044         path->leave_spinning = 1;
8045         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8046                                       ins, size);
8047         if (ret) {
8048                 btrfs_free_path(path);
8049                 return ret;
8050         }
8051
8052         leaf = path->nodes[0];
8053         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8054                                      struct btrfs_extent_item);
8055         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
8056         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8057         btrfs_set_extent_flags(leaf, extent_item,
8058                                flags | BTRFS_EXTENT_FLAG_DATA);
8059
8060         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8061         btrfs_set_extent_inline_ref_type(leaf, iref, type);
8062         if (parent > 0) {
8063                 struct btrfs_shared_data_ref *ref;
8064                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
8065                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8066                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
8067         } else {
8068                 struct btrfs_extent_data_ref *ref;
8069                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
8070                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
8071                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
8072                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
8073                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
8074         }
8075
8076         btrfs_mark_buffer_dirty(path->nodes[0]);
8077         btrfs_free_path(path);
8078
8079         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8080                                           ins->offset);
8081         if (ret)
8082                 return ret;
8083
8084         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
8085         if (ret) { /* -ENOENT, logic error */
8086                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8087                         ins->objectid, ins->offset);
8088                 BUG();
8089         }
8090         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
8091         return ret;
8092 }
8093
8094 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
8095                                      struct btrfs_root *root,
8096                                      u64 parent, u64 root_objectid,
8097                                      u64 flags, struct btrfs_disk_key *key,
8098                                      int level, struct btrfs_key *ins)
8099 {
8100         int ret;
8101         struct btrfs_fs_info *fs_info = root->fs_info;
8102         struct btrfs_extent_item *extent_item;
8103         struct btrfs_tree_block_info *block_info;
8104         struct btrfs_extent_inline_ref *iref;
8105         struct btrfs_path *path;
8106         struct extent_buffer *leaf;
8107         u32 size = sizeof(*extent_item) + sizeof(*iref);
8108         u64 num_bytes = ins->offset;
8109         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8110                                                  SKINNY_METADATA);
8111
8112         if (!skinny_metadata)
8113                 size += sizeof(*block_info);
8114
8115         path = btrfs_alloc_path();
8116         if (!path) {
8117                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
8118                                                    root->nodesize);
8119                 return -ENOMEM;
8120         }
8121
8122         path->leave_spinning = 1;
8123         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
8124                                       ins, size);
8125         if (ret) {
8126                 btrfs_free_path(path);
8127                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
8128                                                    root->nodesize);
8129                 return ret;
8130         }
8131
8132         leaf = path->nodes[0];
8133         extent_item = btrfs_item_ptr(leaf, path->slots[0],
8134                                      struct btrfs_extent_item);
8135         btrfs_set_extent_refs(leaf, extent_item, 1);
8136         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
8137         btrfs_set_extent_flags(leaf, extent_item,
8138                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
8139
8140         if (skinny_metadata) {
8141                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
8142                 num_bytes = root->nodesize;
8143         } else {
8144                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
8145                 btrfs_set_tree_block_key(leaf, block_info, key);
8146                 btrfs_set_tree_block_level(leaf, block_info, level);
8147                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
8148         }
8149
8150         if (parent > 0) {
8151                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
8152                 btrfs_set_extent_inline_ref_type(leaf, iref,
8153                                                  BTRFS_SHARED_BLOCK_REF_KEY);
8154                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
8155         } else {
8156                 btrfs_set_extent_inline_ref_type(leaf, iref,
8157                                                  BTRFS_TREE_BLOCK_REF_KEY);
8158                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
8159         }
8160
8161         btrfs_mark_buffer_dirty(leaf);
8162         btrfs_free_path(path);
8163
8164         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
8165                                           num_bytes);
8166         if (ret)
8167                 return ret;
8168
8169         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
8170                                  1);
8171         if (ret) { /* -ENOENT, logic error */
8172                 btrfs_err(fs_info, "update block group failed for %llu %llu",
8173                         ins->objectid, ins->offset);
8174                 BUG();
8175         }
8176
8177         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
8178         return ret;
8179 }
8180
8181 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
8182                                      struct btrfs_root *root,
8183                                      u64 root_objectid, u64 owner,
8184                                      u64 offset, u64 ram_bytes,
8185                                      struct btrfs_key *ins)
8186 {
8187         int ret;
8188
8189         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
8190
8191         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
8192                                          ins->offset, 0,
8193                                          root_objectid, owner, offset,
8194                                          ram_bytes, BTRFS_ADD_DELAYED_EXTENT,
8195                                          NULL);
8196         return ret;
8197 }
8198
8199 /*
8200  * this is used by the tree logging recovery code.  It records that
8201  * an extent has been allocated and makes sure to clear the free
8202  * space cache bits as well
8203  */
8204 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
8205                                    struct btrfs_root *root,
8206                                    u64 root_objectid, u64 owner, u64 offset,
8207                                    struct btrfs_key *ins)
8208 {
8209         int ret;
8210         struct btrfs_block_group_cache *block_group;
8211
8212         /*
8213          * Mixed block groups will exclude before processing the log so we only
8214          * need to do the exclude dance if this fs isn't mixed.
8215          */
8216         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
8217                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
8218                 if (ret)
8219                         return ret;
8220         }
8221
8222         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
8223         if (!block_group)
8224                 return -EINVAL;
8225
8226         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
8227                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
8228         BUG_ON(ret); /* logic error */
8229         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
8230                                          0, owner, offset, ins, 1);
8231         btrfs_put_block_group(block_group);
8232         return ret;
8233 }
8234
8235 static struct extent_buffer *
8236 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
8237                       u64 bytenr, int level)
8238 {
8239         struct extent_buffer *buf;
8240
8241         buf = btrfs_find_create_tree_block(root, bytenr);
8242         if (IS_ERR(buf))
8243                 return buf;
8244
8245         btrfs_set_header_generation(buf, trans->transid);
8246         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
8247         btrfs_tree_lock(buf);
8248         clean_tree_block(trans, root->fs_info, buf);
8249         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
8250
8251         btrfs_set_lock_blocking(buf);
8252         set_extent_buffer_uptodate(buf);
8253
8254         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
8255                 buf->log_index = root->log_transid % 2;
8256                 /*
8257                  * we allow two log transactions at a time, use different
8258                  * EXENT bit to differentiate dirty pages.
8259                  */
8260                 if (buf->log_index == 0)
8261                         set_extent_dirty(&root->dirty_log_pages, buf->start,
8262                                         buf->start + buf->len - 1, GFP_NOFS);
8263                 else
8264                         set_extent_new(&root->dirty_log_pages, buf->start,
8265                                         buf->start + buf->len - 1);
8266         } else {
8267                 buf->log_index = -1;
8268                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
8269                          buf->start + buf->len - 1, GFP_NOFS);
8270         }
8271         trans->dirty = true;
8272         /* this returns a buffer locked for blocking */
8273         return buf;
8274 }
8275
8276 static struct btrfs_block_rsv *
8277 use_block_rsv(struct btrfs_trans_handle *trans,
8278               struct btrfs_root *root, u32 blocksize)
8279 {
8280         struct btrfs_block_rsv *block_rsv;
8281         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
8282         int ret;
8283         bool global_updated = false;
8284
8285         block_rsv = get_block_rsv(trans, root);
8286
8287         if (unlikely(block_rsv->size == 0))
8288                 goto try_reserve;
8289 again:
8290         ret = block_rsv_use_bytes(block_rsv, blocksize);
8291         if (!ret)
8292                 return block_rsv;
8293
8294         if (block_rsv->failfast)
8295                 return ERR_PTR(ret);
8296
8297         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
8298                 global_updated = true;
8299                 update_global_block_rsv(root->fs_info);
8300                 goto again;
8301         }
8302
8303         if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
8304                 static DEFINE_RATELIMIT_STATE(_rs,
8305                                 DEFAULT_RATELIMIT_INTERVAL * 10,
8306                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
8307                 if (__ratelimit(&_rs))
8308                         WARN(1, KERN_DEBUG
8309                                 "BTRFS: block rsv returned %d\n", ret);
8310         }
8311 try_reserve:
8312         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
8313                                      BTRFS_RESERVE_NO_FLUSH);
8314         if (!ret)
8315                 return block_rsv;
8316         /*
8317          * If we couldn't reserve metadata bytes try and use some from
8318          * the global reserve if its space type is the same as the global
8319          * reservation.
8320          */
8321         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
8322             block_rsv->space_info == global_rsv->space_info) {
8323                 ret = block_rsv_use_bytes(global_rsv, blocksize);
8324                 if (!ret)
8325                         return global_rsv;
8326         }
8327         return ERR_PTR(ret);
8328 }
8329
8330 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
8331                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
8332 {
8333         block_rsv_add_bytes(block_rsv, blocksize, 0);
8334         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
8335 }
8336
8337 /*
8338  * finds a free extent and does all the dirty work required for allocation
8339  * returns the tree buffer or an ERR_PTR on error.
8340  */
8341 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
8342                                         struct btrfs_root *root,
8343                                         u64 parent, u64 root_objectid,
8344                                         struct btrfs_disk_key *key, int level,
8345                                         u64 hint, u64 empty_size)
8346 {
8347         struct btrfs_key ins;
8348         struct btrfs_block_rsv *block_rsv;
8349         struct extent_buffer *buf;
8350         struct btrfs_delayed_extent_op *extent_op;
8351         u64 flags = 0;
8352         int ret;
8353         u32 blocksize = root->nodesize;
8354         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
8355                                                  SKINNY_METADATA);
8356
8357 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
8358         if (btrfs_is_testing(root->fs_info)) {
8359                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
8360                                             level);
8361                 if (!IS_ERR(buf))
8362                         root->alloc_bytenr += blocksize;
8363                 return buf;
8364         }
8365 #endif
8366
8367         block_rsv = use_block_rsv(trans, root, blocksize);
8368         if (IS_ERR(block_rsv))
8369                 return ERR_CAST(block_rsv);
8370
8371         ret = btrfs_reserve_extent(root, blocksize, blocksize,
8372                                    empty_size, hint, &ins, 0, 0);
8373         if (ret)
8374                 goto out_unuse;
8375
8376         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8377         if (IS_ERR(buf)) {
8378                 ret = PTR_ERR(buf);
8379                 goto out_free_reserved;
8380         }
8381
8382         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8383                 if (parent == 0)
8384                         parent = ins.objectid;
8385                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8386         } else
8387                 BUG_ON(parent > 0);
8388
8389         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8390                 extent_op = btrfs_alloc_delayed_extent_op();
8391                 if (!extent_op) {
8392                         ret = -ENOMEM;
8393                         goto out_free_buf;
8394                 }
8395                 if (key)
8396                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8397                 else
8398                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8399                 extent_op->flags_to_set = flags;
8400                 extent_op->update_key = skinny_metadata ? false : true;
8401                 extent_op->update_flags = true;
8402                 extent_op->is_data = false;
8403                 extent_op->level = level;
8404
8405                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
8406                                                  ins.objectid, ins.offset,
8407                                                  parent, root_objectid, level,
8408                                                  BTRFS_ADD_DELAYED_EXTENT,
8409                                                  extent_op);
8410                 if (ret)
8411                         goto out_free_delayed;
8412         }
8413         return buf;
8414
8415 out_free_delayed:
8416         btrfs_free_delayed_extent_op(extent_op);
8417 out_free_buf:
8418         free_extent_buffer(buf);
8419 out_free_reserved:
8420         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
8421 out_unuse:
8422         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
8423         return ERR_PTR(ret);
8424 }
8425
8426 struct walk_control {
8427         u64 refs[BTRFS_MAX_LEVEL];
8428         u64 flags[BTRFS_MAX_LEVEL];
8429         struct btrfs_key update_progress;
8430         int stage;
8431         int level;
8432         int shared_level;
8433         int update_ref;
8434         int keep_locks;
8435         int reada_slot;
8436         int reada_count;
8437         int for_reloc;
8438 };
8439
8440 #define DROP_REFERENCE  1
8441 #define UPDATE_BACKREF  2
8442
8443 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8444                                      struct btrfs_root *root,
8445                                      struct walk_control *wc,
8446                                      struct btrfs_path *path)
8447 {
8448         u64 bytenr;
8449         u64 generation;
8450         u64 refs;
8451         u64 flags;
8452         u32 nritems;
8453         u32 blocksize;
8454         struct btrfs_key key;
8455         struct extent_buffer *eb;
8456         int ret;
8457         int slot;
8458         int nread = 0;
8459
8460         if (path->slots[wc->level] < wc->reada_slot) {
8461                 wc->reada_count = wc->reada_count * 2 / 3;
8462                 wc->reada_count = max(wc->reada_count, 2);
8463         } else {
8464                 wc->reada_count = wc->reada_count * 3 / 2;
8465                 wc->reada_count = min_t(int, wc->reada_count,
8466                                         BTRFS_NODEPTRS_PER_BLOCK(root));
8467         }
8468
8469         eb = path->nodes[wc->level];
8470         nritems = btrfs_header_nritems(eb);
8471         blocksize = root->nodesize;
8472
8473         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8474                 if (nread >= wc->reada_count)
8475                         break;
8476
8477                 cond_resched();
8478                 bytenr = btrfs_node_blockptr(eb, slot);
8479                 generation = btrfs_node_ptr_generation(eb, slot);
8480
8481                 if (slot == path->slots[wc->level])
8482                         goto reada;
8483
8484                 if (wc->stage == UPDATE_BACKREF &&
8485                     generation <= root->root_key.offset)
8486                         continue;
8487
8488                 /* We don't lock the tree block, it's OK to be racy here */
8489                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
8490                                                wc->level - 1, 1, &refs,
8491                                                &flags);
8492                 /* We don't care about errors in readahead. */
8493                 if (ret < 0)
8494                         continue;
8495                 BUG_ON(refs == 0);
8496
8497                 if (wc->stage == DROP_REFERENCE) {
8498                         if (refs == 1)
8499                                 goto reada;
8500
8501                         if (wc->level == 1 &&
8502                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8503                                 continue;
8504                         if (!wc->update_ref ||
8505                             generation <= root->root_key.offset)
8506                                 continue;
8507                         btrfs_node_key_to_cpu(eb, &key, slot);
8508                         ret = btrfs_comp_cpu_keys(&key,
8509                                                   &wc->update_progress);
8510                         if (ret < 0)
8511                                 continue;
8512                 } else {
8513                         if (wc->level == 1 &&
8514                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8515                                 continue;
8516                 }
8517 reada:
8518                 readahead_tree_block(root, bytenr);
8519                 nread++;
8520         }
8521         wc->reada_slot = slot;
8522 }
8523
8524 /*
8525  * These may not be seen by the usual inc/dec ref code so we have to
8526  * add them here.
8527  */
8528 static int record_one_subtree_extent(struct btrfs_trans_handle *trans,
8529                                      struct btrfs_root *root, u64 bytenr,
8530                                      u64 num_bytes)
8531 {
8532         struct btrfs_qgroup_extent_record *qrecord;
8533         struct btrfs_delayed_ref_root *delayed_refs;
8534
8535         qrecord = kmalloc(sizeof(*qrecord), GFP_NOFS);
8536         if (!qrecord)
8537                 return -ENOMEM;
8538
8539         qrecord->bytenr = bytenr;
8540         qrecord->num_bytes = num_bytes;
8541         qrecord->old_roots = NULL;
8542
8543         delayed_refs = &trans->transaction->delayed_refs;
8544         spin_lock(&delayed_refs->lock);
8545         if (btrfs_qgroup_insert_dirty_extent(trans->fs_info,
8546                                              delayed_refs, qrecord))
8547                 kfree(qrecord);
8548         spin_unlock(&delayed_refs->lock);
8549
8550         return 0;
8551 }
8552
8553 static int account_leaf_items(struct btrfs_trans_handle *trans,
8554                               struct btrfs_root *root,
8555                               struct extent_buffer *eb)
8556 {
8557         int nr = btrfs_header_nritems(eb);
8558         int i, extent_type, ret;
8559         struct btrfs_key key;
8560         struct btrfs_file_extent_item *fi;
8561         u64 bytenr, num_bytes;
8562
8563         /* We can be called directly from walk_up_proc() */
8564         if (!root->fs_info->quota_enabled)
8565                 return 0;
8566
8567         for (i = 0; i < nr; i++) {
8568                 btrfs_item_key_to_cpu(eb, &key, i);
8569
8570                 if (key.type != BTRFS_EXTENT_DATA_KEY)
8571                         continue;
8572
8573                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
8574                 /* filter out non qgroup-accountable extents  */
8575                 extent_type = btrfs_file_extent_type(eb, fi);
8576
8577                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
8578                         continue;
8579
8580                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
8581                 if (!bytenr)
8582                         continue;
8583
8584                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
8585
8586                 ret = record_one_subtree_extent(trans, root, bytenr, num_bytes);
8587                 if (ret)
8588                         return ret;
8589         }
8590         return 0;
8591 }
8592
8593 /*
8594  * Walk up the tree from the bottom, freeing leaves and any interior
8595  * nodes which have had all slots visited. If a node (leaf or
8596  * interior) is freed, the node above it will have it's slot
8597  * incremented. The root node will never be freed.
8598  *
8599  * At the end of this function, we should have a path which has all
8600  * slots incremented to the next position for a search. If we need to
8601  * read a new node it will be NULL and the node above it will have the
8602  * correct slot selected for a later read.
8603  *
8604  * If we increment the root nodes slot counter past the number of
8605  * elements, 1 is returned to signal completion of the search.
8606  */
8607 static int adjust_slots_upwards(struct btrfs_root *root,
8608                                 struct btrfs_path *path, int root_level)
8609 {
8610         int level = 0;
8611         int nr, slot;
8612         struct extent_buffer *eb;
8613
8614         if (root_level == 0)
8615                 return 1;
8616
8617         while (level <= root_level) {
8618                 eb = path->nodes[level];
8619                 nr = btrfs_header_nritems(eb);
8620                 path->slots[level]++;
8621                 slot = path->slots[level];
8622                 if (slot >= nr || level == 0) {
8623                         /*
8624                          * Don't free the root -  we will detect this
8625                          * condition after our loop and return a
8626                          * positive value for caller to stop walking the tree.
8627                          */
8628                         if (level != root_level) {
8629                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8630                                 path->locks[level] = 0;
8631
8632                                 free_extent_buffer(eb);
8633                                 path->nodes[level] = NULL;
8634                                 path->slots[level] = 0;
8635                         }
8636                 } else {
8637                         /*
8638                          * We have a valid slot to walk back down
8639                          * from. Stop here so caller can process these
8640                          * new nodes.
8641                          */
8642                         break;
8643                 }
8644
8645                 level++;
8646         }
8647
8648         eb = path->nodes[root_level];
8649         if (path->slots[root_level] >= btrfs_header_nritems(eb))
8650                 return 1;
8651
8652         return 0;
8653 }
8654
8655 /*
8656  * root_eb is the subtree root and is locked before this function is called.
8657  */
8658 static int account_shared_subtree(struct btrfs_trans_handle *trans,
8659                                   struct btrfs_root *root,
8660                                   struct extent_buffer *root_eb,
8661                                   u64 root_gen,
8662                                   int root_level)
8663 {
8664         int ret = 0;
8665         int level;
8666         struct extent_buffer *eb = root_eb;
8667         struct btrfs_path *path = NULL;
8668
8669         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
8670         BUG_ON(root_eb == NULL);
8671
8672         if (!root->fs_info->quota_enabled)
8673                 return 0;
8674
8675         if (!extent_buffer_uptodate(root_eb)) {
8676                 ret = btrfs_read_buffer(root_eb, root_gen);
8677                 if (ret)
8678                         goto out;
8679         }
8680
8681         if (root_level == 0) {
8682                 ret = account_leaf_items(trans, root, root_eb);
8683                 goto out;
8684         }
8685
8686         path = btrfs_alloc_path();
8687         if (!path)
8688                 return -ENOMEM;
8689
8690         /*
8691          * Walk down the tree.  Missing extent blocks are filled in as
8692          * we go. Metadata is accounted every time we read a new
8693          * extent block.
8694          *
8695          * When we reach a leaf, we account for file extent items in it,
8696          * walk back up the tree (adjusting slot pointers as we go)
8697          * and restart the search process.
8698          */
8699         extent_buffer_get(root_eb); /* For path */
8700         path->nodes[root_level] = root_eb;
8701         path->slots[root_level] = 0;
8702         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
8703 walk_down:
8704         level = root_level;
8705         while (level >= 0) {
8706                 if (path->nodes[level] == NULL) {
8707                         int parent_slot;
8708                         u64 child_gen;
8709                         u64 child_bytenr;
8710
8711                         /* We need to get child blockptr/gen from
8712                          * parent before we can read it. */
8713                         eb = path->nodes[level + 1];
8714                         parent_slot = path->slots[level + 1];
8715                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8716                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8717
8718                         eb = read_tree_block(root, child_bytenr, child_gen);
8719                         if (IS_ERR(eb)) {
8720                                 ret = PTR_ERR(eb);
8721                                 goto out;
8722                         } else if (!extent_buffer_uptodate(eb)) {
8723                                 free_extent_buffer(eb);
8724                                 ret = -EIO;
8725                                 goto out;
8726                         }
8727
8728                         path->nodes[level] = eb;
8729                         path->slots[level] = 0;
8730
8731                         btrfs_tree_read_lock(eb);
8732                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8733                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8734
8735                         ret = record_one_subtree_extent(trans, root, child_bytenr,
8736                                                         root->nodesize);
8737                         if (ret)
8738                                 goto out;
8739                 }
8740
8741                 if (level == 0) {
8742                         ret = account_leaf_items(trans, root, path->nodes[level]);
8743                         if (ret)
8744                                 goto out;
8745
8746                         /* Nonzero return here means we completed our search */
8747                         ret = adjust_slots_upwards(root, path, root_level);
8748                         if (ret)
8749                                 break;
8750
8751                         /* Restart search with new slots */
8752                         goto walk_down;
8753                 }
8754
8755                 level--;
8756         }
8757
8758         ret = 0;
8759 out:
8760         btrfs_free_path(path);
8761
8762         return ret;
8763 }
8764
8765 /*
8766  * helper to process tree block while walking down the tree.
8767  *
8768  * when wc->stage == UPDATE_BACKREF, this function updates
8769  * back refs for pointers in the block.
8770  *
8771  * NOTE: return value 1 means we should stop walking down.
8772  */
8773 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8774                                    struct btrfs_root *root,
8775                                    struct btrfs_path *path,
8776                                    struct walk_control *wc, int lookup_info)
8777 {
8778         int level = wc->level;
8779         struct extent_buffer *eb = path->nodes[level];
8780         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8781         int ret;
8782
8783         if (wc->stage == UPDATE_BACKREF &&
8784             btrfs_header_owner(eb) != root->root_key.objectid)
8785                 return 1;
8786
8787         /*
8788          * when reference count of tree block is 1, it won't increase
8789          * again. once full backref flag is set, we never clear it.
8790          */
8791         if (lookup_info &&
8792             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8793              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8794                 BUG_ON(!path->locks[level]);
8795                 ret = btrfs_lookup_extent_info(trans, root,
8796                                                eb->start, level, 1,
8797                                                &wc->refs[level],
8798                                                &wc->flags[level]);
8799                 BUG_ON(ret == -ENOMEM);
8800                 if (ret)
8801                         return ret;
8802                 BUG_ON(wc->refs[level] == 0);
8803         }
8804
8805         if (wc->stage == DROP_REFERENCE) {
8806                 if (wc->refs[level] > 1)
8807                         return 1;
8808
8809                 if (path->locks[level] && !wc->keep_locks) {
8810                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8811                         path->locks[level] = 0;
8812                 }
8813                 return 0;
8814         }
8815
8816         /* wc->stage == UPDATE_BACKREF */
8817         if (!(wc->flags[level] & flag)) {
8818                 BUG_ON(!path->locks[level]);
8819                 ret = btrfs_inc_ref(trans, root, eb, 1);
8820                 BUG_ON(ret); /* -ENOMEM */
8821                 ret = btrfs_dec_ref(trans, root, eb, 0);
8822                 BUG_ON(ret); /* -ENOMEM */
8823                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8824                                                   eb->len, flag,
8825                                                   btrfs_header_level(eb), 0);
8826                 BUG_ON(ret); /* -ENOMEM */
8827                 wc->flags[level] |= flag;
8828         }
8829
8830         /*
8831          * the block is shared by multiple trees, so it's not good to
8832          * keep the tree lock
8833          */
8834         if (path->locks[level] && level > 0) {
8835                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8836                 path->locks[level] = 0;
8837         }
8838         return 0;
8839 }
8840
8841 /*
8842  * helper to process tree block pointer.
8843  *
8844  * when wc->stage == DROP_REFERENCE, this function checks
8845  * reference count of the block pointed to. if the block
8846  * is shared and we need update back refs for the subtree
8847  * rooted at the block, this function changes wc->stage to
8848  * UPDATE_BACKREF. if the block is shared and there is no
8849  * need to update back, this function drops the reference
8850  * to the block.
8851  *
8852  * NOTE: return value 1 means we should stop walking down.
8853  */
8854 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8855                                  struct btrfs_root *root,
8856                                  struct btrfs_path *path,
8857                                  struct walk_control *wc, int *lookup_info)
8858 {
8859         u64 bytenr;
8860         u64 generation;
8861         u64 parent;
8862         u32 blocksize;
8863         struct btrfs_key key;
8864         struct extent_buffer *next;
8865         int level = wc->level;
8866         int reada = 0;
8867         int ret = 0;
8868         bool need_account = false;
8869
8870         generation = btrfs_node_ptr_generation(path->nodes[level],
8871                                                path->slots[level]);
8872         /*
8873          * if the lower level block was created before the snapshot
8874          * was created, we know there is no need to update back refs
8875          * for the subtree
8876          */
8877         if (wc->stage == UPDATE_BACKREF &&
8878             generation <= root->root_key.offset) {
8879                 *lookup_info = 1;
8880                 return 1;
8881         }
8882
8883         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8884         blocksize = root->nodesize;
8885
8886         next = btrfs_find_tree_block(root->fs_info, bytenr);
8887         if (!next) {
8888                 next = btrfs_find_create_tree_block(root, bytenr);
8889                 if (IS_ERR(next))
8890                         return PTR_ERR(next);
8891
8892                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8893                                                level - 1);
8894                 reada = 1;
8895         }
8896         btrfs_tree_lock(next);
8897         btrfs_set_lock_blocking(next);
8898
8899         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8900                                        &wc->refs[level - 1],
8901                                        &wc->flags[level - 1]);
8902         if (ret < 0) {
8903                 btrfs_tree_unlock(next);
8904                 return ret;
8905         }
8906
8907         if (unlikely(wc->refs[level - 1] == 0)) {
8908                 btrfs_err(root->fs_info, "Missing references.");
8909                 BUG();
8910         }
8911         *lookup_info = 0;
8912
8913         if (wc->stage == DROP_REFERENCE) {
8914                 if (wc->refs[level - 1] > 1) {
8915                         need_account = true;
8916                         if (level == 1 &&
8917                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8918                                 goto skip;
8919
8920                         if (!wc->update_ref ||
8921                             generation <= root->root_key.offset)
8922                                 goto skip;
8923
8924                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8925                                               path->slots[level]);
8926                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8927                         if (ret < 0)
8928                                 goto skip;
8929
8930                         wc->stage = UPDATE_BACKREF;
8931                         wc->shared_level = level - 1;
8932                 }
8933         } else {
8934                 if (level == 1 &&
8935                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8936                         goto skip;
8937         }
8938
8939         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8940                 btrfs_tree_unlock(next);
8941                 free_extent_buffer(next);
8942                 next = NULL;
8943                 *lookup_info = 1;
8944         }
8945
8946         if (!next) {
8947                 if (reada && level == 1)
8948                         reada_walk_down(trans, root, wc, path);
8949                 next = read_tree_block(root, bytenr, generation);
8950                 if (IS_ERR(next)) {
8951                         return PTR_ERR(next);
8952                 } else if (!extent_buffer_uptodate(next)) {
8953                         free_extent_buffer(next);
8954                         return -EIO;
8955                 }
8956                 btrfs_tree_lock(next);
8957                 btrfs_set_lock_blocking(next);
8958         }
8959
8960         level--;
8961         BUG_ON(level != btrfs_header_level(next));
8962         path->nodes[level] = next;
8963         path->slots[level] = 0;
8964         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8965         wc->level = level;
8966         if (wc->level == 1)
8967                 wc->reada_slot = 0;
8968         return 0;
8969 skip:
8970         wc->refs[level - 1] = 0;
8971         wc->flags[level - 1] = 0;
8972         if (wc->stage == DROP_REFERENCE) {
8973                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8974                         parent = path->nodes[level]->start;
8975                 } else {
8976                         BUG_ON(root->root_key.objectid !=
8977                                btrfs_header_owner(path->nodes[level]));
8978                         parent = 0;
8979                 }
8980
8981                 if (need_account) {
8982                         ret = account_shared_subtree(trans, root, next,
8983                                                      generation, level - 1);
8984                         if (ret) {
8985                                 btrfs_err_rl(root->fs_info,
8986                                         "Error "
8987                                         "%d accounting shared subtree. Quota "
8988                                         "is out of sync, rescan required.",
8989                                         ret);
8990                         }
8991                 }
8992                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8993                                 root->root_key.objectid, level - 1, 0);
8994                 BUG_ON(ret); /* -ENOMEM */
8995         }
8996         btrfs_tree_unlock(next);
8997         free_extent_buffer(next);
8998         *lookup_info = 1;
8999         return 1;
9000 }
9001
9002 /*
9003  * helper to process tree block while walking up the tree.
9004  *
9005  * when wc->stage == DROP_REFERENCE, this function drops
9006  * reference count on the block.
9007  *
9008  * when wc->stage == UPDATE_BACKREF, this function changes
9009  * wc->stage back to DROP_REFERENCE if we changed wc->stage
9010  * to UPDATE_BACKREF previously while processing the block.
9011  *
9012  * NOTE: return value 1 means we should stop walking up.
9013  */
9014 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
9015                                  struct btrfs_root *root,
9016                                  struct btrfs_path *path,
9017                                  struct walk_control *wc)
9018 {
9019         int ret;
9020         int level = wc->level;
9021         struct extent_buffer *eb = path->nodes[level];
9022         u64 parent = 0;
9023
9024         if (wc->stage == UPDATE_BACKREF) {
9025                 BUG_ON(wc->shared_level < level);
9026                 if (level < wc->shared_level)
9027                         goto out;
9028
9029                 ret = find_next_key(path, level + 1, &wc->update_progress);
9030                 if (ret > 0)
9031                         wc->update_ref = 0;
9032
9033                 wc->stage = DROP_REFERENCE;
9034                 wc->shared_level = -1;
9035                 path->slots[level] = 0;
9036
9037                 /*
9038                  * check reference count again if the block isn't locked.
9039                  * we should start walking down the tree again if reference
9040                  * count is one.
9041                  */
9042                 if (!path->locks[level]) {
9043                         BUG_ON(level == 0);
9044                         btrfs_tree_lock(eb);
9045                         btrfs_set_lock_blocking(eb);
9046                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9047
9048                         ret = btrfs_lookup_extent_info(trans, root,
9049                                                        eb->start, level, 1,
9050                                                        &wc->refs[level],
9051                                                        &wc->flags[level]);
9052                         if (ret < 0) {
9053                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9054                                 path->locks[level] = 0;
9055                                 return ret;
9056                         }
9057                         BUG_ON(wc->refs[level] == 0);
9058                         if (wc->refs[level] == 1) {
9059                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
9060                                 path->locks[level] = 0;
9061                                 return 1;
9062                         }
9063                 }
9064         }
9065
9066         /* wc->stage == DROP_REFERENCE */
9067         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
9068
9069         if (wc->refs[level] == 1) {
9070                 if (level == 0) {
9071                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
9072                                 ret = btrfs_dec_ref(trans, root, eb, 1);
9073                         else
9074                                 ret = btrfs_dec_ref(trans, root, eb, 0);
9075                         BUG_ON(ret); /* -ENOMEM */
9076                         ret = account_leaf_items(trans, root, eb);
9077                         if (ret) {
9078                                 btrfs_err_rl(root->fs_info,
9079                                         "error "
9080                                         "%d accounting leaf items. Quota "
9081                                         "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_path *path;
9199         struct btrfs_trans_handle *trans;
9200         struct btrfs_root *tree_root = root->fs_info->tree_root;
9201         struct btrfs_root_item *root_item = &root->root_item;
9202         struct walk_control *wc;
9203         struct btrfs_key key;
9204         int err = 0;
9205         int ret;
9206         int level;
9207         bool root_dropped = false;
9208
9209         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
9210
9211         path = btrfs_alloc_path();
9212         if (!path) {
9213                 err = -ENOMEM;
9214                 goto out;
9215         }
9216
9217         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9218         if (!wc) {
9219                 btrfs_free_path(path);
9220                 err = -ENOMEM;
9221                 goto out;
9222         }
9223
9224         trans = btrfs_start_transaction(tree_root, 0);
9225         if (IS_ERR(trans)) {
9226                 err = PTR_ERR(trans);
9227                 goto out_free;
9228         }
9229
9230         if (block_rsv)
9231                 trans->block_rsv = block_rsv;
9232
9233         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
9234                 level = btrfs_header_level(root->node);
9235                 path->nodes[level] = btrfs_lock_root_node(root);
9236                 btrfs_set_lock_blocking(path->nodes[level]);
9237                 path->slots[level] = 0;
9238                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9239                 memset(&wc->update_progress, 0,
9240                        sizeof(wc->update_progress));
9241         } else {
9242                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
9243                 memcpy(&wc->update_progress, &key,
9244                        sizeof(wc->update_progress));
9245
9246                 level = root_item->drop_level;
9247                 BUG_ON(level == 0);
9248                 path->lowest_level = level;
9249                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
9250                 path->lowest_level = 0;
9251                 if (ret < 0) {
9252                         err = ret;
9253                         goto out_end_trans;
9254                 }
9255                 WARN_ON(ret > 0);
9256
9257                 /*
9258                  * unlock our path, this is safe because only this
9259                  * function is allowed to delete this snapshot
9260                  */
9261                 btrfs_unlock_up_safe(path, 0);
9262
9263                 level = btrfs_header_level(root->node);
9264                 while (1) {
9265                         btrfs_tree_lock(path->nodes[level]);
9266                         btrfs_set_lock_blocking(path->nodes[level]);
9267                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9268
9269                         ret = btrfs_lookup_extent_info(trans, root,
9270                                                 path->nodes[level]->start,
9271                                                 level, 1, &wc->refs[level],
9272                                                 &wc->flags[level]);
9273                         if (ret < 0) {
9274                                 err = ret;
9275                                 goto out_end_trans;
9276                         }
9277                         BUG_ON(wc->refs[level] == 0);
9278
9279                         if (level == root_item->drop_level)
9280                                 break;
9281
9282                         btrfs_tree_unlock(path->nodes[level]);
9283                         path->locks[level] = 0;
9284                         WARN_ON(wc->refs[level] != 1);
9285                         level--;
9286                 }
9287         }
9288
9289         wc->level = level;
9290         wc->shared_level = -1;
9291         wc->stage = DROP_REFERENCE;
9292         wc->update_ref = update_ref;
9293         wc->keep_locks = 0;
9294         wc->for_reloc = for_reloc;
9295         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9296
9297         while (1) {
9298
9299                 ret = walk_down_tree(trans, root, path, wc);
9300                 if (ret < 0) {
9301                         err = ret;
9302                         break;
9303                 }
9304
9305                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
9306                 if (ret < 0) {
9307                         err = ret;
9308                         break;
9309                 }
9310
9311                 if (ret > 0) {
9312                         BUG_ON(wc->stage != DROP_REFERENCE);
9313                         break;
9314                 }
9315
9316                 if (wc->stage == DROP_REFERENCE) {
9317                         level = wc->level;
9318                         btrfs_node_key(path->nodes[level],
9319                                        &root_item->drop_progress,
9320                                        path->slots[level]);
9321                         root_item->drop_level = level;
9322                 }
9323
9324                 BUG_ON(wc->level == 0);
9325                 if (btrfs_should_end_transaction(trans, tree_root) ||
9326                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
9327                         ret = btrfs_update_root(trans, tree_root,
9328                                                 &root->root_key,
9329                                                 root_item);
9330                         if (ret) {
9331                                 btrfs_abort_transaction(trans, ret);
9332                                 err = ret;
9333                                 goto out_end_trans;
9334                         }
9335
9336                         btrfs_end_transaction_throttle(trans, tree_root);
9337                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
9338                                 pr_debug("BTRFS: drop snapshot early exit\n");
9339                                 err = -EAGAIN;
9340                                 goto out_free;
9341                         }
9342
9343                         trans = btrfs_start_transaction(tree_root, 0);
9344                         if (IS_ERR(trans)) {
9345                                 err = PTR_ERR(trans);
9346                                 goto out_free;
9347                         }
9348                         if (block_rsv)
9349                                 trans->block_rsv = block_rsv;
9350                 }
9351         }
9352         btrfs_release_path(path);
9353         if (err)
9354                 goto out_end_trans;
9355
9356         ret = btrfs_del_root(trans, tree_root, &root->root_key);
9357         if (ret) {
9358                 btrfs_abort_transaction(trans, ret);
9359                 goto out_end_trans;
9360         }
9361
9362         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
9363                 ret = btrfs_find_root(tree_root, &root->root_key, path,
9364                                       NULL, NULL);
9365                 if (ret < 0) {
9366                         btrfs_abort_transaction(trans, ret);
9367                         err = ret;
9368                         goto out_end_trans;
9369                 } else if (ret > 0) {
9370                         /* if we fail to delete the orphan item this time
9371                          * around, it'll get picked up the next time.
9372                          *
9373                          * The most common failure here is just -ENOENT.
9374                          */
9375                         btrfs_del_orphan_item(trans, tree_root,
9376                                               root->root_key.objectid);
9377                 }
9378         }
9379
9380         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9381                 btrfs_add_dropped_root(trans, root);
9382         } else {
9383                 free_extent_buffer(root->node);
9384                 free_extent_buffer(root->commit_root);
9385                 btrfs_put_fs_root(root);
9386         }
9387         root_dropped = true;
9388 out_end_trans:
9389         btrfs_end_transaction_throttle(trans, tree_root);
9390 out_free:
9391         kfree(wc);
9392         btrfs_free_path(path);
9393 out:
9394         /*
9395          * So if we need to stop dropping the snapshot for whatever reason we
9396          * need to make sure to add it back to the dead root list so that we
9397          * keep trying to do the work later.  This also cleans up roots if we
9398          * don't have it in the radix (like when we recover after a power fail
9399          * or unmount) so we don't leak memory.
9400          */
9401         if (!for_reloc && root_dropped == false)
9402                 btrfs_add_dead_root(root);
9403         if (err && err != -EAGAIN)
9404                 btrfs_handle_fs_error(root->fs_info, err, NULL);
9405         return err;
9406 }
9407
9408 /*
9409  * drop subtree rooted at tree block 'node'.
9410  *
9411  * NOTE: this function will unlock and release tree block 'node'
9412  * only used by relocation code
9413  */
9414 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9415                         struct btrfs_root *root,
9416                         struct extent_buffer *node,
9417                         struct extent_buffer *parent)
9418 {
9419         struct btrfs_path *path;
9420         struct walk_control *wc;
9421         int level;
9422         int parent_level;
9423         int ret = 0;
9424         int wret;
9425
9426         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9427
9428         path = btrfs_alloc_path();
9429         if (!path)
9430                 return -ENOMEM;
9431
9432         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9433         if (!wc) {
9434                 btrfs_free_path(path);
9435                 return -ENOMEM;
9436         }
9437
9438         btrfs_assert_tree_locked(parent);
9439         parent_level = btrfs_header_level(parent);
9440         extent_buffer_get(parent);
9441         path->nodes[parent_level] = parent;
9442         path->slots[parent_level] = btrfs_header_nritems(parent);
9443
9444         btrfs_assert_tree_locked(node);
9445         level = btrfs_header_level(node);
9446         path->nodes[level] = node;
9447         path->slots[level] = 0;
9448         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9449
9450         wc->refs[parent_level] = 1;
9451         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9452         wc->level = level;
9453         wc->shared_level = -1;
9454         wc->stage = DROP_REFERENCE;
9455         wc->update_ref = 0;
9456         wc->keep_locks = 1;
9457         wc->for_reloc = 1;
9458         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9459
9460         while (1) {
9461                 wret = walk_down_tree(trans, root, path, wc);
9462                 if (wret < 0) {
9463                         ret = wret;
9464                         break;
9465                 }
9466
9467                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9468                 if (wret < 0)
9469                         ret = wret;
9470                 if (wret != 0)
9471                         break;
9472         }
9473
9474         kfree(wc);
9475         btrfs_free_path(path);
9476         return ret;
9477 }
9478
9479 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
9480 {
9481         u64 num_devices;
9482         u64 stripped;
9483
9484         /*
9485          * if restripe for this chunk_type is on pick target profile and
9486          * return, otherwise do the usual balance
9487          */
9488         stripped = get_restripe_target(root->fs_info, flags);
9489         if (stripped)
9490                 return extended_to_chunk(stripped);
9491
9492         num_devices = root->fs_info->fs_devices->rw_devices;
9493
9494         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9495                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9496                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9497
9498         if (num_devices == 1) {
9499                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9500                 stripped = flags & ~stripped;
9501
9502                 /* turn raid0 into single device chunks */
9503                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9504                         return stripped;
9505
9506                 /* turn mirroring into duplication */
9507                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9508                              BTRFS_BLOCK_GROUP_RAID10))
9509                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9510         } else {
9511                 /* they already had raid on here, just return */
9512                 if (flags & stripped)
9513                         return flags;
9514
9515                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9516                 stripped = flags & ~stripped;
9517
9518                 /* switch duplicated blocks with raid1 */
9519                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9520                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9521
9522                 /* this is drive concat, leave it alone */
9523         }
9524
9525         return flags;
9526 }
9527
9528 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9529 {
9530         struct btrfs_space_info *sinfo = cache->space_info;
9531         u64 num_bytes;
9532         u64 min_allocable_bytes;
9533         int ret = -ENOSPC;
9534
9535         /*
9536          * We need some metadata space and system metadata space for
9537          * allocating chunks in some corner cases until we force to set
9538          * it to be readonly.
9539          */
9540         if ((sinfo->flags &
9541              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9542             !force)
9543                 min_allocable_bytes = SZ_1M;
9544         else
9545                 min_allocable_bytes = 0;
9546
9547         spin_lock(&sinfo->lock);
9548         spin_lock(&cache->lock);
9549
9550         if (cache->ro) {
9551                 cache->ro++;
9552                 ret = 0;
9553                 goto out;
9554         }
9555
9556         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9557                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9558
9559         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
9560             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
9561             min_allocable_bytes <= sinfo->total_bytes) {
9562                 sinfo->bytes_readonly += num_bytes;
9563                 cache->ro++;
9564                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9565                 ret = 0;
9566         }
9567 out:
9568         spin_unlock(&cache->lock);
9569         spin_unlock(&sinfo->lock);
9570         return ret;
9571 }
9572
9573 int btrfs_inc_block_group_ro(struct btrfs_root *root,
9574                              struct btrfs_block_group_cache *cache)
9575
9576 {
9577         struct btrfs_trans_handle *trans;
9578         u64 alloc_flags;
9579         int ret;
9580
9581 again:
9582         trans = btrfs_join_transaction(root);
9583         if (IS_ERR(trans))
9584                 return PTR_ERR(trans);
9585
9586         /*
9587          * we're not allowed to set block groups readonly after the dirty
9588          * block groups cache has started writing.  If it already started,
9589          * back off and let this transaction commit
9590          */
9591         mutex_lock(&root->fs_info->ro_block_group_mutex);
9592         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9593                 u64 transid = trans->transid;
9594
9595                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
9596                 btrfs_end_transaction(trans, root);
9597
9598                 ret = btrfs_wait_for_commit(root, transid);
9599                 if (ret)
9600                         return ret;
9601                 goto again;
9602         }
9603
9604         /*
9605          * if we are changing raid levels, try to allocate a corresponding
9606          * block group with the new raid level.
9607          */
9608         alloc_flags = update_block_group_flags(root, cache->flags);
9609         if (alloc_flags != cache->flags) {
9610                 ret = do_chunk_alloc(trans, root, alloc_flags,
9611                                      CHUNK_ALLOC_FORCE);
9612                 /*
9613                  * ENOSPC is allowed here, we may have enough space
9614                  * already allocated at the new raid level to
9615                  * carry on
9616                  */
9617                 if (ret == -ENOSPC)
9618                         ret = 0;
9619                 if (ret < 0)
9620                         goto out;
9621         }
9622
9623         ret = inc_block_group_ro(cache, 0);
9624         if (!ret)
9625                 goto out;
9626         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
9627         ret = do_chunk_alloc(trans, root, alloc_flags,
9628                              CHUNK_ALLOC_FORCE);
9629         if (ret < 0)
9630                 goto out;
9631         ret = inc_block_group_ro(cache, 0);
9632 out:
9633         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9634                 alloc_flags = update_block_group_flags(root, cache->flags);
9635                 lock_chunks(root->fs_info->chunk_root);
9636                 check_system_chunk(trans, root, alloc_flags);
9637                 unlock_chunks(root->fs_info->chunk_root);
9638         }
9639         mutex_unlock(&root->fs_info->ro_block_group_mutex);
9640
9641         btrfs_end_transaction(trans, root);
9642         return ret;
9643 }
9644
9645 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9646                             struct btrfs_root *root, u64 type)
9647 {
9648         u64 alloc_flags = get_alloc_profile(root, type);
9649         return do_chunk_alloc(trans, root, alloc_flags,
9650                               CHUNK_ALLOC_FORCE);
9651 }
9652
9653 /*
9654  * helper to account the unused space of all the readonly block group in the
9655  * space_info. takes mirrors into account.
9656  */
9657 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9658 {
9659         struct btrfs_block_group_cache *block_group;
9660         u64 free_bytes = 0;
9661         int factor;
9662
9663         /* It's df, we don't care if it's racy */
9664         if (list_empty(&sinfo->ro_bgs))
9665                 return 0;
9666
9667         spin_lock(&sinfo->lock);
9668         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9669                 spin_lock(&block_group->lock);
9670
9671                 if (!block_group->ro) {
9672                         spin_unlock(&block_group->lock);
9673                         continue;
9674                 }
9675
9676                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9677                                           BTRFS_BLOCK_GROUP_RAID10 |
9678                                           BTRFS_BLOCK_GROUP_DUP))
9679                         factor = 2;
9680                 else
9681                         factor = 1;
9682
9683                 free_bytes += (block_group->key.offset -
9684                                btrfs_block_group_used(&block_group->item)) *
9685                                factor;
9686
9687                 spin_unlock(&block_group->lock);
9688         }
9689         spin_unlock(&sinfo->lock);
9690
9691         return free_bytes;
9692 }
9693
9694 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9695                               struct btrfs_block_group_cache *cache)
9696 {
9697         struct btrfs_space_info *sinfo = cache->space_info;
9698         u64 num_bytes;
9699
9700         BUG_ON(!cache->ro);
9701
9702         spin_lock(&sinfo->lock);
9703         spin_lock(&cache->lock);
9704         if (!--cache->ro) {
9705                 num_bytes = cache->key.offset - cache->reserved -
9706                             cache->pinned - cache->bytes_super -
9707                             btrfs_block_group_used(&cache->item);
9708                 sinfo->bytes_readonly -= num_bytes;
9709                 list_del_init(&cache->ro_list);
9710         }
9711         spin_unlock(&cache->lock);
9712         spin_unlock(&sinfo->lock);
9713 }
9714
9715 /*
9716  * checks to see if its even possible to relocate this block group.
9717  *
9718  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9719  * ok to go ahead and try.
9720  */
9721 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9722 {
9723         struct btrfs_block_group_cache *block_group;
9724         struct btrfs_space_info *space_info;
9725         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9726         struct btrfs_device *device;
9727         struct btrfs_trans_handle *trans;
9728         u64 min_free;
9729         u64 dev_min = 1;
9730         u64 dev_nr = 0;
9731         u64 target;
9732         int debug;
9733         int index;
9734         int full = 0;
9735         int ret = 0;
9736
9737         debug = btrfs_test_opt(root->fs_info, ENOSPC_DEBUG);
9738
9739         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9740
9741         /* odd, couldn't find the block group, leave it alone */
9742         if (!block_group) {
9743                 if (debug)
9744                         btrfs_warn(root->fs_info,
9745                                    "can't find block group for bytenr %llu",
9746                                    bytenr);
9747                 return -1;
9748         }
9749
9750         min_free = btrfs_block_group_used(&block_group->item);
9751
9752         /* no bytes used, we're good */
9753         if (!min_free)
9754                 goto out;
9755
9756         space_info = block_group->space_info;
9757         spin_lock(&space_info->lock);
9758
9759         full = space_info->full;
9760
9761         /*
9762          * if this is the last block group we have in this space, we can't
9763          * relocate it unless we're able to allocate a new chunk below.
9764          *
9765          * Otherwise, we need to make sure we have room in the space to handle
9766          * all of the extents from this block group.  If we can, we're good
9767          */
9768         if ((space_info->total_bytes != block_group->key.offset) &&
9769             (space_info->bytes_used + space_info->bytes_reserved +
9770              space_info->bytes_pinned + space_info->bytes_readonly +
9771              min_free < space_info->total_bytes)) {
9772                 spin_unlock(&space_info->lock);
9773                 goto out;
9774         }
9775         spin_unlock(&space_info->lock);
9776
9777         /*
9778          * ok we don't have enough space, but maybe we have free space on our
9779          * devices to allocate new chunks for relocation, so loop through our
9780          * alloc devices and guess if we have enough space.  if this block
9781          * group is going to be restriped, run checks against the target
9782          * profile instead of the current one.
9783          */
9784         ret = -1;
9785
9786         /*
9787          * index:
9788          *      0: raid10
9789          *      1: raid1
9790          *      2: dup
9791          *      3: raid0
9792          *      4: single
9793          */
9794         target = get_restripe_target(root->fs_info, block_group->flags);
9795         if (target) {
9796                 index = __get_raid_index(extended_to_chunk(target));
9797         } else {
9798                 /*
9799                  * this is just a balance, so if we were marked as full
9800                  * we know there is no space for a new chunk
9801                  */
9802                 if (full) {
9803                         if (debug)
9804                                 btrfs_warn(root->fs_info,
9805                                         "no space to alloc new chunk for block group %llu",
9806                                         block_group->key.objectid);
9807                         goto out;
9808                 }
9809
9810                 index = get_block_group_index(block_group);
9811         }
9812
9813         if (index == BTRFS_RAID_RAID10) {
9814                 dev_min = 4;
9815                 /* Divide by 2 */
9816                 min_free >>= 1;
9817         } else if (index == BTRFS_RAID_RAID1) {
9818                 dev_min = 2;
9819         } else if (index == BTRFS_RAID_DUP) {
9820                 /* Multiply by 2 */
9821                 min_free <<= 1;
9822         } else if (index == BTRFS_RAID_RAID0) {
9823                 dev_min = fs_devices->rw_devices;
9824                 min_free = div64_u64(min_free, dev_min);
9825         }
9826
9827         /* We need to do this so that we can look at pending chunks */
9828         trans = btrfs_join_transaction(root);
9829         if (IS_ERR(trans)) {
9830                 ret = PTR_ERR(trans);
9831                 goto out;
9832         }
9833
9834         mutex_lock(&root->fs_info->chunk_mutex);
9835         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9836                 u64 dev_offset;
9837
9838                 /*
9839                  * check to make sure we can actually find a chunk with enough
9840                  * space to fit our block group in.
9841                  */
9842                 if (device->total_bytes > device->bytes_used + min_free &&
9843                     !device->is_tgtdev_for_dev_replace) {
9844                         ret = find_free_dev_extent(trans, device, min_free,
9845                                                    &dev_offset, NULL);
9846                         if (!ret)
9847                                 dev_nr++;
9848
9849                         if (dev_nr >= dev_min)
9850                                 break;
9851
9852                         ret = -1;
9853                 }
9854         }
9855         if (debug && ret == -1)
9856                 btrfs_warn(root->fs_info,
9857                         "no space to allocate a new chunk for block group %llu",
9858                         block_group->key.objectid);
9859         mutex_unlock(&root->fs_info->chunk_mutex);
9860         btrfs_end_transaction(trans, root);
9861 out:
9862         btrfs_put_block_group(block_group);
9863         return ret;
9864 }
9865
9866 static int find_first_block_group(struct btrfs_root *root,
9867                 struct btrfs_path *path, struct btrfs_key *key)
9868 {
9869         int ret = 0;
9870         struct btrfs_key found_key;
9871         struct extent_buffer *leaf;
9872         int slot;
9873
9874         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9875         if (ret < 0)
9876                 goto out;
9877
9878         while (1) {
9879                 slot = path->slots[0];
9880                 leaf = path->nodes[0];
9881                 if (slot >= btrfs_header_nritems(leaf)) {
9882                         ret = btrfs_next_leaf(root, path);
9883                         if (ret == 0)
9884                                 continue;
9885                         if (ret < 0)
9886                                 goto out;
9887                         break;
9888                 }
9889                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9890
9891                 if (found_key.objectid >= key->objectid &&
9892                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9893                         struct extent_map_tree *em_tree;
9894                         struct extent_map *em;
9895
9896                         em_tree = &root->fs_info->mapping_tree.map_tree;
9897                         read_lock(&em_tree->lock);
9898                         em = lookup_extent_mapping(em_tree, found_key.objectid,
9899                                                    found_key.offset);
9900                         read_unlock(&em_tree->lock);
9901                         if (!em) {
9902                                 btrfs_err(root->fs_info,
9903                         "logical %llu len %llu found bg but no related chunk",
9904                                           found_key.objectid, found_key.offset);
9905                                 ret = -ENOENT;
9906                         } else {
9907                                 ret = 0;
9908                         }
9909                         goto out;
9910                 }
9911                 path->slots[0]++;
9912         }
9913 out:
9914         return ret;
9915 }
9916
9917 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9918 {
9919         struct btrfs_block_group_cache *block_group;
9920         u64 last = 0;
9921
9922         while (1) {
9923                 struct inode *inode;
9924
9925                 block_group = btrfs_lookup_first_block_group(info, last);
9926                 while (block_group) {
9927                         spin_lock(&block_group->lock);
9928                         if (block_group->iref)
9929                                 break;
9930                         spin_unlock(&block_group->lock);
9931                         block_group = next_block_group(info->tree_root,
9932                                                        block_group);
9933                 }
9934                 if (!block_group) {
9935                         if (last == 0)
9936                                 break;
9937                         last = 0;
9938                         continue;
9939                 }
9940
9941                 inode = block_group->inode;
9942                 block_group->iref = 0;
9943                 block_group->inode = NULL;
9944                 spin_unlock(&block_group->lock);
9945                 iput(inode);
9946                 last = block_group->key.objectid + block_group->key.offset;
9947                 btrfs_put_block_group(block_group);
9948         }
9949 }
9950
9951 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9952 {
9953         struct btrfs_block_group_cache *block_group;
9954         struct btrfs_space_info *space_info;
9955         struct btrfs_caching_control *caching_ctl;
9956         struct rb_node *n;
9957
9958         down_write(&info->commit_root_sem);
9959         while (!list_empty(&info->caching_block_groups)) {
9960                 caching_ctl = list_entry(info->caching_block_groups.next,
9961                                          struct btrfs_caching_control, list);
9962                 list_del(&caching_ctl->list);
9963                 put_caching_control(caching_ctl);
9964         }
9965         up_write(&info->commit_root_sem);
9966
9967         spin_lock(&info->unused_bgs_lock);
9968         while (!list_empty(&info->unused_bgs)) {
9969                 block_group = list_first_entry(&info->unused_bgs,
9970                                                struct btrfs_block_group_cache,
9971                                                bg_list);
9972                 list_del_init(&block_group->bg_list);
9973                 btrfs_put_block_group(block_group);
9974         }
9975         spin_unlock(&info->unused_bgs_lock);
9976
9977         spin_lock(&info->block_group_cache_lock);
9978         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9979                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9980                                        cache_node);
9981                 rb_erase(&block_group->cache_node,
9982                          &info->block_group_cache_tree);
9983                 RB_CLEAR_NODE(&block_group->cache_node);
9984                 spin_unlock(&info->block_group_cache_lock);
9985
9986                 down_write(&block_group->space_info->groups_sem);
9987                 list_del(&block_group->list);
9988                 up_write(&block_group->space_info->groups_sem);
9989
9990                 if (block_group->cached == BTRFS_CACHE_STARTED)
9991                         wait_block_group_cache_done(block_group);
9992
9993                 /*
9994                  * We haven't cached this block group, which means we could
9995                  * possibly have excluded extents on this block group.
9996                  */
9997                 if (block_group->cached == BTRFS_CACHE_NO ||
9998                     block_group->cached == BTRFS_CACHE_ERROR)
9999                         free_excluded_extents(info->extent_root, block_group);
10000
10001                 btrfs_remove_free_space_cache(block_group);
10002                 btrfs_put_block_group(block_group);
10003
10004                 spin_lock(&info->block_group_cache_lock);
10005         }
10006         spin_unlock(&info->block_group_cache_lock);
10007
10008         /* now that all the block groups are freed, go through and
10009          * free all the space_info structs.  This is only called during
10010          * the final stages of unmount, and so we know nobody is
10011          * using them.  We call synchronize_rcu() once before we start,
10012          * just to be on the safe side.
10013          */
10014         synchronize_rcu();
10015
10016         release_global_block_rsv(info);
10017
10018         while (!list_empty(&info->space_info)) {
10019                 int i;
10020
10021                 space_info = list_entry(info->space_info.next,
10022                                         struct btrfs_space_info,
10023                                         list);
10024
10025                 /*
10026                  * Do not hide this behind enospc_debug, this is actually
10027                  * important and indicates a real bug if this happens.
10028                  */
10029                 if (WARN_ON(space_info->bytes_pinned > 0 ||
10030                             space_info->bytes_reserved > 0 ||
10031                             space_info->bytes_may_use > 0))
10032                         dump_space_info(space_info, 0, 0);
10033                 list_del(&space_info->list);
10034                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
10035                         struct kobject *kobj;
10036                         kobj = space_info->block_group_kobjs[i];
10037                         space_info->block_group_kobjs[i] = NULL;
10038                         if (kobj) {
10039                                 kobject_del(kobj);
10040                                 kobject_put(kobj);
10041                         }
10042                 }
10043                 kobject_del(&space_info->kobj);
10044                 kobject_put(&space_info->kobj);
10045         }
10046         return 0;
10047 }
10048
10049 static void __link_block_group(struct btrfs_space_info *space_info,
10050                                struct btrfs_block_group_cache *cache)
10051 {
10052         int index = get_block_group_index(cache);
10053         bool first = false;
10054
10055         down_write(&space_info->groups_sem);
10056         if (list_empty(&space_info->block_groups[index]))
10057                 first = true;
10058         list_add_tail(&cache->list, &space_info->block_groups[index]);
10059         up_write(&space_info->groups_sem);
10060
10061         if (first) {
10062                 struct raid_kobject *rkobj;
10063                 int ret;
10064
10065                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
10066                 if (!rkobj)
10067                         goto out_err;
10068                 rkobj->raid_type = index;
10069                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
10070                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
10071                                   "%s", get_raid_name(index));
10072                 if (ret) {
10073                         kobject_put(&rkobj->kobj);
10074                         goto out_err;
10075                 }
10076                 space_info->block_group_kobjs[index] = &rkobj->kobj;
10077         }
10078
10079         return;
10080 out_err:
10081         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
10082 }
10083
10084 static struct btrfs_block_group_cache *
10085 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
10086 {
10087         struct btrfs_block_group_cache *cache;
10088
10089         cache = kzalloc(sizeof(*cache), GFP_NOFS);
10090         if (!cache)
10091                 return NULL;
10092
10093         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
10094                                         GFP_NOFS);
10095         if (!cache->free_space_ctl) {
10096                 kfree(cache);
10097                 return NULL;
10098         }
10099
10100         cache->key.objectid = start;
10101         cache->key.offset = size;
10102         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10103
10104         cache->sectorsize = root->sectorsize;
10105         cache->fs_info = root->fs_info;
10106         cache->full_stripe_len = btrfs_full_stripe_len(root,
10107                                                &root->fs_info->mapping_tree,
10108                                                start);
10109         set_free_space_tree_thresholds(cache);
10110
10111         atomic_set(&cache->count, 1);
10112         spin_lock_init(&cache->lock);
10113         init_rwsem(&cache->data_rwsem);
10114         INIT_LIST_HEAD(&cache->list);
10115         INIT_LIST_HEAD(&cache->cluster_list);
10116         INIT_LIST_HEAD(&cache->bg_list);
10117         INIT_LIST_HEAD(&cache->ro_list);
10118         INIT_LIST_HEAD(&cache->dirty_list);
10119         INIT_LIST_HEAD(&cache->io_list);
10120         btrfs_init_free_space_ctl(cache);
10121         atomic_set(&cache->trimming, 0);
10122         mutex_init(&cache->free_space_lock);
10123
10124         return cache;
10125 }
10126
10127 int btrfs_read_block_groups(struct btrfs_root *root)
10128 {
10129         struct btrfs_path *path;
10130         int ret;
10131         struct btrfs_block_group_cache *cache;
10132         struct btrfs_fs_info *info = root->fs_info;
10133         struct btrfs_space_info *space_info;
10134         struct btrfs_key key;
10135         struct btrfs_key found_key;
10136         struct extent_buffer *leaf;
10137         int need_clear = 0;
10138         u64 cache_gen;
10139
10140         root = info->extent_root;
10141         key.objectid = 0;
10142         key.offset = 0;
10143         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
10144         path = btrfs_alloc_path();
10145         if (!path)
10146                 return -ENOMEM;
10147         path->reada = READA_FORWARD;
10148
10149         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
10150         if (btrfs_test_opt(root->fs_info, SPACE_CACHE) &&
10151             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
10152                 need_clear = 1;
10153         if (btrfs_test_opt(root->fs_info, CLEAR_CACHE))
10154                 need_clear = 1;
10155
10156         while (1) {
10157                 ret = find_first_block_group(root, path, &key);
10158                 if (ret > 0)
10159                         break;
10160                 if (ret != 0)
10161                         goto error;
10162
10163                 leaf = path->nodes[0];
10164                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
10165
10166                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
10167                                                        found_key.offset);
10168                 if (!cache) {
10169                         ret = -ENOMEM;
10170                         goto error;
10171                 }
10172
10173                 if (need_clear) {
10174                         /*
10175                          * When we mount with old space cache, we need to
10176                          * set BTRFS_DC_CLEAR and set dirty flag.
10177                          *
10178                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
10179                          *    truncate the old free space cache inode and
10180                          *    setup a new one.
10181                          * b) Setting 'dirty flag' makes sure that we flush
10182                          *    the new space cache info onto disk.
10183                          */
10184                         if (btrfs_test_opt(root->fs_info, SPACE_CACHE))
10185                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
10186                 }
10187
10188                 read_extent_buffer(leaf, &cache->item,
10189                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
10190                                    sizeof(cache->item));
10191                 cache->flags = btrfs_block_group_flags(&cache->item);
10192
10193                 key.objectid = found_key.objectid + found_key.offset;
10194                 btrfs_release_path(path);
10195
10196                 /*
10197                  * We need to exclude the super stripes now so that the space
10198                  * info has super bytes accounted for, otherwise we'll think
10199                  * we have more space than we actually do.
10200                  */
10201                 ret = exclude_super_stripes(root, cache);
10202                 if (ret) {
10203                         /*
10204                          * We may have excluded something, so call this just in
10205                          * case.
10206                          */
10207                         free_excluded_extents(root, cache);
10208                         btrfs_put_block_group(cache);
10209                         goto error;
10210                 }
10211
10212                 /*
10213                  * check for two cases, either we are full, and therefore
10214                  * don't need to bother with the caching work since we won't
10215                  * find any space, or we are empty, and we can just add all
10216                  * the space in and be done with it.  This saves us _alot_ of
10217                  * time, particularly in the full case.
10218                  */
10219                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
10220                         cache->last_byte_to_unpin = (u64)-1;
10221                         cache->cached = BTRFS_CACHE_FINISHED;
10222                         free_excluded_extents(root, cache);
10223                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10224                         cache->last_byte_to_unpin = (u64)-1;
10225                         cache->cached = BTRFS_CACHE_FINISHED;
10226                         add_new_free_space(cache, root->fs_info,
10227                                            found_key.objectid,
10228                                            found_key.objectid +
10229                                            found_key.offset);
10230                         free_excluded_extents(root, cache);
10231                 }
10232
10233                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
10234                 if (ret) {
10235                         btrfs_remove_free_space_cache(cache);
10236                         btrfs_put_block_group(cache);
10237                         goto error;
10238                 }
10239
10240                 trace_btrfs_add_block_group(root->fs_info, cache, 0);
10241                 ret = update_space_info(info, cache->flags, found_key.offset,
10242                                         btrfs_block_group_used(&cache->item),
10243                                         cache->bytes_super, &space_info);
10244                 if (ret) {
10245                         btrfs_remove_free_space_cache(cache);
10246                         spin_lock(&info->block_group_cache_lock);
10247                         rb_erase(&cache->cache_node,
10248                                  &info->block_group_cache_tree);
10249                         RB_CLEAR_NODE(&cache->cache_node);
10250                         spin_unlock(&info->block_group_cache_lock);
10251                         btrfs_put_block_group(cache);
10252                         goto error;
10253                 }
10254
10255                 cache->space_info = space_info;
10256
10257                 __link_block_group(space_info, cache);
10258
10259                 set_avail_alloc_bits(root->fs_info, cache->flags);
10260                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
10261                         inc_block_group_ro(cache, 1);
10262                 } else if (btrfs_block_group_used(&cache->item) == 0) {
10263                         spin_lock(&info->unused_bgs_lock);
10264                         /* Should always be true but just in case. */
10265                         if (list_empty(&cache->bg_list)) {
10266                                 btrfs_get_block_group(cache);
10267                                 list_add_tail(&cache->bg_list,
10268                                               &info->unused_bgs);
10269                         }
10270                         spin_unlock(&info->unused_bgs_lock);
10271                 }
10272         }
10273
10274         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
10275                 if (!(get_alloc_profile(root, space_info->flags) &
10276                       (BTRFS_BLOCK_GROUP_RAID10 |
10277                        BTRFS_BLOCK_GROUP_RAID1 |
10278                        BTRFS_BLOCK_GROUP_RAID5 |
10279                        BTRFS_BLOCK_GROUP_RAID6 |
10280                        BTRFS_BLOCK_GROUP_DUP)))
10281                         continue;
10282                 /*
10283                  * avoid allocating from un-mirrored block group if there are
10284                  * mirrored block groups.
10285                  */
10286                 list_for_each_entry(cache,
10287                                 &space_info->block_groups[BTRFS_RAID_RAID0],
10288                                 list)
10289                         inc_block_group_ro(cache, 1);
10290                 list_for_each_entry(cache,
10291                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
10292                                 list)
10293                         inc_block_group_ro(cache, 1);
10294         }
10295
10296         init_global_block_rsv(info);
10297         ret = 0;
10298 error:
10299         btrfs_free_path(path);
10300         return ret;
10301 }
10302
10303 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
10304                                        struct btrfs_root *root)
10305 {
10306         struct btrfs_block_group_cache *block_group, *tmp;
10307         struct btrfs_root *extent_root = root->fs_info->extent_root;
10308         struct btrfs_block_group_item item;
10309         struct btrfs_key key;
10310         int ret = 0;
10311         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
10312
10313         trans->can_flush_pending_bgs = false;
10314         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
10315                 if (ret)
10316                         goto next;
10317
10318                 spin_lock(&block_group->lock);
10319                 memcpy(&item, &block_group->item, sizeof(item));
10320                 memcpy(&key, &block_group->key, sizeof(key));
10321                 spin_unlock(&block_group->lock);
10322
10323                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
10324                                         sizeof(item));
10325                 if (ret)
10326                         btrfs_abort_transaction(trans, ret);
10327                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
10328                                                key.objectid, key.offset);
10329                 if (ret)
10330                         btrfs_abort_transaction(trans, ret);
10331                 add_block_group_free_space(trans, root->fs_info, block_group);
10332                 /* already aborted the transaction if it failed. */
10333 next:
10334                 list_del_init(&block_group->bg_list);
10335         }
10336         trans->can_flush_pending_bgs = can_flush_pending_bgs;
10337 }
10338
10339 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
10340                            struct btrfs_root *root, u64 bytes_used,
10341                            u64 type, u64 chunk_objectid, u64 chunk_offset,
10342                            u64 size)
10343 {
10344         int ret;
10345         struct btrfs_root *extent_root;
10346         struct btrfs_block_group_cache *cache;
10347         extent_root = root->fs_info->extent_root;
10348
10349         btrfs_set_log_full_commit(root->fs_info, trans);
10350
10351         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
10352         if (!cache)
10353                 return -ENOMEM;
10354
10355         btrfs_set_block_group_used(&cache->item, bytes_used);
10356         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
10357         btrfs_set_block_group_flags(&cache->item, type);
10358
10359         cache->flags = type;
10360         cache->last_byte_to_unpin = (u64)-1;
10361         cache->cached = BTRFS_CACHE_FINISHED;
10362         cache->needs_free_space = 1;
10363         ret = exclude_super_stripes(root, cache);
10364         if (ret) {
10365                 /*
10366                  * We may have excluded something, so call this just in
10367                  * case.
10368                  */
10369                 free_excluded_extents(root, cache);
10370                 btrfs_put_block_group(cache);
10371                 return ret;
10372         }
10373
10374         add_new_free_space(cache, root->fs_info, chunk_offset,
10375                            chunk_offset + size);
10376
10377         free_excluded_extents(root, cache);
10378
10379 #ifdef CONFIG_BTRFS_DEBUG
10380         if (btrfs_should_fragment_free_space(root, cache)) {
10381                 u64 new_bytes_used = size - bytes_used;
10382
10383                 bytes_used += new_bytes_used >> 1;
10384                 fragment_free_space(root, cache);
10385         }
10386 #endif
10387         /*
10388          * Call to ensure the corresponding space_info object is created and
10389          * assigned to our block group, but don't update its counters just yet.
10390          * We want our bg to be added to the rbtree with its ->space_info set.
10391          */
10392         ret = update_space_info(root->fs_info, cache->flags, 0, 0, 0,
10393                                 &cache->space_info);
10394         if (ret) {
10395                 btrfs_remove_free_space_cache(cache);
10396                 btrfs_put_block_group(cache);
10397                 return ret;
10398         }
10399
10400         ret = btrfs_add_block_group_cache(root->fs_info, cache);
10401         if (ret) {
10402                 btrfs_remove_free_space_cache(cache);
10403                 btrfs_put_block_group(cache);
10404                 return ret;
10405         }
10406
10407         /*
10408          * Now that our block group has its ->space_info set and is inserted in
10409          * the rbtree, update the space info's counters.
10410          */
10411         trace_btrfs_add_block_group(root->fs_info, cache, 1);
10412         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
10413                                 cache->bytes_super, &cache->space_info);
10414         if (ret) {
10415                 btrfs_remove_free_space_cache(cache);
10416                 spin_lock(&root->fs_info->block_group_cache_lock);
10417                 rb_erase(&cache->cache_node,
10418                          &root->fs_info->block_group_cache_tree);
10419                 RB_CLEAR_NODE(&cache->cache_node);
10420                 spin_unlock(&root->fs_info->block_group_cache_lock);
10421                 btrfs_put_block_group(cache);
10422                 return ret;
10423         }
10424         update_global_block_rsv(root->fs_info);
10425
10426         __link_block_group(cache->space_info, cache);
10427
10428         list_add_tail(&cache->bg_list, &trans->new_bgs);
10429
10430         set_avail_alloc_bits(extent_root->fs_info, type);
10431         return 0;
10432 }
10433
10434 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10435 {
10436         u64 extra_flags = chunk_to_extended(flags) &
10437                                 BTRFS_EXTENDED_PROFILE_MASK;
10438
10439         write_seqlock(&fs_info->profiles_lock);
10440         if (flags & BTRFS_BLOCK_GROUP_DATA)
10441                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10442         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10443                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10444         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10445                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10446         write_sequnlock(&fs_info->profiles_lock);
10447 }
10448
10449 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10450                              struct btrfs_root *root, u64 group_start,
10451                              struct extent_map *em)
10452 {
10453         struct btrfs_path *path;
10454         struct btrfs_block_group_cache *block_group;
10455         struct btrfs_free_cluster *cluster;
10456         struct btrfs_root *tree_root = root->fs_info->tree_root;
10457         struct btrfs_key key;
10458         struct inode *inode;
10459         struct kobject *kobj = NULL;
10460         int ret;
10461         int index;
10462         int factor;
10463         struct btrfs_caching_control *caching_ctl = NULL;
10464         bool remove_em;
10465
10466         root = root->fs_info->extent_root;
10467
10468         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
10469         BUG_ON(!block_group);
10470         BUG_ON(!block_group->ro);
10471
10472         /*
10473          * Free the reserved super bytes from this block group before
10474          * remove it.
10475          */
10476         free_excluded_extents(root, block_group);
10477
10478         memcpy(&key, &block_group->key, sizeof(key));
10479         index = get_block_group_index(block_group);
10480         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10481                                   BTRFS_BLOCK_GROUP_RAID1 |
10482                                   BTRFS_BLOCK_GROUP_RAID10))
10483                 factor = 2;
10484         else
10485                 factor = 1;
10486
10487         /* make sure this block group isn't part of an allocation cluster */
10488         cluster = &root->fs_info->data_alloc_cluster;
10489         spin_lock(&cluster->refill_lock);
10490         btrfs_return_cluster_to_free_space(block_group, cluster);
10491         spin_unlock(&cluster->refill_lock);
10492
10493         /*
10494          * make sure this block group isn't part of a metadata
10495          * allocation cluster
10496          */
10497         cluster = &root->fs_info->meta_alloc_cluster;
10498         spin_lock(&cluster->refill_lock);
10499         btrfs_return_cluster_to_free_space(block_group, cluster);
10500         spin_unlock(&cluster->refill_lock);
10501
10502         path = btrfs_alloc_path();
10503         if (!path) {
10504                 ret = -ENOMEM;
10505                 goto out;
10506         }
10507
10508         /*
10509          * get the inode first so any iput calls done for the io_list
10510          * aren't the final iput (no unlinks allowed now)
10511          */
10512         inode = lookup_free_space_inode(tree_root, block_group, path);
10513
10514         mutex_lock(&trans->transaction->cache_write_mutex);
10515         /*
10516          * make sure our free spache cache IO is done before remove the
10517          * free space inode
10518          */
10519         spin_lock(&trans->transaction->dirty_bgs_lock);
10520         if (!list_empty(&block_group->io_list)) {
10521                 list_del_init(&block_group->io_list);
10522
10523                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10524
10525                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10526                 btrfs_wait_cache_io(root, trans, block_group,
10527                                     &block_group->io_ctl, path,
10528                                     block_group->key.objectid);
10529                 btrfs_put_block_group(block_group);
10530                 spin_lock(&trans->transaction->dirty_bgs_lock);
10531         }
10532
10533         if (!list_empty(&block_group->dirty_list)) {
10534                 list_del_init(&block_group->dirty_list);
10535                 btrfs_put_block_group(block_group);
10536         }
10537         spin_unlock(&trans->transaction->dirty_bgs_lock);
10538         mutex_unlock(&trans->transaction->cache_write_mutex);
10539
10540         if (!IS_ERR(inode)) {
10541                 ret = btrfs_orphan_add(trans, inode);
10542                 if (ret) {
10543                         btrfs_add_delayed_iput(inode);
10544                         goto out;
10545                 }
10546                 clear_nlink(inode);
10547                 /* One for the block groups ref */
10548                 spin_lock(&block_group->lock);
10549                 if (block_group->iref) {
10550                         block_group->iref = 0;
10551                         block_group->inode = NULL;
10552                         spin_unlock(&block_group->lock);
10553                         iput(inode);
10554                 } else {
10555                         spin_unlock(&block_group->lock);
10556                 }
10557                 /* One for our lookup ref */
10558                 btrfs_add_delayed_iput(inode);
10559         }
10560
10561         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10562         key.offset = block_group->key.objectid;
10563         key.type = 0;
10564
10565         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10566         if (ret < 0)
10567                 goto out;
10568         if (ret > 0)
10569                 btrfs_release_path(path);
10570         if (ret == 0) {
10571                 ret = btrfs_del_item(trans, tree_root, path);
10572                 if (ret)
10573                         goto out;
10574                 btrfs_release_path(path);
10575         }
10576
10577         spin_lock(&root->fs_info->block_group_cache_lock);
10578         rb_erase(&block_group->cache_node,
10579                  &root->fs_info->block_group_cache_tree);
10580         RB_CLEAR_NODE(&block_group->cache_node);
10581
10582         if (root->fs_info->first_logical_byte == block_group->key.objectid)
10583                 root->fs_info->first_logical_byte = (u64)-1;
10584         spin_unlock(&root->fs_info->block_group_cache_lock);
10585
10586         down_write(&block_group->space_info->groups_sem);
10587         /*
10588          * we must use list_del_init so people can check to see if they
10589          * are still on the list after taking the semaphore
10590          */
10591         list_del_init(&block_group->list);
10592         if (list_empty(&block_group->space_info->block_groups[index])) {
10593                 kobj = block_group->space_info->block_group_kobjs[index];
10594                 block_group->space_info->block_group_kobjs[index] = NULL;
10595                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
10596         }
10597         up_write(&block_group->space_info->groups_sem);
10598         if (kobj) {
10599                 kobject_del(kobj);
10600                 kobject_put(kobj);
10601         }
10602
10603         if (block_group->has_caching_ctl)
10604                 caching_ctl = get_caching_control(block_group);
10605         if (block_group->cached == BTRFS_CACHE_STARTED)
10606                 wait_block_group_cache_done(block_group);
10607         if (block_group->has_caching_ctl) {
10608                 down_write(&root->fs_info->commit_root_sem);
10609                 if (!caching_ctl) {
10610                         struct btrfs_caching_control *ctl;
10611
10612                         list_for_each_entry(ctl,
10613                                     &root->fs_info->caching_block_groups, list)
10614                                 if (ctl->block_group == block_group) {
10615                                         caching_ctl = ctl;
10616                                         atomic_inc(&caching_ctl->count);
10617                                         break;
10618                                 }
10619                 }
10620                 if (caching_ctl)
10621                         list_del_init(&caching_ctl->list);
10622                 up_write(&root->fs_info->commit_root_sem);
10623                 if (caching_ctl) {
10624                         /* Once for the caching bgs list and once for us. */
10625                         put_caching_control(caching_ctl);
10626                         put_caching_control(caching_ctl);
10627                 }
10628         }
10629
10630         spin_lock(&trans->transaction->dirty_bgs_lock);
10631         if (!list_empty(&block_group->dirty_list)) {
10632                 WARN_ON(1);
10633         }
10634         if (!list_empty(&block_group->io_list)) {
10635                 WARN_ON(1);
10636         }
10637         spin_unlock(&trans->transaction->dirty_bgs_lock);
10638         btrfs_remove_free_space_cache(block_group);
10639
10640         spin_lock(&block_group->space_info->lock);
10641         list_del_init(&block_group->ro_list);
10642
10643         if (btrfs_test_opt(root->fs_info, ENOSPC_DEBUG)) {
10644                 WARN_ON(block_group->space_info->total_bytes
10645                         < block_group->key.offset);
10646                 WARN_ON(block_group->space_info->bytes_readonly
10647                         < block_group->key.offset);
10648                 WARN_ON(block_group->space_info->disk_total
10649                         < block_group->key.offset * factor);
10650         }
10651         block_group->space_info->total_bytes -= block_group->key.offset;
10652         block_group->space_info->bytes_readonly -= block_group->key.offset;
10653         block_group->space_info->disk_total -= block_group->key.offset * factor;
10654
10655         spin_unlock(&block_group->space_info->lock);
10656
10657         memcpy(&key, &block_group->key, sizeof(key));
10658
10659         lock_chunks(root);
10660         if (!list_empty(&em->list)) {
10661                 /* We're in the transaction->pending_chunks list. */
10662                 free_extent_map(em);
10663         }
10664         spin_lock(&block_group->lock);
10665         block_group->removed = 1;
10666         /*
10667          * At this point trimming can't start on this block group, because we
10668          * removed the block group from the tree fs_info->block_group_cache_tree
10669          * so no one can't find it anymore and even if someone already got this
10670          * block group before we removed it from the rbtree, they have already
10671          * incremented block_group->trimming - if they didn't, they won't find
10672          * any free space entries because we already removed them all when we
10673          * called btrfs_remove_free_space_cache().
10674          *
10675          * And we must not remove the extent map from the fs_info->mapping_tree
10676          * to prevent the same logical address range and physical device space
10677          * ranges from being reused for a new block group. This is because our
10678          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10679          * completely transactionless, so while it is trimming a range the
10680          * currently running transaction might finish and a new one start,
10681          * allowing for new block groups to be created that can reuse the same
10682          * physical device locations unless we take this special care.
10683          *
10684          * There may also be an implicit trim operation if the file system
10685          * is mounted with -odiscard. The same protections must remain
10686          * in place until the extents have been discarded completely when
10687          * the transaction commit has completed.
10688          */
10689         remove_em = (atomic_read(&block_group->trimming) == 0);
10690         /*
10691          * Make sure a trimmer task always sees the em in the pinned_chunks list
10692          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10693          * before checking block_group->removed).
10694          */
10695         if (!remove_em) {
10696                 /*
10697                  * Our em might be in trans->transaction->pending_chunks which
10698                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10699                  * and so is the fs_info->pinned_chunks list.
10700                  *
10701                  * So at this point we must be holding the chunk_mutex to avoid
10702                  * any races with chunk allocation (more specifically at
10703                  * volumes.c:contains_pending_extent()), to ensure it always
10704                  * sees the em, either in the pending_chunks list or in the
10705                  * pinned_chunks list.
10706                  */
10707                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10708         }
10709         spin_unlock(&block_group->lock);
10710
10711         if (remove_em) {
10712                 struct extent_map_tree *em_tree;
10713
10714                 em_tree = &root->fs_info->mapping_tree.map_tree;
10715                 write_lock(&em_tree->lock);
10716                 /*
10717                  * The em might be in the pending_chunks list, so make sure the
10718                  * chunk mutex is locked, since remove_extent_mapping() will
10719                  * delete us from that list.
10720                  */
10721                 remove_extent_mapping(em_tree, em);
10722                 write_unlock(&em_tree->lock);
10723                 /* once for the tree */
10724                 free_extent_map(em);
10725         }
10726
10727         unlock_chunks(root);
10728
10729         ret = remove_block_group_free_space(trans, root->fs_info, block_group);
10730         if (ret)
10731                 goto out;
10732
10733         btrfs_put_block_group(block_group);
10734         btrfs_put_block_group(block_group);
10735
10736         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10737         if (ret > 0)
10738                 ret = -EIO;
10739         if (ret < 0)
10740                 goto out;
10741
10742         ret = btrfs_del_item(trans, root, path);
10743 out:
10744         btrfs_free_path(path);
10745         return ret;
10746 }
10747
10748 struct btrfs_trans_handle *
10749 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10750                                      const u64 chunk_offset)
10751 {
10752         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10753         struct extent_map *em;
10754         struct map_lookup *map;
10755         unsigned int num_items;
10756
10757         read_lock(&em_tree->lock);
10758         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10759         read_unlock(&em_tree->lock);
10760         ASSERT(em && em->start == chunk_offset);
10761
10762         /*
10763          * We need to reserve 3 + N units from the metadata space info in order
10764          * to remove a block group (done at btrfs_remove_chunk() and at
10765          * btrfs_remove_block_group()), which are used for:
10766          *
10767          * 1 unit for adding the free space inode's orphan (located in the tree
10768          * of tree roots).
10769          * 1 unit for deleting the block group item (located in the extent
10770          * tree).
10771          * 1 unit for deleting the free space item (located in tree of tree
10772          * roots).
10773          * N units for deleting N device extent items corresponding to each
10774          * stripe (located in the device tree).
10775          *
10776          * In order to remove a block group we also need to reserve units in the
10777          * system space info in order to update the chunk tree (update one or
10778          * more device items and remove one chunk item), but this is done at
10779          * btrfs_remove_chunk() through a call to check_system_chunk().
10780          */
10781         map = em->map_lookup;
10782         num_items = 3 + map->num_stripes;
10783         free_extent_map(em);
10784
10785         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10786                                                            num_items, 1);
10787 }
10788
10789 /*
10790  * Process the unused_bgs list and remove any that don't have any allocated
10791  * space inside of them.
10792  */
10793 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10794 {
10795         struct btrfs_block_group_cache *block_group;
10796         struct btrfs_space_info *space_info;
10797         struct btrfs_root *root = fs_info->extent_root;
10798         struct btrfs_trans_handle *trans;
10799         int ret = 0;
10800
10801         if (!fs_info->open)
10802                 return;
10803
10804         spin_lock(&fs_info->unused_bgs_lock);
10805         while (!list_empty(&fs_info->unused_bgs)) {
10806                 u64 start, end;
10807                 int trimming;
10808
10809                 block_group = list_first_entry(&fs_info->unused_bgs,
10810                                                struct btrfs_block_group_cache,
10811                                                bg_list);
10812                 list_del_init(&block_group->bg_list);
10813
10814                 space_info = block_group->space_info;
10815
10816                 if (ret || btrfs_mixed_space_info(space_info)) {
10817                         btrfs_put_block_group(block_group);
10818                         continue;
10819                 }
10820                 spin_unlock(&fs_info->unused_bgs_lock);
10821
10822                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10823
10824                 /* Don't want to race with allocators so take the groups_sem */
10825                 down_write(&space_info->groups_sem);
10826                 spin_lock(&block_group->lock);
10827                 if (block_group->reserved ||
10828                     btrfs_block_group_used(&block_group->item) ||
10829                     block_group->ro ||
10830                     list_is_singular(&block_group->list)) {
10831                         /*
10832                          * We want to bail if we made new allocations or have
10833                          * outstanding allocations in this block group.  We do
10834                          * the ro check in case balance is currently acting on
10835                          * this block group.
10836                          */
10837                         spin_unlock(&block_group->lock);
10838                         up_write(&space_info->groups_sem);
10839                         goto next;
10840                 }
10841                 spin_unlock(&block_group->lock);
10842
10843                 /* We don't want to force the issue, only flip if it's ok. */
10844                 ret = inc_block_group_ro(block_group, 0);
10845                 up_write(&space_info->groups_sem);
10846                 if (ret < 0) {
10847                         ret = 0;
10848                         goto next;
10849                 }
10850
10851                 /*
10852                  * Want to do this before we do anything else so we can recover
10853                  * properly if we fail to join the transaction.
10854                  */
10855                 trans = btrfs_start_trans_remove_block_group(fs_info,
10856                                                      block_group->key.objectid);
10857                 if (IS_ERR(trans)) {
10858                         btrfs_dec_block_group_ro(root, block_group);
10859                         ret = PTR_ERR(trans);
10860                         goto next;
10861                 }
10862
10863                 /*
10864                  * We could have pending pinned extents for this block group,
10865                  * just delete them, we don't care about them anymore.
10866                  */
10867                 start = block_group->key.objectid;
10868                 end = start + block_group->key.offset - 1;
10869                 /*
10870                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10871                  * btrfs_finish_extent_commit(). If we are at transaction N,
10872                  * another task might be running finish_extent_commit() for the
10873                  * previous transaction N - 1, and have seen a range belonging
10874                  * to the block group in freed_extents[] before we were able to
10875                  * clear the whole block group range from freed_extents[]. This
10876                  * means that task can lookup for the block group after we
10877                  * unpinned it from freed_extents[] and removed it, leading to
10878                  * a BUG_ON() at btrfs_unpin_extent_range().
10879                  */
10880                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10881                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10882                                   EXTENT_DIRTY);
10883                 if (ret) {
10884                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10885                         btrfs_dec_block_group_ro(root, block_group);
10886                         goto end_trans;
10887                 }
10888                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10889                                   EXTENT_DIRTY);
10890                 if (ret) {
10891                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10892                         btrfs_dec_block_group_ro(root, block_group);
10893                         goto end_trans;
10894                 }
10895                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10896
10897                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10898                 spin_lock(&space_info->lock);
10899                 spin_lock(&block_group->lock);
10900
10901                 space_info->bytes_pinned -= block_group->pinned;
10902                 space_info->bytes_readonly += block_group->pinned;
10903                 percpu_counter_add(&space_info->total_bytes_pinned,
10904                                    -block_group->pinned);
10905                 block_group->pinned = 0;
10906
10907                 spin_unlock(&block_group->lock);
10908                 spin_unlock(&space_info->lock);
10909
10910                 /* DISCARD can flip during remount */
10911                 trimming = btrfs_test_opt(root->fs_info, DISCARD);
10912
10913                 /* Implicit trim during transaction commit. */
10914                 if (trimming)
10915                         btrfs_get_block_group_trimming(block_group);
10916
10917                 /*
10918                  * Btrfs_remove_chunk will abort the transaction if things go
10919                  * horribly wrong.
10920                  */
10921                 ret = btrfs_remove_chunk(trans, root,
10922                                          block_group->key.objectid);
10923
10924                 if (ret) {
10925                         if (trimming)
10926                                 btrfs_put_block_group_trimming(block_group);
10927                         goto end_trans;
10928                 }
10929
10930                 /*
10931                  * If we're not mounted with -odiscard, we can just forget
10932                  * about this block group. Otherwise we'll need to wait
10933                  * until transaction commit to do the actual discard.
10934                  */
10935                 if (trimming) {
10936                         spin_lock(&fs_info->unused_bgs_lock);
10937                         /*
10938                          * A concurrent scrub might have added us to the list
10939                          * fs_info->unused_bgs, so use a list_move operation
10940                          * to add the block group to the deleted_bgs list.
10941                          */
10942                         list_move(&block_group->bg_list,
10943                                   &trans->transaction->deleted_bgs);
10944                         spin_unlock(&fs_info->unused_bgs_lock);
10945                         btrfs_get_block_group(block_group);
10946                 }
10947 end_trans:
10948                 btrfs_end_transaction(trans, root);
10949 next:
10950                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10951                 btrfs_put_block_group(block_group);
10952                 spin_lock(&fs_info->unused_bgs_lock);
10953         }
10954         spin_unlock(&fs_info->unused_bgs_lock);
10955 }
10956
10957 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10958 {
10959         struct btrfs_space_info *space_info;
10960         struct btrfs_super_block *disk_super;
10961         u64 features;
10962         u64 flags;
10963         int mixed = 0;
10964         int ret;
10965
10966         disk_super = fs_info->super_copy;
10967         if (!btrfs_super_root(disk_super))
10968                 return -EINVAL;
10969
10970         features = btrfs_super_incompat_flags(disk_super);
10971         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10972                 mixed = 1;
10973
10974         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10975         ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10976         if (ret)
10977                 goto out;
10978
10979         if (mixed) {
10980                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10981                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10982         } else {
10983                 flags = BTRFS_BLOCK_GROUP_METADATA;
10984                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10985                 if (ret)
10986                         goto out;
10987
10988                 flags = BTRFS_BLOCK_GROUP_DATA;
10989                 ret = update_space_info(fs_info, flags, 0, 0, 0, &space_info);
10990         }
10991 out:
10992         return ret;
10993 }
10994
10995 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10996 {
10997         return unpin_extent_range(root, start, end, false);
10998 }
10999
11000 /*
11001  * It used to be that old block groups would be left around forever.
11002  * Iterating over them would be enough to trim unused space.  Since we
11003  * now automatically remove them, we also need to iterate over unallocated
11004  * space.
11005  *
11006  * We don't want a transaction for this since the discard may take a
11007  * substantial amount of time.  We don't require that a transaction be
11008  * running, but we do need to take a running transaction into account
11009  * to ensure that we're not discarding chunks that were released in
11010  * the current transaction.
11011  *
11012  * Holding the chunks lock will prevent other threads from allocating
11013  * or releasing chunks, but it won't prevent a running transaction
11014  * from committing and releasing the memory that the pending chunks
11015  * list head uses.  For that, we need to take a reference to the
11016  * transaction.
11017  */
11018 static int btrfs_trim_free_extents(struct btrfs_device *device,
11019                                    u64 minlen, u64 *trimmed)
11020 {
11021         u64 start = 0, len = 0;
11022         int ret;
11023
11024         *trimmed = 0;
11025
11026         /* Not writeable = nothing to do. */
11027         if (!device->writeable)
11028                 return 0;
11029
11030         /* No free space = nothing to do. */
11031         if (device->total_bytes <= device->bytes_used)
11032                 return 0;
11033
11034         ret = 0;
11035
11036         while (1) {
11037                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
11038                 struct btrfs_transaction *trans;
11039                 u64 bytes;
11040
11041                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
11042                 if (ret)
11043                         return ret;
11044
11045                 down_read(&fs_info->commit_root_sem);
11046
11047                 spin_lock(&fs_info->trans_lock);
11048                 trans = fs_info->running_transaction;
11049                 if (trans)
11050                         atomic_inc(&trans->use_count);
11051                 spin_unlock(&fs_info->trans_lock);
11052
11053                 ret = find_free_dev_extent_start(trans, device, minlen, start,
11054                                                  &start, &len);
11055                 if (trans)
11056                         btrfs_put_transaction(trans);
11057
11058                 if (ret) {
11059                         up_read(&fs_info->commit_root_sem);
11060                         mutex_unlock(&fs_info->chunk_mutex);
11061                         if (ret == -ENOSPC)
11062                                 ret = 0;
11063                         break;
11064                 }
11065
11066                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
11067                 up_read(&fs_info->commit_root_sem);
11068                 mutex_unlock(&fs_info->chunk_mutex);
11069
11070                 if (ret)
11071                         break;
11072
11073                 start += len;
11074                 *trimmed += bytes;
11075
11076                 if (fatal_signal_pending(current)) {
11077                         ret = -ERESTARTSYS;
11078                         break;
11079                 }
11080
11081                 cond_resched();
11082         }
11083
11084         return ret;
11085 }
11086
11087 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
11088 {
11089         struct btrfs_fs_info *fs_info = root->fs_info;
11090         struct btrfs_block_group_cache *cache = NULL;
11091         struct btrfs_device *device;
11092         struct list_head *devices;
11093         u64 group_trimmed;
11094         u64 start;
11095         u64 end;
11096         u64 trimmed = 0;
11097         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
11098         int ret = 0;
11099
11100         /*
11101          * try to trim all FS space, our block group may start from non-zero.
11102          */
11103         if (range->len == total_bytes)
11104                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
11105         else
11106                 cache = btrfs_lookup_block_group(fs_info, range->start);
11107
11108         while (cache) {
11109                 if (cache->key.objectid >= (range->start + range->len)) {
11110                         btrfs_put_block_group(cache);
11111                         break;
11112                 }
11113
11114                 start = max(range->start, cache->key.objectid);
11115                 end = min(range->start + range->len,
11116                                 cache->key.objectid + cache->key.offset);
11117
11118                 if (end - start >= range->minlen) {
11119                         if (!block_group_cache_done(cache)) {
11120                                 ret = cache_block_group(cache, 0);
11121                                 if (ret) {
11122                                         btrfs_put_block_group(cache);
11123                                         break;
11124                                 }
11125                                 ret = wait_block_group_cache_done(cache);
11126                                 if (ret) {
11127                                         btrfs_put_block_group(cache);
11128                                         break;
11129                                 }
11130                         }
11131                         ret = btrfs_trim_block_group(cache,
11132                                                      &group_trimmed,
11133                                                      start,
11134                                                      end,
11135                                                      range->minlen);
11136
11137                         trimmed += group_trimmed;
11138                         if (ret) {
11139                                 btrfs_put_block_group(cache);
11140                                 break;
11141                         }
11142                 }
11143
11144                 cache = next_block_group(fs_info->tree_root, cache);
11145         }
11146
11147         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
11148         devices = &root->fs_info->fs_devices->alloc_list;
11149         list_for_each_entry(device, devices, dev_alloc_list) {
11150                 ret = btrfs_trim_free_extents(device, range->minlen,
11151                                               &group_trimmed);
11152                 if (ret)
11153                         break;
11154
11155                 trimmed += group_trimmed;
11156         }
11157         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
11158
11159         range->len = trimmed;
11160         return ret;
11161 }
11162
11163 /*
11164  * btrfs_{start,end}_write_no_snapshoting() are similar to
11165  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
11166  * data into the page cache through nocow before the subvolume is snapshoted,
11167  * but flush the data into disk after the snapshot creation, or to prevent
11168  * operations while snapshoting is ongoing and that cause the snapshot to be
11169  * inconsistent (writes followed by expanding truncates for example).
11170  */
11171 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
11172 {
11173         percpu_counter_dec(&root->subv_writers->counter);
11174         /*
11175          * Make sure counter is updated before we wake up waiters.
11176          */
11177         smp_mb();
11178         if (waitqueue_active(&root->subv_writers->wait))
11179                 wake_up(&root->subv_writers->wait);
11180 }
11181
11182 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
11183 {
11184         if (atomic_read(&root->will_be_snapshoted))
11185                 return 0;
11186
11187         percpu_counter_inc(&root->subv_writers->counter);
11188         /*
11189          * Make sure counter is updated before we check for snapshot creation.
11190          */
11191         smp_mb();
11192         if (atomic_read(&root->will_be_snapshoted)) {
11193                 btrfs_end_write_no_snapshoting(root);
11194                 return 0;
11195         }
11196         return 1;
11197 }
11198
11199 static int wait_snapshoting_atomic_t(atomic_t *a)
11200 {
11201         schedule();
11202         return 0;
11203 }
11204
11205 void btrfs_wait_for_snapshot_creation(struct btrfs_root *root)
11206 {
11207         while (true) {
11208                 int ret;
11209
11210                 ret = btrfs_start_write_no_snapshoting(root);
11211                 if (ret)
11212                         break;
11213                 wait_on_atomic_t(&root->will_be_snapshoted,
11214                                  wait_snapshoting_atomic_t,
11215                                  TASK_UNINTERRUPTIBLE);
11216         }
11217 }