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 "math.h"
  37 #include "sysfs.h"
  38 #include "qgroup.h"
  39
  40 #undef SCRAMBLE_DELAYED_REFS
  41
  42 /*
  43  * control flags for do_chunk_alloc's force field
  44  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  45  * if we really need one.
  46  *
  47  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  48  * if we have very few chunks already allocated.  This is
  49  * used as part of the clustering code to help make sure
  50  * we have a good pool of storage to cluster in, without
  51  * filling the FS with empty chunks
  52  *
  53  * CHUNK_ALLOC_FORCE means it must try to allocate one
  54  *
  55  */
  56 enum {
  57         CHUNK_ALLOC_NO_FORCE = 0,
  58         CHUNK_ALLOC_LIMITED = 1,
  59         CHUNK_ALLOC_FORCE = 2,
  60 };
  61
  62 /*
  63  * Control how reservations are dealt with.
  64  *
  65  * RESERVE_FREE - freeing a reservation.
  66  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  67  *   ENOSPC accounting
  68  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  69  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  70  */
  71 enum {
  72         RESERVE_FREE = 0,
  73         RESERVE_ALLOC = 1,
  74         RESERVE_ALLOC_NO_ACCOUNT = 2,
  75 };
  76
  77 static int update_block_group(struct btrfs_trans_handle *trans,
  78                               struct btrfs_root *root, u64 bytenr,
  79                               u64 num_bytes, int alloc);
  80 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  81                                 struct btrfs_root *root,
  82                                 struct btrfs_delayed_ref_node *node, u64 parent,
  83                                 u64 root_objectid, u64 owner_objectid,
  84                                 u64 owner_offset, int refs_to_drop,
  85                                 struct btrfs_delayed_extent_op *extra_op);
  86 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  87                                     struct extent_buffer *leaf,
  88                                     struct btrfs_extent_item *ei);
  89 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  90                                       struct btrfs_root *root,
  91                                       u64 parent, u64 root_objectid,
  92                                       u64 flags, u64 owner, u64 offset,
  93                                       struct btrfs_key *ins, int ref_mod);
  94 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  95                                      struct btrfs_root *root,
  96                                      u64 parent, u64 root_objectid,
  97                                      u64 flags, struct btrfs_disk_key *key,
  98                                      int level, struct btrfs_key *ins,
  99                                      int no_quota);
 100 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
 101                           struct btrfs_root *extent_root, u64 flags,
 102                           int force);
 103 static int find_next_key(struct btrfs_path *path, int level,
 104                          struct btrfs_key *key);
 105 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 106                             int dump_block_groups);
 107 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 108                                        u64 num_bytes, int reserve,
 109                                        int delalloc);
 110 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 111                                u64 num_bytes);
 112 int btrfs_pin_extent(struct btrfs_root *root,
 113                      u64 bytenr, u64 num_bytes, int reserved);
 114
 115 static noinline int
 116 block_group_cache_done(struct btrfs_block_group_cache *cache)
 117 {
 118         smp_mb();
 119         return cache->cached == BTRFS_CACHE_FINISHED ||
 120                 cache->cached == BTRFS_CACHE_ERROR;
 121 }
 122
 123 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 124 {
 125         return (cache->flags & bits) == bits;
 126 }
 127
 128 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 129 {
 130         atomic_inc(&cache->count);
 131 }
 132
 133 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 134 {
 135         if (atomic_dec_and_test(&cache->count)) {
 136                 WARN_ON(cache->pinned > 0);
 137                 WARN_ON(cache->reserved > 0);
 138                 kfree(cache->free_space_ctl);
 139                 kfree(cache);
 140         }
 141 }
 142
 143 /*
 144  * this adds the block group to the fs_info rb tree for the block group
 145  * cache
 146  */
 147 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 148                                 struct btrfs_block_group_cache *block_group)
 149 {
 150         struct rb_node **p;
 151         struct rb_node *parent = NULL;
 152         struct btrfs_block_group_cache *cache;
 153
 154         spin_lock(&info->block_group_cache_lock);
 155         p = &info->block_group_cache_tree.rb_node;
 156
 157         while (*p) {
 158                 parent = *p;
 159                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 160                                  cache_node);
 161                 if (block_group->key.objectid < cache->key.objectid) {
 162                         p = &(*p)->rb_left;
 163                 } else if (block_group->key.objectid > cache->key.objectid) {
 164                         p = &(*p)->rb_right;
 165                 } else {
 166                         spin_unlock(&info->block_group_cache_lock);
 167                         return -EEXIST;
 168                 }
 169         }
 170
 171         rb_link_node(&block_group->cache_node, parent, p);
 172         rb_insert_color(&block_group->cache_node,
 173                         &info->block_group_cache_tree);
 174
 175         if (info->first_logical_byte > block_group->key.objectid)
 176                 info->first_logical_byte = block_group->key.objectid;
 177
 178         spin_unlock(&info->block_group_cache_lock);
 179
 180         return 0;
 181 }
 182
 183 /*
 184  * This will return the block group at or after bytenr if contains is 0, else
 185  * it will return the block group that contains the bytenr
 186  */
 187 static struct btrfs_block_group_cache *
 188 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 189                               int contains)
 190 {
 191         struct btrfs_block_group_cache *cache, *ret = NULL;
 192         struct rb_node *n;
 193         u64 end, start;
 194
 195         spin_lock(&info->block_group_cache_lock);
 196         n = info->block_group_cache_tree.rb_node;
 197
 198         while (n) {
 199                 cache = rb_entry(n, struct btrfs_block_group_cache,
 200                                  cache_node);
 201                 end = cache->key.objectid + cache->key.offset - 1;
 202                 start = cache->key.objectid;
 203
 204                 if (bytenr < start) {
 205                         if (!contains && (!ret || start < ret->key.objectid))
 206                                 ret = cache;
 207                         n = n->rb_left;
 208                 } else if (bytenr > start) {
 209                         if (contains && bytenr <= end) {
 210                                 ret = cache;
 211                                 break;
 212                         }
 213                         n = n->rb_right;
 214                 } else {
 215                         ret = cache;
 216                         break;
 217                 }
 218         }
 219         if (ret) {
 220                 btrfs_get_block_group(ret);
 221                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 222                         info->first_logical_byte = ret->key.objectid;
 223         }
 224         spin_unlock(&info->block_group_cache_lock);
 225
 226         return ret;
 227 }
 228
 229 static int add_excluded_extent(struct btrfs_root *root,
 230                                u64 start, u64 num_bytes)
 231 {
 232         u64 end = start + num_bytes - 1;
 233         set_extent_bits(&root->fs_info->freed_extents[0],
 234                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 235         set_extent_bits(&root->fs_info->freed_extents[1],
 236                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 237         return 0;
 238 }
 239
 240 static void free_excluded_extents(struct btrfs_root *root,
 241                                   struct btrfs_block_group_cache *cache)
 242 {
 243         u64 start, end;
 244
 245         start = cache->key.objectid;
 246         end = start + cache->key.offset - 1;
 247
 248         clear_extent_bits(&root->fs_info->freed_extents[0],
 249                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 250         clear_extent_bits(&root->fs_info->freed_extents[1],
 251                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 252 }
 253
 254 static int exclude_super_stripes(struct btrfs_root *root,
 255                                  struct btrfs_block_group_cache *cache)
 256 {
 257         u64 bytenr;
 258         u64 *logical;
 259         int stripe_len;
 260         int i, nr, ret;
 261
 262         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 263                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 264                 cache->bytes_super += stripe_len;
 265                 ret = add_excluded_extent(root, cache->key.objectid,
 266                                           stripe_len);
 267                 if (ret)
 268                         return ret;
 269         }
 270
 271         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 272                 bytenr = btrfs_sb_offset(i);
 273                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 274                                        cache->key.objectid, bytenr,
 275                                        0, &logical, &nr, &stripe_len);
 276                 if (ret)
 277                         return ret;
 278
 279                 while (nr--) {
 280                         u64 start, len;
 281
 282                         if (logical[nr] > cache->key.objectid +
 283                             cache->key.offset)
 284                                 continue;
 285
 286                         if (logical[nr] + stripe_len <= cache->key.objectid)
 287                                 continue;
 288
 289                         start = logical[nr];
 290                         if (start < cache->key.objectid) {
 291                                 start = cache->key.objectid;
 292                                 len = (logical[nr] + stripe_len) - start;
 293                         } else {
 294                                 len = min_t(u64, stripe_len,
 295                                             cache->key.objectid +
 296                                             cache->key.offset - start);
 297                         }
 298
 299                         cache->bytes_super += len;
 300                         ret = add_excluded_extent(root, start, len);
 301                         if (ret) {
 302                                 kfree(logical);
 303                                 return ret;
 304                         }
 305                 }
 306
 307                 kfree(logical);
 308         }
 309         return 0;
 310 }
 311
 312 static struct btrfs_caching_control *
 313 get_caching_control(struct btrfs_block_group_cache *cache)
 314 {
 315         struct btrfs_caching_control *ctl;
 316
 317         spin_lock(&cache->lock);
 318         if (!cache->caching_ctl) {
 319                 spin_unlock(&cache->lock);
 320                 return NULL;
 321         }
 322
 323         ctl = cache->caching_ctl;
 324         atomic_inc(&ctl->count);
 325         spin_unlock(&cache->lock);
 326         return ctl;
 327 }
 328
 329 static void put_caching_control(struct btrfs_caching_control *ctl)
 330 {
 331         if (atomic_dec_and_test(&ctl->count))
 332                 kfree(ctl);
 333 }
 334
 335 #ifdef CONFIG_BTRFS_DEBUG
 336 static void fragment_free_space(struct btrfs_root *root,
 337                                 struct btrfs_block_group_cache *block_group)
 338 {
 339         u64 start = block_group->key.objectid;
 340         u64 len = block_group->key.offset;
 341         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 342                 root->nodesize : root->sectorsize;
 343         u64 step = chunk << 1;
 344
 345         while (len > chunk) {
 346                 btrfs_remove_free_space(block_group, start, chunk);
 347                 start += step;
 348                 if (len < step)
 349                         len = 0;
 350                 else
 351                         len -= step;
 352         }
 353 }
 354 #endif
 355
 356 /*
 357  * this is only called by cache_block_group, since we could have freed extents
 358  * we need to check the pinned_extents for any extents that can't be used yet
 359  * since their free space will be released as soon as the transaction commits.
 360  */
 361 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 362                               struct btrfs_fs_info *info, u64 start, u64 end)
 363 {
 364         u64 extent_start, extent_end, size, total_added = 0;
 365         int ret;
 366
 367         while (start < end) {
 368                 ret = find_first_extent_bit(info->pinned_extents, start,
 369                                             &extent_start, &extent_end,
 370                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 371                                             NULL);
 372                 if (ret)
 373                         break;
 374
 375                 if (extent_start <= start) {
 376                         start = extent_end + 1;
 377                 } else if (extent_start > start && extent_start < end) {
 378                         size = extent_start - start;
 379                         total_added += size;
 380                         ret = btrfs_add_free_space(block_group, start,
 381                                                    size);
 382                         BUG_ON(ret); /* -ENOMEM or logic error */
 383                         start = extent_end + 1;
 384                 } else {
 385                         break;
 386                 }
 387         }
 388
 389         if (start < end) {
 390                 size = end - start;
 391                 total_added += size;
 392                 ret = btrfs_add_free_space(block_group, start, size);
 393                 BUG_ON(ret); /* -ENOMEM or logic error */
 394         }
 395
 396         return total_added;
 397 }
 398
 399 static noinline void caching_thread(struct btrfs_work *work)
 400 {
 401         struct btrfs_block_group_cache *block_group;
 402         struct btrfs_fs_info *fs_info;
 403         struct btrfs_caching_control *caching_ctl;
 404         struct btrfs_root *extent_root;
 405         struct btrfs_path *path;
 406         struct extent_buffer *leaf;
 407         struct btrfs_key key;
 408         u64 total_found = 0;
 409         u64 last = 0;
 410         u32 nritems;
 411         int ret = -ENOMEM;
 412         bool wakeup = true;
 413
 414         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 415         block_group = caching_ctl->block_group;
 416         fs_info = block_group->fs_info;
 417         extent_root = fs_info->extent_root;
 418
 419         path = btrfs_alloc_path();
 420         if (!path)
 421                 goto out;
 422
 423         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 424
 425 #ifdef CONFIG_BTRFS_DEBUG
 426         /*
 427          * If we're fragmenting we don't want to make anybody think we can
 428          * allocate from this block group until we've had a chance to fragment
 429          * the free space.
 430          */
 431         if (btrfs_should_fragment_free_space(extent_root, block_group))
 432                 wakeup = false;
 433 #endif
 434         /*
 435          * We don't want to deadlock with somebody trying to allocate a new
 436          * extent for the extent root while also trying to search the extent
 437          * root to add free space.  So we skip locking and search the commit
 438          * root, since its read-only
 439          */
 440         path->skip_locking = 1;
 441         path->search_commit_root = 1;
 442         path->reada = 1;
 443
 444         key.objectid = last;
 445         key.offset = 0;
 446         key.type = BTRFS_EXTENT_ITEM_KEY;
 447 again:
 448         mutex_lock(&caching_ctl->mutex);
 449         /* need to make sure the commit_root doesn't disappear */
 450         down_read(&fs_info->commit_root_sem);
 451
 452 next:
 453         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 454         if (ret < 0)
 455                 goto err;
 456
 457         leaf = path->nodes[0];
 458         nritems = btrfs_header_nritems(leaf);
 459
 460         while (1) {
 461                 if (btrfs_fs_closing(fs_info) > 1) {
 462                         last = (u64)-1;
 463                         break;
 464                 }
 465
 466                 if (path->slots[0] < nritems) {
 467                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 468                 } else {
 469                         ret = find_next_key(path, 0, &key);
 470                         if (ret)
 471                                 break;
 472
 473                         if (need_resched() ||
 474                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 475                                 if (wakeup)
 476                                         caching_ctl->progress = last;
 477                                 btrfs_release_path(path);
 478                                 up_read(&fs_info->commit_root_sem);
 479                                 mutex_unlock(&caching_ctl->mutex);
 480                                 cond_resched();
 481                                 goto again;
 482                         }
 483
 484                         ret = btrfs_next_leaf(extent_root, path);
 485                         if (ret < 0)
 486                                 goto err;
 487                         if (ret)
 488                                 break;
 489                         leaf = path->nodes[0];
 490                         nritems = btrfs_header_nritems(leaf);
 491                         continue;
 492                 }
 493
 494                 if (key.objectid < last) {
 495                         key.objectid = last;
 496                         key.offset = 0;
 497                         key.type = BTRFS_EXTENT_ITEM_KEY;
 498
 499                         if (wakeup)
 500                                 caching_ctl->progress = last;
 501                         btrfs_release_path(path);
 502                         goto next;
 503                 }
 504
 505                 if (key.objectid < block_group->key.objectid) {
 506                         path->slots[0]++;
 507                         continue;
 508                 }
 509
 510                 if (key.objectid >= block_group->key.objectid +
 511                     block_group->key.offset)
 512                         break;
 513
 514                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 515                     key.type == BTRFS_METADATA_ITEM_KEY) {
 516                         total_found += add_new_free_space(block_group,
 517                                                           fs_info, last,
 518                                                           key.objectid);
 519                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 520                                 last = key.objectid +
 521                                         fs_info->tree_root->nodesize;
 522                         else
 523                                 last = key.objectid + key.offset;
 524
 525                         if (total_found > (1024 * 1024 * 2)) {
 526                                 total_found = 0;
 527                                 if (wakeup)
 528                                         wake_up(&caching_ctl->wait);
 529                         }
 530                 }
 531                 path->slots[0]++;
 532         }
 533         ret = 0;
 534
 535         total_found += add_new_free_space(block_group, fs_info, last,
 536                                           block_group->key.objectid +
 537                                           block_group->key.offset);
 538         spin_lock(&block_group->lock);
 539         block_group->caching_ctl = NULL;
 540         block_group->cached = BTRFS_CACHE_FINISHED;
 541         spin_unlock(&block_group->lock);
 542
 543 #ifdef CONFIG_BTRFS_DEBUG
 544         if (btrfs_should_fragment_free_space(extent_root, block_group)) {
 545                 u64 bytes_used;
 546
 547                 spin_lock(&block_group->space_info->lock);
 548                 spin_lock(&block_group->lock);
 549                 bytes_used = block_group->key.offset -
 550                         btrfs_block_group_used(&block_group->item);
 551                 block_group->space_info->bytes_used += bytes_used >> 1;
 552                 spin_unlock(&block_group->lock);
 553                 spin_unlock(&block_group->space_info->lock);
 554                 fragment_free_space(extent_root, block_group);
 555         }
 556 #endif
 557
 558         caching_ctl->progress = (u64)-1;
 559 err:
 560         btrfs_free_path(path);
 561         up_read(&fs_info->commit_root_sem);
 562
 563         free_excluded_extents(extent_root, block_group);
 564
 565         mutex_unlock(&caching_ctl->mutex);
 566 out:
 567         if (ret) {
 568                 spin_lock(&block_group->lock);
 569                 block_group->caching_ctl = NULL;
 570                 block_group->cached = BTRFS_CACHE_ERROR;
 571                 spin_unlock(&block_group->lock);
 572         }
 573         wake_up(&caching_ctl->wait);
 574
 575         put_caching_control(caching_ctl);
 576         btrfs_put_block_group(block_group);
 577 }
 578
 579 static int cache_block_group(struct btrfs_block_group_cache *cache,
 580                              int load_cache_only)
 581 {
 582         DEFINE_WAIT(wait);
 583         struct btrfs_fs_info *fs_info = cache->fs_info;
 584         struct btrfs_caching_control *caching_ctl;
 585         int ret = 0;
 586
 587         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 588         if (!caching_ctl)
 589                 return -ENOMEM;
 590
 591         INIT_LIST_HEAD(&caching_ctl->list);
 592         mutex_init(&caching_ctl->mutex);
 593         init_waitqueue_head(&caching_ctl->wait);
 594         caching_ctl->block_group = cache;
 595         caching_ctl->progress = cache->key.objectid;
 596         atomic_set(&caching_ctl->count, 1);
 597         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 598                         caching_thread, NULL, NULL);
 599
 600         spin_lock(&cache->lock);
 601         /*
 602          * This should be a rare occasion, but this could happen I think in the
 603          * case where one thread starts to load the space cache info, and then
 604          * some other thread starts a transaction commit which tries to do an
 605          * allocation while the other thread is still loading the space cache
 606          * info.  The previous loop should have kept us from choosing this block
 607          * group, but if we've moved to the state where we will wait on caching
 608          * block groups we need to first check if we're doing a fast load here,
 609          * so we can wait for it to finish, otherwise we could end up allocating
 610          * from a block group who's cache gets evicted for one reason or
 611          * another.
 612          */
 613         while (cache->cached == BTRFS_CACHE_FAST) {
 614                 struct btrfs_caching_control *ctl;
 615
 616                 ctl = cache->caching_ctl;
 617                 atomic_inc(&ctl->count);
 618                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 619                 spin_unlock(&cache->lock);
 620
 621                 schedule();
 622
 623                 finish_wait(&ctl->wait, &wait);
 624                 put_caching_control(ctl);
 625                 spin_lock(&cache->lock);
 626         }
 627
 628         if (cache->cached != BTRFS_CACHE_NO) {
 629                 spin_unlock(&cache->lock);
 630                 kfree(caching_ctl);
 631                 return 0;
 632         }
 633         WARN_ON(cache->caching_ctl);
 634         cache->caching_ctl = caching_ctl;
 635         cache->cached = BTRFS_CACHE_FAST;
 636         spin_unlock(&cache->lock);
 637
 638         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 639                 mutex_lock(&caching_ctl->mutex);
 640                 ret = load_free_space_cache(fs_info, cache);
 641
 642                 spin_lock(&cache->lock);
 643                 if (ret == 1) {
 644                         cache->caching_ctl = NULL;
 645                         cache->cached = BTRFS_CACHE_FINISHED;
 646                         cache->last_byte_to_unpin = (u64)-1;
 647                         caching_ctl->progress = (u64)-1;
 648                 } else {
 649                         if (load_cache_only) {
 650                                 cache->caching_ctl = NULL;
 651                                 cache->cached = BTRFS_CACHE_NO;
 652                         } else {
 653                                 cache->cached = BTRFS_CACHE_STARTED;
 654                                 cache->has_caching_ctl = 1;
 655                         }
 656                 }
 657                 spin_unlock(&cache->lock);
 658 #ifdef CONFIG_BTRFS_DEBUG
 659                 if (ret == 1 &&
 660                     btrfs_should_fragment_free_space(fs_info->extent_root,
 661                                                      cache)) {
 662                         u64 bytes_used;
 663
 664                         spin_lock(&cache->space_info->lock);
 665                         spin_lock(&cache->lock);
 666                         bytes_used = cache->key.offset -
 667                                 btrfs_block_group_used(&cache->item);
 668                         cache->space_info->bytes_used += bytes_used >> 1;
 669                         spin_unlock(&cache->lock);
 670                         spin_unlock(&cache->space_info->lock);
 671                         fragment_free_space(fs_info->extent_root, cache);
 672                 }
 673 #endif
 674                 mutex_unlock(&caching_ctl->mutex);
 675
 676                 wake_up(&caching_ctl->wait);
 677                 if (ret == 1) {
 678                         put_caching_control(caching_ctl);
 679                         free_excluded_extents(fs_info->extent_root, cache);
 680                         return 0;
 681                 }
 682         } else {
 683                 /*
 684                  * We are not going to do the fast caching, set cached to the
 685                  * appropriate value and wakeup any waiters.
 686                  */
 687                 spin_lock(&cache->lock);
 688                 if (load_cache_only) {
 689                         cache->caching_ctl = NULL;
 690                         cache->cached = BTRFS_CACHE_NO;
 691                 } else {
 692                         cache->cached = BTRFS_CACHE_STARTED;
 693                         cache->has_caching_ctl = 1;
 694                 }
 695                 spin_unlock(&cache->lock);
 696                 wake_up(&caching_ctl->wait);
 697         }
 698
 699         if (load_cache_only) {
 700                 put_caching_control(caching_ctl);
 701                 return 0;
 702         }
 703
 704         down_write(&fs_info->commit_root_sem);
 705         atomic_inc(&caching_ctl->count);
 706         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 707         up_write(&fs_info->commit_root_sem);
 708
 709         btrfs_get_block_group(cache);
 710
 711         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 712
 713         return ret;
 714 }
 715
 716 /*
 717  * return the block group that starts at or after bytenr
 718  */
 719 static struct btrfs_block_group_cache *
 720 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 721 {
 722         struct btrfs_block_group_cache *cache;
 723
 724         cache = block_group_cache_tree_search(info, bytenr, 0);
 725
 726         return cache;
 727 }
 728
 729 /*
 730  * return the block group that contains the given bytenr
 731  */
 732 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 733                                                  struct btrfs_fs_info *info,
 734                                                  u64 bytenr)
 735 {
 736         struct btrfs_block_group_cache *cache;
 737
 738         cache = block_group_cache_tree_search(info, bytenr, 1);
 739
 740         return cache;
 741 }
 742
 743 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 744                                                   u64 flags)
 745 {
 746         struct list_head *head = &info->space_info;
 747         struct btrfs_space_info *found;
 748
 749         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 750
 751         rcu_read_lock();
 752         list_for_each_entry_rcu(found, head, list) {
 753                 if (found->flags & flags) {
 754                         rcu_read_unlock();
 755                         return found;
 756                 }
 757         }
 758         rcu_read_unlock();
 759         return NULL;
 760 }
 761
 762 /*
 763  * after adding space to the filesystem, we need to clear the full flags
 764  * on all the space infos.
 765  */
 766 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 767 {
 768         struct list_head *head = &info->space_info;
 769         struct btrfs_space_info *found;
 770
 771         rcu_read_lock();
 772         list_for_each_entry_rcu(found, head, list)
 773                 found->full = 0;
 774         rcu_read_unlock();
 775 }
 776
 777 /* simple helper to search for an existing data extent at a given offset */
 778 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 779 {
 780         int ret;
 781         struct btrfs_key key;
 782         struct btrfs_path *path;
 783
 784         path = btrfs_alloc_path();
 785         if (!path)
 786                 return -ENOMEM;
 787
 788         key.objectid = start;
 789         key.offset = len;
 790         key.type = BTRFS_EXTENT_ITEM_KEY;
 791         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 792                                 0, 0);
 793         btrfs_free_path(path);
 794         return ret;
 795 }
 796
 797 /*
 798  * helper function to lookup reference count and flags of a tree block.
 799  *
 800  * the head node for delayed ref is used to store the sum of all the
 801  * reference count modifications queued up in the rbtree. the head
 802  * node may also store the extent flags to set. This way you can check
 803  * to see what the reference count and extent flags would be if all of
 804  * the delayed refs are not processed.
 805  */
 806 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 807                              struct btrfs_root *root, u64 bytenr,
 808                              u64 offset, int metadata, u64 *refs, u64 *flags)
 809 {
 810         struct btrfs_delayed_ref_head *head;
 811         struct btrfs_delayed_ref_root *delayed_refs;
 812         struct btrfs_path *path;
 813         struct btrfs_extent_item *ei;
 814         struct extent_buffer *leaf;
 815         struct btrfs_key key;
 816         u32 item_size;
 817         u64 num_refs;
 818         u64 extent_flags;
 819         int ret;
 820
 821         /*
 822          * If we don't have skinny metadata, don't bother doing anything
 823          * different
 824          */
 825         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 826                 offset = root->nodesize;
 827                 metadata = 0;
 828         }
 829
 830         path = btrfs_alloc_path();
 831         if (!path)
 832                 return -ENOMEM;
 833
 834         if (!trans) {
 835                 path->skip_locking = 1;
 836                 path->search_commit_root = 1;
 837         }
 838
 839 search_again:
 840         key.objectid = bytenr;
 841         key.offset = offset;
 842         if (metadata)
 843                 key.type = BTRFS_METADATA_ITEM_KEY;
 844         else
 845                 key.type = BTRFS_EXTENT_ITEM_KEY;
 846
 847         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 848                                 &key, path, 0, 0);
 849         if (ret < 0)
 850                 goto out_free;
 851
 852         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 853                 if (path->slots[0]) {
 854                         path->slots[0]--;
 855                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 856                                               path->slots[0]);
 857                         if (key.objectid == bytenr &&
 858                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 859                             key.offset == root->nodesize)
 860                                 ret = 0;
 861                 }
 862         }
 863
 864         if (ret == 0) {
 865                 leaf = path->nodes[0];
 866                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 867                 if (item_size >= sizeof(*ei)) {
 868                         ei = btrfs_item_ptr(leaf, path->slots[0],
 869                                             struct btrfs_extent_item);
 870                         num_refs = btrfs_extent_refs(leaf, ei);
 871                         extent_flags = btrfs_extent_flags(leaf, ei);
 872                 } else {
 873 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 874                         struct btrfs_extent_item_v0 *ei0;
 875                         BUG_ON(item_size != sizeof(*ei0));
 876                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 877                                              struct btrfs_extent_item_v0);
 878                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 879                         /* FIXME: this isn't correct for data */
 880                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 881 #else
 882                         BUG();
 883 #endif
 884                 }
 885                 BUG_ON(num_refs == 0);
 886         } else {
 887                 num_refs = 0;
 888                 extent_flags = 0;
 889                 ret = 0;
 890         }
 891
 892         if (!trans)
 893                 goto out;
 894
 895         delayed_refs = &trans->transaction->delayed_refs;
 896         spin_lock(&delayed_refs->lock);
 897         head = btrfs_find_delayed_ref_head(trans, bytenr);
 898         if (head) {
 899                 if (!mutex_trylock(&head->mutex)) {
 900                         atomic_inc(&head->node.refs);
 901                         spin_unlock(&delayed_refs->lock);
 902
 903                         btrfs_release_path(path);
 904
 905                         /*
 906                          * Mutex was contended, block until it's released and try
 907                          * again
 908                          */
 909                         mutex_lock(&head->mutex);
 910                         mutex_unlock(&head->mutex);
 911                         btrfs_put_delayed_ref(&head->node);
 912                         goto search_again;
 913                 }
 914                 spin_lock(&head->lock);
 915                 if (head->extent_op && head->extent_op->update_flags)
 916                         extent_flags |= head->extent_op->flags_to_set;
 917                 else
 918                         BUG_ON(num_refs == 0);
 919
 920                 num_refs += head->node.ref_mod;
 921                 spin_unlock(&head->lock);
 922                 mutex_unlock(&head->mutex);
 923         }
 924         spin_unlock(&delayed_refs->lock);
 925 out:
 926         WARN_ON(num_refs == 0);
 927         if (refs)
 928                 *refs = num_refs;
 929         if (flags)
 930                 *flags = extent_flags;
 931 out_free:
 932         btrfs_free_path(path);
 933         return ret;
 934 }
 935
 936 /*
 937  * Back reference rules.  Back refs have three main goals:
 938  *
 939  * 1) differentiate between all holders of references to an extent so that
 940  *    when a reference is dropped we can make sure it was a valid reference
 941  *    before freeing the extent.
 942  *
 943  * 2) Provide enough information to quickly find the holders of an extent
 944  *    if we notice a given block is corrupted or bad.
 945  *
 946  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 947  *    maintenance.  This is actually the same as #2, but with a slightly
 948  *    different use case.
 949  *
 950  * There are two kinds of back refs. The implicit back refs is optimized
 951  * for pointers in non-shared tree blocks. For a given pointer in a block,
 952  * back refs of this kind provide information about the block's owner tree
 953  * and the pointer's key. These information allow us to find the block by
 954  * b-tree searching. The full back refs is for pointers in tree blocks not
 955  * referenced by their owner trees. The location of tree block is recorded
 956  * in the back refs. Actually the full back refs is generic, and can be
 957  * used in all cases the implicit back refs is used. The major shortcoming
 958  * of the full back refs is its overhead. Every time a tree block gets
 959  * COWed, we have to update back refs entry for all pointers in it.
 960  *
 961  * For a newly allocated tree block, we use implicit back refs for
 962  * pointers in it. This means most tree related operations only involve
 963  * implicit back refs. For a tree block created in old transaction, the
 964  * only way to drop a reference to it is COW it. So we can detect the
 965  * event that tree block loses its owner tree's reference and do the
 966  * back refs conversion.
 967  *
 968  * When a tree block is COW'd through a tree, there are four cases:
 969  *
 970  * The reference count of the block is one and the tree is the block's
 971  * owner tree. Nothing to do in this case.
 972  *
 973  * The reference count of the block is one and the tree is not the
 974  * block's owner tree. In this case, full back refs is used for pointers
 975  * in the block. Remove these full back refs, add implicit back refs for
 976  * every pointers in the new block.
 977  *
 978  * The reference count of the block is greater than one and the tree is
 979  * the block's owner tree. In this case, implicit back refs is used for
 980  * pointers in the block. Add full back refs for every pointers in the
 981  * block, increase lower level extents' reference counts. The original
 982  * implicit back refs are entailed to the new block.
 983  *
 984  * The reference count of the block is greater than one and the tree is
 985  * not the block's owner tree. Add implicit back refs for every pointer in
 986  * the new block, increase lower level extents' reference count.
 987  *
 988  * Back Reference Key composing:
 989  *
 990  * The key objectid corresponds to the first byte in the extent,
 991  * The key type is used to differentiate between types of back refs.
 992  * There are different meanings of the key offset for different types
 993  * of back refs.
 994  *
 995  * File extents can be referenced by:
 996  *
 997  * - multiple snapshots, subvolumes, or different generations in one subvol
 998  * - different files inside a single subvolume
 999  * - different offsets inside a file (bookend extents in file.c)
1000  *
1001  * The extent ref structure for the implicit back refs has fields for:
1002  *
1003  * - Objectid of the subvolume root
1004  * - objectid of the file holding the reference
1005  * - original offset in the file
1006  * - how many bookend extents
1007  *
1008  * The key offset for the implicit back refs is hash of the first
1009  * three fields.
1010  *
1011  * The extent ref structure for the full back refs has field for:
1012  *
1013  * - number of pointers in the tree leaf
1014  *
1015  * The key offset for the implicit back refs is the first byte of
1016  * the tree leaf
1017  *
1018  * When a file extent is allocated, The implicit back refs is used.
1019  * the fields are filled in:
1020  *
1021  *     (root_key.objectid, inode objectid, offset in file, 1)
1022  *
1023  * When a file extent is removed file truncation, we find the
1024  * corresponding implicit back refs and check the following fields:
1025  *
1026  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1027  *
1028  * Btree extents can be referenced by:
1029  *
1030  * - Different subvolumes
1031  *
1032  * Both the implicit back refs and the full back refs for tree blocks
1033  * only consist of key. The key offset for the implicit back refs is
1034  * objectid of block's owner tree. The key offset for the full back refs
1035  * is the first byte of parent block.
1036  *
1037  * When implicit back refs is used, information about the lowest key and
1038  * level of the tree block are required. These information are stored in
1039  * tree block info structure.
1040  */
1041
1042 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1043 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1044                                   struct btrfs_root *root,
1045                                   struct btrfs_path *path,
1046                                   u64 owner, u32 extra_size)
1047 {
1048         struct btrfs_extent_item *item;
1049         struct btrfs_extent_item_v0 *ei0;
1050         struct btrfs_extent_ref_v0 *ref0;
1051         struct btrfs_tree_block_info *bi;
1052         struct extent_buffer *leaf;
1053         struct btrfs_key key;
1054         struct btrfs_key found_key;
1055         u32 new_size = sizeof(*item);
1056         u64 refs;
1057         int ret;
1058
1059         leaf = path->nodes[0];
1060         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1061
1062         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1063         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1064                              struct btrfs_extent_item_v0);
1065         refs = btrfs_extent_refs_v0(leaf, ei0);
1066
1067         if (owner == (u64)-1) {
1068                 while (1) {
1069                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1070                                 ret = btrfs_next_leaf(root, path);
1071                                 if (ret < 0)
1072                                         return ret;
1073                                 BUG_ON(ret > 0); /* Corruption */
1074                                 leaf = path->nodes[0];
1075                         }
1076                         btrfs_item_key_to_cpu(leaf, &found_key,
1077                                               path->slots[0]);
1078                         BUG_ON(key.objectid != found_key.objectid);
1079                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1080                                 path->slots[0]++;
1081                                 continue;
1082                         }
1083                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1084                                               struct btrfs_extent_ref_v0);
1085                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1086                         break;
1087                 }
1088         }
1089         btrfs_release_path(path);
1090
1091         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1092                 new_size += sizeof(*bi);
1093
1094         new_size -= sizeof(*ei0);
1095         ret = btrfs_search_slot(trans, root, &key, path,
1096                                 new_size + extra_size, 1);
1097         if (ret < 0)
1098                 return ret;
1099         BUG_ON(ret); /* Corruption */
1100
1101         btrfs_extend_item(root, path, new_size);
1102
1103         leaf = path->nodes[0];
1104         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1105         btrfs_set_extent_refs(leaf, item, refs);
1106         /* FIXME: get real generation */
1107         btrfs_set_extent_generation(leaf, item, 0);
1108         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1109                 btrfs_set_extent_flags(leaf, item,
1110                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1111                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1112                 bi = (struct btrfs_tree_block_info *)(item + 1);
1113                 /* FIXME: get first key of the block */
1114                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1115                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1116         } else {
1117                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1118         }
1119         btrfs_mark_buffer_dirty(leaf);
1120         return 0;
1121 }
1122 #endif
1123
1124 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1125 {
1126         u32 high_crc = ~(u32)0;
1127         u32 low_crc = ~(u32)0;
1128         __le64 lenum;
1129
1130         lenum = cpu_to_le64(root_objectid);
1131         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1132         lenum = cpu_to_le64(owner);
1133         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1134         lenum = cpu_to_le64(offset);
1135         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1136
1137         return ((u64)high_crc << 31) ^ (u64)low_crc;
1138 }
1139
1140 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1141                                      struct btrfs_extent_data_ref *ref)
1142 {
1143         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1144                                     btrfs_extent_data_ref_objectid(leaf, ref),
1145                                     btrfs_extent_data_ref_offset(leaf, ref));
1146 }
1147
1148 static int match_extent_data_ref(struct extent_buffer *leaf,
1149                                  struct btrfs_extent_data_ref *ref,
1150                                  u64 root_objectid, u64 owner, u64 offset)
1151 {
1152         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1153             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1154             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1155                 return 0;
1156         return 1;
1157 }
1158
1159 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1160                                            struct btrfs_root *root,
1161                                            struct btrfs_path *path,
1162                                            u64 bytenr, u64 parent,
1163                                            u64 root_objectid,
1164                                            u64 owner, u64 offset)
1165 {
1166         struct btrfs_key key;
1167         struct btrfs_extent_data_ref *ref;
1168         struct extent_buffer *leaf;
1169         u32 nritems;
1170         int ret;
1171         int recow;
1172         int err = -ENOENT;
1173
1174         key.objectid = bytenr;
1175         if (parent) {
1176                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1177                 key.offset = parent;
1178         } else {
1179                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1180                 key.offset = hash_extent_data_ref(root_objectid,
1181                                                   owner, offset);
1182         }
1183 again:
1184         recow = 0;
1185         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1186         if (ret < 0) {
1187                 err = ret;
1188                 goto fail;
1189         }
1190
1191         if (parent) {
1192                 if (!ret)
1193                         return 0;
1194 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1195                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1196                 btrfs_release_path(path);
1197                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1198                 if (ret < 0) {
1199                         err = ret;
1200                         goto fail;
1201                 }
1202                 if (!ret)
1203                         return 0;
1204 #endif
1205                 goto fail;
1206         }
1207
1208         leaf = path->nodes[0];
1209         nritems = btrfs_header_nritems(leaf);
1210         while (1) {
1211                 if (path->slots[0] >= nritems) {
1212                         ret = btrfs_next_leaf(root, path);
1213                         if (ret < 0)
1214                                 err = ret;
1215                         if (ret)
1216                                 goto fail;
1217
1218                         leaf = path->nodes[0];
1219                         nritems = btrfs_header_nritems(leaf);
1220                         recow = 1;
1221                 }
1222
1223                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1224                 if (key.objectid != bytenr ||
1225                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1226                         goto fail;
1227
1228                 ref = btrfs_item_ptr(leaf, path->slots[0],
1229                                      struct btrfs_extent_data_ref);
1230
1231                 if (match_extent_data_ref(leaf, ref, root_objectid,
1232                                           owner, offset)) {
1233                         if (recow) {
1234                                 btrfs_release_path(path);
1235                                 goto again;
1236                         }
1237                         err = 0;
1238                         break;
1239                 }
1240                 path->slots[0]++;
1241         }
1242 fail:
1243         return err;
1244 }
1245
1246 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1247                                            struct btrfs_root *root,
1248                                            struct btrfs_path *path,
1249                                            u64 bytenr, u64 parent,
1250                                            u64 root_objectid, u64 owner,
1251                                            u64 offset, int refs_to_add)
1252 {
1253         struct btrfs_key key;
1254         struct extent_buffer *leaf;
1255         u32 size;
1256         u32 num_refs;
1257         int ret;
1258
1259         key.objectid = bytenr;
1260         if (parent) {
1261                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1262                 key.offset = parent;
1263                 size = sizeof(struct btrfs_shared_data_ref);
1264         } else {
1265                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1266                 key.offset = hash_extent_data_ref(root_objectid,
1267                                                   owner, offset);
1268                 size = sizeof(struct btrfs_extent_data_ref);
1269         }
1270
1271         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1272         if (ret && ret != -EEXIST)
1273                 goto fail;
1274
1275         leaf = path->nodes[0];
1276         if (parent) {
1277                 struct btrfs_shared_data_ref *ref;
1278                 ref = btrfs_item_ptr(leaf, path->slots[0],
1279                                      struct btrfs_shared_data_ref);
1280                 if (ret == 0) {
1281                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1282                 } else {
1283                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1284                         num_refs += refs_to_add;
1285                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1286                 }
1287         } else {
1288                 struct btrfs_extent_data_ref *ref;
1289                 while (ret == -EEXIST) {
1290                         ref = btrfs_item_ptr(leaf, path->slots[0],
1291                                              struct btrfs_extent_data_ref);
1292                         if (match_extent_data_ref(leaf, ref, root_objectid,
1293                                                   owner, offset))
1294                                 break;
1295                         btrfs_release_path(path);
1296                         key.offset++;
1297                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1298                                                       size);
1299                         if (ret && ret != -EEXIST)
1300                                 goto fail;
1301
1302                         leaf = path->nodes[0];
1303                 }
1304                 ref = btrfs_item_ptr(leaf, path->slots[0],
1305                                      struct btrfs_extent_data_ref);
1306                 if (ret == 0) {
1307                         btrfs_set_extent_data_ref_root(leaf, ref,
1308                                                        root_objectid);
1309                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1310                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1311                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1312                 } else {
1313                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1314                         num_refs += refs_to_add;
1315                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1316                 }
1317         }
1318         btrfs_mark_buffer_dirty(leaf);
1319         ret = 0;
1320 fail:
1321         btrfs_release_path(path);
1322         return ret;
1323 }
1324
1325 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1326                                            struct btrfs_root *root,
1327                                            struct btrfs_path *path,
1328                                            int refs_to_drop, int *last_ref)
1329 {
1330         struct btrfs_key key;
1331         struct btrfs_extent_data_ref *ref1 = NULL;
1332         struct btrfs_shared_data_ref *ref2 = NULL;
1333         struct extent_buffer *leaf;
1334         u32 num_refs = 0;
1335         int ret = 0;
1336
1337         leaf = path->nodes[0];
1338         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1339
1340         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1341                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1342                                       struct btrfs_extent_data_ref);
1343                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1344         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1345                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1346                                       struct btrfs_shared_data_ref);
1347                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1348 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1349         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1350                 struct btrfs_extent_ref_v0 *ref0;
1351                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1352                                       struct btrfs_extent_ref_v0);
1353                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1354 #endif
1355         } else {
1356                 BUG();
1357         }
1358
1359         BUG_ON(num_refs < refs_to_drop);
1360         num_refs -= refs_to_drop;
1361
1362         if (num_refs == 0) {
1363                 ret = btrfs_del_item(trans, root, path);
1364                 *last_ref = 1;
1365         } else {
1366                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1367                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1368                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1369                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1370 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1371                 else {
1372                         struct btrfs_extent_ref_v0 *ref0;
1373                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1374                                         struct btrfs_extent_ref_v0);
1375                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1376                 }
1377 #endif
1378                 btrfs_mark_buffer_dirty(leaf);
1379         }
1380         return ret;
1381 }
1382
1383 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1384                                           struct btrfs_extent_inline_ref *iref)
1385 {
1386         struct btrfs_key key;
1387         struct extent_buffer *leaf;
1388         struct btrfs_extent_data_ref *ref1;
1389         struct btrfs_shared_data_ref *ref2;
1390         u32 num_refs = 0;
1391
1392         leaf = path->nodes[0];
1393         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1394         if (iref) {
1395                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1396                     BTRFS_EXTENT_DATA_REF_KEY) {
1397                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1398                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1399                 } else {
1400                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1401                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1402                 }
1403         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1404                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1405                                       struct btrfs_extent_data_ref);
1406                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1407         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1408                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1409                                       struct btrfs_shared_data_ref);
1410                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1411 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1412         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1413                 struct btrfs_extent_ref_v0 *ref0;
1414                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1415                                       struct btrfs_extent_ref_v0);
1416                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1417 #endif
1418         } else {
1419                 WARN_ON(1);
1420         }
1421         return num_refs;
1422 }
1423
1424 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1425                                           struct btrfs_root *root,
1426                                           struct btrfs_path *path,
1427                                           u64 bytenr, u64 parent,
1428                                           u64 root_objectid)
1429 {
1430         struct btrfs_key key;
1431         int ret;
1432
1433         key.objectid = bytenr;
1434         if (parent) {
1435                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1436                 key.offset = parent;
1437         } else {
1438                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1439                 key.offset = root_objectid;
1440         }
1441
1442         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1443         if (ret > 0)
1444                 ret = -ENOENT;
1445 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1446         if (ret == -ENOENT && parent) {
1447                 btrfs_release_path(path);
1448                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1449                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1450                 if (ret > 0)
1451                         ret = -ENOENT;
1452         }
1453 #endif
1454         return ret;
1455 }
1456
1457 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1458                                           struct btrfs_root *root,
1459                                           struct btrfs_path *path,
1460                                           u64 bytenr, u64 parent,
1461                                           u64 root_objectid)
1462 {
1463         struct btrfs_key key;
1464         int ret;
1465
1466         key.objectid = bytenr;
1467         if (parent) {
1468                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1469                 key.offset = parent;
1470         } else {
1471                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1472                 key.offset = root_objectid;
1473         }
1474
1475         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1476         btrfs_release_path(path);
1477         return ret;
1478 }
1479
1480 static inline int extent_ref_type(u64 parent, u64 owner)
1481 {
1482         int type;
1483         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1484                 if (parent > 0)
1485                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1486                 else
1487                         type = BTRFS_TREE_BLOCK_REF_KEY;
1488         } else {
1489                 if (parent > 0)
1490                         type = BTRFS_SHARED_DATA_REF_KEY;
1491                 else
1492                         type = BTRFS_EXTENT_DATA_REF_KEY;
1493         }
1494         return type;
1495 }
1496
1497 static int find_next_key(struct btrfs_path *path, int level,
1498                          struct btrfs_key *key)
1499
1500 {
1501         for (; level < BTRFS_MAX_LEVEL; level++) {
1502                 if (!path->nodes[level])
1503                         break;
1504                 if (path->slots[level] + 1 >=
1505                     btrfs_header_nritems(path->nodes[level]))
1506                         continue;
1507                 if (level == 0)
1508                         btrfs_item_key_to_cpu(path->nodes[level], key,
1509                                               path->slots[level] + 1);
1510                 else
1511                         btrfs_node_key_to_cpu(path->nodes[level], key,
1512                                               path->slots[level] + 1);
1513                 return 0;
1514         }
1515         return 1;
1516 }
1517
1518 /*
1519  * look for inline back ref. if back ref is found, *ref_ret is set
1520  * to the address of inline back ref, and 0 is returned.
1521  *
1522  * if back ref isn't found, *ref_ret is set to the address where it
1523  * should be inserted, and -ENOENT is returned.
1524  *
1525  * if insert is true and there are too many inline back refs, the path
1526  * points to the extent item, and -EAGAIN is returned.
1527  *
1528  * NOTE: inline back refs are ordered in the same way that back ref
1529  *       items in the tree are ordered.
1530  */
1531 static noinline_for_stack
1532 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1533                                  struct btrfs_root *root,
1534                                  struct btrfs_path *path,
1535                                  struct btrfs_extent_inline_ref **ref_ret,
1536                                  u64 bytenr, u64 num_bytes,
1537                                  u64 parent, u64 root_objectid,
1538                                  u64 owner, u64 offset, int insert)
1539 {
1540         struct btrfs_key key;
1541         struct extent_buffer *leaf;
1542         struct btrfs_extent_item *ei;
1543         struct btrfs_extent_inline_ref *iref;
1544         u64 flags;
1545         u64 item_size;
1546         unsigned long ptr;
1547         unsigned long end;
1548         int extra_size;
1549         int type;
1550         int want;
1551         int ret;
1552         int err = 0;
1553         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1554                                                  SKINNY_METADATA);
1555
1556         key.objectid = bytenr;
1557         key.type = BTRFS_EXTENT_ITEM_KEY;
1558         key.offset = num_bytes;
1559
1560         want = extent_ref_type(parent, owner);
1561         if (insert) {
1562                 extra_size = btrfs_extent_inline_ref_size(want);
1563                 path->keep_locks = 1;
1564         } else
1565                 extra_size = -1;
1566
1567         /*
1568          * Owner is our parent level, so we can just add one to get the level
1569          * for the block we are interested in.
1570          */
1571         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1572                 key.type = BTRFS_METADATA_ITEM_KEY;
1573                 key.offset = owner;
1574         }
1575
1576 again:
1577         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1578         if (ret < 0) {
1579                 err = ret;
1580                 goto out;
1581         }
1582
1583         /*
1584          * We may be a newly converted file system which still has the old fat
1585          * extent entries for metadata, so try and see if we have one of those.
1586          */
1587         if (ret > 0 && skinny_metadata) {
1588                 skinny_metadata = false;
1589                 if (path->slots[0]) {
1590                         path->slots[0]--;
1591                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1592                                               path->slots[0]);
1593                         if (key.objectid == bytenr &&
1594                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1595                             key.offset == num_bytes)
1596                                 ret = 0;
1597                 }
1598                 if (ret) {
1599                         key.objectid = bytenr;
1600                         key.type = BTRFS_EXTENT_ITEM_KEY;
1601                         key.offset = num_bytes;
1602                         btrfs_release_path(path);
1603                         goto again;
1604                 }
1605         }
1606
1607         if (ret && !insert) {
1608                 err = -ENOENT;
1609                 goto out;
1610         } else if (WARN_ON(ret)) {
1611                 err = -EIO;
1612                 goto out;
1613         }
1614
1615         leaf = path->nodes[0];
1616         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1617 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1618         if (item_size < sizeof(*ei)) {
1619                 if (!insert) {
1620                         err = -ENOENT;
1621                         goto out;
1622                 }
1623                 ret = convert_extent_item_v0(trans, root, path, owner,
1624                                              extra_size);
1625                 if (ret < 0) {
1626                         err = ret;
1627                         goto out;
1628                 }
1629                 leaf = path->nodes[0];
1630                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1631         }
1632 #endif
1633         BUG_ON(item_size < sizeof(*ei));
1634
1635         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1636         flags = btrfs_extent_flags(leaf, ei);
1637
1638         ptr = (unsigned long)(ei + 1);
1639         end = (unsigned long)ei + item_size;
1640
1641         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1642                 ptr += sizeof(struct btrfs_tree_block_info);
1643                 BUG_ON(ptr > end);
1644         }
1645
1646         err = -ENOENT;
1647         while (1) {
1648                 if (ptr >= end) {
1649                         WARN_ON(ptr > end);
1650                         break;
1651                 }
1652                 iref = (struct btrfs_extent_inline_ref *)ptr;
1653                 type = btrfs_extent_inline_ref_type(leaf, iref);
1654                 if (want < type)
1655                         break;
1656                 if (want > type) {
1657                         ptr += btrfs_extent_inline_ref_size(type);
1658                         continue;
1659                 }
1660
1661                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1662                         struct btrfs_extent_data_ref *dref;
1663                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1664                         if (match_extent_data_ref(leaf, dref, root_objectid,
1665                                                   owner, offset)) {
1666                                 err = 0;
1667                                 break;
1668                         }
1669                         if (hash_extent_data_ref_item(leaf, dref) <
1670                             hash_extent_data_ref(root_objectid, owner, offset))
1671                                 break;
1672                 } else {
1673                         u64 ref_offset;
1674                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1675                         if (parent > 0) {
1676                                 if (parent == ref_offset) {
1677                                         err = 0;
1678                                         break;
1679                                 }
1680                                 if (ref_offset < parent)
1681                                         break;
1682                         } else {
1683                                 if (root_objectid == ref_offset) {
1684                                         err = 0;
1685                                         break;
1686                                 }
1687                                 if (ref_offset < root_objectid)
1688                                         break;
1689                         }
1690                 }
1691                 ptr += btrfs_extent_inline_ref_size(type);
1692         }
1693         if (err == -ENOENT && insert) {
1694                 if (item_size + extra_size >=
1695                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1696                         err = -EAGAIN;
1697                         goto out;
1698                 }
1699                 /*
1700                  * To add new inline back ref, we have to make sure
1701                  * there is no corresponding back ref item.
1702                  * For simplicity, we just do not add new inline back
1703                  * ref if there is any kind of item for this block
1704                  */
1705                 if (find_next_key(path, 0, &key) == 0 &&
1706                     key.objectid == bytenr &&
1707                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1708                         err = -EAGAIN;
1709                         goto out;
1710                 }
1711         }
1712         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1713 out:
1714         if (insert) {
1715                 path->keep_locks = 0;
1716                 btrfs_unlock_up_safe(path, 1);
1717         }
1718         return err;
1719 }
1720
1721 /*
1722  * helper to add new inline back ref
1723  */
1724 static noinline_for_stack
1725 void setup_inline_extent_backref(struct btrfs_root *root,
1726                                  struct btrfs_path *path,
1727                                  struct btrfs_extent_inline_ref *iref,
1728                                  u64 parent, u64 root_objectid,
1729                                  u64 owner, u64 offset, int refs_to_add,
1730                                  struct btrfs_delayed_extent_op *extent_op)
1731 {
1732         struct extent_buffer *leaf;
1733         struct btrfs_extent_item *ei;
1734         unsigned long ptr;
1735         unsigned long end;
1736         unsigned long item_offset;
1737         u64 refs;
1738         int size;
1739         int type;
1740
1741         leaf = path->nodes[0];
1742         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1743         item_offset = (unsigned long)iref - (unsigned long)ei;
1744
1745         type = extent_ref_type(parent, owner);
1746         size = btrfs_extent_inline_ref_size(type);
1747
1748         btrfs_extend_item(root, path, size);
1749
1750         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1751         refs = btrfs_extent_refs(leaf, ei);
1752         refs += refs_to_add;
1753         btrfs_set_extent_refs(leaf, ei, refs);
1754         if (extent_op)
1755                 __run_delayed_extent_op(extent_op, leaf, ei);
1756
1757         ptr = (unsigned long)ei + item_offset;
1758         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1759         if (ptr < end - size)
1760                 memmove_extent_buffer(leaf, ptr + size, ptr,
1761                                       end - size - ptr);
1762
1763         iref = (struct btrfs_extent_inline_ref *)ptr;
1764         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1765         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1766                 struct btrfs_extent_data_ref *dref;
1767                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1768                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1769                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1770                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1771                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1772         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1773                 struct btrfs_shared_data_ref *sref;
1774                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1775                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1776                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1777         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1778                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1779         } else {
1780                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1781         }
1782         btrfs_mark_buffer_dirty(leaf);
1783 }
1784
1785 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1786                                  struct btrfs_root *root,
1787                                  struct btrfs_path *path,
1788                                  struct btrfs_extent_inline_ref **ref_ret,
1789                                  u64 bytenr, u64 num_bytes, u64 parent,
1790                                  u64 root_objectid, u64 owner, u64 offset)
1791 {
1792         int ret;
1793
1794         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1795                                            bytenr, num_bytes, parent,
1796                                            root_objectid, owner, offset, 0);
1797         if (ret != -ENOENT)
1798                 return ret;
1799
1800         btrfs_release_path(path);
1801         *ref_ret = NULL;
1802
1803         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1804                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1805                                             root_objectid);
1806         } else {
1807                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1808                                              root_objectid, owner, offset);
1809         }
1810         return ret;
1811 }
1812
1813 /*
1814  * helper to update/remove inline back ref
1815  */
1816 static noinline_for_stack
1817 void update_inline_extent_backref(struct btrfs_root *root,
1818                                   struct btrfs_path *path,
1819                                   struct btrfs_extent_inline_ref *iref,
1820                                   int refs_to_mod,
1821                                   struct btrfs_delayed_extent_op *extent_op,
1822                                   int *last_ref)
1823 {
1824         struct extent_buffer *leaf;
1825         struct btrfs_extent_item *ei;
1826         struct btrfs_extent_data_ref *dref = NULL;
1827         struct btrfs_shared_data_ref *sref = NULL;
1828         unsigned long ptr;
1829         unsigned long end;
1830         u32 item_size;
1831         int size;
1832         int type;
1833         u64 refs;
1834
1835         leaf = path->nodes[0];
1836         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1837         refs = btrfs_extent_refs(leaf, ei);
1838         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1839         refs += refs_to_mod;
1840         btrfs_set_extent_refs(leaf, ei, refs);
1841         if (extent_op)
1842                 __run_delayed_extent_op(extent_op, leaf, ei);
1843
1844         type = btrfs_extent_inline_ref_type(leaf, iref);
1845
1846         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1847                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1848                 refs = btrfs_extent_data_ref_count(leaf, dref);
1849         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1850                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1851                 refs = btrfs_shared_data_ref_count(leaf, sref);
1852         } else {
1853                 refs = 1;
1854                 BUG_ON(refs_to_mod != -1);
1855         }
1856
1857         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1858         refs += refs_to_mod;
1859
1860         if (refs > 0) {
1861                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1862                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1863                 else
1864                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1865         } else {
1866                 *last_ref = 1;
1867                 size =  btrfs_extent_inline_ref_size(type);
1868                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1869                 ptr = (unsigned long)iref;
1870                 end = (unsigned long)ei + item_size;
1871                 if (ptr + size < end)
1872                         memmove_extent_buffer(leaf, ptr, ptr + size,
1873                                               end - ptr - size);
1874                 item_size -= size;
1875                 btrfs_truncate_item(root, path, item_size, 1);
1876         }
1877         btrfs_mark_buffer_dirty(leaf);
1878 }
1879
1880 static noinline_for_stack
1881 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1882                                  struct btrfs_root *root,
1883                                  struct btrfs_path *path,
1884                                  u64 bytenr, u64 num_bytes, u64 parent,
1885                                  u64 root_objectid, u64 owner,
1886                                  u64 offset, int refs_to_add,
1887                                  struct btrfs_delayed_extent_op *extent_op)
1888 {
1889         struct btrfs_extent_inline_ref *iref;
1890         int ret;
1891
1892         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1893                                            bytenr, num_bytes, parent,
1894                                            root_objectid, owner, offset, 1);
1895         if (ret == 0) {
1896                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1897                 update_inline_extent_backref(root, path, iref,
1898                                              refs_to_add, extent_op, NULL);
1899         } else if (ret == -ENOENT) {
1900                 setup_inline_extent_backref(root, path, iref, parent,
1901                                             root_objectid, owner, offset,
1902                                             refs_to_add, extent_op);
1903                 ret = 0;
1904         }
1905         return ret;
1906 }
1907
1908 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1909                                  struct btrfs_root *root,
1910                                  struct btrfs_path *path,
1911                                  u64 bytenr, u64 parent, u64 root_objectid,
1912                                  u64 owner, u64 offset, int refs_to_add)
1913 {
1914         int ret;
1915         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1916                 BUG_ON(refs_to_add != 1);
1917                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1918                                             parent, root_objectid);
1919         } else {
1920                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1921                                              parent, root_objectid,
1922                                              owner, offset, refs_to_add);
1923         }
1924         return ret;
1925 }
1926
1927 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1928                                  struct btrfs_root *root,
1929                                  struct btrfs_path *path,
1930                                  struct btrfs_extent_inline_ref *iref,
1931                                  int refs_to_drop, int is_data, int *last_ref)
1932 {
1933         int ret = 0;
1934
1935         BUG_ON(!is_data && refs_to_drop != 1);
1936         if (iref) {
1937                 update_inline_extent_backref(root, path, iref,
1938                                              -refs_to_drop, NULL, last_ref);
1939         } else if (is_data) {
1940                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1941                                              last_ref);
1942         } else {
1943                 *last_ref = 1;
1944                 ret = btrfs_del_item(trans, root, path);
1945         }
1946         return ret;
1947 }
1948
1949 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1950 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1951                                u64 *discarded_bytes)
1952 {
1953         int j, ret = 0;
1954         u64 bytes_left, end;
1955         u64 aligned_start = ALIGN(start, 1 << 9);
1956
1957         if (WARN_ON(start != aligned_start)) {
1958                 len -= aligned_start - start;
1959                 len = round_down(len, 1 << 9);
1960                 start = aligned_start;
1961         }
1962
1963         *discarded_bytes = 0;
1964
1965         if (!len)
1966                 return 0;
1967
1968         end = start + len;
1969         bytes_left = len;
1970
1971         /* Skip any superblocks on this device. */
1972         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1973                 u64 sb_start = btrfs_sb_offset(j);
1974                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1975                 u64 size = sb_start - start;
1976
1977                 if (!in_range(sb_start, start, bytes_left) &&
1978                     !in_range(sb_end, start, bytes_left) &&
1979                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1980                         continue;
1981
1982                 /*
1983                  * Superblock spans beginning of range.  Adjust start and
1984                  * try again.
1985                  */
1986                 if (sb_start <= start) {
1987                         start += sb_end - start;
1988                         if (start > end) {
1989                                 bytes_left = 0;
1990                                 break;
1991                         }
1992                         bytes_left = end - start;
1993                         continue;
1994                 }
1995
1996                 if (size) {
1997                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
1998                                                    GFP_NOFS, 0);
1999                         if (!ret)
2000                                 *discarded_bytes += size;
2001                         else if (ret != -EOPNOTSUPP)
2002                                 return ret;
2003                 }
2004
2005                 start = sb_end;
2006                 if (start > end) {
2007                         bytes_left = 0;
2008                         break;
2009                 }
2010                 bytes_left = end - start;
2011         }
2012
2013         if (bytes_left) {
2014                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2015                                            GFP_NOFS, 0);
2016                 if (!ret)
2017                         *discarded_bytes += bytes_left;
2018         }
2019         return ret;
2020 }
2021
2022 int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
2023                          u64 num_bytes, u64 *actual_bytes)
2024 {
2025         int ret;
2026         u64 discarded_bytes = 0;
2027         struct btrfs_bio *bbio = NULL;
2028
2029
2030         /* Tell the block device(s) that the sectors can be discarded */
2031         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
2032                               bytenr, &num_bytes, &bbio, 0);
2033         /* Error condition is -ENOMEM */
2034         if (!ret) {
2035                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2036                 int i;
2037
2038
2039                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2040                         u64 bytes;
2041                         if (!stripe->dev->can_discard)
2042                                 continue;
2043
2044                         ret = btrfs_issue_discard(stripe->dev->bdev,
2045                                                   stripe->physical,
2046                                                   stripe->length,
2047                                                   &bytes);
2048                         if (!ret)
2049                                 discarded_bytes += bytes;
2050                         else if (ret != -EOPNOTSUPP)
2051                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2052
2053                         /*
2054                          * Just in case we get back EOPNOTSUPP for some reason,
2055                          * just ignore the return value so we don't screw up
2056                          * people calling discard_extent.
2057                          */
2058                         ret = 0;
2059                 }
2060                 btrfs_put_bbio(bbio);
2061         }
2062
2063         if (actual_bytes)
2064                 *actual_bytes = discarded_bytes;
2065
2066
2067         if (ret == -EOPNOTSUPP)
2068                 ret = 0;
2069         return ret;
2070 }
2071
2072 /* Can return -ENOMEM */
2073 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2074                          struct btrfs_root *root,
2075                          u64 bytenr, u64 num_bytes, u64 parent,
2076                          u64 root_objectid, u64 owner, u64 offset,
2077                          int no_quota)
2078 {
2079         int ret;
2080         struct btrfs_fs_info *fs_info = root->fs_info;
2081
2082         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2083                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2084
2085         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2086                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2087                                         num_bytes,
2088                                         parent, root_objectid, (int)owner,
2089                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
2090         } else {
2091                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2092                                         num_bytes,
2093                                         parent, root_objectid, owner, offset,
2094                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
2095         }
2096         return ret;
2097 }
2098
2099 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2100                                   struct btrfs_root *root,
2101                                   struct btrfs_delayed_ref_node *node,
2102                                   u64 parent, u64 root_objectid,
2103                                   u64 owner, u64 offset, int refs_to_add,
2104                                   struct btrfs_delayed_extent_op *extent_op)
2105 {
2106         struct btrfs_fs_info *fs_info = root->fs_info;
2107         struct btrfs_path *path;
2108         struct extent_buffer *leaf;
2109         struct btrfs_extent_item *item;
2110         struct btrfs_key key;
2111         u64 bytenr = node->bytenr;
2112         u64 num_bytes = node->num_bytes;
2113         u64 refs;
2114         int ret;
2115         int no_quota = node->no_quota;
2116
2117         path = btrfs_alloc_path();
2118         if (!path)
2119                 return -ENOMEM;
2120
2121         if (!is_fstree(root_objectid) || !root->fs_info->quota_enabled)
2122                 no_quota = 1;
2123
2124         path->reada = 1;
2125         path->leave_spinning = 1;
2126         /* this will setup the path even if it fails to insert the back ref */
2127         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2128                                            bytenr, num_bytes, parent,
2129                                            root_objectid, owner, offset,
2130                                            refs_to_add, extent_op);
2131         if ((ret < 0 && ret != -EAGAIN) || !ret)
2132                 goto out;
2133
2134         /*
2135          * Ok we had -EAGAIN which means we didn't have space to insert and
2136          * inline extent ref, so just update the reference count and add a
2137          * normal backref.
2138          */
2139         leaf = path->nodes[0];
2140         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2141         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2142         refs = btrfs_extent_refs(leaf, item);
2143         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2144         if (extent_op)
2145                 __run_delayed_extent_op(extent_op, leaf, item);
2146
2147         btrfs_mark_buffer_dirty(leaf);
2148         btrfs_release_path(path);
2149
2150         path->reada = 1;
2151         path->leave_spinning = 1;
2152         /* now insert the actual backref */
2153         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2154                                     path, bytenr, parent, root_objectid,
2155                                     owner, offset, refs_to_add);
2156         if (ret)
2157                 btrfs_abort_transaction(trans, root, ret);
2158 out:
2159         btrfs_free_path(path);
2160         return ret;
2161 }
2162
2163 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2164                                 struct btrfs_root *root,
2165                                 struct btrfs_delayed_ref_node *node,
2166                                 struct btrfs_delayed_extent_op *extent_op,
2167                                 int insert_reserved)
2168 {
2169         int ret = 0;
2170         struct btrfs_delayed_data_ref *ref;
2171         struct btrfs_key ins;
2172         u64 parent = 0;
2173         u64 ref_root = 0;
2174         u64 flags = 0;
2175
2176         ins.objectid = node->bytenr;
2177         ins.offset = node->num_bytes;
2178         ins.type = BTRFS_EXTENT_ITEM_KEY;
2179
2180         ref = btrfs_delayed_node_to_data_ref(node);
2181         trace_run_delayed_data_ref(node, ref, node->action);
2182
2183         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2184                 parent = ref->parent;
2185         ref_root = ref->root;
2186
2187         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2188                 if (extent_op)
2189                         flags |= extent_op->flags_to_set;
2190                 ret = alloc_reserved_file_extent(trans, root,
2191                                                  parent, ref_root, flags,
2192                                                  ref->objectid, ref->offset,
2193                                                  &ins, node->ref_mod);
2194         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2195                 ret = __btrfs_inc_extent_ref(trans, root, node, parent,
2196                                              ref_root, ref->objectid,
2197                                              ref->offset, node->ref_mod,
2198                                              extent_op);
2199         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2200                 ret = __btrfs_free_extent(trans, root, node, parent,
2201                                           ref_root, ref->objectid,
2202                                           ref->offset, node->ref_mod,
2203                                           extent_op);
2204         } else {
2205                 BUG();
2206         }
2207         return ret;
2208 }
2209
2210 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2211                                     struct extent_buffer *leaf,
2212                                     struct btrfs_extent_item *ei)
2213 {
2214         u64 flags = btrfs_extent_flags(leaf, ei);
2215         if (extent_op->update_flags) {
2216                 flags |= extent_op->flags_to_set;
2217                 btrfs_set_extent_flags(leaf, ei, flags);
2218         }
2219
2220         if (extent_op->update_key) {
2221                 struct btrfs_tree_block_info *bi;
2222                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2223                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2224                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2225         }
2226 }
2227
2228 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2229                                  struct btrfs_root *root,
2230                                  struct btrfs_delayed_ref_node *node,
2231                                  struct btrfs_delayed_extent_op *extent_op)
2232 {
2233         struct btrfs_key key;
2234         struct btrfs_path *path;
2235         struct btrfs_extent_item *ei;
2236         struct extent_buffer *leaf;
2237         u32 item_size;
2238         int ret;
2239         int err = 0;
2240         int metadata = !extent_op->is_data;
2241
2242         if (trans->aborted)
2243                 return 0;
2244
2245         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2246                 metadata = 0;
2247
2248         path = btrfs_alloc_path();
2249         if (!path)
2250                 return -ENOMEM;
2251
2252         key.objectid = node->bytenr;
2253
2254         if (metadata) {
2255                 key.type = BTRFS_METADATA_ITEM_KEY;
2256                 key.offset = extent_op->level;
2257         } else {
2258                 key.type = BTRFS_EXTENT_ITEM_KEY;
2259                 key.offset = node->num_bytes;
2260         }
2261
2262 again:
2263         path->reada = 1;
2264         path->leave_spinning = 1;
2265         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2266                                 path, 0, 1);
2267         if (ret < 0) {
2268                 err = ret;
2269                 goto out;
2270         }
2271         if (ret > 0) {
2272                 if (metadata) {
2273                         if (path->slots[0] > 0) {
2274                                 path->slots[0]--;
2275                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2276                                                       path->slots[0]);
2277                                 if (key.objectid == node->bytenr &&
2278                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2279                                     key.offset == node->num_bytes)
2280                                         ret = 0;
2281                         }
2282                         if (ret > 0) {
2283                                 btrfs_release_path(path);
2284                                 metadata = 0;
2285
2286                                 key.objectid = node->bytenr;
2287                                 key.offset = node->num_bytes;
2288                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2289                                 goto again;
2290                         }
2291                 } else {
2292                         err = -EIO;
2293                         goto out;
2294                 }
2295         }
2296
2297         leaf = path->nodes[0];
2298         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2299 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2300         if (item_size < sizeof(*ei)) {
2301                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2302                                              path, (u64)-1, 0);
2303                 if (ret < 0) {
2304                         err = ret;
2305                         goto out;
2306                 }
2307                 leaf = path->nodes[0];
2308                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2309         }
2310 #endif
2311         BUG_ON(item_size < sizeof(*ei));
2312         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2313         __run_delayed_extent_op(extent_op, leaf, ei);
2314
2315         btrfs_mark_buffer_dirty(leaf);
2316 out:
2317         btrfs_free_path(path);
2318         return err;
2319 }
2320
2321 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2322                                 struct btrfs_root *root,
2323                                 struct btrfs_delayed_ref_node *node,
2324                                 struct btrfs_delayed_extent_op *extent_op,
2325                                 int insert_reserved)
2326 {
2327         int ret = 0;
2328         struct btrfs_delayed_tree_ref *ref;
2329         struct btrfs_key ins;
2330         u64 parent = 0;
2331         u64 ref_root = 0;
2332         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2333                                                  SKINNY_METADATA);
2334
2335         ref = btrfs_delayed_node_to_tree_ref(node);
2336         trace_run_delayed_tree_ref(node, ref, node->action);
2337
2338         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2339                 parent = ref->parent;
2340         ref_root = ref->root;
2341
2342         ins.objectid = node->bytenr;
2343         if (skinny_metadata) {
2344                 ins.offset = ref->level;
2345                 ins.type = BTRFS_METADATA_ITEM_KEY;
2346         } else {
2347                 ins.offset = node->num_bytes;
2348                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2349         }
2350
2351         BUG_ON(node->ref_mod != 1);
2352         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2353                 BUG_ON(!extent_op || !extent_op->update_flags);
2354                 ret = alloc_reserved_tree_block(trans, root,
2355                                                 parent, ref_root,
2356                                                 extent_op->flags_to_set,
2357                                                 &extent_op->key,
2358                                                 ref->level, &ins,
2359                                                 node->no_quota);
2360         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2361                 ret = __btrfs_inc_extent_ref(trans, root, node,
2362                                              parent, ref_root,
2363                                              ref->level, 0, 1,
2364                                              extent_op);
2365         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2366                 ret = __btrfs_free_extent(trans, root, node,
2367                                           parent, ref_root,
2368                                           ref->level, 0, 1, extent_op);
2369         } else {
2370                 BUG();
2371         }
2372         return ret;
2373 }
2374
2375 /* helper function to actually process a single delayed ref entry */
2376 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2377                                struct btrfs_root *root,
2378                                struct btrfs_delayed_ref_node *node,
2379                                struct btrfs_delayed_extent_op *extent_op,
2380                                int insert_reserved)
2381 {
2382         int ret = 0;
2383
2384         if (trans->aborted) {
2385                 if (insert_reserved)
2386                         btrfs_pin_extent(root, node->bytenr,
2387                                          node->num_bytes, 1);
2388                 return 0;
2389         }
2390
2391         if (btrfs_delayed_ref_is_head(node)) {
2392                 struct btrfs_delayed_ref_head *head;
2393                 /*
2394                  * we've hit the end of the chain and we were supposed
2395                  * to insert this extent into the tree.  But, it got
2396                  * deleted before we ever needed to insert it, so all
2397                  * we have to do is clean up the accounting
2398                  */
2399                 BUG_ON(extent_op);
2400                 head = btrfs_delayed_node_to_head(node);
2401                 trace_run_delayed_ref_head(node, head, node->action);
2402
2403                 if (insert_reserved) {
2404                         btrfs_pin_extent(root, node->bytenr,
2405                                          node->num_bytes, 1);
2406                         if (head->is_data) {
2407                                 ret = btrfs_del_csums(trans, root,
2408                                                       node->bytenr,
2409                                                       node->num_bytes);
2410                         }
2411                 }
2412                 return ret;
2413         }
2414
2415         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2416             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2417                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2418                                            insert_reserved);
2419         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2420                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2421                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2422                                            insert_reserved);
2423         else
2424                 BUG();
2425         return ret;
2426 }
2427
2428 static inline struct btrfs_delayed_ref_node *
2429 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2430 {
2431         struct btrfs_delayed_ref_node *ref;
2432
2433         if (list_empty(&head->ref_list))
2434                 return NULL;
2435
2436         /*
2437          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2438          * This is to prevent a ref count from going down to zero, which deletes
2439          * the extent item from the extent tree, when there still are references
2440          * to add, which would fail because they would not find the extent item.
2441          */
2442         list_for_each_entry(ref, &head->ref_list, list) {
2443                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2444                         return ref;
2445         }
2446
2447         return list_entry(head->ref_list.next, struct btrfs_delayed_ref_node,
2448                           list);
2449 }
2450
2451 /*
2452  * Returns 0 on success or if called with an already aborted transaction.
2453  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2454  */
2455 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2456                                              struct btrfs_root *root,
2457                                              unsigned long nr)
2458 {
2459         struct btrfs_delayed_ref_root *delayed_refs;
2460         struct btrfs_delayed_ref_node *ref;
2461         struct btrfs_delayed_ref_head *locked_ref = NULL;
2462         struct btrfs_delayed_extent_op *extent_op;
2463         struct btrfs_fs_info *fs_info = root->fs_info;
2464         ktime_t start = ktime_get();
2465         int ret;
2466         unsigned long count = 0;
2467         unsigned long actual_count = 0;
2468         int must_insert_reserved = 0;
2469
2470         delayed_refs = &trans->transaction->delayed_refs;
2471         while (1) {
2472                 if (!locked_ref) {
2473                         if (count >= nr)
2474                                 break;
2475
2476                         spin_lock(&delayed_refs->lock);
2477                         locked_ref = btrfs_select_ref_head(trans);
2478                         if (!locked_ref) {
2479                                 spin_unlock(&delayed_refs->lock);
2480                                 break;
2481                         }
2482
2483                         /* grab the lock that says we are going to process
2484                          * all the refs for this head */
2485                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2486                         spin_unlock(&delayed_refs->lock);
2487                         /*
2488                          * we may have dropped the spin lock to get the head
2489                          * mutex lock, and that might have given someone else
2490                          * time to free the head.  If that's true, it has been
2491                          * removed from our list and we can move on.
2492                          */
2493                         if (ret == -EAGAIN) {
2494                                 locked_ref = NULL;
2495                                 count++;
2496                                 continue;
2497                         }
2498                 }
2499
2500                 spin_lock(&locked_ref->lock);
2501
2502                 /*
2503                  * locked_ref is the head node, so we have to go one
2504                  * node back for any delayed ref updates
2505                  */
2506                 ref = select_delayed_ref(locked_ref);
2507
2508                 if (ref && ref->seq &&
2509                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2510                         spin_unlock(&locked_ref->lock);
2511                         btrfs_delayed_ref_unlock(locked_ref);
2512                         spin_lock(&delayed_refs->lock);
2513                         locked_ref->processing = 0;
2514                         delayed_refs->num_heads_ready++;
2515                         spin_unlock(&delayed_refs->lock);
2516                         locked_ref = NULL;
2517                         cond_resched();
2518                         count++;
2519                         continue;
2520                 }
2521
2522                 /*
2523                  * record the must insert reserved flag before we
2524                  * drop the spin lock.
2525                  */
2526                 must_insert_reserved = locked_ref->must_insert_reserved;
2527                 locked_ref->must_insert_reserved = 0;
2528
2529                 extent_op = locked_ref->extent_op;
2530                 locked_ref->extent_op = NULL;
2531
2532                 if (!ref) {
2533
2534
2535                         /* All delayed refs have been processed, Go ahead
2536                          * and send the head node to run_one_delayed_ref,
2537                          * so that any accounting fixes can happen
2538                          */
2539                         ref = &locked_ref->node;
2540
2541                         if (extent_op && must_insert_reserved) {
2542                                 btrfs_free_delayed_extent_op(extent_op);
2543                                 extent_op = NULL;
2544                         }
2545
2546                         if (extent_op) {
2547                                 spin_unlock(&locked_ref->lock);
2548                                 ret = run_delayed_extent_op(trans, root,
2549                                                             ref, extent_op);
2550                                 btrfs_free_delayed_extent_op(extent_op);
2551
2552                                 if (ret) {
2553                                         /*
2554                                          * Need to reset must_insert_reserved if
2555                                          * there was an error so the abort stuff
2556                                          * can cleanup the reserved space
2557                                          * properly.
2558                                          */
2559                                         if (must_insert_reserved)
2560                                                 locked_ref->must_insert_reserved = 1;
2561                                         locked_ref->processing = 0;
2562                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2563                                         btrfs_delayed_ref_unlock(locked_ref);
2564                                         return ret;
2565                                 }
2566                                 continue;
2567                         }
2568
2569                         /*
2570                          * Need to drop our head ref lock and re-aqcuire the
2571                          * delayed ref lock and then re-check to make sure
2572                          * nobody got added.
2573                          */
2574                         spin_unlock(&locked_ref->lock);
2575                         spin_lock(&delayed_refs->lock);
2576                         spin_lock(&locked_ref->lock);
2577                         if (!list_empty(&locked_ref->ref_list) ||
2578                             locked_ref->extent_op) {
2579                                 spin_unlock(&locked_ref->lock);
2580                                 spin_unlock(&delayed_refs->lock);
2581                                 continue;
2582                         }
2583                         ref->in_tree = 0;
2584                         delayed_refs->num_heads--;
2585                         rb_erase(&locked_ref->href_node,
2586                                  &delayed_refs->href_root);
2587                         spin_unlock(&delayed_refs->lock);
2588                 } else {
2589                         actual_count++;
2590                         ref->in_tree = 0;
2591                         list_del(&ref->list);
2592                 }
2593                 atomic_dec(&delayed_refs->num_entries);
2594
2595                 if (!btrfs_delayed_ref_is_head(ref)) {
2596                         /*
2597                          * when we play the delayed ref, also correct the
2598                          * ref_mod on head
2599                          */
2600                         switch (ref->action) {
2601                         case BTRFS_ADD_DELAYED_REF:
2602                         case BTRFS_ADD_DELAYED_EXTENT:
2603                                 locked_ref->node.ref_mod -= ref->ref_mod;
2604                                 break;
2605                         case BTRFS_DROP_DELAYED_REF:
2606                                 locked_ref->node.ref_mod += ref->ref_mod;
2607                                 break;
2608                         default:
2609                                 WARN_ON(1);
2610                         }
2611                 }
2612                 spin_unlock(&locked_ref->lock);
2613
2614                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2615                                           must_insert_reserved);
2616
2617                 btrfs_free_delayed_extent_op(extent_op);
2618                 if (ret) {
2619                         locked_ref->processing = 0;
2620                         btrfs_delayed_ref_unlock(locked_ref);
2621                         btrfs_put_delayed_ref(ref);
2622                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2623                         return ret;
2624                 }
2625
2626                 /*
2627                  * If this node is a head, that means all the refs in this head
2628                  * have been dealt with, and we will pick the next head to deal
2629                  * with, so we must unlock the head and drop it from the cluster
2630                  * list before we release it.
2631                  */
2632                 if (btrfs_delayed_ref_is_head(ref)) {
2633                         if (locked_ref->is_data &&
2634                             locked_ref->total_ref_mod < 0) {
2635                                 spin_lock(&delayed_refs->lock);
2636                                 delayed_refs->pending_csums -= ref->num_bytes;
2637                                 spin_unlock(&delayed_refs->lock);
2638                         }
2639                         btrfs_delayed_ref_unlock(locked_ref);
2640                         locked_ref = NULL;
2641                 }
2642                 btrfs_put_delayed_ref(ref);
2643                 count++;
2644                 cond_resched();
2645         }
2646
2647         /*
2648          * We don't want to include ref heads since we can have empty ref heads
2649          * and those will drastically skew our runtime down since we just do
2650          * accounting, no actual extent tree updates.
2651          */
2652         if (actual_count > 0) {
2653                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2654                 u64 avg;
2655
2656                 /*
2657                  * We weigh the current average higher than our current runtime
2658                  * to avoid large swings in the average.
2659                  */
2660                 spin_lock(&delayed_refs->lock);
2661                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2662                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2663                 spin_unlock(&delayed_refs->lock);
2664         }
2665         return 0;
2666 }
2667
2668 #ifdef SCRAMBLE_DELAYED_REFS
2669 /*
2670  * Normally delayed refs get processed in ascending bytenr order. This
2671  * correlates in most cases to the order added. To expose dependencies on this
2672  * order, we start to process the tree in the middle instead of the beginning
2673  */
2674 static u64 find_middle(struct rb_root *root)
2675 {
2676         struct rb_node *n = root->rb_node;
2677         struct btrfs_delayed_ref_node *entry;
2678         int alt = 1;
2679         u64 middle;
2680         u64 first = 0, last = 0;
2681
2682         n = rb_first(root);
2683         if (n) {
2684                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2685                 first = entry->bytenr;
2686         }
2687         n = rb_last(root);
2688         if (n) {
2689                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2690                 last = entry->bytenr;
2691         }
2692         n = root->rb_node;
2693
2694         while (n) {
2695                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2696                 WARN_ON(!entry->in_tree);
2697
2698                 middle = entry->bytenr;
2699
2700                 if (alt)
2701                         n = n->rb_left;
2702                 else
2703                         n = n->rb_right;
2704
2705                 alt = 1 - alt;
2706         }
2707         return middle;
2708 }
2709 #endif
2710
2711 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2712 {
2713         u64 num_bytes;
2714
2715         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2716                              sizeof(struct btrfs_extent_inline_ref));
2717         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2718                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2719
2720         /*
2721          * We don't ever fill up leaves all the way so multiply by 2 just to be
2722          * closer to what we're really going to want to ouse.
2723          */
2724         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2725 }
2726
2727 /*
2728  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2729  * would require to store the csums for that many bytes.
2730  */
2731 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2732 {
2733         u64 csum_size;
2734         u64 num_csums_per_leaf;
2735         u64 num_csums;
2736
2737         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
2738         num_csums_per_leaf = div64_u64(csum_size,
2739                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2740         num_csums = div64_u64(csum_bytes, root->sectorsize);
2741         num_csums += num_csums_per_leaf - 1;
2742         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2743         return num_csums;
2744 }
2745
2746 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2747                                        struct btrfs_root *root)
2748 {
2749         struct btrfs_block_rsv *global_rsv;
2750         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2751         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2752         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2753         u64 num_bytes, num_dirty_bgs_bytes;
2754         int ret = 0;
2755
2756         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2757         num_heads = heads_to_leaves(root, num_heads);
2758         if (num_heads > 1)
2759                 num_bytes += (num_heads - 1) * root->nodesize;
2760         num_bytes <<= 1;
2761         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2762         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2763                                                              num_dirty_bgs);
2764         global_rsv = &root->fs_info->global_block_rsv;
2765
2766         /*
2767          * If we can't allocate any more chunks lets make sure we have _lots_ of
2768          * wiggle room since running delayed refs can create more delayed refs.
2769          */
2770         if (global_rsv->space_info->full) {
2771                 num_dirty_bgs_bytes <<= 1;
2772                 num_bytes <<= 1;
2773         }
2774
2775         spin_lock(&global_rsv->lock);
2776         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2777                 ret = 1;
2778         spin_unlock(&global_rsv->lock);
2779         return ret;
2780 }
2781
2782 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2783                                        struct btrfs_root *root)
2784 {
2785         struct btrfs_fs_info *fs_info = root->fs_info;
2786         u64 num_entries =
2787                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2788         u64 avg_runtime;
2789         u64 val;
2790
2791         smp_mb();
2792         avg_runtime = fs_info->avg_delayed_ref_runtime;
2793         val = num_entries * avg_runtime;
2794         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2795                 return 1;
2796         if (val >= NSEC_PER_SEC / 2)
2797                 return 2;
2798
2799         return btrfs_check_space_for_delayed_refs(trans, root);
2800 }
2801
2802 struct async_delayed_refs {
2803         struct btrfs_root *root;
2804         int count;
2805         int error;
2806         int sync;
2807         struct completion wait;
2808         struct btrfs_work work;
2809 };
2810
2811 static void delayed_ref_async_start(struct btrfs_work *work)
2812 {
2813         struct async_delayed_refs *async;
2814         struct btrfs_trans_handle *trans;
2815         int ret;
2816
2817         async = container_of(work, struct async_delayed_refs, work);
2818
2819         trans = btrfs_join_transaction(async->root);
2820         if (IS_ERR(trans)) {
2821                 async->error = PTR_ERR(trans);
2822                 goto done;
2823         }
2824
2825         /*
2826          * trans->sync means that when we call end_transaciton, we won't
2827          * wait on delayed refs
2828          */
2829         trans->sync = true;
2830         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2831         if (ret)
2832                 async->error = ret;
2833
2834         ret = btrfs_end_transaction(trans, async->root);
2835         if (ret && !async->error)
2836                 async->error = ret;
2837 done:
2838         if (async->sync)
2839                 complete(&async->wait);
2840         else
2841                 kfree(async);
2842 }
2843
2844 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2845                                  unsigned long count, int wait)
2846 {
2847         struct async_delayed_refs *async;
2848         int ret;
2849
2850         async = kmalloc(sizeof(*async), GFP_NOFS);
2851         if (!async)
2852                 return -ENOMEM;
2853
2854         async->root = root->fs_info->tree_root;
2855         async->count = count;
2856         async->error = 0;
2857         if (wait)
2858                 async->sync = 1;
2859         else
2860                 async->sync = 0;
2861         init_completion(&async->wait);
2862
2863         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2864                         delayed_ref_async_start, NULL, NULL);
2865
2866         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2867
2868         if (wait) {
2869                 wait_for_completion(&async->wait);
2870                 ret = async->error;
2871                 kfree(async);
2872                 return ret;
2873         }
2874         return 0;
2875 }
2876
2877 /*
2878  * this starts processing the delayed reference count updates and
2879  * extent insertions we have queued up so far.  count can be
2880  * 0, which means to process everything in the tree at the start
2881  * of the run (but not newly added entries), or it can be some target
2882  * number you'd like to process.
2883  *
2884  * Returns 0 on success or if called with an aborted transaction
2885  * Returns <0 on error and aborts the transaction
2886  */
2887 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2888                            struct btrfs_root *root, unsigned long count)
2889 {
2890         struct rb_node *node;
2891         struct btrfs_delayed_ref_root *delayed_refs;
2892         struct btrfs_delayed_ref_head *head;
2893         int ret;
2894         int run_all = count == (unsigned long)-1;
2895         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2896
2897         /* We'll clean this up in btrfs_cleanup_transaction */
2898         if (trans->aborted)
2899                 return 0;
2900
2901         if (root == root->fs_info->extent_root)
2902                 root = root->fs_info->tree_root;
2903
2904         delayed_refs = &trans->transaction->delayed_refs;
2905         if (count == 0)
2906                 count = atomic_read(&delayed_refs->num_entries) * 2;
2907
2908 again:
2909 #ifdef SCRAMBLE_DELAYED_REFS
2910         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2911 #endif
2912         trans->can_flush_pending_bgs = false;
2913         ret = __btrfs_run_delayed_refs(trans, root, count);
2914         if (ret < 0) {
2915                 btrfs_abort_transaction(trans, root, ret);
2916                 return ret;
2917         }
2918
2919         if (run_all) {
2920                 if (!list_empty(&trans->new_bgs))
2921                         btrfs_create_pending_block_groups(trans, root);
2922
2923                 spin_lock(&delayed_refs->lock);
2924                 node = rb_first(&delayed_refs->href_root);
2925                 if (!node) {
2926                         spin_unlock(&delayed_refs->lock);
2927                         goto out;
2928                 }
2929                 count = (unsigned long)-1;
2930
2931                 while (node) {
2932                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2933                                         href_node);
2934                         if (btrfs_delayed_ref_is_head(&head->node)) {
2935                                 struct btrfs_delayed_ref_node *ref;
2936
2937                                 ref = &head->node;
2938                                 atomic_inc(&ref->refs);
2939
2940                                 spin_unlock(&delayed_refs->lock);
2941                                 /*
2942                                  * Mutex was contended, block until it's
2943                                  * released and try again
2944                                  */
2945                                 mutex_lock(&head->mutex);
2946                                 mutex_unlock(&head->mutex);
2947
2948                                 btrfs_put_delayed_ref(ref);
2949                                 cond_resched();
2950                                 goto again;
2951                         } else {
2952                                 WARN_ON(1);
2953                         }
2954                         node = rb_next(node);
2955                 }
2956                 spin_unlock(&delayed_refs->lock);
2957                 cond_resched();
2958                 goto again;
2959         }
2960 out:
2961         assert_qgroups_uptodate(trans);
2962         trans->can_flush_pending_bgs = can_flush_pending_bgs;
2963         return 0;
2964 }
2965
2966 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2967                                 struct btrfs_root *root,
2968                                 u64 bytenr, u64 num_bytes, u64 flags,
2969                                 int level, int is_data)
2970 {
2971         struct btrfs_delayed_extent_op *extent_op;
2972         int ret;
2973
2974         extent_op = btrfs_alloc_delayed_extent_op();
2975         if (!extent_op)
2976                 return -ENOMEM;
2977
2978         extent_op->flags_to_set = flags;
2979         extent_op->update_flags = 1;
2980         extent_op->update_key = 0;
2981         extent_op->is_data = is_data ? 1 : 0;
2982         extent_op->level = level;
2983
2984         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2985                                           num_bytes, extent_op);
2986         if (ret)
2987                 btrfs_free_delayed_extent_op(extent_op);
2988         return ret;
2989 }
2990
2991 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2992                                       struct btrfs_root *root,
2993                                       struct btrfs_path *path,
2994                                       u64 objectid, u64 offset, u64 bytenr)
2995 {
2996         struct btrfs_delayed_ref_head *head;
2997         struct btrfs_delayed_ref_node *ref;
2998         struct btrfs_delayed_data_ref *data_ref;
2999         struct btrfs_delayed_ref_root *delayed_refs;
3000         int ret = 0;
3001
3002         delayed_refs = &trans->transaction->delayed_refs;
3003         spin_lock(&delayed_refs->lock);
3004         head = btrfs_find_delayed_ref_head(trans, bytenr);
3005         if (!head) {
3006                 spin_unlock(&delayed_refs->lock);
3007                 return 0;
3008         }
3009
3010         if (!mutex_trylock(&head->mutex)) {
3011                 atomic_inc(&head->node.refs);
3012                 spin_unlock(&delayed_refs->lock);
3013
3014                 btrfs_release_path(path);
3015
3016                 /*
3017                  * Mutex was contended, block until it's released and let
3018                  * caller try again
3019                  */
3020                 mutex_lock(&head->mutex);
3021                 mutex_unlock(&head->mutex);
3022                 btrfs_put_delayed_ref(&head->node);
3023                 return -EAGAIN;
3024         }
3025         spin_unlock(&delayed_refs->lock);
3026
3027         spin_lock(&head->lock);
3028         list_for_each_entry(ref, &head->ref_list, list) {
3029                 /* If it's a shared ref we know a cross reference exists */
3030                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3031                         ret = 1;
3032                         break;
3033                 }
3034
3035                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3036
3037                 /*
3038                  * If our ref doesn't match the one we're currently looking at
3039                  * then we have a cross reference.
3040                  */
3041                 if (data_ref->root != root->root_key.objectid ||
3042                     data_ref->objectid != objectid ||
3043                     data_ref->offset != offset) {
3044                         ret = 1;
3045                         break;
3046                 }
3047         }
3048         spin_unlock(&head->lock);
3049         mutex_unlock(&head->mutex);
3050         return ret;
3051 }
3052
3053 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
3054                                         struct btrfs_root *root,
3055                                         struct btrfs_path *path,
3056                                         u64 objectid, u64 offset, u64 bytenr)
3057 {
3058         struct btrfs_root *extent_root = root->fs_info->extent_root;
3059         struct extent_buffer *leaf;
3060         struct btrfs_extent_data_ref *ref;
3061         struct btrfs_extent_inline_ref *iref;
3062         struct btrfs_extent_item *ei;
3063         struct btrfs_key key;
3064         u32 item_size;
3065         int ret;
3066
3067         key.objectid = bytenr;
3068         key.offset = (u64)-1;
3069         key.type = BTRFS_EXTENT_ITEM_KEY;
3070
3071         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3072         if (ret < 0)
3073                 goto out;
3074         BUG_ON(ret == 0); /* Corruption */
3075
3076         ret = -ENOENT;
3077         if (path->slots[0] == 0)
3078                 goto out;
3079
3080         path->slots[0]--;
3081         leaf = path->nodes[0];
3082         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3083
3084         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3085                 goto out;
3086
3087         ret = 1;
3088         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3089 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3090         if (item_size < sizeof(*ei)) {
3091                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3092                 goto out;
3093         }
3094 #endif
3095         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3096
3097         if (item_size != sizeof(*ei) +
3098             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3099                 goto out;
3100
3101         if (btrfs_extent_generation(leaf, ei) <=
3102             btrfs_root_last_snapshot(&root->root_item))
3103                 goto out;
3104
3105         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3106         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3107             BTRFS_EXTENT_DATA_REF_KEY)
3108                 goto out;
3109
3110         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3111         if (btrfs_extent_refs(leaf, ei) !=
3112             btrfs_extent_data_ref_count(leaf, ref) ||
3113             btrfs_extent_data_ref_root(leaf, ref) !=
3114             root->root_key.objectid ||
3115             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3116             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3117                 goto out;
3118
3119         ret = 0;
3120 out:
3121         return ret;
3122 }
3123
3124 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3125                           struct btrfs_root *root,
3126                           u64 objectid, u64 offset, u64 bytenr)
3127 {
3128         struct btrfs_path *path;
3129         int ret;
3130         int ret2;
3131
3132         path = btrfs_alloc_path();
3133         if (!path)
3134                 return -ENOENT;
3135
3136         do {
3137                 ret = check_committed_ref(trans, root, path, objectid,
3138                                           offset, bytenr);
3139                 if (ret && ret != -ENOENT)
3140                         goto out;
3141
3142                 ret2 = check_delayed_ref(trans, root, path, objectid,
3143                                          offset, bytenr);
3144         } while (ret2 == -EAGAIN);
3145
3146         if (ret2 && ret2 != -ENOENT) {
3147                 ret = ret2;
3148                 goto out;
3149         }
3150
3151         if (ret != -ENOENT || ret2 != -ENOENT)
3152                 ret = 0;
3153 out:
3154         btrfs_free_path(path);
3155         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3156                 WARN_ON(ret > 0);
3157         return ret;
3158 }
3159
3160 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3161                            struct btrfs_root *root,
3162                            struct extent_buffer *buf,
3163                            int full_backref, int inc)
3164 {
3165         u64 bytenr;
3166         u64 num_bytes;
3167         u64 parent;
3168         u64 ref_root;
3169         u32 nritems;
3170         struct btrfs_key key;
3171         struct btrfs_file_extent_item *fi;
3172         int i;
3173         int level;
3174         int ret = 0;
3175         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3176                             u64, u64, u64, u64, u64, u64, int);
3177
3178
3179         if (btrfs_test_is_dummy_root(root))
3180                 return 0;
3181
3182         ref_root = btrfs_header_owner(buf);
3183         nritems = btrfs_header_nritems(buf);
3184         level = btrfs_header_level(buf);
3185
3186         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3187                 return 0;
3188
3189         if (inc)
3190                 process_func = btrfs_inc_extent_ref;
3191         else
3192                 process_func = btrfs_free_extent;
3193
3194         if (full_backref)
3195                 parent = buf->start;
3196         else
3197                 parent = 0;
3198
3199         for (i = 0; i < nritems; i++) {
3200                 if (level == 0) {
3201                         btrfs_item_key_to_cpu(buf, &key, i);
3202                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3203                                 continue;
3204                         fi = btrfs_item_ptr(buf, i,
3205                                             struct btrfs_file_extent_item);
3206                         if (btrfs_file_extent_type(buf, fi) ==
3207                             BTRFS_FILE_EXTENT_INLINE)
3208                                 continue;
3209                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3210                         if (bytenr == 0)
3211                                 continue;
3212
3213                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3214                         key.offset -= btrfs_file_extent_offset(buf, fi);
3215                         ret = process_func(trans, root, bytenr, num_bytes,
3216                                            parent, ref_root, key.objectid,
3217                                            key.offset, 1);
3218                         if (ret)
3219                                 goto fail;
3220                 } else {
3221                         bytenr = btrfs_node_blockptr(buf, i);
3222                         num_bytes = root->nodesize;
3223                         ret = process_func(trans, root, bytenr, num_bytes,
3224                                            parent, ref_root, level - 1, 0,
3225                                            1);
3226                         if (ret)
3227                                 goto fail;
3228                 }
3229         }
3230         return 0;
3231 fail:
3232         return ret;
3233 }
3234
3235 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3236                   struct extent_buffer *buf, int full_backref)
3237 {
3238         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3239 }
3240
3241 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3242                   struct extent_buffer *buf, int full_backref)
3243 {
3244         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3245 }
3246
3247 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3248                                  struct btrfs_root *root,
3249                                  struct btrfs_path *path,
3250                                  struct btrfs_block_group_cache *cache)
3251 {
3252         int ret;
3253         struct btrfs_root *extent_root = root->fs_info->extent_root;
3254         unsigned long bi;
3255         struct extent_buffer *leaf;
3256
3257         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3258         if (ret) {
3259                 if (ret > 0)
3260                         ret = -ENOENT;
3261                 goto fail;
3262         }
3263
3264         leaf = path->nodes[0];
3265         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3266         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3267         btrfs_mark_buffer_dirty(leaf);
3268 fail:
3269         btrfs_release_path(path);
3270         return ret;
3271
3272 }
3273
3274 static struct btrfs_block_group_cache *
3275 next_block_group(struct btrfs_root *root,
3276                  struct btrfs_block_group_cache *cache)
3277 {
3278         struct rb_node *node;
3279
3280         spin_lock(&root->fs_info->block_group_cache_lock);
3281
3282         /* If our block group was removed, we need a full search. */
3283         if (RB_EMPTY_NODE(&cache->cache_node)) {
3284                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3285
3286                 spin_unlock(&root->fs_info->block_group_cache_lock);
3287                 btrfs_put_block_group(cache);
3288                 cache = btrfs_lookup_first_block_group(root->fs_info,
3289                                                        next_bytenr);
3290                 return cache;
3291         }
3292         node = rb_next(&cache->cache_node);
3293         btrfs_put_block_group(cache);
3294         if (node) {
3295                 cache = rb_entry(node, struct btrfs_block_group_cache,
3296                                  cache_node);
3297                 btrfs_get_block_group(cache);
3298         } else
3299                 cache = NULL;
3300         spin_unlock(&root->fs_info->block_group_cache_lock);
3301         return cache;
3302 }
3303
3304 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3305                             struct btrfs_trans_handle *trans,
3306                             struct btrfs_path *path)
3307 {
3308         struct btrfs_root *root = block_group->fs_info->tree_root;
3309         struct inode *inode = NULL;
3310         u64 alloc_hint = 0;
3311         int dcs = BTRFS_DC_ERROR;
3312         u64 num_pages = 0;
3313         int retries = 0;
3314         int ret = 0;
3315
3316         /*
3317          * If this block group is smaller than 100 megs don't bother caching the
3318          * block group.
3319          */
3320         if (block_group->key.offset < (100 * 1024 * 1024)) {
3321                 spin_lock(&block_group->lock);
3322                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3323                 spin_unlock(&block_group->lock);
3324                 return 0;
3325         }
3326
3327         if (trans->aborted)
3328                 return 0;
3329 again:
3330         inode = lookup_free_space_inode(root, block_group, path);
3331         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3332                 ret = PTR_ERR(inode);
3333                 btrfs_release_path(path);
3334                 goto out;
3335         }
3336
3337         if (IS_ERR(inode)) {
3338                 BUG_ON(retries);
3339                 retries++;
3340
3341                 if (block_group->ro)
3342                         goto out_free;
3343
3344                 ret = create_free_space_inode(root, trans, block_group, path);
3345                 if (ret)
3346                         goto out_free;
3347                 goto again;
3348         }
3349
3350         /* We've already setup this transaction, go ahead and exit */
3351         if (block_group->cache_generation == trans->transid &&
3352             i_size_read(inode)) {
3353                 dcs = BTRFS_DC_SETUP;
3354                 goto out_put;
3355         }
3356
3357         /*
3358          * We want to set the generation to 0, that way if anything goes wrong
3359          * from here on out we know not to trust this cache when we load up next
3360          * time.
3361          */
3362         BTRFS_I(inode)->generation = 0;
3363         ret = btrfs_update_inode(trans, root, inode);
3364         if (ret) {
3365                 /*
3366                  * So theoretically we could recover from this, simply set the
3367                  * super cache generation to 0 so we know to invalidate the
3368                  * cache, but then we'd have to keep track of the block groups
3369                  * that fail this way so we know we _have_ to reset this cache
3370                  * before the next commit or risk reading stale cache.  So to
3371                  * limit our exposure to horrible edge cases lets just abort the
3372                  * transaction, this only happens in really bad situations
3373                  * anyway.
3374                  */
3375                 btrfs_abort_transaction(trans, root, ret);
3376                 goto out_put;
3377         }
3378         WARN_ON(ret);
3379
3380         if (i_size_read(inode) > 0) {
3381                 ret = btrfs_check_trunc_cache_free_space(root,
3382                                         &root->fs_info->global_block_rsv);
3383                 if (ret)
3384                         goto out_put;
3385
3386                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3387                 if (ret)
3388                         goto out_put;
3389         }
3390
3391         spin_lock(&block_group->lock);
3392         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3393             !btrfs_test_opt(root, SPACE_CACHE)) {
3394                 /*
3395                  * don't bother trying to write stuff out _if_
3396                  * a) we're not cached,
3397                  * b) we're with nospace_cache mount option.
3398                  */
3399                 dcs = BTRFS_DC_WRITTEN;
3400                 spin_unlock(&block_group->lock);
3401                 goto out_put;
3402         }
3403         spin_unlock(&block_group->lock);
3404
3405         /*
3406          * Try to preallocate enough space based on how big the block group is.
3407          * Keep in mind this has to include any pinned space which could end up
3408          * taking up quite a bit since it's not folded into the other space
3409          * cache.
3410          */
3411         num_pages = div_u64(block_group->key.offset, 256 * 1024 * 1024);
3412         if (!num_pages)
3413                 num_pages = 1;
3414
3415         num_pages *= 16;
3416         num_pages *= PAGE_CACHE_SIZE;
3417
3418         ret = btrfs_check_data_free_space(inode, num_pages, num_pages);
3419         if (ret)
3420                 goto out_put;
3421
3422         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3423                                               num_pages, num_pages,
3424                                               &alloc_hint);
3425         if (!ret)
3426                 dcs = BTRFS_DC_SETUP;
3427         btrfs_free_reserved_data_space(inode, num_pages);
3428
3429 out_put:
3430         iput(inode);
3431 out_free:
3432         btrfs_release_path(path);
3433 out:
3434         spin_lock(&block_group->lock);
3435         if (!ret && dcs == BTRFS_DC_SETUP)
3436                 block_group->cache_generation = trans->transid;
3437         block_group->disk_cache_state = dcs;
3438         spin_unlock(&block_group->lock);
3439
3440         return ret;
3441 }
3442
3443 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3444                             struct btrfs_root *root)
3445 {
3446         struct btrfs_block_group_cache *cache, *tmp;
3447         struct btrfs_transaction *cur_trans = trans->transaction;
3448         struct btrfs_path *path;
3449
3450         if (list_empty(&cur_trans->dirty_bgs) ||
3451             !btrfs_test_opt(root, SPACE_CACHE))
3452                 return 0;
3453
3454         path = btrfs_alloc_path();
3455         if (!path)
3456                 return -ENOMEM;
3457
3458         /* Could add new block groups, use _safe just in case */
3459         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3460                                  dirty_list) {
3461                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3462                         cache_save_setup(cache, trans, path);
3463         }
3464
3465         btrfs_free_path(path);
3466         return 0;
3467 }
3468
3469 /*
3470  * transaction commit does final block group cache writeback during a
3471  * critical section where nothing is allowed to change the FS.  This is
3472  * required in order for the cache to actually match the block group,
3473  * but can introduce a lot of latency into the commit.
3474  *
3475  * So, btrfs_start_dirty_block_groups is here to kick off block group
3476  * cache IO.  There's a chance we'll have to redo some of it if the
3477  * block group changes again during the commit, but it greatly reduces
3478  * the commit latency by getting rid of the easy block groups while
3479  * we're still allowing others to join the commit.
3480  */
3481 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3482                                    struct btrfs_root *root)
3483 {
3484         struct btrfs_block_group_cache *cache;
3485         struct btrfs_transaction *cur_trans = trans->transaction;
3486         int ret = 0;
3487         int should_put;
3488         struct btrfs_path *path = NULL;
3489         LIST_HEAD(dirty);
3490         struct list_head *io = &cur_trans->io_bgs;
3491         int num_started = 0;
3492         int loops = 0;
3493
3494         spin_lock(&cur_trans->dirty_bgs_lock);
3495         if (list_empty(&cur_trans->dirty_bgs)) {
3496                 spin_unlock(&cur_trans->dirty_bgs_lock);
3497                 return 0;
3498         }
3499         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3500         spin_unlock(&cur_trans->dirty_bgs_lock);
3501
3502 again:
3503         /*
3504          * make sure all the block groups on our dirty list actually
3505          * exist
3506          */
3507         btrfs_create_pending_block_groups(trans, root);
3508
3509         if (!path) {
3510                 path = btrfs_alloc_path();
3511                 if (!path)
3512                         return -ENOMEM;
3513         }
3514
3515         /*
3516          * cache_write_mutex is here only to save us from balance or automatic
3517          * removal of empty block groups deleting this block group while we are
3518          * writing out the cache
3519          */
3520         mutex_lock(&trans->transaction->cache_write_mutex);
3521         while (!list_empty(&dirty)) {
3522                 cache = list_first_entry(&dirty,
3523                                          struct btrfs_block_group_cache,
3524                                          dirty_list);
3525                 /*
3526                  * this can happen if something re-dirties a block
3527                  * group that is already under IO.  Just wait for it to
3528                  * finish and then do it all again
3529                  */
3530                 if (!list_empty(&cache->io_list)) {
3531                         list_del_init(&cache->io_list);
3532                         btrfs_wait_cache_io(root, trans, cache,
3533                                             &cache->io_ctl, path,
3534                                             cache->key.objectid);
3535                         btrfs_put_block_group(cache);
3536                 }
3537
3538
3539                 /*
3540                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3541                  * if it should update the cache_state.  Don't delete
3542                  * until after we wait.
3543                  *
3544                  * Since we're not running in the commit critical section
3545                  * we need the dirty_bgs_lock to protect from update_block_group
3546                  */
3547                 spin_lock(&cur_trans->dirty_bgs_lock);
3548                 list_del_init(&cache->dirty_list);
3549                 spin_unlock(&cur_trans->dirty_bgs_lock);
3550
3551                 should_put = 1;
3552
3553                 cache_save_setup(cache, trans, path);
3554
3555                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3556                         cache->io_ctl.inode = NULL;
3557                         ret = btrfs_write_out_cache(root, trans, cache, path);
3558                         if (ret == 0 && cache->io_ctl.inode) {
3559                                 num_started++;
3560                                 should_put = 0;
3561
3562                                 /*
3563                                  * the cache_write_mutex is protecting
3564                                  * the io_list
3565                                  */
3566                                 list_add_tail(&cache->io_list, io);
3567                         } else {
3568                                 /*
3569                                  * if we failed to write the cache, the
3570                                  * generation will be bad and life goes on
3571                                  */
3572                                 ret = 0;
3573                         }
3574                 }
3575                 if (!ret) {
3576                         ret = write_one_cache_group(trans, root, path, cache);
3577                         /*
3578                          * Our block group might still be attached to the list
3579                          * of new block groups in the transaction handle of some
3580                          * other task (struct btrfs_trans_handle->new_bgs). This
3581                          * means its block group item isn't yet in the extent
3582                          * tree. If this happens ignore the error, as we will
3583                          * try again later in the critical section of the
3584                          * transaction commit.
3585                          */
3586                         if (ret == -ENOENT) {
3587                                 ret = 0;
3588                                 spin_lock(&cur_trans->dirty_bgs_lock);
3589                                 if (list_empty(&cache->dirty_list)) {
3590                                         list_add_tail(&cache->dirty_list,
3591                                                       &cur_trans->dirty_bgs);
3592                                         btrfs_get_block_group(cache);
3593                                 }
3594                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3595                         } else if (ret) {
3596                                 btrfs_abort_transaction(trans, root, ret);
3597                         }
3598                 }
3599
3600                 /* if its not on the io list, we need to put the block group */
3601                 if (should_put)
3602                         btrfs_put_block_group(cache);
3603
3604                 if (ret)
3605                         break;
3606
3607                 /*
3608                  * Avoid blocking other tasks for too long. It might even save
3609                  * us from writing caches for block groups that are going to be
3610                  * removed.
3611                  */
3612                 mutex_unlock(&trans->transaction->cache_write_mutex);
3613                 mutex_lock(&trans->transaction->cache_write_mutex);
3614         }
3615         mutex_unlock(&trans->transaction->cache_write_mutex);
3616
3617         /*
3618          * go through delayed refs for all the stuff we've just kicked off
3619          * and then loop back (just once)
3620          */
3621         ret = btrfs_run_delayed_refs(trans, root, 0);
3622         if (!ret && loops == 0) {
3623                 loops++;
3624                 spin_lock(&cur_trans->dirty_bgs_lock);
3625                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3626                 /*
3627                  * dirty_bgs_lock protects us from concurrent block group
3628                  * deletes too (not just cache_write_mutex).
3629                  */
3630                 if (!list_empty(&dirty)) {
3631                         spin_unlock(&cur_trans->dirty_bgs_lock);
3632                         goto again;
3633                 }
3634                 spin_unlock(&cur_trans->dirty_bgs_lock);
3635         }
3636
3637         btrfs_free_path(path);
3638         return ret;
3639 }
3640
3641 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3642                                    struct btrfs_root *root)
3643 {
3644         struct btrfs_block_group_cache *cache;
3645         struct btrfs_transaction *cur_trans = trans->transaction;
3646         int ret = 0;
3647         int should_put;
3648         struct btrfs_path *path;
3649         struct list_head *io = &cur_trans->io_bgs;
3650         int num_started = 0;
3651
3652         path = btrfs_alloc_path();
3653         if (!path)
3654                 return -ENOMEM;
3655
3656         /*
3657          * We don't need the lock here since we are protected by the transaction
3658          * commit.  We want to do the cache_save_setup first and then run the
3659          * delayed refs to make sure we have the best chance at doing this all
3660          * in one shot.
3661          */
3662         while (!list_empty(&cur_trans->dirty_bgs)) {
3663                 cache = list_first_entry(&cur_trans->dirty_bgs,
3664                                          struct btrfs_block_group_cache,
3665                                          dirty_list);
3666
3667                 /*
3668                  * this can happen if cache_save_setup re-dirties a block
3669                  * group that is already under IO.  Just wait for it to
3670                  * finish and then do it all again
3671                  */
3672                 if (!list_empty(&cache->io_list)) {
3673                         list_del_init(&cache->io_list);
3674                         btrfs_wait_cache_io(root, trans, cache,
3675                                             &cache->io_ctl, path,
3676                                             cache->key.objectid);
3677                         btrfs_put_block_group(cache);
3678                 }
3679
3680                 /*
3681                  * don't remove from the dirty list until after we've waited
3682                  * on any pending IO
3683                  */
3684                 list_del_init(&cache->dirty_list);
3685                 should_put = 1;
3686
3687                 cache_save_setup(cache, trans, path);
3688
3689                 if (!ret)
3690                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3691
3692                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3693                         cache->io_ctl.inode = NULL;
3694                         ret = btrfs_write_out_cache(root, trans, cache, path);
3695                         if (ret == 0 && cache->io_ctl.inode) {
3696                                 num_started++;
3697                                 should_put = 0;
3698                                 list_add_tail(&cache->io_list, io);
3699                         } else {
3700                                 /*
3701                                  * if we failed to write the cache, the
3702                                  * generation will be bad and life goes on
3703                                  */
3704                                 ret = 0;
3705                         }
3706                 }
3707                 if (!ret) {
3708                         ret = write_one_cache_group(trans, root, path, cache);
3709                         if (ret)
3710                                 btrfs_abort_transaction(trans, root, ret);
3711                 }
3712
3713                 /* if its not on the io list, we need to put the block group */
3714                 if (should_put)
3715                         btrfs_put_block_group(cache);
3716         }
3717
3718         while (!list_empty(io)) {
3719                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3720                                          io_list);
3721                 list_del_init(&cache->io_list);
3722                 btrfs_wait_cache_io(root, trans, cache,
3723                                     &cache->io_ctl, path, cache->key.objectid);
3724                 btrfs_put_block_group(cache);
3725         }
3726
3727         btrfs_free_path(path);
3728         return ret;
3729 }
3730
3731 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3732 {
3733         struct btrfs_block_group_cache *block_group;
3734         int readonly = 0;
3735
3736         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3737         if (!block_group || block_group->ro)
3738                 readonly = 1;
3739         if (block_group)
3740                 btrfs_put_block_group(block_group);
3741         return readonly;
3742 }
3743
3744 static const char *alloc_name(u64 flags)
3745 {
3746         switch (flags) {
3747         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3748                 return "mixed";
3749         case BTRFS_BLOCK_GROUP_METADATA:
3750                 return "metadata";
3751         case BTRFS_BLOCK_GROUP_DATA:
3752                 return "data";
3753         case BTRFS_BLOCK_GROUP_SYSTEM:
3754                 return "system";
3755         default:
3756                 WARN_ON(1);
3757                 return "invalid-combination";
3758         };
3759 }
3760
3761 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3762                              u64 total_bytes, u64 bytes_used,
3763                              struct btrfs_space_info **space_info)
3764 {
3765         struct btrfs_space_info *found;
3766         int i;
3767         int factor;
3768         int ret;
3769
3770         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3771                      BTRFS_BLOCK_GROUP_RAID10))
3772                 factor = 2;
3773         else
3774                 factor = 1;
3775
3776         found = __find_space_info(info, flags);
3777         if (found) {
3778                 spin_lock(&found->lock);
3779                 found->total_bytes += total_bytes;
3780                 found->disk_total += total_bytes * factor;
3781                 found->bytes_used += bytes_used;
3782                 found->disk_used += bytes_used * factor;
3783                 if (total_bytes > 0)
3784                         found->full = 0;
3785                 spin_unlock(&found->lock);
3786                 *space_info = found;
3787                 return 0;
3788         }
3789         found = kzalloc(sizeof(*found), GFP_NOFS);
3790         if (!found)
3791                 return -ENOMEM;
3792
3793         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3794         if (ret) {
3795                 kfree(found);
3796                 return ret;
3797         }
3798
3799         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3800                 INIT_LIST_HEAD(&found->block_groups[i]);
3801         init_rwsem(&found->groups_sem);
3802         spin_lock_init(&found->lock);
3803         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3804         found->total_bytes = total_bytes;
3805         found->disk_total = total_bytes * factor;
3806         found->bytes_used = bytes_used;
3807         found->disk_used = bytes_used * factor;
3808         found->bytes_pinned = 0;
3809         found->bytes_reserved = 0;
3810         found->bytes_readonly = 0;
3811         found->bytes_may_use = 0;
3812         found->full = 0;
3813         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3814         found->chunk_alloc = 0;
3815         found->flush = 0;
3816         init_waitqueue_head(&found->wait);
3817         INIT_LIST_HEAD(&found->ro_bgs);
3818
3819         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3820                                     info->space_info_kobj, "%s",
3821                                     alloc_name(found->flags));
3822         if (ret) {
3823                 kfree(found);
3824                 return ret;
3825         }
3826
3827         *space_info = found;
3828         list_add_rcu(&found->list, &info->space_info);
3829         if (flags & BTRFS_BLOCK_GROUP_DATA)
3830                 info->data_sinfo = found;
3831
3832         return ret;
3833 }
3834
3835 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3836 {
3837         u64 extra_flags = chunk_to_extended(flags) &
3838                                 BTRFS_EXTENDED_PROFILE_MASK;
3839
3840         write_seqlock(&fs_info->profiles_lock);
3841         if (flags & BTRFS_BLOCK_GROUP_DATA)
3842                 fs_info->avail_data_alloc_bits |= extra_flags;
3843         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3844                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3845         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3846                 fs_info->avail_system_alloc_bits |= extra_flags;
3847         write_sequnlock(&fs_info->profiles_lock);
3848 }
3849
3850 /*
3851  * returns target flags in extended format or 0 if restripe for this
3852  * chunk_type is not in progress
3853  *
3854  * should be called with either volume_mutex or balance_lock held
3855  */
3856 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3857 {
3858         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3859         u64 target = 0;
3860
3861         if (!bctl)
3862                 return 0;
3863
3864         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3865             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3866                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3867         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3868                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3869                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3870         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3871                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3872                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3873         }
3874
3875         return target;
3876 }
3877
3878 /*
3879  * @flags: available profiles in extended format (see ctree.h)
3880  *
3881  * Returns reduced profile in chunk format.  If profile changing is in
3882  * progress (either running or paused) picks the target profile (if it's
3883  * already available), otherwise falls back to plain reducing.
3884  */
3885 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3886 {
3887         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3888         u64 target;
3889         u64 raid_type;
3890         u64 allowed = 0;
3891
3892         /*
3893          * see if restripe for this chunk_type is in progress, if so
3894          * try to reduce to the target profile
3895          */
3896         spin_lock(&root->fs_info->balance_lock);
3897         target = get_restripe_target(root->fs_info, flags);
3898         if (target) {
3899                 /* pick target profile only if it's already available */
3900                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3901                         spin_unlock(&root->fs_info->balance_lock);
3902                         return extended_to_chunk(target);
3903                 }
3904         }
3905         spin_unlock(&root->fs_info->balance_lock);
3906
3907         /* First, mask out the RAID levels which aren't possible */
3908         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
3909                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
3910                         allowed |= btrfs_raid_group[raid_type];
3911         }
3912         allowed &= flags;
3913
3914         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
3915                 allowed = BTRFS_BLOCK_GROUP_RAID6;
3916         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
3917                 allowed = BTRFS_BLOCK_GROUP_RAID5;
3918         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
3919                 allowed = BTRFS_BLOCK_GROUP_RAID10;
3920         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
3921                 allowed = BTRFS_BLOCK_GROUP_RAID1;
3922         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
3923                 allowed = BTRFS_BLOCK_GROUP_RAID0;
3924
3925         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
3926
3927         return extended_to_chunk(flags | allowed);
3928 }
3929
3930 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3931 {
3932         unsigned seq;
3933         u64 flags;
3934
3935         do {
3936                 flags = orig_flags;
3937                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3938
3939                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3940                         flags |= root->fs_info->avail_data_alloc_bits;
3941                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3942                         flags |= root->fs_info->avail_system_alloc_bits;
3943                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3944                         flags |= root->fs_info->avail_metadata_alloc_bits;
3945         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3946
3947         return btrfs_reduce_alloc_profile(root, flags);
3948 }
3949
3950 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3951 {
3952         u64 flags;
3953         u64 ret;
3954
3955         if (data)
3956                 flags = BTRFS_BLOCK_GROUP_DATA;
3957         else if (root == root->fs_info->chunk_root)
3958                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3959         else
3960                 flags = BTRFS_BLOCK_GROUP_METADATA;
3961
3962         ret = get_alloc_profile(root, flags);
3963         return ret;
3964 }
3965
3966 /*
3967  * This will check the space that the inode allocates from to make sure we have
3968  * enough space for bytes.
3969  */
3970 int btrfs_check_data_free_space(struct inode *inode, u64 bytes, u64 write_bytes)
3971 {
3972         struct btrfs_space_info *data_sinfo;
3973         struct btrfs_root *root = BTRFS_I(inode)->root;
3974         struct btrfs_fs_info *fs_info = root->fs_info;
3975         u64 used;
3976         int ret = 0;
3977         int need_commit = 2;
3978         int have_pinned_space;
3979
3980         /* make sure bytes are sectorsize aligned */
3981         bytes = ALIGN(bytes, root->sectorsize);
3982
3983         if (btrfs_is_free_space_inode(inode)) {
3984                 need_commit = 0;
3985                 ASSERT(current->journal_info);
3986         }
3987
3988         data_sinfo = fs_info->data_sinfo;
3989         if (!data_sinfo)
3990                 goto alloc;
3991
3992 again:
3993         /* make sure we have enough space to handle the data first */
3994         spin_lock(&data_sinfo->lock);
3995         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3996                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3997                 data_sinfo->bytes_may_use;
3998
3999         if (used + bytes > data_sinfo->total_bytes) {
4000                 struct btrfs_trans_handle *trans;
4001
4002                 /*
4003                  * if we don't have enough free bytes in this space then we need
4004                  * to alloc a new chunk.
4005                  */
4006                 if (!data_sinfo->full) {
4007                         u64 alloc_target;
4008
4009                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4010                         spin_unlock(&data_sinfo->lock);
4011 alloc:
4012                         alloc_target = btrfs_get_alloc_profile(root, 1);
4013                         /*
4014                          * It is ugly that we don't call nolock join
4015                          * transaction for the free space inode case here.
4016                          * But it is safe because we only do the data space
4017                          * reservation for the free space cache in the
4018                          * transaction context, the common join transaction
4019                          * just increase the counter of the current transaction
4020                          * handler, doesn't try to acquire the trans_lock of
4021                          * the fs.
4022                          */
4023                         trans = btrfs_join_transaction(root);
4024                         if (IS_ERR(trans))
4025                                 return PTR_ERR(trans);
4026
4027                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4028                                              alloc_target,
4029                                              CHUNK_ALLOC_NO_FORCE);
4030                         btrfs_end_transaction(trans, root);
4031                         if (ret < 0) {
4032                                 if (ret != -ENOSPC)
4033                                         return ret;
4034                                 else {
4035                                         have_pinned_space = 1;
4036                                         goto commit_trans;
4037                                 }
4038                         }
4039
4040                         if (!data_sinfo)
4041                                 data_sinfo = fs_info->data_sinfo;
4042
4043                         goto again;
4044                 }
4045
4046                 /*
4047                  * If we don't have enough pinned space to deal with this
4048                  * allocation, and no removed chunk in current transaction,
4049                  * don't bother committing the transaction.
4050                  */
4051                 have_pinned_space = percpu_counter_compare(
4052                         &data_sinfo->total_bytes_pinned,
4053                         used + bytes - data_sinfo->total_bytes);
4054                 spin_unlock(&data_sinfo->lock);
4055
4056                 /* commit the current transaction and try again */
4057 commit_trans:
4058                 if (need_commit &&
4059                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
4060                         need_commit--;
4061
4062                         if (need_commit > 0)
4063                                 btrfs_wait_ordered_roots(fs_info, -1);
4064
4065                         trans = btrfs_join_transaction(root);
4066                         if (IS_ERR(trans))
4067                                 return PTR_ERR(trans);
4068                         if (have_pinned_space >= 0 ||
4069                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4070                                      &trans->transaction->flags) ||
4071                             need_commit > 0) {
4072                                 ret = btrfs_commit_transaction(trans, root);
4073                                 if (ret)
4074                                         return ret;
4075                                 /*
4076                                  * make sure that all running delayed iput are
4077                                  * done
4078                                  */
4079                                 down_write(&root->fs_info->delayed_iput_sem);
4080                                 up_write(&root->fs_info->delayed_iput_sem);
4081                                 goto again;
4082                         } else {
4083                                 btrfs_end_transaction(trans, root);
4084                         }
4085                 }
4086
4087                 trace_btrfs_space_reservation(root->fs_info,
4088                                               "space_info:enospc",
4089                                               data_sinfo->flags, bytes, 1);
4090                 return -ENOSPC;
4091         }
4092         ret = btrfs_qgroup_reserve(root, write_bytes);
4093         if (ret)
4094                 goto out;
4095         data_sinfo->bytes_may_use += bytes;
4096         trace_btrfs_space_reservation(root->fs_info, "space_info",
4097                                       data_sinfo->flags, bytes, 1);
4098 out:
4099         spin_unlock(&data_sinfo->lock);
4100
4101         return ret;
4102 }
4103
4104 /*
4105  * Called if we need to clear a data reservation for this inode.
4106  */
4107 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
4108 {
4109         struct btrfs_root *root = BTRFS_I(inode)->root;
4110         struct btrfs_space_info *data_sinfo;
4111
4112         /* make sure bytes are sectorsize aligned */
4113         bytes = ALIGN(bytes, root->sectorsize);
4114
4115         data_sinfo = root->fs_info->data_sinfo;
4116         spin_lock(&data_sinfo->lock);
4117         WARN_ON(data_sinfo->bytes_may_use < bytes);
4118         data_sinfo->bytes_may_use -= bytes;
4119         trace_btrfs_space_reservation(root->fs_info, "space_info",
4120                                       data_sinfo->flags, bytes, 0);
4121         spin_unlock(&data_sinfo->lock);
4122 }
4123
4124 static void force_metadata_allocation(struct btrfs_fs_info *info)
4125 {
4126         struct list_head *head = &info->space_info;
4127         struct btrfs_space_info *found;
4128
4129         rcu_read_lock();
4130         list_for_each_entry_rcu(found, head, list) {
4131                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4132                         found->force_alloc = CHUNK_ALLOC_FORCE;
4133         }
4134         rcu_read_unlock();
4135 }
4136
4137 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4138 {
4139         return (global->size << 1);
4140 }
4141
4142 static int should_alloc_chunk(struct btrfs_root *root,
4143                               struct btrfs_space_info *sinfo, int force)
4144 {
4145         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4146         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4147         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4148         u64 thresh;
4149
4150         if (force == CHUNK_ALLOC_FORCE)
4151                 return 1;
4152
4153         /*
4154          * We need to take into account the global rsv because for all intents
4155          * and purposes it's used space.  Don't worry about locking the
4156          * global_rsv, it doesn't change except when the transaction commits.
4157          */
4158         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4159                 num_allocated += calc_global_rsv_need_space(global_rsv);
4160
4161         /*
4162          * in limited mode, we want to have some free space up to
4163          * about 1% of the FS size.
4164          */
4165         if (force == CHUNK_ALLOC_LIMITED) {
4166                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4167                 thresh = max_t(u64, 64 * 1024 * 1024,
4168                                div_factor_fine(thresh, 1));
4169
4170                 if (num_bytes - num_allocated < thresh)
4171                         return 1;
4172         }
4173
4174         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
4175                 return 0;
4176         return 1;
4177 }
4178
4179 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4180 {
4181         u64 num_dev;
4182
4183         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4184                     BTRFS_BLOCK_GROUP_RAID0 |
4185                     BTRFS_BLOCK_GROUP_RAID5 |
4186                     BTRFS_BLOCK_GROUP_RAID6))
4187                 num_dev = root->fs_info->fs_devices->rw_devices;
4188         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4189                 num_dev = 2;
4190         else
4191                 num_dev = 1;    /* DUP or single */
4192
4193         return num_dev;
4194 }
4195
4196 /*
4197  * If @is_allocation is true, reserve space in the system space info necessary
4198  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4199  * removing a chunk.
4200  */
4201 void check_system_chunk(struct btrfs_trans_handle *trans,
4202                         struct btrfs_root *root,
4203                         u64 type)
4204 {
4205         struct btrfs_space_info *info;
4206         u64 left;
4207         u64 thresh;
4208         int ret = 0;
4209         u64 num_devs;
4210
4211         /*
4212          * Needed because we can end up allocating a system chunk and for an
4213          * atomic and race free space reservation in the chunk block reserve.
4214          */
4215         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4216
4217         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4218         spin_lock(&info->lock);
4219         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4220                 info->bytes_reserved - info->bytes_readonly -
4221                 info->bytes_may_use;
4222         spin_unlock(&info->lock);
4223
4224         num_devs = get_profile_num_devs(root, type);
4225
4226         /* num_devs device items to update and 1 chunk item to add or remove */
4227         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4228                 btrfs_calc_trans_metadata_size(root, 1);
4229
4230         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4231                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4232                         left, thresh, type);
4233                 dump_space_info(info, 0, 0);
4234         }
4235
4236         if (left < thresh) {
4237                 u64 flags;
4238
4239                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4240                 /*
4241                  * Ignore failure to create system chunk. We might end up not
4242                  * needing it, as we might not need to COW all nodes/leafs from
4243                  * the paths we visit in the chunk tree (they were already COWed
4244                  * or created in the current transaction for example).
4245                  */
4246                 ret = btrfs_alloc_chunk(trans, root, flags);
4247         }
4248
4249         if (!ret) {
4250                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4251                                           &root->fs_info->chunk_block_rsv,
4252                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4253                 if (!ret)
4254                         trans->chunk_bytes_reserved += thresh;
4255         }
4256 }
4257
4258 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4259                           struct btrfs_root *extent_root, u64 flags, int force)
4260 {
4261         struct btrfs_space_info *space_info;
4262         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4263         int wait_for_alloc = 0;
4264         int ret = 0;
4265
4266         /* Don't re-enter if we're already allocating a chunk */
4267         if (trans->allocating_chunk)
4268                 return -ENOSPC;
4269
4270         space_info = __find_space_info(extent_root->fs_info, flags);
4271         if (!space_info) {
4272                 ret = update_space_info(extent_root->fs_info, flags,
4273                                         0, 0, &space_info);
4274                 BUG_ON(ret); /* -ENOMEM */
4275         }
4276         BUG_ON(!space_info); /* Logic error */
4277
4278 again:
4279         spin_lock(&space_info->lock);
4280         if (force < space_info->force_alloc)
4281                 force = space_info->force_alloc;
4282         if (space_info->full) {
4283                 if (should_alloc_chunk(extent_root, space_info, force))
4284                         ret = -ENOSPC;
4285                 else
4286                         ret = 0;
4287                 spin_unlock(&space_info->lock);
4288                 return ret;
4289         }
4290
4291         if (!should_alloc_chunk(extent_root, space_info, force)) {
4292                 spin_unlock(&space_info->lock);
4293                 return 0;
4294         } else if (space_info->chunk_alloc) {
4295                 wait_for_alloc = 1;
4296         } else {
4297                 space_info->chunk_alloc = 1;
4298         }
4299
4300         spin_unlock(&space_info->lock);
4301
4302         mutex_lock(&fs_info->chunk_mutex);
4303
4304         /*
4305          * The chunk_mutex is held throughout the entirety of a chunk
4306          * allocation, so once we've acquired the chunk_mutex we know that the
4307          * other guy is done and we need to recheck and see if we should
4308          * allocate.
4309          */
4310         if (wait_for_alloc) {
4311                 mutex_unlock(&fs_info->chunk_mutex);
4312                 wait_for_alloc = 0;
4313                 goto again;
4314         }
4315
4316         trans->allocating_chunk = true;
4317
4318         /*
4319          * If we have mixed data/metadata chunks we want to make sure we keep
4320          * allocating mixed chunks instead of individual chunks.
4321          */
4322         if (btrfs_mixed_space_info(space_info))
4323                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4324
4325         /*
4326          * if we're doing a data chunk, go ahead and make sure that
4327          * we keep a reasonable number of metadata chunks allocated in the
4328          * FS as well.
4329          */
4330         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4331                 fs_info->data_chunk_allocations++;
4332                 if (!(fs_info->data_chunk_allocations %
4333                       fs_info->metadata_ratio))
4334                         force_metadata_allocation(fs_info);
4335         }
4336
4337         /*
4338          * Check if we have enough space in SYSTEM chunk because we may need
4339          * to update devices.
4340          */
4341         check_system_chunk(trans, extent_root, flags);
4342
4343         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4344         trans->allocating_chunk = false;
4345
4346         spin_lock(&space_info->lock);
4347         if (ret < 0 && ret != -ENOSPC)
4348                 goto out;
4349         if (ret)
4350                 space_info->full = 1;
4351         else
4352                 ret = 1;
4353
4354         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4355 out:
4356         space_info->chunk_alloc = 0;
4357         spin_unlock(&space_info->lock);
4358         mutex_unlock(&fs_info->chunk_mutex);
4359         /*
4360          * When we allocate a new chunk we reserve space in the chunk block
4361          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4362          * add new nodes/leafs to it if we end up needing to do it when
4363          * inserting the chunk item and updating device items as part of the
4364          * second phase of chunk allocation, performed by
4365          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4366          * large number of new block groups to create in our transaction
4367          * handle's new_bgs list to avoid exhausting the chunk block reserve
4368          * in extreme cases - like having a single transaction create many new
4369          * block groups when starting to write out the free space caches of all
4370          * the block groups that were made dirty during the lifetime of the
4371          * transaction.
4372          */
4373         if (trans->can_flush_pending_bgs &&
4374             trans->chunk_bytes_reserved >= (2 * 1024 * 1024ull)) {
4375                 btrfs_create_pending_block_groups(trans, trans->root);
4376                 btrfs_trans_release_chunk_metadata(trans);
4377         }
4378         return ret;
4379 }
4380
4381 static int can_overcommit(struct btrfs_root *root,
4382                           struct btrfs_space_info *space_info, u64 bytes,
4383                           enum btrfs_reserve_flush_enum flush)
4384 {
4385         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4386         u64 profile = btrfs_get_alloc_profile(root, 0);
4387         u64 space_size;
4388         u64 avail;
4389         u64 used;
4390
4391         used = space_info->bytes_used + space_info->bytes_reserved +
4392                 space_info->bytes_pinned + space_info->bytes_readonly;
4393
4394         /*
4395          * We only want to allow over committing if we have lots of actual space
4396          * free, but if we don't have enough space to handle the global reserve
4397          * space then we could end up having a real enospc problem when trying
4398          * to allocate a chunk or some other such important allocation.
4399          */
4400         spin_lock(&global_rsv->lock);
4401         space_size = calc_global_rsv_need_space(global_rsv);
4402         spin_unlock(&global_rsv->lock);
4403         if (used + space_size >= space_info->total_bytes)
4404                 return 0;
4405
4406         used += space_info->bytes_may_use;
4407
4408         spin_lock(&root->fs_info->free_chunk_lock);
4409         avail = root->fs_info->free_chunk_space;
4410         spin_unlock(&root->fs_info->free_chunk_lock);
4411
4412         /*
4413          * If we have dup, raid1 or raid10 then only half of the free
4414          * space is actually useable.  For raid56, the space info used
4415          * doesn't include the parity drive, so we don't have to
4416          * change the math
4417          */
4418         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4419                        BTRFS_BLOCK_GROUP_RAID1 |
4420                        BTRFS_BLOCK_GROUP_RAID10))
4421                 avail >>= 1;
4422
4423         /*
4424          * If we aren't flushing all things, let us overcommit up to
4425          * 1/2th of the space. If we can flush, don't let us overcommit
4426          * too much, let it overcommit up to 1/8 of the space.
4427          */
4428         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4429                 avail >>= 3;
4430         else
4431                 avail >>= 1;
4432
4433         if (used + bytes < space_info->total_bytes + avail)
4434                 return 1;
4435         return 0;
4436 }
4437
4438 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4439                                          unsigned long nr_pages, int nr_items)
4440 {
4441         struct super_block *sb = root->fs_info->sb;
4442
4443         if (down_read_trylock(&sb->s_umount)) {
4444                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4445                 up_read(&sb->s_umount);
4446         } else {
4447                 /*
4448                  * We needn't worry the filesystem going from r/w to r/o though
4449                  * we don't acquire ->s_umount mutex, because the filesystem
4450                  * should guarantee the delalloc inodes list be empty after
4451                  * the filesystem is readonly(all dirty pages are written to
4452                  * the disk).
4453                  */
4454                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4455                 if (!current->journal_info)
4456                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4457         }
4458 }
4459
4460 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4461 {
4462         u64 bytes;
4463         int nr;
4464
4465         bytes = btrfs_calc_trans_metadata_size(root, 1);
4466         nr = (int)div64_u64(to_reclaim, bytes);
4467         if (!nr)
4468                 nr = 1;
4469         return nr;
4470 }
4471
4472 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4473
4474 /*
4475  * shrink metadata reservation for delalloc
4476  */
4477 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4478                             bool wait_ordered)
4479 {
4480         struct btrfs_block_rsv *block_rsv;
4481         struct btrfs_space_info *space_info;
4482         struct btrfs_trans_handle *trans;
4483         u64 delalloc_bytes;
4484         u64 max_reclaim;
4485         long time_left;
4486         unsigned long nr_pages;
4487         int loops;
4488         int items;
4489         enum btrfs_reserve_flush_enum flush;
4490
4491         /* Calc the number of the pages we need flush for space reservation */
4492         items = calc_reclaim_items_nr(root, to_reclaim);
4493         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4494
4495         trans = (struct btrfs_trans_handle *)current->journal_info;
4496         block_rsv = &root->fs_info->delalloc_block_rsv;
4497         space_info = block_rsv->space_info;
4498
4499         delalloc_bytes = percpu_counter_sum_positive(
4500                                                 &root->fs_info->delalloc_bytes);
4501         if (delalloc_bytes == 0) {
4502                 if (trans)
4503                         return;
4504                 if (wait_ordered)
4505                         btrfs_wait_ordered_roots(root->fs_info, items);
4506                 return;
4507         }
4508
4509         loops = 0;
4510         while (delalloc_bytes && loops < 3) {
4511                 max_reclaim = min(delalloc_bytes, to_reclaim);
4512                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4513                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4514                 /*
4515                  * We need to wait for the async pages to actually start before
4516                  * we do anything.
4517                  */
4518                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4519                 if (!max_reclaim)
4520                         goto skip_async;
4521
4522                 if (max_reclaim <= nr_pages)
4523                         max_reclaim = 0;
4524                 else
4525                         max_reclaim -= nr_pages;
4526
4527                 wait_event(root->fs_info->async_submit_wait,
4528                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4529                            (int)max_reclaim);
4530 skip_async:
4531                 if (!trans)
4532                         flush = BTRFS_RESERVE_FLUSH_ALL;
4533                 else
4534                         flush = BTRFS_RESERVE_NO_FLUSH;
4535                 spin_lock(&space_info->lock);
4536                 if (can_overcommit(root, space_info, orig, flush)) {
4537                         spin_unlock(&space_info->lock);
4538                         break;
4539                 }
4540                 spin_unlock(&space_info->lock);
4541
4542                 loops++;
4543                 if (wait_ordered && !trans) {
4544                         btrfs_wait_ordered_roots(root->fs_info, items);
4545                 } else {
4546                         time_left = schedule_timeout_killable(1);
4547                         if (time_left)
4548                                 break;
4549                 }
4550                 delalloc_bytes = percpu_counter_sum_positive(
4551                                                 &root->fs_info->delalloc_bytes);
4552         }
4553 }
4554
4555 /**
4556  * maybe_commit_transaction - possibly commit the transaction if its ok to
4557  * @root - the root we're allocating for
4558  * @bytes - the number of bytes we want to reserve
4559  * @force - force the commit
4560  *
4561  * This will check to make sure that committing the transaction will actually
4562  * get us somewhere and then commit the transaction if it does.  Otherwise it
4563  * will return -ENOSPC.
4564  */
4565 static int may_commit_transaction(struct btrfs_root *root,
4566                                   struct btrfs_space_info *space_info,
4567                                   u64 bytes, int force)
4568 {
4569         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4570         struct btrfs_trans_handle *trans;
4571
4572         trans = (struct btrfs_trans_handle *)current->journal_info;
4573         if (trans)
4574                 return -EAGAIN;
4575
4576         if (force)
4577                 goto commit;
4578
4579         /* See if there is enough pinned space to make this reservation */
4580         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4581                                    bytes) >= 0)
4582                 goto commit;
4583
4584         /*
4585          * See if there is some space in the delayed insertion reservation for
4586          * this reservation.
4587          */
4588         if (space_info != delayed_rsv->space_info)
4589                 return -ENOSPC;
4590
4591         spin_lock(&delayed_rsv->lock);
4592         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4593                                    bytes - delayed_rsv->size) >= 0) {
4594                 spin_unlock(&delayed_rsv->lock);
4595                 return -ENOSPC;
4596         }
4597         spin_unlock(&delayed_rsv->lock);
4598
4599 commit:
4600         trans = btrfs_join_transaction(root);
4601         if (IS_ERR(trans))
4602                 return -ENOSPC;
4603
4604         return btrfs_commit_transaction(trans, root);
4605 }
4606
4607 enum flush_state {
4608         FLUSH_DELAYED_ITEMS_NR  =       1,
4609         FLUSH_DELAYED_ITEMS     =       2,
4610         FLUSH_DELALLOC          =       3,
4611         FLUSH_DELALLOC_WAIT     =       4,
4612         ALLOC_CHUNK             =       5,
4613         COMMIT_TRANS            =       6,
4614 };
4615
4616 static int flush_space(struct btrfs_root *root,
4617                        struct btrfs_space_info *space_info, u64 num_bytes,
4618                        u64 orig_bytes, int state)
4619 {
4620         struct btrfs_trans_handle *trans;
4621         int nr;
4622         int ret = 0;
4623
4624         switch (state) {
4625         case FLUSH_DELAYED_ITEMS_NR:
4626         case FLUSH_DELAYED_ITEMS:
4627                 if (state == FLUSH_DELAYED_ITEMS_NR)
4628                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4629                 else
4630                         nr = -1;
4631
4632                 trans = btrfs_join_transaction(root);
4633                 if (IS_ERR(trans)) {
4634                         ret = PTR_ERR(trans);
4635                         break;
4636                 }
4637                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4638                 btrfs_end_transaction(trans, root);
4639                 break;
4640         case FLUSH_DELALLOC:
4641         case FLUSH_DELALLOC_WAIT:
4642                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4643                                 state == FLUSH_DELALLOC_WAIT);
4644                 break;
4645         case ALLOC_CHUNK:
4646                 trans = btrfs_join_transaction(root);
4647                 if (IS_ERR(trans)) {
4648                         ret = PTR_ERR(trans);
4649                         break;
4650                 }
4651                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4652                                      btrfs_get_alloc_profile(root, 0),
4653                                      CHUNK_ALLOC_NO_FORCE);
4654                 btrfs_end_transaction(trans, root);
4655                 if (ret == -ENOSPC)
4656                         ret = 0;
4657                 break;
4658         case COMMIT_TRANS:
4659                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4660                 break;
4661         default:
4662                 ret = -ENOSPC;
4663                 break;
4664         }
4665
4666         return ret;
4667 }
4668
4669 static inline u64
4670 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4671                                  struct btrfs_space_info *space_info)
4672 {
4673         u64 used;
4674         u64 expected;
4675         u64 to_reclaim;
4676
4677         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4678                                 16 * 1024 * 1024);
4679         spin_lock(&space_info->lock);
4680         if (can_overcommit(root, space_info, to_reclaim,
4681                            BTRFS_RESERVE_FLUSH_ALL)) {
4682                 to_reclaim = 0;
4683                 goto out;
4684         }
4685
4686         used = space_info->bytes_used + space_info->bytes_reserved +
4687                space_info->bytes_pinned + space_info->bytes_readonly +
4688                space_info->bytes_may_use;
4689         if (can_overcommit(root, space_info, 1024 * 1024,
4690                            BTRFS_RESERVE_FLUSH_ALL))
4691                 expected = div_factor_fine(space_info->total_bytes, 95);
4692         else
4693                 expected = div_factor_fine(space_info->total_bytes, 90);
4694
4695         if (used > expected)
4696                 to_reclaim = used - expected;
4697         else
4698                 to_reclaim = 0;
4699         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4700                                      space_info->bytes_reserved);
4701 out:
4702         spin_unlock(&space_info->lock);
4703
4704         return to_reclaim;
4705 }
4706
4707 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4708                                         struct btrfs_fs_info *fs_info, u64 used)
4709 {
4710         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4711
4712         /* If we're just plain full then async reclaim just slows us down. */
4713         if (space_info->bytes_used >= thresh)
4714                 return 0;
4715
4716         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4717                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4718 }
4719
4720 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4721                                        struct btrfs_fs_info *fs_info,
4722                                        int flush_state)
4723 {
4724         u64 used;
4725
4726         spin_lock(&space_info->lock);
4727         /*
4728          * We run out of space and have not got any free space via flush_space,
4729          * so don't bother doing async reclaim.
4730          */
4731         if (flush_state > COMMIT_TRANS && space_info->full) {
4732                 spin_unlock(&space_info->lock);
4733                 return 0;
4734         }
4735
4736         used = space_info->bytes_used + space_info->bytes_reserved +
4737                space_info->bytes_pinned + space_info->bytes_readonly +
4738                space_info->bytes_may_use;
4739         if (need_do_async_reclaim(space_info, fs_info, used)) {
4740                 spin_unlock(&space_info->lock);
4741                 return 1;
4742         }
4743         spin_unlock(&space_info->lock);
4744
4745         return 0;
4746 }
4747
4748 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4749 {
4750         struct btrfs_fs_info *fs_info;
4751         struct btrfs_space_info *space_info;
4752         u64 to_reclaim;
4753         int flush_state;
4754
4755         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4756         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4757
4758         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4759                                                       space_info);
4760         if (!to_reclaim)
4761                 return;
4762
4763         flush_state = FLUSH_DELAYED_ITEMS_NR;
4764         do {
4765                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4766                             to_reclaim, flush_state);
4767                 flush_state++;
4768                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4769                                                  flush_state))
4770                         return;
4771         } while (flush_state < COMMIT_TRANS);
4772 }
4773
4774 void btrfs_init_async_reclaim_work(struct work_struct *work)
4775 {
4776         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4777 }
4778
4779 /**
4780  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4781  * @root - the root we're allocating for
4782  * @block_rsv - the block_rsv we're allocating for
4783  * @orig_bytes - the number of bytes we want
4784  * @flush - whether or not we can flush to make our reservation
4785  *
4786  * This will reserve orgi_bytes number of bytes from the space info associated
4787  * with the block_rsv.  If there is not enough space it will make an attempt to
4788  * flush out space to make room.  It will do this by flushing delalloc if
4789  * possible or committing the transaction.  If flush is 0 then no attempts to
4790  * regain reservations will be made and this will fail if there is not enough
4791  * space already.
4792  */
4793 static int reserve_metadata_bytes(struct btrfs_root *root,
4794                                   struct btrfs_block_rsv *block_rsv,
4795                                   u64 orig_bytes,
4796                                   enum btrfs_reserve_flush_enum flush)
4797 {
4798         struct btrfs_space_info *space_info = block_rsv->space_info;
4799         u64 used;
4800         u64 num_bytes = orig_bytes;
4801         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4802         int ret = 0;
4803         bool flushing = false;
4804
4805 again:
4806         ret = 0;
4807         spin_lock(&space_info->lock);
4808         /*
4809          * We only want to wait if somebody other than us is flushing and we
4810          * are actually allowed to flush all things.
4811          */
4812         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4813                space_info->flush) {
4814                 spin_unlock(&space_info->lock);
4815                 /*
4816                  * If we have a trans handle we can't wait because the flusher
4817                  * may have to commit the transaction, which would mean we would
4818                  * deadlock since we are waiting for the flusher to finish, but
4819                  * hold the current transaction open.
4820                  */
4821                 if (current->journal_info)
4822                         return -EAGAIN;
4823                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4824                 /* Must have been killed, return */
4825                 if (ret)
4826                         return -EINTR;
4827
4828                 spin_lock(&space_info->lock);
4829         }
4830
4831         ret = -ENOSPC;
4832         used = space_info->bytes_used + space_info->bytes_reserved +
4833                 space_info->bytes_pinned + space_info->bytes_readonly +
4834                 space_info->bytes_may_use;
4835
4836         /*
4837          * The idea here is that we've not already over-reserved the block group
4838          * then we can go ahead and save our reservation first and then start
4839          * flushing if we need to.  Otherwise if we've already overcommitted
4840          * lets start flushing stuff first and then come back and try to make
4841          * our reservation.
4842          */
4843         if (used <= space_info->total_bytes) {
4844                 if (used + orig_bytes <= space_info->total_bytes) {
4845                         space_info->bytes_may_use += orig_bytes;
4846                         trace_btrfs_space_reservation(root->fs_info,
4847                                 "space_info", space_info->flags, orig_bytes, 1);
4848                         ret = 0;
4849                 } else {
4850                         /*
4851                          * Ok set num_bytes to orig_bytes since we aren't
4852                          * overocmmitted, this way we only try and reclaim what
4853                          * we need.
4854                          */
4855                         num_bytes = orig_bytes;
4856                 }
4857         } else {
4858                 /*
4859                  * Ok we're over committed, set num_bytes to the overcommitted
4860                  * amount plus the amount of bytes that we need for this
4861                  * reservation.
4862                  */
4863                 num_bytes = used - space_info->total_bytes +
4864                         (orig_bytes * 2);
4865         }
4866
4867         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4868                 space_info->bytes_may_use += orig_bytes;
4869                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4870                                               space_info->flags, orig_bytes,
4871                                               1);
4872                 ret = 0;
4873         }
4874
4875         /*
4876          * Couldn't make our reservation, save our place so while we're trying
4877          * to reclaim space we can actually use it instead of somebody else
4878          * stealing it from us.
4879          *
4880          * We make the other tasks wait for the flush only when we can flush
4881          * all things.
4882          */
4883         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4884                 flushing = true;
4885                 space_info->flush = 1;
4886         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4887                 used += orig_bytes;
4888                 /*
4889                  * We will do the space reservation dance during log replay,
4890                  * which means we won't have fs_info->fs_root set, so don't do
4891                  * the async reclaim as we will panic.
4892                  */
4893                 if (!root->fs_info->log_root_recovering &&
4894                     need_do_async_reclaim(space_info, root->fs_info, used) &&
4895                     !work_busy(&root->fs_info->async_reclaim_work))
4896                         queue_work(system_unbound_wq,
4897                                    &root->fs_info->async_reclaim_work);
4898         }
4899         spin_unlock(&space_info->lock);
4900
4901         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4902                 goto out;
4903
4904         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4905                           flush_state);
4906         flush_state++;
4907
4908         /*
4909          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4910          * would happen. So skip delalloc flush.
4911          */
4912         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4913             (flush_state == FLUSH_DELALLOC ||
4914              flush_state == FLUSH_DELALLOC_WAIT))
4915                 flush_state = ALLOC_CHUNK;
4916
4917         if (!ret)
4918                 goto again;
4919         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4920                  flush_state < COMMIT_TRANS)
4921                 goto again;
4922         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4923                  flush_state <= COMMIT_TRANS)
4924                 goto again;
4925
4926 out:
4927         if (ret == -ENOSPC &&
4928             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4929                 struct btrfs_block_rsv *global_rsv =
4930                         &root->fs_info->global_block_rsv;
4931
4932                 if (block_rsv != global_rsv &&
4933                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4934                         ret = 0;
4935         }
4936         if (ret == -ENOSPC)
4937                 trace_btrfs_space_reservation(root->fs_info,
4938                                               "space_info:enospc",
4939                                               space_info->flags, orig_bytes, 1);
4940         if (flushing) {
4941                 spin_lock(&space_info->lock);
4942                 space_info->flush = 0;
4943                 wake_up_all(&space_info->wait);
4944                 spin_unlock(&space_info->lock);
4945         }
4946         return ret;
4947 }
4948
4949 static struct btrfs_block_rsv *get_block_rsv(
4950                                         const struct btrfs_trans_handle *trans,
4951                                         const struct btrfs_root *root)
4952 {
4953         struct btrfs_block_rsv *block_rsv = NULL;
4954
4955         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4956             (root == root->fs_info->csum_root && trans->adding_csums) ||
4957              (root == root->fs_info->uuid_root))
4958                 block_rsv = trans->block_rsv;
4959
4960         if (!block_rsv)
4961                 block_rsv = root->block_rsv;
4962
4963         if (!block_rsv)
4964                 block_rsv = &root->fs_info->empty_block_rsv;
4965
4966         return block_rsv;
4967 }
4968
4969 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4970                                u64 num_bytes)
4971 {
4972         int ret = -ENOSPC;
4973         spin_lock(&block_rsv->lock);
4974         if (block_rsv->reserved >= num_bytes) {
4975                 block_rsv->reserved -= num_bytes;
4976                 if (block_rsv->reserved < block_rsv->size)
4977                         block_rsv->full = 0;
4978                 ret = 0;
4979         }
4980         spin_unlock(&block_rsv->lock);
4981         return ret;
4982 }
4983
4984 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4985                                 u64 num_bytes, int update_size)
4986 {
4987         spin_lock(&block_rsv->lock);
4988         block_rsv->reserved += num_bytes;
4989         if (update_size)
4990                 block_rsv->size += num_bytes;
4991         else if (block_rsv->reserved >= block_rsv->size)
4992                 block_rsv->full = 1;
4993         spin_unlock(&block_rsv->lock);
4994 }
4995
4996 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4997                              struct btrfs_block_rsv *dest, u64 num_bytes,
4998                              int min_factor)
4999 {
5000         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5001         u64 min_bytes;
5002
5003         if (global_rsv->space_info != dest->space_info)
5004                 return -ENOSPC;
5005
5006         spin_lock(&global_rsv->lock);
5007         min_bytes = div_factor(global_rsv->size, min_factor);
5008         if (global_rsv->reserved < min_bytes + num_bytes) {
5009                 spin_unlock(&global_rsv->lock);
5010                 return -ENOSPC;
5011         }
5012         global_rsv->reserved -= num_bytes;
5013         if (global_rsv->reserved < global_rsv->size)
5014                 global_rsv->full = 0;
5015         spin_unlock(&global_rsv->lock);
5016
5017         block_rsv_add_bytes(dest, num_bytes, 1);
5018         return 0;
5019 }
5020
5021 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5022                                     struct btrfs_block_rsv *block_rsv,
5023                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5024 {
5025         struct btrfs_space_info *space_info = block_rsv->space_info;
5026
5027         spin_lock(&block_rsv->lock);
5028         if (num_bytes == (u64)-1)
5029                 num_bytes = block_rsv->size;
5030         block_rsv->size -= num_bytes;
5031         if (block_rsv->reserved >= block_rsv->size) {
5032                 num_bytes = block_rsv->reserved - block_rsv->size;
5033                 block_rsv->reserved = block_rsv->size;
5034                 block_rsv->full = 1;
5035         } else {
5036                 num_bytes = 0;
5037         }
5038         spin_unlock(&block_rsv->lock);
5039
5040         if (num_bytes > 0) {
5041                 if (dest) {
5042                         spin_lock(&dest->lock);
5043                         if (!dest->full) {
5044                                 u64 bytes_to_add;
5045
5046                                 bytes_to_add = dest->size - dest->reserved;
5047                                 bytes_to_add = min(num_bytes, bytes_to_add);
5048                                 dest->reserved += bytes_to_add;
5049                                 if (dest->reserved >= dest->size)
5050                                         dest->full = 1;
5051                                 num_bytes -= bytes_to_add;
5052                         }
5053                         spin_unlock(&dest->lock);
5054                 }
5055                 if (num_bytes) {
5056                         spin_lock(&space_info->lock);
5057                         space_info->bytes_may_use -= num_bytes;
5058                         trace_btrfs_space_reservation(fs_info, "space_info",
5059                                         space_info->flags, num_bytes, 0);
5060                         spin_unlock(&space_info->lock);
5061                 }
5062         }
5063 }
5064
5065 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
5066                                    struct btrfs_block_rsv *dst, u64 num_bytes)
5067 {
5068         int ret;
5069
5070         ret = block_rsv_use_bytes(src, num_bytes);
5071         if (ret)
5072                 return ret;
5073
5074         block_rsv_add_bytes(dst, num_bytes, 1);
5075         return 0;
5076 }
5077
5078 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5079 {
5080         memset(rsv, 0, sizeof(*rsv));
5081         spin_lock_init(&rsv->lock);
5082         rsv->type = type;
5083 }
5084
5085 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5086                                               unsigned short type)
5087 {
5088         struct btrfs_block_rsv *block_rsv;
5089         struct btrfs_fs_info *fs_info = root->fs_info;
5090
5091         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5092         if (!block_rsv)
5093                 return NULL;
5094
5095         btrfs_init_block_rsv(block_rsv, type);
5096         block_rsv->space_info = __find_space_info(fs_info,
5097                                                   BTRFS_BLOCK_GROUP_METADATA);
5098         return block_rsv;
5099 }
5100
5101 void btrfs_free_block_rsv(struct btrfs_root *root,
5102                           struct btrfs_block_rsv *rsv)
5103 {
5104         if (!rsv)
5105                 return;
5106         btrfs_block_rsv_release(root, rsv, (u64)-1);
5107         kfree(rsv);
5108 }
5109
5110 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5111 {
5112         kfree(rsv);
5113 }
5114
5115 int btrfs_block_rsv_add(struct btrfs_root *root,
5116                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5117                         enum btrfs_reserve_flush_enum flush)
5118 {
5119         int ret;
5120
5121         if (num_bytes == 0)
5122                 return 0;
5123
5124         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5125         if (!ret) {
5126                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5127                 return 0;
5128         }
5129
5130         return ret;
5131 }
5132
5133 int btrfs_block_rsv_check(struct btrfs_root *root,
5134                           struct btrfs_block_rsv *block_rsv, int min_factor)
5135 {
5136         u64 num_bytes = 0;
5137         int ret = -ENOSPC;
5138
5139         if (!block_rsv)
5140                 return 0;
5141
5142         spin_lock(&block_rsv->lock);
5143         num_bytes = div_factor(block_rsv->size, min_factor);
5144         if (block_rsv->reserved >= num_bytes)
5145                 ret = 0;
5146         spin_unlock(&block_rsv->lock);
5147
5148         return ret;
5149 }
5150
5151 int btrfs_block_rsv_refill(struct btrfs_root *root,
5152                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5153                            enum btrfs_reserve_flush_enum flush)
5154 {
5155         u64 num_bytes = 0;
5156         int ret = -ENOSPC;
5157
5158         if (!block_rsv)
5159                 return 0;
5160
5161         spin_lock(&block_rsv->lock);
5162         num_bytes = min_reserved;
5163         if (block_rsv->reserved >= num_bytes)
5164                 ret = 0;
5165         else
5166                 num_bytes -= block_rsv->reserved;
5167         spin_unlock(&block_rsv->lock);
5168
5169         if (!ret)
5170                 return 0;
5171
5172         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5173         if (!ret) {
5174                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5175                 return 0;
5176         }
5177
5178         return ret;
5179 }
5180
5181 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5182                             struct btrfs_block_rsv *dst_rsv,
5183                             u64 num_bytes)
5184 {
5185         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5186 }
5187
5188 void btrfs_block_rsv_release(struct btrfs_root *root,
5189                              struct btrfs_block_rsv *block_rsv,
5190                              u64 num_bytes)
5191 {
5192         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5193         if (global_rsv == block_rsv ||
5194             block_rsv->space_info != global_rsv->space_info)
5195                 global_rsv = NULL;
5196         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5197                                 num_bytes);
5198 }
5199
5200 /*
5201  * helper to calculate size of global block reservation.
5202  * the desired value is sum of space used by extent tree,
5203  * checksum tree and root tree
5204  */
5205 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5206 {
5207         struct btrfs_space_info *sinfo;
5208         u64 num_bytes;
5209         u64 meta_used;
5210         u64 data_used;
5211         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5212
5213         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5214         spin_lock(&sinfo->lock);
5215         data_used = sinfo->bytes_used;
5216         spin_unlock(&sinfo->lock);
5217
5218         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5219         spin_lock(&sinfo->lock);
5220         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5221                 data_used = 0;
5222         meta_used = sinfo->bytes_used;
5223         spin_unlock(&sinfo->lock);
5224
5225         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5226                     csum_size * 2;
5227         num_bytes += div_u64(data_used + meta_used, 50);
5228
5229         if (num_bytes * 3 > meta_used)
5230                 num_bytes = div_u64(meta_used, 3);
5231
5232         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5233 }
5234
5235 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5236 {
5237         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5238         struct btrfs_space_info *sinfo = block_rsv->space_info;
5239         u64 num_bytes;
5240
5241         num_bytes = calc_global_metadata_size(fs_info);
5242
5243         spin_lock(&sinfo->lock);
5244         spin_lock(&block_rsv->lock);
5245
5246         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
5247
5248         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5249                     sinfo->bytes_reserved + sinfo->bytes_readonly +
5250                     sinfo->bytes_may_use;
5251
5252         if (sinfo->total_bytes > num_bytes) {
5253                 num_bytes = sinfo->total_bytes - num_bytes;
5254                 block_rsv->reserved += num_bytes;
5255                 sinfo->bytes_may_use += num_bytes;
5256                 trace_btrfs_space_reservation(fs_info, "space_info",
5257                                       sinfo->flags, num_bytes, 1);
5258         }
5259
5260         if (block_rsv->reserved >= block_rsv->size) {
5261                 num_bytes = block_rsv->reserved - block_rsv->size;
5262                 sinfo->bytes_may_use -= num_bytes;
5263                 trace_btrfs_space_reservation(fs_info, "space_info",
5264                                       sinfo->flags, num_bytes, 0);
5265                 block_rsv->reserved = block_rsv->size;
5266                 block_rsv->full = 1;
5267         }
5268
5269         spin_unlock(&block_rsv->lock);
5270         spin_unlock(&sinfo->lock);
5271 }
5272
5273 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5274 {
5275         struct btrfs_space_info *space_info;
5276
5277         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5278         fs_info->chunk_block_rsv.space_info = space_info;
5279
5280         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5281         fs_info->global_block_rsv.space_info = space_info;
5282         fs_info->delalloc_block_rsv.space_info = space_info;
5283         fs_info->trans_block_rsv.space_info = space_info;
5284         fs_info->empty_block_rsv.space_info = space_info;
5285         fs_info->delayed_block_rsv.space_info = space_info;
5286
5287         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5288         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5289         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5290         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5291         if (fs_info->quota_root)
5292                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5293         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5294
5295         update_global_block_rsv(fs_info);
5296 }
5297
5298 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5299 {
5300         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5301                                 (u64)-1);
5302         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5303         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5304         WARN_ON(fs_info->trans_block_rsv.size > 0);
5305         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5306         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5307         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5308         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5309         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5310 }
5311
5312 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5313                                   struct btrfs_root *root)
5314 {
5315         if (!trans->block_rsv)
5316                 return;
5317
5318         if (!trans->bytes_reserved)
5319                 return;
5320
5321         trace_btrfs_space_reservation(root->fs_info, "transaction",
5322                                       trans->transid, trans->bytes_reserved, 0);
5323         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5324         trans->bytes_reserved = 0;
5325 }
5326
5327 /*
5328  * To be called after all the new block groups attached to the transaction
5329  * handle have been created (btrfs_create_pending_block_groups()).
5330  */
5331 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5332 {
5333         struct btrfs_fs_info *fs_info = trans->root->fs_info;
5334
5335         if (!trans->chunk_bytes_reserved)
5336                 return;
5337
5338         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5339
5340         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5341                                 trans->chunk_bytes_reserved);
5342         trans->chunk_bytes_reserved = 0;
5343 }
5344
5345 /* Can only return 0 or -ENOSPC */
5346 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5347                                   struct inode *inode)
5348 {
5349         struct btrfs_root *root = BTRFS_I(inode)->root;
5350         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5351         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5352
5353         /*
5354          * We need to hold space in order to delete our orphan item once we've
5355          * added it, so this takes the reservation so we can release it later
5356          * when we are truly done with the orphan item.
5357          */
5358         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5359         trace_btrfs_space_reservation(root->fs_info, "orphan",
5360                                       btrfs_ino(inode), num_bytes, 1);
5361         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5362 }
5363
5364 void btrfs_orphan_release_metadata(struct inode *inode)
5365 {
5366         struct btrfs_root *root = BTRFS_I(inode)->root;
5367         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5368         trace_btrfs_space_reservation(root->fs_info, "orphan",
5369                                       btrfs_ino(inode), num_bytes, 0);
5370         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5371 }
5372
5373 /*
5374  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5375  * root: the root of the parent directory
5376  * rsv: block reservation
5377  * items: the number of items that we need do reservation
5378  * qgroup_reserved: used to return the reserved size in qgroup
5379  *
5380  * This function is used to reserve the space for snapshot/subvolume
5381  * creation and deletion. Those operations are different with the
5382  * common file/directory operations, they change two fs/file trees
5383  * and root tree, the number of items that the qgroup reserves is
5384  * different with the free space reservation. So we can not use
5385  * the space reseravtion mechanism in start_transaction().
5386  */
5387 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5388                                      struct btrfs_block_rsv *rsv,
5389                                      int items,
5390                                      u64 *qgroup_reserved,
5391                                      bool use_global_rsv)
5392 {
5393         u64 num_bytes;
5394         int ret;
5395         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5396
5397         if (root->fs_info->quota_enabled) {
5398                 /* One for parent inode, two for dir entries */
5399                 num_bytes = 3 * root->nodesize;
5400                 ret = btrfs_qgroup_reserve(root, num_bytes);
5401                 if (ret)
5402                         return ret;
5403         } else {
5404                 num_bytes = 0;
5405         }
5406
5407         *qgroup_reserved = num_bytes;
5408
5409         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5410         rsv->space_info = __find_space_info(root->fs_info,
5411                                             BTRFS_BLOCK_GROUP_METADATA);
5412         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5413                                   BTRFS_RESERVE_FLUSH_ALL);
5414
5415         if (ret == -ENOSPC && use_global_rsv)
5416                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5417
5418         if (ret) {
5419                 if (*qgroup_reserved)
5420                         btrfs_qgroup_free(root, *qgroup_reserved);
5421         }
5422
5423         return ret;
5424 }
5425
5426 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5427                                       struct btrfs_block_rsv *rsv,
5428                                       u64 qgroup_reserved)
5429 {
5430         btrfs_block_rsv_release(root, rsv, (u64)-1);
5431 }
5432
5433 /**
5434  * drop_outstanding_extent - drop an outstanding extent
5435  * @inode: the inode we're dropping the extent for
5436  * @num_bytes: the number of bytes we're relaseing.
5437  *
5438  * This is called when we are freeing up an outstanding extent, either called
5439  * after an error or after an extent is written.  This will return the number of
5440  * reserved extents that need to be freed.  This must be called with
5441  * BTRFS_I(inode)->lock held.
5442  */
5443 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5444 {
5445         unsigned drop_inode_space = 0;
5446         unsigned dropped_extents = 0;
5447         unsigned num_extents = 0;
5448
5449         num_extents = (unsigned)div64_u64(num_bytes +
5450                                           BTRFS_MAX_EXTENT_SIZE - 1,
5451                                           BTRFS_MAX_EXTENT_SIZE);
5452         ASSERT(num_extents);
5453         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5454         BTRFS_I(inode)->outstanding_extents -= num_extents;
5455
5456         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5457             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5458                                &BTRFS_I(inode)->runtime_flags))
5459                 drop_inode_space = 1;
5460
5461         /*
5462          * If we have more or the same amount of outsanding extents than we have
5463          * reserved then we need to leave the reserved extents count alone.
5464          */
5465         if (BTRFS_I(inode)->outstanding_extents >=
5466             BTRFS_I(inode)->reserved_extents)
5467                 return drop_inode_space;
5468
5469         dropped_extents = BTRFS_I(inode)->reserved_extents -
5470                 BTRFS_I(inode)->outstanding_extents;
5471         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5472         return dropped_extents + drop_inode_space;
5473 }
5474
5475 /**
5476  * calc_csum_metadata_size - return the amount of metada space that must be
5477  *      reserved/free'd for the given bytes.
5478  * @inode: the inode we're manipulating
5479  * @num_bytes: the number of bytes in question
5480  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5481  *
5482  * This adjusts the number of csum_bytes in the inode and then returns the
5483  * correct amount of metadata that must either be reserved or freed.  We
5484  * calculate how many checksums we can fit into one leaf and then divide the
5485  * number of bytes that will need to be checksumed by this value to figure out
5486  * how many checksums will be required.  If we are adding bytes then the number
5487  * may go up and we will return the number of additional bytes that must be
5488  * reserved.  If it is going down we will return the number of bytes that must
5489  * be freed.
5490  *
5491  * This must be called with BTRFS_I(inode)->lock held.
5492  */
5493 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5494                                    int reserve)
5495 {
5496         struct btrfs_root *root = BTRFS_I(inode)->root;
5497         u64 old_csums, num_csums;
5498
5499         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5500             BTRFS_I(inode)->csum_bytes == 0)
5501                 return 0;
5502
5503         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5504         if (reserve)
5505                 BTRFS_I(inode)->csum_bytes += num_bytes;
5506         else
5507                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5508         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5509
5510         /* No change, no need to reserve more */
5511         if (old_csums == num_csums)
5512                 return 0;
5513
5514         if (reserve)
5515                 return btrfs_calc_trans_metadata_size(root,
5516                                                       num_csums - old_csums);
5517
5518         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5519 }
5520
5521 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5522 {
5523         struct btrfs_root *root = BTRFS_I(inode)->root;
5524         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5525         u64 to_reserve = 0;
5526         u64 csum_bytes;
5527         unsigned nr_extents = 0;
5528         int extra_reserve = 0;
5529         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5530         int ret = 0;
5531         bool delalloc_lock = true;
5532         u64 to_free = 0;
5533         unsigned dropped;
5534
5535         /* If we are a free space inode we need to not flush since we will be in
5536          * the middle of a transaction commit.  We also don't need the delalloc
5537          * mutex since we won't race with anybody.  We need this mostly to make
5538          * lockdep shut its filthy mouth.
5539          */
5540         if (btrfs_is_free_space_inode(inode)) {
5541                 flush = BTRFS_RESERVE_NO_FLUSH;
5542                 delalloc_lock = false;
5543         }
5544
5545         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5546             btrfs_transaction_in_commit(root->fs_info))
5547                 schedule_timeout(1);
5548
5549         if (delalloc_lock)
5550                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5551
5552         num_bytes = ALIGN(num_bytes, root->sectorsize);
5553
5554         spin_lock(&BTRFS_I(inode)->lock);
5555         nr_extents = (unsigned)div64_u64(num_bytes +
5556                                          BTRFS_MAX_EXTENT_SIZE - 1,
5557                                          BTRFS_MAX_EXTENT_SIZE);
5558         BTRFS_I(inode)->outstanding_extents += nr_extents;
5559         nr_extents = 0;
5560
5561         if (BTRFS_I(inode)->outstanding_extents >
5562             BTRFS_I(inode)->reserved_extents)
5563                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5564                         BTRFS_I(inode)->reserved_extents;
5565
5566         /*
5567          * Add an item to reserve for updating the inode when we complete the
5568          * delalloc io.
5569          */
5570         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5571                       &BTRFS_I(inode)->runtime_flags)) {
5572                 nr_extents++;
5573                 extra_reserve = 1;
5574         }
5575
5576         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5577         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5578         csum_bytes = BTRFS_I(inode)->csum_bytes;
5579         spin_unlock(&BTRFS_I(inode)->lock);
5580
5581         if (root->fs_info->quota_enabled) {
5582                 ret = btrfs_qgroup_reserve(root, nr_extents * root->nodesize);
5583                 if (ret)
5584                         goto out_fail;
5585         }
5586
5587         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5588         if (unlikely(ret)) {
5589                 if (root->fs_info->quota_enabled)
5590                         btrfs_qgroup_free(root, nr_extents * root->nodesize);
5591                 goto out_fail;
5592         }
5593
5594         spin_lock(&BTRFS_I(inode)->lock);
5595         if (extra_reserve) {
5596                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5597                         &BTRFS_I(inode)->runtime_flags);
5598                 nr_extents--;
5599         }
5600         BTRFS_I(inode)->reserved_extents += nr_extents;
5601         spin_unlock(&BTRFS_I(inode)->lock);
5602
5603         if (delalloc_lock)
5604                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5605
5606         if (to_reserve)
5607                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5608                                               btrfs_ino(inode), to_reserve, 1);
5609         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5610
5611         return 0;
5612
5613 out_fail:
5614         spin_lock(&BTRFS_I(inode)->lock);
5615         dropped = drop_outstanding_extent(inode, num_bytes);
5616         /*
5617          * If the inodes csum_bytes is the same as the original
5618          * csum_bytes then we know we haven't raced with any free()ers
5619          * so we can just reduce our inodes csum bytes and carry on.
5620          */
5621         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5622                 calc_csum_metadata_size(inode, num_bytes, 0);
5623         } else {
5624                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5625                 u64 bytes;
5626
5627                 /*
5628                  * This is tricky, but first we need to figure out how much we
5629                  * free'd from any free-ers that occured during this
5630                  * reservation, so we reset ->csum_bytes to the csum_bytes
5631                  * before we dropped our lock, and then call the free for the
5632                  * number of bytes that were freed while we were trying our
5633                  * reservation.
5634                  */
5635                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5636                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5637                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5638
5639
5640                 /*
5641                  * Now we need to see how much we would have freed had we not
5642                  * been making this reservation and our ->csum_bytes were not
5643                  * artificially inflated.
5644                  */
5645                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5646                 bytes = csum_bytes - orig_csum_bytes;
5647                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5648
5649                 /*
5650                  * Now reset ->csum_bytes to what it should be.  If bytes is
5651                  * more than to_free then we would have free'd more space had we
5652                  * not had an artificially high ->csum_bytes, so we need to free
5653                  * the remainder.  If bytes is the same or less then we don't
5654                  * need to do anything, the other free-ers did the correct
5655                  * thing.
5656                  */
5657                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5658                 if (bytes > to_free)
5659                         to_free = bytes - to_free;
5660                 else
5661                         to_free = 0;
5662         }
5663         spin_unlock(&BTRFS_I(inode)->lock);
5664         if (dropped)
5665                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5666
5667         if (to_free) {
5668                 btrfs_block_rsv_release(root, block_rsv, to_free);
5669                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5670                                               btrfs_ino(inode), to_free, 0);
5671         }
5672         if (delalloc_lock)
5673                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5674         return ret;
5675 }
5676
5677 /**
5678  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5679  * @inode: the inode to release the reservation for
5680  * @num_bytes: the number of bytes we're releasing
5681  *
5682  * This will release the metadata reservation for an inode.  This can be called
5683  * once we complete IO for a given set of bytes to release their metadata
5684  * reservations.
5685  */
5686 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5687 {
5688         struct btrfs_root *root = BTRFS_I(inode)->root;
5689         u64 to_free = 0;
5690         unsigned dropped;
5691
5692         num_bytes = ALIGN(num_bytes, root->sectorsize);
5693         spin_lock(&BTRFS_I(inode)->lock);
5694         dropped = drop_outstanding_extent(inode, num_bytes);
5695
5696         if (num_bytes)
5697                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5698         spin_unlock(&BTRFS_I(inode)->lock);
5699         if (dropped > 0)
5700                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5701
5702         if (btrfs_test_is_dummy_root(root))
5703                 return;
5704
5705         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5706                                       btrfs_ino(inode), to_free, 0);
5707
5708         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5709                                 to_free);
5710 }
5711
5712 /**
5713  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5714  * @inode: inode we're writing to
5715  * @num_bytes: the number of bytes we want to allocate
5716  *
5717  * This will do the following things
5718  *
5719  * o reserve space in the data space info for num_bytes
5720  * o reserve space in the metadata space info based on number of outstanding
5721  *   extents and how much csums will be needed
5722  * o add to the inodes ->delalloc_bytes
5723  * o add it to the fs_info's delalloc inodes list.
5724  *
5725  * This will return 0 for success and -ENOSPC if there is no space left.
5726  */
5727 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5728 {
5729         int ret;
5730
5731         ret = btrfs_check_data_free_space(inode, num_bytes, num_bytes);
5732         if (ret)
5733                 return ret;
5734
5735         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5736         if (ret) {
5737                 btrfs_free_reserved_data_space(inode, num_bytes);
5738                 return ret;
5739         }
5740
5741         return 0;
5742 }
5743
5744 /**
5745  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5746  * @inode: inode we're releasing space for
5747  * @num_bytes: the number of bytes we want to free up
5748  *
5749  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5750  * called in the case that we don't need the metadata AND data reservations
5751  * anymore.  So if there is an error or we insert an inline extent.
5752  *
5753  * This function will release the metadata space that was not used and will
5754  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5755  * list if there are no delalloc bytes left.
5756  */
5757 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5758 {
5759         btrfs_delalloc_release_metadata(inode, num_bytes);
5760         btrfs_free_reserved_data_space(inode, num_bytes);
5761 }
5762
5763 static int update_block_group(struct btrfs_trans_handle *trans,
5764                               struct btrfs_root *root, u64 bytenr,
5765                               u64 num_bytes, int alloc)
5766 {
5767         struct btrfs_block_group_cache *cache = NULL;
5768         struct btrfs_fs_info *info = root->fs_info;
5769         u64 total = num_bytes;
5770         u64 old_val;
5771         u64 byte_in_group;
5772         int factor;
5773
5774         /* block accounting for super block */
5775         spin_lock(&info->delalloc_root_lock);
5776         old_val = btrfs_super_bytes_used(info->super_copy);
5777         if (alloc)
5778                 old_val += num_bytes;
5779         else
5780                 old_val -= num_bytes;
5781         btrfs_set_super_bytes_used(info->super_copy, old_val);
5782         spin_unlock(&info->delalloc_root_lock);
5783
5784         while (total) {
5785                 cache = btrfs_lookup_block_group(info, bytenr);
5786                 if (!cache)
5787                         return -ENOENT;
5788                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5789                                     BTRFS_BLOCK_GROUP_RAID1 |
5790                                     BTRFS_BLOCK_GROUP_RAID10))
5791                         factor = 2;
5792                 else
5793                         factor = 1;
5794                 /*
5795                  * If this block group has free space cache written out, we
5796                  * need to make sure to load it if we are removing space.  This
5797                  * is because we need the unpinning stage to actually add the
5798                  * space back to the block group, otherwise we will leak space.
5799                  */
5800                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5801                         cache_block_group(cache, 1);
5802
5803                 byte_in_group = bytenr - cache->key.objectid;
5804                 WARN_ON(byte_in_group > cache->key.offset);
5805
5806                 spin_lock(&cache->space_info->lock);
5807                 spin_lock(&cache->lock);
5808
5809                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5810                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5811                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5812
5813                 old_val = btrfs_block_group_used(&cache->item);
5814                 num_bytes = min(total, cache->key.offset - byte_in_group);
5815                 if (alloc) {
5816                         old_val += num_bytes;
5817                         btrfs_set_block_group_used(&cache->item, old_val);
5818                         cache->reserved -= num_bytes;
5819                         cache->space_info->bytes_reserved -= num_bytes;
5820                         cache->space_info->bytes_used += num_bytes;
5821                         cache->space_info->disk_used += num_bytes * factor;
5822                         spin_unlock(&cache->lock);
5823                         spin_unlock(&cache->space_info->lock);
5824                 } else {
5825                         old_val -= num_bytes;
5826                         btrfs_set_block_group_used(&cache->item, old_val);
5827                         cache->pinned += num_bytes;
5828                         cache->space_info->bytes_pinned += num_bytes;
5829                         cache->space_info->bytes_used -= num_bytes;
5830                         cache->space_info->disk_used -= num_bytes * factor;
5831                         spin_unlock(&cache->lock);
5832                         spin_unlock(&cache->space_info->lock);
5833
5834                         set_extent_dirty(info->pinned_extents,
5835                                          bytenr, bytenr + num_bytes - 1,
5836                                          GFP_NOFS | __GFP_NOFAIL);
5837                         /*
5838                          * No longer have used bytes in this block group, queue
5839                          * it for deletion.
5840                          */
5841                         if (old_val == 0) {
5842                                 spin_lock(&info->unused_bgs_lock);
5843                                 if (list_empty(&cache->bg_list)) {
5844                                         btrfs_get_block_group(cache);
5845                                         list_add_tail(&cache->bg_list,
5846                                                       &info->unused_bgs);
5847                                 }
5848                                 spin_unlock(&info->unused_bgs_lock);
5849                         }
5850                 }
5851
5852                 spin_lock(&trans->transaction->dirty_bgs_lock);
5853                 if (list_empty(&cache->dirty_list)) {
5854                         list_add_tail(&cache->dirty_list,
5855                                       &trans->transaction->dirty_bgs);
5856                                 trans->transaction->num_dirty_bgs++;
5857                         btrfs_get_block_group(cache);
5858                 }
5859                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5860
5861                 btrfs_put_block_group(cache);
5862                 total -= num_bytes;
5863                 bytenr += num_bytes;
5864         }
5865         return 0;
5866 }
5867
5868 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5869 {
5870         struct btrfs_block_group_cache *cache;
5871         u64 bytenr;
5872
5873         spin_lock(&root->fs_info->block_group_cache_lock);
5874         bytenr = root->fs_info->first_logical_byte;
5875         spin_unlock(&root->fs_info->block_group_cache_lock);
5876
5877         if (bytenr < (u64)-1)
5878                 return bytenr;
5879
5880         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5881         if (!cache)
5882                 return 0;
5883
5884         bytenr = cache->key.objectid;
5885         btrfs_put_block_group(cache);
5886
5887         return bytenr;
5888 }
5889
5890 static int pin_down_extent(struct btrfs_root *root,
5891                            struct btrfs_block_group_cache *cache,
5892                            u64 bytenr, u64 num_bytes, int reserved)
5893 {
5894         spin_lock(&cache->space_info->lock);
5895         spin_lock(&cache->lock);
5896         cache->pinned += num_bytes;
5897         cache->space_info->bytes_pinned += num_bytes;
5898         if (reserved) {
5899                 cache->reserved -= num_bytes;
5900                 cache->space_info->bytes_reserved -= num_bytes;
5901         }
5902         spin_unlock(&cache->lock);
5903         spin_unlock(&cache->space_info->lock);
5904
5905         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5906                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5907         if (reserved)
5908                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5909         return 0;
5910 }
5911
5912 /*
5913  * this function must be called within transaction
5914  */
5915 int btrfs_pin_extent(struct btrfs_root *root,
5916                      u64 bytenr, u64 num_bytes, int reserved)
5917 {
5918         struct btrfs_block_group_cache *cache;
5919
5920         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5921         BUG_ON(!cache); /* Logic error */
5922
5923         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5924
5925         btrfs_put_block_group(cache);
5926         return 0;
5927 }
5928
5929 /*
5930  * this function must be called within transaction
5931  */
5932 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5933                                     u64 bytenr, u64 num_bytes)
5934 {
5935         struct btrfs_block_group_cache *cache;
5936         int ret;
5937
5938         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5939         if (!cache)
5940                 return -EINVAL;
5941
5942         /*
5943          * pull in the free space cache (if any) so that our pin
5944          * removes the free space from the cache.  We have load_only set
5945          * to one because the slow code to read in the free extents does check
5946          * the pinned extents.
5947          */
5948         cache_block_group(cache, 1);
5949
5950         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5951
5952         /* remove us from the free space cache (if we're there at all) */
5953         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5954         btrfs_put_block_group(cache);
5955         return ret;
5956 }
5957
5958 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5959 {
5960         int ret;
5961         struct btrfs_block_group_cache *block_group;
5962         struct btrfs_caching_control *caching_ctl;
5963
5964         block_group = btrfs_lookup_block_group(root->fs_info, start);
5965         if (!block_group)
5966                 return -EINVAL;
5967
5968         cache_block_group(block_group, 0);
5969         caching_ctl = get_caching_control(block_group);
5970
5971         if (!caching_ctl) {
5972                 /* Logic error */
5973                 BUG_ON(!block_group_cache_done(block_group));
5974                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5975         } else {
5976                 mutex_lock(&caching_ctl->mutex);
5977
5978                 if (start >= caching_ctl->progress) {
5979                         ret = add_excluded_extent(root, start, num_bytes);
5980                 } else if (start + num_bytes <= caching_ctl->progress) {
5981                         ret = btrfs_remove_free_space(block_group,
5982                                                       start, num_bytes);
5983                 } else {
5984                         num_bytes = caching_ctl->progress - start;
5985                         ret = btrfs_remove_free_space(block_group,
5986                                                       start, num_bytes);
5987                         if (ret)
5988                                 goto out_lock;
5989
5990                         num_bytes = (start + num_bytes) -
5991                                 caching_ctl->progress;
5992                         start = caching_ctl->progress;
5993                         ret = add_excluded_extent(root, start, num_bytes);
5994                 }
5995 out_lock:
5996                 mutex_unlock(&caching_ctl->mutex);
5997                 put_caching_control(caching_ctl);
5998         }
5999         btrfs_put_block_group(block_group);
6000         return ret;
6001 }
6002
6003 int btrfs_exclude_logged_extents(struct btrfs_root *log,
6004                                  struct extent_buffer *eb)
6005 {
6006         struct btrfs_file_extent_item *item;
6007         struct btrfs_key key;
6008         int found_type;
6009         int i;
6010
6011         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
6012                 return 0;
6013
6014         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6015                 btrfs_item_key_to_cpu(eb, &key, i);
6016                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6017                         continue;
6018                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6019                 found_type = btrfs_file_extent_type(eb, item);
6020                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6021                         continue;
6022                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6023                         continue;
6024                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6025                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6026                 __exclude_logged_extent(log, key.objectid, key.offset);
6027         }
6028
6029         return 0;
6030 }
6031
6032 /**
6033  * btrfs_update_reserved_bytes - update the block_group and space info counters
6034  * @cache:      The cache we are manipulating
6035  * @num_bytes:  The number of bytes in question
6036  * @reserve:    One of the reservation enums
6037  * @delalloc:   The blocks are allocated for the delalloc write
6038  *
6039  * This is called by the allocator when it reserves space, or by somebody who is
6040  * freeing space that was never actually used on disk.  For example if you
6041  * reserve some space for a new leaf in transaction A and before transaction A
6042  * commits you free that leaf, you call this with reserve set to 0 in order to
6043  * clear the reservation.
6044  *
6045  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
6046  * ENOSPC accounting.  For data we handle the reservation through clearing the
6047  * delalloc bits in the io_tree.  We have to do this since we could end up
6048  * allocating less disk space for the amount of data we have reserved in the
6049  * case of compression.
6050  *
6051  * If this is a reservation and the block group has become read only we cannot
6052  * make the reservation and return -EAGAIN, otherwise this function always
6053  * succeeds.
6054  */
6055 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
6056                                        u64 num_bytes, int reserve, int delalloc)
6057 {
6058         struct btrfs_space_info *space_info = cache->space_info;
6059         int ret = 0;
6060
6061         spin_lock(&space_info->lock);
6062         spin_lock(&cache->lock);
6063         if (reserve != RESERVE_FREE) {
6064                 if (cache->ro) {
6065                         ret = -EAGAIN;
6066                 } else {
6067                         cache->reserved += num_bytes;
6068                         space_info->bytes_reserved += num_bytes;
6069                         if (reserve == RESERVE_ALLOC) {
6070                                 trace_btrfs_space_reservation(cache->fs_info,
6071                                                 "space_info", space_info->flags,
6072                                                 num_bytes, 0);
6073                                 space_info->bytes_may_use -= num_bytes;
6074                         }
6075
6076                         if (delalloc)
6077                                 cache->delalloc_bytes += num_bytes;
6078                 }
6079         } else {
6080                 if (cache->ro)
6081                         space_info->bytes_readonly += num_bytes;
6082                 cache->reserved -= num_bytes;
6083                 space_info->bytes_reserved -= num_bytes;
6084
6085                 if (delalloc)
6086                         cache->delalloc_bytes -= num_bytes;
6087         }
6088         spin_unlock(&cache->lock);
6089         spin_unlock(&space_info->lock);
6090         return ret;
6091 }
6092
6093 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6094                                 struct btrfs_root *root)
6095 {
6096         struct btrfs_fs_info *fs_info = root->fs_info;
6097         struct btrfs_caching_control *next;
6098         struct btrfs_caching_control *caching_ctl;
6099         struct btrfs_block_group_cache *cache;
6100
6101         down_write(&fs_info->commit_root_sem);
6102
6103         list_for_each_entry_safe(caching_ctl, next,
6104                                  &fs_info->caching_block_groups, list) {
6105                 cache = caching_ctl->block_group;
6106                 if (block_group_cache_done(cache)) {
6107                         cache->last_byte_to_unpin = (u64)-1;
6108                         list_del_init(&caching_ctl->list);
6109                         put_caching_control(caching_ctl);
6110                 } else {
6111                         cache->last_byte_to_unpin = caching_ctl->progress;
6112                 }
6113         }
6114
6115         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6116                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6117         else
6118                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6119
6120         up_write(&fs_info->commit_root_sem);
6121
6122         update_global_block_rsv(fs_info);
6123 }
6124
6125 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
6126                               const bool return_free_space)
6127 {
6128         struct btrfs_fs_info *fs_info = root->fs_info;
6129         struct btrfs_block_group_cache *cache = NULL;
6130         struct btrfs_space_info *space_info;
6131         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6132         u64 len;
6133         bool readonly;
6134
6135         while (start <= end) {
6136                 readonly = false;
6137                 if (!cache ||
6138                     start >= cache->key.objectid + cache->key.offset) {
6139                         if (cache)
6140                                 btrfs_put_block_group(cache);
6141                         cache = btrfs_lookup_block_group(fs_info, start);
6142                         BUG_ON(!cache); /* Logic error */
6143                 }
6144
6145                 len = cache->key.objectid + cache->key.offset - start;
6146                 len = min(len, end + 1 - start);
6147
6148                 if (start < cache->last_byte_to_unpin) {
6149                         len = min(len, cache->last_byte_to_unpin - start);
6150                         if (return_free_space)
6151                                 btrfs_add_free_space(cache, start, len);
6152                 }
6153
6154                 start += len;
6155                 space_info = cache->space_info;
6156
6157                 spin_lock(&space_info->lock);
6158                 spin_lock(&cache->lock);
6159                 cache->pinned -= len;
6160                 space_info->bytes_pinned -= len;
6161                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6162                 if (cache->ro) {
6163                         space_info->bytes_readonly += len;
6164                         readonly = true;
6165                 }
6166                 spin_unlock(&cache->lock);
6167                 if (!readonly && global_rsv->space_info == space_info) {
6168                         spin_lock(&global_rsv->lock);
6169                         if (!global_rsv->full) {
6170                                 len = min(len, global_rsv->size -
6171                                           global_rsv->reserved);
6172                                 global_rsv->reserved += len;
6173                                 space_info->bytes_may_use += len;
6174                                 if (global_rsv->reserved >= global_rsv->size)
6175                                         global_rsv->full = 1;
6176                         }
6177                         spin_unlock(&global_rsv->lock);
6178                 }
6179                 spin_unlock(&space_info->lock);
6180         }
6181
6182         if (cache)
6183                 btrfs_put_block_group(cache);
6184         return 0;
6185 }
6186
6187 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6188                                struct btrfs_root *root)
6189 {
6190         struct btrfs_fs_info *fs_info = root->fs_info;
6191         struct btrfs_block_group_cache *block_group, *tmp;
6192         struct list_head *deleted_bgs;
6193         struct extent_io_tree *unpin;
6194         u64 start;
6195         u64 end;
6196         int ret;
6197
6198         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6199                 unpin = &fs_info->freed_extents[1];
6200         else
6201                 unpin = &fs_info->freed_extents[0];
6202
6203         while (!trans->aborted) {
6204                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6205                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6206                                             EXTENT_DIRTY, NULL);
6207                 if (ret) {
6208                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6209                         break;
6210                 }
6211
6212                 if (btrfs_test_opt(root, DISCARD))
6213                         ret = btrfs_discard_extent(root, start,
6214                                                    end + 1 - start, NULL);
6215
6216                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
6217                 unpin_extent_range(root, start, end, true);
6218                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6219                 cond_resched();
6220         }
6221
6222         /*
6223          * Transaction is finished.  We don't need the lock anymore.  We
6224          * do need to clean up the block groups in case of a transaction
6225          * abort.
6226          */
6227         deleted_bgs = &trans->transaction->deleted_bgs;
6228         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6229                 u64 trimmed = 0;
6230
6231                 ret = -EROFS;
6232                 if (!trans->aborted)
6233                         ret = btrfs_discard_extent(root,
6234                                                    block_group->key.objectid,
6235                                                    block_group->key.offset,
6236                                                    &trimmed);
6237
6238                 list_del_init(&block_group->bg_list);
6239                 btrfs_put_block_group_trimming(block_group);
6240                 btrfs_put_block_group(block_group);
6241
6242                 if (ret) {
6243                         const char *errstr = btrfs_decode_error(ret);
6244                         btrfs_warn(fs_info,
6245                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6246                                    ret, errstr);
6247                 }
6248         }
6249
6250         return 0;
6251 }
6252
6253 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6254                              u64 owner, u64 root_objectid)
6255 {
6256         struct btrfs_space_info *space_info;
6257         u64 flags;
6258
6259         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6260                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6261                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6262                 else
6263                         flags = BTRFS_BLOCK_GROUP_METADATA;
6264         } else {
6265                 flags = BTRFS_BLOCK_GROUP_DATA;
6266         }
6267
6268         space_info = __find_space_info(fs_info, flags);
6269         BUG_ON(!space_info); /* Logic bug */
6270         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6271 }
6272
6273
6274 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6275                                 struct btrfs_root *root,
6276                                 struct btrfs_delayed_ref_node *node, u64 parent,
6277                                 u64 root_objectid, u64 owner_objectid,
6278                                 u64 owner_offset, int refs_to_drop,
6279                                 struct btrfs_delayed_extent_op *extent_op)
6280 {
6281         struct btrfs_key key;
6282         struct btrfs_path *path;
6283         struct btrfs_fs_info *info = root->fs_info;
6284         struct btrfs_root *extent_root = info->extent_root;
6285         struct extent_buffer *leaf;
6286         struct btrfs_extent_item *ei;
6287         struct btrfs_extent_inline_ref *iref;
6288         int ret;
6289         int is_data;
6290         int extent_slot = 0;
6291         int found_extent = 0;
6292         int num_to_del = 1;
6293         int no_quota = node->no_quota;
6294         u32 item_size;
6295         u64 refs;
6296         u64 bytenr = node->bytenr;
6297         u64 num_bytes = node->num_bytes;
6298         int last_ref = 0;
6299         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6300                                                  SKINNY_METADATA);
6301
6302         if (!info->quota_enabled || !is_fstree(root_objectid))
6303                 no_quota = 1;
6304
6305         path = btrfs_alloc_path();
6306         if (!path)
6307                 return -ENOMEM;
6308
6309         path->reada = 1;
6310         path->leave_spinning = 1;
6311
6312         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6313         BUG_ON(!is_data && refs_to_drop != 1);
6314
6315         if (is_data)
6316                 skinny_metadata = 0;
6317
6318         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6319                                     bytenr, num_bytes, parent,
6320                                     root_objectid, owner_objectid,
6321                                     owner_offset);
6322         if (ret == 0) {
6323                 extent_slot = path->slots[0];
6324                 while (extent_slot >= 0) {
6325                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6326                                               extent_slot);
6327                         if (key.objectid != bytenr)
6328                                 break;
6329                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6330                             key.offset == num_bytes) {
6331                                 found_extent = 1;
6332                                 break;
6333                         }
6334                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6335                             key.offset == owner_objectid) {
6336                                 found_extent = 1;
6337                                 break;
6338                         }
6339                         if (path->slots[0] - extent_slot > 5)
6340                                 break;
6341                         extent_slot--;
6342                 }
6343 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6344                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6345                 if (found_extent && item_size < sizeof(*ei))
6346                         found_extent = 0;
6347 #endif
6348                 if (!found_extent) {
6349                         BUG_ON(iref);
6350                         ret = remove_extent_backref(trans, extent_root, path,
6351                                                     NULL, refs_to_drop,
6352                                                     is_data, &last_ref);
6353                         if (ret) {
6354                                 btrfs_abort_transaction(trans, extent_root, ret);
6355                                 goto out;
6356                         }
6357                         btrfs_release_path(path);
6358                         path->leave_spinning = 1;
6359
6360                         key.objectid = bytenr;
6361                         key.type = BTRFS_EXTENT_ITEM_KEY;
6362                         key.offset = num_bytes;
6363
6364                         if (!is_data && skinny_metadata) {
6365                                 key.type = BTRFS_METADATA_ITEM_KEY;
6366                                 key.offset = owner_objectid;
6367                         }
6368
6369                         ret = btrfs_search_slot(trans, extent_root,
6370                                                 &key, path, -1, 1);
6371                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6372                                 /*
6373                                  * Couldn't find our skinny metadata item,
6374                                  * see if we have ye olde extent item.
6375                                  */
6376                                 path->slots[0]--;
6377                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6378                                                       path->slots[0]);
6379                                 if (key.objectid == bytenr &&
6380                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6381                                     key.offset == num_bytes)
6382                                         ret = 0;
6383                         }
6384
6385                         if (ret > 0 && skinny_metadata) {
6386                                 skinny_metadata = false;
6387                                 key.objectid = bytenr;
6388                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6389                                 key.offset = num_bytes;
6390                                 btrfs_release_path(path);
6391                                 ret = btrfs_search_slot(trans, extent_root,
6392                                                         &key, path, -1, 1);
6393                         }
6394
6395                         if (ret) {
6396                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6397                                         ret, bytenr);
6398                                 if (ret > 0)
6399                                         btrfs_print_leaf(extent_root,
6400                                                          path->nodes[0]);
6401                         }
6402                         if (ret < 0) {
6403                                 btrfs_abort_transaction(trans, extent_root, ret);
6404                                 goto out;
6405                         }
6406                         extent_slot = path->slots[0];
6407                 }
6408         } else if (WARN_ON(ret == -ENOENT)) {
6409                 btrfs_print_leaf(extent_root, path->nodes[0]);
6410                 btrfs_err(info,
6411                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6412                         bytenr, parent, root_objectid, owner_objectid,
6413                         owner_offset);
6414                 btrfs_abort_transaction(trans, extent_root, ret);
6415                 goto out;
6416         } else {
6417                 btrfs_abort_transaction(trans, extent_root, ret);
6418                 goto out;
6419         }
6420
6421         leaf = path->nodes[0];
6422         item_size = btrfs_item_size_nr(leaf, extent_slot);
6423 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6424         if (item_size < sizeof(*ei)) {
6425                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6426                 ret = convert_extent_item_v0(trans, extent_root, path,
6427                                              owner_objectid, 0);
6428                 if (ret < 0) {
6429                         btrfs_abort_transaction(trans, extent_root, ret);
6430                         goto out;
6431                 }
6432
6433                 btrfs_release_path(path);
6434                 path->leave_spinning = 1;
6435
6436                 key.objectid = bytenr;
6437                 key.type = BTRFS_EXTENT_ITEM_KEY;
6438                 key.offset = num_bytes;
6439
6440                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6441                                         -1, 1);
6442                 if (ret) {
6443                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6444                                 ret, bytenr);
6445                         btrfs_print_leaf(extent_root, path->nodes[0]);
6446                 }
6447                 if (ret < 0) {
6448                         btrfs_abort_transaction(trans, extent_root, ret);
6449                         goto out;
6450                 }
6451
6452                 extent_slot = path->slots[0];
6453                 leaf = path->nodes[0];
6454                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6455         }
6456 #endif
6457         BUG_ON(item_size < sizeof(*ei));
6458         ei = btrfs_item_ptr(leaf, extent_slot,
6459                             struct btrfs_extent_item);
6460         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6461             key.type == BTRFS_EXTENT_ITEM_KEY) {
6462                 struct btrfs_tree_block_info *bi;
6463                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6464                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6465                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6466         }
6467
6468         refs = btrfs_extent_refs(leaf, ei);
6469         if (refs < refs_to_drop) {
6470                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6471                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6472                 ret = -EINVAL;
6473                 btrfs_abort_transaction(trans, extent_root, ret);
6474                 goto out;
6475         }
6476         refs -= refs_to_drop;
6477
6478         if (refs > 0) {
6479                 if (extent_op)
6480                         __run_delayed_extent_op(extent_op, leaf, ei);
6481                 /*
6482                  * In the case of inline back ref, reference count will
6483                  * be updated by remove_extent_backref
6484                  */
6485                 if (iref) {
6486                         BUG_ON(!found_extent);
6487                 } else {
6488                         btrfs_set_extent_refs(leaf, ei, refs);
6489                         btrfs_mark_buffer_dirty(leaf);
6490                 }
6491                 if (found_extent) {
6492                         ret = remove_extent_backref(trans, extent_root, path,
6493                                                     iref, refs_to_drop,
6494                                                     is_data, &last_ref);
6495                         if (ret) {
6496                                 btrfs_abort_transaction(trans, extent_root, ret);
6497                                 goto out;
6498                         }
6499                 }
6500                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6501                                  root_objectid);
6502         } else {
6503                 if (found_extent) {
6504                         BUG_ON(is_data && refs_to_drop !=
6505                                extent_data_ref_count(path, iref));
6506                         if (iref) {
6507                                 BUG_ON(path->slots[0] != extent_slot);
6508                         } else {
6509                                 BUG_ON(path->slots[0] != extent_slot + 1);
6510                                 path->slots[0] = extent_slot;
6511                                 num_to_del = 2;
6512                         }
6513                 }
6514
6515                 last_ref = 1;
6516                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6517                                       num_to_del);
6518                 if (ret) {
6519                         btrfs_abort_transaction(trans, extent_root, ret);
6520                         goto out;
6521                 }
6522                 btrfs_release_path(path);
6523
6524                 if (is_data) {
6525                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6526                         if (ret) {
6527                                 btrfs_abort_transaction(trans, extent_root, ret);
6528                                 goto out;
6529                         }
6530                 }
6531
6532                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6533                 if (ret) {
6534                         btrfs_abort_transaction(trans, extent_root, ret);
6535                         goto out;
6536                 }
6537         }
6538         btrfs_release_path(path);
6539
6540 out:
6541         btrfs_free_path(path);
6542         return ret;
6543 }
6544
6545 /*
6546  * when we free an block, it is possible (and likely) that we free the last
6547  * delayed ref for that extent as well.  This searches the delayed ref tree for
6548  * a given extent, and if there are no other delayed refs to be processed, it
6549  * removes it from the tree.
6550  */
6551 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6552                                       struct btrfs_root *root, u64 bytenr)
6553 {
6554         struct btrfs_delayed_ref_head *head;
6555         struct btrfs_delayed_ref_root *delayed_refs;
6556         int ret = 0;
6557
6558         delayed_refs = &trans->transaction->delayed_refs;
6559         spin_lock(&delayed_refs->lock);
6560         head = btrfs_find_delayed_ref_head(trans, bytenr);
6561         if (!head)
6562                 goto out_delayed_unlock;
6563
6564         spin_lock(&head->lock);
6565         if (!list_empty(&head->ref_list))
6566                 goto out;
6567
6568         if (head->extent_op) {
6569                 if (!head->must_insert_reserved)
6570                         goto out;
6571                 btrfs_free_delayed_extent_op(head->extent_op);
6572                 head->extent_op = NULL;
6573         }
6574
6575         /*
6576          * waiting for the lock here would deadlock.  If someone else has it
6577          * locked they are already in the process of dropping it anyway
6578          */
6579         if (!mutex_trylock(&head->mutex))
6580                 goto out;
6581
6582         /*
6583          * at this point we have a head with no other entries.  Go
6584          * ahead and process it.
6585          */
6586         head->node.in_tree = 0;
6587         rb_erase(&head->href_node, &delayed_refs->href_root);
6588
6589         atomic_dec(&delayed_refs->num_entries);
6590
6591         /*
6592          * we don't take a ref on the node because we're removing it from the
6593          * tree, so we just steal the ref the tree was holding.
6594          */
6595         delayed_refs->num_heads--;
6596         if (head->processing == 0)
6597                 delayed_refs->num_heads_ready--;
6598         head->processing = 0;
6599         spin_unlock(&head->lock);
6600         spin_unlock(&delayed_refs->lock);
6601
6602         BUG_ON(head->extent_op);
6603         if (head->must_insert_reserved)
6604                 ret = 1;
6605
6606         mutex_unlock(&head->mutex);
6607         btrfs_put_delayed_ref(&head->node);
6608         return ret;
6609 out:
6610         spin_unlock(&head->lock);
6611
6612 out_delayed_unlock:
6613         spin_unlock(&delayed_refs->lock);
6614         return 0;
6615 }
6616
6617 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6618                            struct btrfs_root *root,
6619                            struct extent_buffer *buf,
6620                            u64 parent, int last_ref)
6621 {
6622         int pin = 1;
6623         int ret;
6624
6625         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6626                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6627                                         buf->start, buf->len,
6628                                         parent, root->root_key.objectid,
6629                                         btrfs_header_level(buf),
6630                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6631                 BUG_ON(ret); /* -ENOMEM */
6632         }
6633
6634         if (!last_ref)
6635                 return;
6636
6637         if (btrfs_header_generation(buf) == trans->transid) {
6638                 struct btrfs_block_group_cache *cache;
6639
6640                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6641                         ret = check_ref_cleanup(trans, root, buf->start);
6642                         if (!ret)
6643                                 goto out;
6644                 }
6645
6646                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6647
6648                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6649                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6650                         btrfs_put_block_group(cache);
6651                         goto out;
6652                 }
6653
6654                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6655
6656                 btrfs_add_free_space(cache, buf->start, buf->len);
6657                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6658                 btrfs_put_block_group(cache);
6659                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6660                 pin = 0;
6661         }
6662 out:
6663         if (pin)
6664                 add_pinned_bytes(root->fs_info, buf->len,
6665                                  btrfs_header_level(buf),
6666                                  root->root_key.objectid);
6667
6668         /*
6669          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6670          * anymore.
6671          */
6672         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6673 }
6674
6675 /* Can return -ENOMEM */
6676 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6677                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6678                       u64 owner, u64 offset, int no_quota)
6679 {
6680         int ret;
6681         struct btrfs_fs_info *fs_info = root->fs_info;
6682
6683         if (btrfs_test_is_dummy_root(root))
6684                 return 0;
6685
6686         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6687
6688         /*
6689          * tree log blocks never actually go into the extent allocation
6690          * tree, just update pinning info and exit early.
6691          */
6692         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6693                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6694                 /* unlocks the pinned mutex */
6695                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6696                 ret = 0;
6697         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6698                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6699                                         num_bytes,
6700                                         parent, root_objectid, (int)owner,
6701                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6702         } else {
6703                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6704                                                 num_bytes,
6705                                                 parent, root_objectid, owner,
6706                                                 offset, BTRFS_DROP_DELAYED_REF,
6707                                                 NULL, no_quota);
6708         }
6709         return ret;
6710 }
6711
6712 /*
6713  * when we wait for progress in the block group caching, its because
6714  * our allocation attempt failed at least once.  So, we must sleep
6715  * and let some progress happen before we try again.
6716  *
6717  * This function will sleep at least once waiting for new free space to
6718  * show up, and then it will check the block group free space numbers
6719  * for our min num_bytes.  Another option is to have it go ahead
6720  * and look in the rbtree for a free extent of a given size, but this
6721  * is a good start.
6722  *
6723  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6724  * any of the information in this block group.
6725  */
6726 static noinline void
6727 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6728                                 u64 num_bytes)
6729 {
6730         struct btrfs_caching_control *caching_ctl;
6731
6732         caching_ctl = get_caching_control(cache);
6733         if (!caching_ctl)
6734                 return;
6735
6736         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6737                    (cache->free_space_ctl->free_space >= num_bytes));
6738
6739         put_caching_control(caching_ctl);
6740 }
6741
6742 static noinline int
6743 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6744 {
6745         struct btrfs_caching_control *caching_ctl;
6746         int ret = 0;
6747
6748         caching_ctl = get_caching_control(cache);
6749         if (!caching_ctl)
6750                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6751
6752         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6753         if (cache->cached == BTRFS_CACHE_ERROR)
6754                 ret = -EIO;
6755         put_caching_control(caching_ctl);
6756         return ret;
6757 }
6758
6759 int __get_raid_index(u64 flags)
6760 {
6761         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6762                 return BTRFS_RAID_RAID10;
6763         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6764                 return BTRFS_RAID_RAID1;
6765         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6766                 return BTRFS_RAID_DUP;
6767         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6768                 return BTRFS_RAID_RAID0;
6769         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6770                 return BTRFS_RAID_RAID5;
6771         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6772                 return BTRFS_RAID_RAID6;
6773
6774         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6775 }
6776
6777 int get_block_group_index(struct btrfs_block_group_cache *cache)
6778 {
6779         return __get_raid_index(cache->flags);
6780 }
6781
6782 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6783         [BTRFS_RAID_RAID10]     = "raid10",
6784         [BTRFS_RAID_RAID1]      = "raid1",
6785         [BTRFS_RAID_DUP]        = "dup",
6786         [BTRFS_RAID_RAID0]      = "raid0",
6787         [BTRFS_RAID_SINGLE]     = "single",
6788         [BTRFS_RAID_RAID5]      = "raid5",
6789         [BTRFS_RAID_RAID6]      = "raid6",
6790 };
6791
6792 static const char *get_raid_name(enum btrfs_raid_types type)
6793 {
6794         if (type >= BTRFS_NR_RAID_TYPES)
6795                 return NULL;
6796
6797         return btrfs_raid_type_names[type];
6798 }
6799
6800 enum btrfs_loop_type {
6801         LOOP_CACHING_NOWAIT = 0,
6802         LOOP_CACHING_WAIT = 1,
6803         LOOP_ALLOC_CHUNK = 2,
6804         LOOP_NO_EMPTY_SIZE = 3,
6805 };
6806
6807 static inline void
6808 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6809                        int delalloc)
6810 {
6811         if (delalloc)
6812                 down_read(&cache->data_rwsem);
6813 }
6814
6815 static inline void
6816 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6817                        int delalloc)
6818 {
6819         btrfs_get_block_group(cache);
6820         if (delalloc)
6821                 down_read(&cache->data_rwsem);
6822 }
6823
6824 static struct btrfs_block_group_cache *
6825 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6826                    struct btrfs_free_cluster *cluster,
6827                    int delalloc)
6828 {
6829         struct btrfs_block_group_cache *used_bg;
6830         bool locked = false;
6831 again:
6832         spin_lock(&cluster->refill_lock);
6833         if (locked) {
6834                 if (used_bg == cluster->block_group)
6835                         return used_bg;
6836
6837                 up_read(&used_bg->data_rwsem);
6838                 btrfs_put_block_group(used_bg);
6839         }
6840
6841         used_bg = cluster->block_group;
6842         if (!used_bg)
6843                 return NULL;
6844
6845         if (used_bg == block_group)
6846                 return used_bg;
6847
6848         btrfs_get_block_group(used_bg);
6849
6850         if (!delalloc)
6851                 return used_bg;
6852
6853         if (down_read_trylock(&used_bg->data_rwsem))
6854                 return used_bg;
6855
6856         spin_unlock(&cluster->refill_lock);
6857         down_read(&used_bg->data_rwsem);
6858         locked = true;
6859         goto again;
6860 }
6861
6862 static inline void
6863 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6864                          int delalloc)
6865 {
6866         if (delalloc)
6867                 up_read(&cache->data_rwsem);
6868         btrfs_put_block_group(cache);
6869 }
6870
6871 /*
6872  * walks the btree of allocated extents and find a hole of a given size.
6873  * The key ins is changed to record the hole:
6874  * ins->objectid == start position
6875  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6876  * ins->offset == the size of the hole.
6877  * Any available blocks before search_start are skipped.
6878  *
6879  * If there is no suitable free space, we will record the max size of
6880  * the free space extent currently.
6881  */
6882 static noinline int find_free_extent(struct btrfs_root *orig_root,
6883                                      u64 num_bytes, u64 empty_size,
6884                                      u64 hint_byte, struct btrfs_key *ins,
6885                                      u64 flags, int delalloc)
6886 {
6887         int ret = 0;
6888         struct btrfs_root *root = orig_root->fs_info->extent_root;
6889         struct btrfs_free_cluster *last_ptr = NULL;
6890         struct btrfs_block_group_cache *block_group = NULL;
6891         u64 search_start = 0;
6892         u64 max_extent_size = 0;
6893         int empty_cluster = 2 * 1024 * 1024;
6894         struct btrfs_space_info *space_info;
6895         int loop = 0;
6896         int index = __get_raid_index(flags);
6897         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6898                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6899         bool failed_cluster_refill = false;
6900         bool failed_alloc = false;
6901         bool use_cluster = true;
6902         bool have_caching_bg = false;
6903
6904         WARN_ON(num_bytes < root->sectorsize);
6905         ins->type = BTRFS_EXTENT_ITEM_KEY;
6906         ins->objectid = 0;
6907         ins->offset = 0;
6908
6909         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6910
6911         space_info = __find_space_info(root->fs_info, flags);
6912         if (!space_info) {
6913                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6914                 return -ENOSPC;
6915         }
6916
6917         /*
6918          * If the space info is for both data and metadata it means we have a
6919          * small filesystem and we can't use the clustering stuff.
6920          */
6921         if (btrfs_mixed_space_info(space_info))
6922                 use_cluster = false;
6923
6924         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6925                 last_ptr = &root->fs_info->meta_alloc_cluster;
6926                 if (!btrfs_test_opt(root, SSD))
6927                         empty_cluster = 64 * 1024;
6928         }
6929
6930         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6931             btrfs_test_opt(root, SSD)) {
6932                 last_ptr = &root->fs_info->data_alloc_cluster;
6933         }
6934
6935         if (last_ptr) {
6936                 spin_lock(&last_ptr->lock);
6937                 if (last_ptr->block_group)
6938                         hint_byte = last_ptr->window_start;
6939                 spin_unlock(&last_ptr->lock);
6940         }
6941
6942         search_start = max(search_start, first_logical_byte(root, 0));
6943         search_start = max(search_start, hint_byte);
6944
6945         if (!last_ptr)
6946                 empty_cluster = 0;
6947
6948         if (search_start == hint_byte) {
6949                 block_group = btrfs_lookup_block_group(root->fs_info,
6950                                                        search_start);
6951                 /*
6952                  * we don't want to use the block group if it doesn't match our
6953                  * allocation bits, or if its not cached.
6954                  *
6955                  * However if we are re-searching with an ideal block group
6956                  * picked out then we don't care that the block group is cached.
6957                  */
6958                 if (block_group && block_group_bits(block_group, flags) &&
6959                     block_group->cached != BTRFS_CACHE_NO) {
6960                         down_read(&space_info->groups_sem);
6961                         if (list_empty(&block_group->list) ||
6962                             block_group->ro) {
6963                                 /*
6964                                  * someone is removing this block group,
6965                                  * we can't jump into the have_block_group
6966                                  * target because our list pointers are not
6967                                  * valid
6968                                  */
6969                                 btrfs_put_block_group(block_group);
6970                                 up_read(&space_info->groups_sem);
6971                         } else {
6972                                 index = get_block_group_index(block_group);
6973                                 btrfs_lock_block_group(block_group, delalloc);
6974                                 goto have_block_group;
6975                         }
6976                 } else if (block_group) {
6977                         btrfs_put_block_group(block_group);
6978                 }
6979         }
6980 search:
6981         have_caching_bg = false;
6982         down_read(&space_info->groups_sem);
6983         list_for_each_entry(block_group, &space_info->block_groups[index],
6984                             list) {
6985                 u64 offset;
6986                 int cached;
6987
6988                 btrfs_grab_block_group(block_group, delalloc);
6989                 search_start = block_group->key.objectid;
6990
6991                 /*
6992                  * this can happen if we end up cycling through all the
6993                  * raid types, but we want to make sure we only allocate
6994                  * for the proper type.
6995                  */
6996                 if (!block_group_bits(block_group, flags)) {
6997                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6998                                 BTRFS_BLOCK_GROUP_RAID1 |
6999                                 BTRFS_BLOCK_GROUP_RAID5 |
7000                                 BTRFS_BLOCK_GROUP_RAID6 |
7001                                 BTRFS_BLOCK_GROUP_RAID10;
7002
7003                         /*
7004                          * if they asked for extra copies and this block group
7005                          * doesn't provide them, bail.  This does allow us to
7006                          * fill raid0 from raid1.
7007                          */
7008                         if ((flags & extra) && !(block_group->flags & extra))
7009                                 goto loop;
7010                 }
7011
7012 have_block_group:
7013                 cached = block_group_cache_done(block_group);
7014                 if (unlikely(!cached)) {
7015                         ret = cache_block_group(block_group, 0);
7016                         BUG_ON(ret < 0);
7017                         ret = 0;
7018                 }
7019
7020                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7021                         goto loop;
7022                 if (unlikely(block_group->ro))
7023                         goto loop;
7024
7025                 /*
7026                  * Ok we want to try and use the cluster allocator, so
7027                  * lets look there
7028                  */
7029                 if (last_ptr) {
7030                         struct btrfs_block_group_cache *used_block_group;
7031                         unsigned long aligned_cluster;
7032                         /*
7033                          * the refill lock keeps out other
7034                          * people trying to start a new cluster
7035                          */
7036                         used_block_group = btrfs_lock_cluster(block_group,
7037                                                               last_ptr,
7038                                                               delalloc);
7039                         if (!used_block_group)
7040                                 goto refill_cluster;
7041
7042                         if (used_block_group != block_group &&
7043                             (used_block_group->ro ||
7044                              !block_group_bits(used_block_group, flags)))
7045                                 goto release_cluster;
7046
7047                         offset = btrfs_alloc_from_cluster(used_block_group,
7048                                                 last_ptr,
7049                                                 num_bytes,
7050                                                 used_block_group->key.objectid,
7051                                                 &max_extent_size);
7052                         if (offset) {
7053                                 /* we have a block, we're done */
7054                                 spin_unlock(&last_ptr->refill_lock);
7055                                 trace_btrfs_reserve_extent_cluster(root,
7056                                                 used_block_group,
7057                                                 search_start, num_bytes);
7058                                 if (used_block_group != block_group) {
7059                                         btrfs_release_block_group(block_group,
7060                                                                   delalloc);
7061                                         block_group = used_block_group;
7062                                 }
7063                                 goto checks;
7064                         }
7065
7066                         WARN_ON(last_ptr->block_group != used_block_group);
7067 release_cluster:
7068                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7069                          * set up a new clusters, so lets just skip it
7070                          * and let the allocator find whatever block
7071                          * it can find.  If we reach this point, we
7072                          * will have tried the cluster allocator
7073                          * plenty of times and not have found
7074                          * anything, so we are likely way too
7075                          * fragmented for the clustering stuff to find
7076                          * anything.
7077                          *
7078                          * However, if the cluster is taken from the
7079                          * current block group, release the cluster
7080                          * first, so that we stand a better chance of
7081                          * succeeding in the unclustered
7082                          * allocation.  */
7083                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7084                             used_block_group != block_group) {
7085                                 spin_unlock(&last_ptr->refill_lock);
7086                                 btrfs_release_block_group(used_block_group,
7087                                                           delalloc);
7088                                 goto unclustered_alloc;
7089                         }
7090
7091                         /*
7092                          * this cluster didn't work out, free it and
7093                          * start over
7094                          */
7095                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7096
7097                         if (used_block_group != block_group)
7098                                 btrfs_release_block_group(used_block_group,
7099                                                           delalloc);
7100 refill_cluster:
7101                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7102                                 spin_unlock(&last_ptr->refill_lock);
7103                                 goto unclustered_alloc;
7104                         }
7105
7106                         aligned_cluster = max_t(unsigned long,
7107                                                 empty_cluster + empty_size,
7108                                               block_group->full_stripe_len);
7109
7110                         /* allocate a cluster in this block group */
7111                         ret = btrfs_find_space_cluster(root, block_group,
7112                                                        last_ptr, search_start,
7113                                                        num_bytes,
7114                                                        aligned_cluster);
7115                         if (ret == 0) {
7116                                 /*
7117                                  * now pull our allocation out of this
7118                                  * cluster
7119                                  */
7120                                 offset = btrfs_alloc_from_cluster(block_group,
7121                                                         last_ptr,
7122                                                         num_bytes,
7123                                                         search_start,
7124                                                         &max_extent_size);
7125                                 if (offset) {
7126                                         /* we found one, proceed */
7127                                         spin_unlock(&last_ptr->refill_lock);
7128                                         trace_btrfs_reserve_extent_cluster(root,
7129                                                 block_group, search_start,
7130                                                 num_bytes);
7131                                         goto checks;
7132                                 }
7133                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7134                                    && !failed_cluster_refill) {
7135                                 spin_unlock(&last_ptr->refill_lock);
7136
7137                                 failed_cluster_refill = true;
7138                                 wait_block_group_cache_progress(block_group,
7139                                        num_bytes + empty_cluster + empty_size);
7140                                 goto have_block_group;
7141                         }
7142
7143                         /*
7144                          * at this point we either didn't find a cluster
7145                          * or we weren't able to allocate a block from our
7146                          * cluster.  Free the cluster we've been trying
7147                          * to use, and go to the next block group
7148                          */
7149                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7150                         spin_unlock(&last_ptr->refill_lock);
7151                         goto loop;
7152                 }
7153
7154 unclustered_alloc:
7155                 spin_lock(&block_group->free_space_ctl->tree_lock);
7156                 if (cached &&
7157                     block_group->free_space_ctl->free_space <
7158                     num_bytes + empty_cluster + empty_size) {
7159                         if (block_group->free_space_ctl->free_space >
7160                             max_extent_size)
7161                                 max_extent_size =
7162                                         block_group->free_space_ctl->free_space;
7163                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7164                         goto loop;
7165                 }
7166                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7167
7168                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7169                                                     num_bytes, empty_size,
7170                                                     &max_extent_size);
7171                 /*
7172                  * If we didn't find a chunk, and we haven't failed on this
7173                  * block group before, and this block group is in the middle of
7174                  * caching and we are ok with waiting, then go ahead and wait
7175                  * for progress to be made, and set failed_alloc to true.
7176                  *
7177                  * If failed_alloc is true then we've already waited on this
7178                  * block group once and should move on to the next block group.
7179                  */
7180                 if (!offset && !failed_alloc && !cached &&
7181                     loop > LOOP_CACHING_NOWAIT) {
7182                         wait_block_group_cache_progress(block_group,
7183                                                 num_bytes + empty_size);
7184                         failed_alloc = true;
7185                         goto have_block_group;
7186                 } else if (!offset) {
7187                         if (!cached)
7188                                 have_caching_bg = true;
7189                         goto loop;
7190                 }
7191 checks:
7192                 search_start = ALIGN(offset, root->stripesize);
7193
7194                 /* move on to the next group */
7195                 if (search_start + num_bytes >
7196                     block_group->key.objectid + block_group->key.offset) {
7197                         btrfs_add_free_space(block_group, offset, num_bytes);
7198                         goto loop;
7199                 }
7200
7201                 if (offset < search_start)
7202                         btrfs_add_free_space(block_group, offset,
7203                                              search_start - offset);
7204                 BUG_ON(offset > search_start);
7205
7206                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7207                                                   alloc_type, delalloc);
7208                 if (ret == -EAGAIN) {
7209                         btrfs_add_free_space(block_group, offset, num_bytes);
7210                         goto loop;
7211                 }
7212
7213                 /* we are all good, lets return */
7214                 ins->objectid = search_start;
7215                 ins->offset = num_bytes;
7216
7217                 trace_btrfs_reserve_extent(orig_root, block_group,
7218                                            search_start, num_bytes);
7219                 btrfs_release_block_group(block_group, delalloc);
7220                 break;
7221 loop:
7222                 failed_cluster_refill = false;
7223                 failed_alloc = false;
7224                 BUG_ON(index != get_block_group_index(block_group));
7225                 btrfs_release_block_group(block_group, delalloc);
7226         }
7227         up_read(&space_info->groups_sem);
7228
7229         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7230                 goto search;
7231
7232         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7233                 goto search;
7234
7235         /*
7236          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7237          *                      caching kthreads as we move along
7238          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7239          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7240          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7241          *                      again
7242          */
7243         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7244                 index = 0;
7245                 loop++;
7246                 if (loop == LOOP_ALLOC_CHUNK) {
7247                         struct btrfs_trans_handle *trans;
7248                         int exist = 0;
7249
7250                         trans = current->journal_info;
7251                         if (trans)
7252                                 exist = 1;
7253                         else
7254                                 trans = btrfs_join_transaction(root);
7255
7256                         if (IS_ERR(trans)) {
7257                                 ret = PTR_ERR(trans);
7258                                 goto out;
7259                         }
7260
7261                         ret = do_chunk_alloc(trans, root, flags,
7262                                              CHUNK_ALLOC_FORCE);
7263                         /*
7264                          * Do not bail out on ENOSPC since we
7265                          * can do more things.
7266                          */
7267                         if (ret < 0 && ret != -ENOSPC)
7268                                 btrfs_abort_transaction(trans,
7269                                                         root, ret);
7270                         else
7271                                 ret = 0;
7272                         if (!exist)
7273                                 btrfs_end_transaction(trans, root);
7274                         if (ret)
7275                                 goto out;
7276                 }
7277
7278                 if (loop == LOOP_NO_EMPTY_SIZE) {
7279                         empty_size = 0;
7280                         empty_cluster = 0;
7281                 }
7282
7283                 goto search;
7284         } else if (!ins->objectid) {
7285                 ret = -ENOSPC;
7286         } else if (ins->objectid) {
7287                 ret = 0;
7288         }
7289 out:
7290         if (ret == -ENOSPC)
7291                 ins->offset = max_extent_size;
7292         return ret;
7293 }
7294
7295 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7296                             int dump_block_groups)
7297 {
7298         struct btrfs_block_group_cache *cache;
7299         int index = 0;
7300
7301         spin_lock(&info->lock);
7302         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7303                info->flags,
7304                info->total_bytes - info->bytes_used - info->bytes_pinned -
7305                info->bytes_reserved - info->bytes_readonly,
7306                (info->full) ? "" : "not ");
7307         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7308                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7309                info->total_bytes, info->bytes_used, info->bytes_pinned,
7310                info->bytes_reserved, info->bytes_may_use,
7311                info->bytes_readonly);
7312         spin_unlock(&info->lock);
7313
7314         if (!dump_block_groups)
7315                 return;
7316
7317         down_read(&info->groups_sem);
7318 again:
7319         list_for_each_entry(cache, &info->block_groups[index], list) {
7320                 spin_lock(&cache->lock);
7321                 printk(KERN_INFO "BTRFS: "
7322                            "block group %llu has %llu bytes, "
7323                            "%llu used %llu pinned %llu reserved %s\n",
7324                        cache->key.objectid, cache->key.offset,
7325                        btrfs_block_group_used(&cache->item), cache->pinned,
7326                        cache->reserved, cache->ro ? "[readonly]" : "");
7327                 btrfs_dump_free_space(cache, bytes);
7328                 spin_unlock(&cache->lock);
7329         }
7330         if (++index < BTRFS_NR_RAID_TYPES)
7331                 goto again;
7332         up_read(&info->groups_sem);
7333 }
7334
7335 int btrfs_reserve_extent(struct btrfs_root *root,
7336                          u64 num_bytes, u64 min_alloc_size,
7337                          u64 empty_size, u64 hint_byte,
7338                          struct btrfs_key *ins, int is_data, int delalloc)
7339 {
7340         bool final_tried = false;
7341         u64 flags;
7342         int ret;
7343
7344         flags = btrfs_get_alloc_profile(root, is_data);
7345 again:
7346         WARN_ON(num_bytes < root->sectorsize);
7347         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7348                                flags, delalloc);
7349
7350         if (ret == -ENOSPC) {
7351                 if (!final_tried && ins->offset) {
7352                         num_bytes = min(num_bytes >> 1, ins->offset);
7353                         num_bytes = round_down(num_bytes, root->sectorsize);
7354                         num_bytes = max(num_bytes, min_alloc_size);
7355                         if (num_bytes == min_alloc_size)
7356                                 final_tried = true;
7357                         goto again;
7358                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7359                         struct btrfs_space_info *sinfo;
7360
7361                         sinfo = __find_space_info(root->fs_info, flags);
7362                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7363                                 flags, num_bytes);
7364                         if (sinfo)
7365                                 dump_space_info(sinfo, num_bytes, 1);
7366                 }
7367         }
7368
7369         return ret;
7370 }
7371
7372 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7373                                         u64 start, u64 len,
7374                                         int pin, int delalloc)
7375 {
7376         struct btrfs_block_group_cache *cache;
7377         int ret = 0;
7378
7379         cache = btrfs_lookup_block_group(root->fs_info, start);
7380         if (!cache) {
7381                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7382                         start);
7383                 return -ENOSPC;
7384         }
7385
7386         if (pin)
7387                 pin_down_extent(root, cache, start, len, 1);
7388         else {
7389                 if (btrfs_test_opt(root, DISCARD))
7390                         ret = btrfs_discard_extent(root, start, len, NULL);
7391                 btrfs_add_free_space(cache, start, len);
7392                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7393         }
7394
7395         btrfs_put_block_group(cache);
7396
7397         trace_btrfs_reserved_extent_free(root, start, len);
7398
7399         return ret;
7400 }
7401
7402 int btrfs_free_reserved_extent(struct btrfs_root *root,
7403                                u64 start, u64 len, int delalloc)
7404 {
7405         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7406 }
7407
7408 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7409                                        u64 start, u64 len)
7410 {
7411         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7412 }
7413
7414 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7415                                       struct btrfs_root *root,
7416                                       u64 parent, u64 root_objectid,
7417                                       u64 flags, u64 owner, u64 offset,
7418                                       struct btrfs_key *ins, int ref_mod)
7419 {
7420         int ret;
7421         struct btrfs_fs_info *fs_info = root->fs_info;
7422         struct btrfs_extent_item *extent_item;
7423         struct btrfs_extent_inline_ref *iref;
7424         struct btrfs_path *path;
7425         struct extent_buffer *leaf;
7426         int type;
7427         u32 size;
7428
7429         if (parent > 0)
7430                 type = BTRFS_SHARED_DATA_REF_KEY;
7431         else
7432                 type = BTRFS_EXTENT_DATA_REF_KEY;
7433
7434         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7435
7436         path = btrfs_alloc_path();
7437         if (!path)
7438                 return -ENOMEM;
7439
7440         path->leave_spinning = 1;
7441         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7442                                       ins, size);
7443         if (ret) {
7444                 btrfs_free_path(path);
7445                 return ret;
7446         }
7447
7448         leaf = path->nodes[0];
7449         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7450                                      struct btrfs_extent_item);
7451         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7452         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7453         btrfs_set_extent_flags(leaf, extent_item,
7454                                flags | BTRFS_EXTENT_FLAG_DATA);
7455
7456         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7457         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7458         if (parent > 0) {
7459                 struct btrfs_shared_data_ref *ref;
7460                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7461                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7462                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7463         } else {
7464                 struct btrfs_extent_data_ref *ref;
7465                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7466                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7467                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7468                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7469                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7470         }
7471
7472         btrfs_mark_buffer_dirty(path->nodes[0]);
7473         btrfs_free_path(path);
7474
7475         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7476         if (ret) { /* -ENOENT, logic error */
7477                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7478                         ins->objectid, ins->offset);
7479                 BUG();
7480         }
7481         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7482         return ret;
7483 }
7484
7485 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7486                                      struct btrfs_root *root,
7487                                      u64 parent, u64 root_objectid,
7488                                      u64 flags, struct btrfs_disk_key *key,
7489                                      int level, struct btrfs_key *ins,
7490                                      int no_quota)
7491 {
7492         int ret;
7493         struct btrfs_fs_info *fs_info = root->fs_info;
7494         struct btrfs_extent_item *extent_item;
7495         struct btrfs_tree_block_info *block_info;
7496         struct btrfs_extent_inline_ref *iref;
7497         struct btrfs_path *path;
7498         struct extent_buffer *leaf;
7499         u32 size = sizeof(*extent_item) + sizeof(*iref);
7500         u64 num_bytes = ins->offset;
7501         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7502                                                  SKINNY_METADATA);
7503
7504         if (!skinny_metadata)
7505                 size += sizeof(*block_info);
7506
7507         path = btrfs_alloc_path();
7508         if (!path) {
7509                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7510                                                    root->nodesize);
7511                 return -ENOMEM;
7512         }
7513
7514         path->leave_spinning = 1;
7515         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7516                                       ins, size);
7517         if (ret) {
7518                 btrfs_free_path(path);
7519                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7520                                                    root->nodesize);
7521                 return ret;
7522         }
7523
7524         leaf = path->nodes[0];
7525         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7526                                      struct btrfs_extent_item);
7527         btrfs_set_extent_refs(leaf, extent_item, 1);
7528         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7529         btrfs_set_extent_flags(leaf, extent_item,
7530                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7531
7532         if (skinny_metadata) {
7533                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7534                 num_bytes = root->nodesize;
7535         } else {
7536                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7537                 btrfs_set_tree_block_key(leaf, block_info, key);
7538                 btrfs_set_tree_block_level(leaf, block_info, level);
7539                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7540         }
7541
7542         if (parent > 0) {
7543                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7544                 btrfs_set_extent_inline_ref_type(leaf, iref,
7545                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7546                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7547         } else {
7548                 btrfs_set_extent_inline_ref_type(leaf, iref,
7549                                                  BTRFS_TREE_BLOCK_REF_KEY);
7550                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7551         }
7552
7553         btrfs_mark_buffer_dirty(leaf);
7554         btrfs_free_path(path);
7555
7556         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7557                                  1);
7558         if (ret) { /* -ENOENT, logic error */
7559                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7560                         ins->objectid, ins->offset);
7561                 BUG();
7562         }
7563
7564         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7565         return ret;
7566 }
7567
7568 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7569                                      struct btrfs_root *root,
7570                                      u64 root_objectid, u64 owner,
7571                                      u64 offset, struct btrfs_key *ins)
7572 {
7573         int ret;
7574
7575         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7576
7577         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7578                                          ins->offset, 0,
7579                                          root_objectid, owner, offset,
7580                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7581         return ret;
7582 }
7583
7584 /*
7585  * this is used by the tree logging recovery code.  It records that
7586  * an extent has been allocated and makes sure to clear the free
7587  * space cache bits as well
7588  */
7589 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7590                                    struct btrfs_root *root,
7591                                    u64 root_objectid, u64 owner, u64 offset,
7592                                    struct btrfs_key *ins)
7593 {
7594         int ret;
7595         struct btrfs_block_group_cache *block_group;
7596
7597         /*
7598          * Mixed block groups will exclude before processing the log so we only
7599          * need to do the exlude dance if this fs isn't mixed.
7600          */
7601         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7602                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7603                 if (ret)
7604                         return ret;
7605         }
7606
7607         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7608         if (!block_group)
7609                 return -EINVAL;
7610
7611         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7612                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7613         BUG_ON(ret); /* logic error */
7614         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7615                                          0, owner, offset, ins, 1);
7616         btrfs_put_block_group(block_group);
7617         return ret;
7618 }
7619
7620 static struct extent_buffer *
7621 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7622                       u64 bytenr, int level)
7623 {
7624         struct extent_buffer *buf;
7625
7626         buf = btrfs_find_create_tree_block(root, bytenr);
7627         if (!buf)
7628                 return ERR_PTR(-ENOMEM);
7629         btrfs_set_header_generation(buf, trans->transid);
7630         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7631         btrfs_tree_lock(buf);
7632         clean_tree_block(trans, root->fs_info, buf);
7633         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7634
7635         btrfs_set_lock_blocking(buf);
7636         btrfs_set_buffer_uptodate(buf);
7637
7638         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7639                 buf->log_index = root->log_transid % 2;
7640                 /*
7641                  * we allow two log transactions at a time, use different
7642                  * EXENT bit to differentiate dirty pages.
7643                  */
7644                 if (buf->log_index == 0)
7645                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7646                                         buf->start + buf->len - 1, GFP_NOFS);
7647                 else
7648                         set_extent_new(&root->dirty_log_pages, buf->start,
7649                                         buf->start + buf->len - 1, GFP_NOFS);
7650         } else {
7651                 buf->log_index = -1;
7652                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7653                          buf->start + buf->len - 1, GFP_NOFS);
7654         }
7655         trans->blocks_used++;
7656         /* this returns a buffer locked for blocking */
7657         return buf;
7658 }
7659
7660 static struct btrfs_block_rsv *
7661 use_block_rsv(struct btrfs_trans_handle *trans,
7662               struct btrfs_root *root, u32 blocksize)
7663 {
7664         struct btrfs_block_rsv *block_rsv;
7665         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7666         int ret;
7667         bool global_updated = false;
7668
7669         block_rsv = get_block_rsv(trans, root);
7670
7671         if (unlikely(block_rsv->size == 0))
7672                 goto try_reserve;
7673 again:
7674         ret = block_rsv_use_bytes(block_rsv, blocksize);
7675         if (!ret)
7676                 return block_rsv;
7677
7678         if (block_rsv->failfast)
7679                 return ERR_PTR(ret);
7680
7681         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7682                 global_updated = true;
7683                 update_global_block_rsv(root->fs_info);
7684                 goto again;
7685         }
7686
7687         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7688                 static DEFINE_RATELIMIT_STATE(_rs,
7689                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7690                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7691                 if (__ratelimit(&_rs))
7692                         WARN(1, KERN_DEBUG
7693                                 "BTRFS: block rsv returned %d\n", ret);
7694         }
7695 try_reserve:
7696         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7697                                      BTRFS_RESERVE_NO_FLUSH);
7698         if (!ret)
7699                 return block_rsv;
7700         /*
7701          * If we couldn't reserve metadata bytes try and use some from
7702          * the global reserve if its space type is the same as the global
7703          * reservation.
7704          */
7705         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7706             block_rsv->space_info == global_rsv->space_info) {
7707                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7708                 if (!ret)
7709                         return global_rsv;
7710         }
7711         return ERR_PTR(ret);
7712 }
7713
7714 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7715                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7716 {
7717         block_rsv_add_bytes(block_rsv, blocksize, 0);
7718         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7719 }
7720
7721 /*
7722  * finds a free extent and does all the dirty work required for allocation
7723  * returns the tree buffer or an ERR_PTR on error.
7724  */
7725 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7726                                         struct btrfs_root *root,
7727                                         u64 parent, u64 root_objectid,
7728                                         struct btrfs_disk_key *key, int level,
7729                                         u64 hint, u64 empty_size)
7730 {
7731         struct btrfs_key ins;
7732         struct btrfs_block_rsv *block_rsv;
7733         struct extent_buffer *buf;
7734         struct btrfs_delayed_extent_op *extent_op;
7735         u64 flags = 0;
7736         int ret;
7737         u32 blocksize = root->nodesize;
7738         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7739                                                  SKINNY_METADATA);
7740
7741         if (btrfs_test_is_dummy_root(root)) {
7742                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7743                                             level);
7744                 if (!IS_ERR(buf))
7745                         root->alloc_bytenr += blocksize;
7746                 return buf;
7747         }
7748
7749         block_rsv = use_block_rsv(trans, root, blocksize);
7750         if (IS_ERR(block_rsv))
7751                 return ERR_CAST(block_rsv);
7752
7753         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7754                                    empty_size, hint, &ins, 0, 0);
7755         if (ret)
7756                 goto out_unuse;
7757
7758         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
7759         if (IS_ERR(buf)) {
7760                 ret = PTR_ERR(buf);
7761                 goto out_free_reserved;
7762         }
7763
7764         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7765                 if (parent == 0)
7766                         parent = ins.objectid;
7767                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7768         } else
7769                 BUG_ON(parent > 0);
7770
7771         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7772                 extent_op = btrfs_alloc_delayed_extent_op();
7773                 if (!extent_op) {
7774                         ret = -ENOMEM;
7775                         goto out_free_buf;
7776                 }
7777                 if (key)
7778                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7779                 else
7780                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7781                 extent_op->flags_to_set = flags;
7782                 if (skinny_metadata)
7783                         extent_op->update_key = 0;
7784                 else
7785                         extent_op->update_key = 1;
7786                 extent_op->update_flags = 1;
7787                 extent_op->is_data = 0;
7788                 extent_op->level = level;
7789
7790                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7791                                                  ins.objectid, ins.offset,
7792                                                  parent, root_objectid, level,
7793                                                  BTRFS_ADD_DELAYED_EXTENT,
7794                                                  extent_op, 0);
7795                 if (ret)
7796                         goto out_free_delayed;
7797         }
7798         return buf;
7799
7800 out_free_delayed:
7801         btrfs_free_delayed_extent_op(extent_op);
7802 out_free_buf:
7803         free_extent_buffer(buf);
7804 out_free_reserved:
7805         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
7806 out_unuse:
7807         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7808         return ERR_PTR(ret);
7809 }
7810
7811 struct walk_control {
7812         u64 refs[BTRFS_MAX_LEVEL];
7813         u64 flags[BTRFS_MAX_LEVEL];
7814         struct btrfs_key update_progress;
7815         int stage;
7816         int level;
7817         int shared_level;
7818         int update_ref;
7819         int keep_locks;
7820         int reada_slot;
7821         int reada_count;
7822         int for_reloc;
7823 };
7824
7825 #define DROP_REFERENCE  1
7826 #define UPDATE_BACKREF  2
7827
7828 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7829                                      struct btrfs_root *root,
7830                                      struct walk_control *wc,
7831                                      struct btrfs_path *path)
7832 {
7833         u64 bytenr;
7834         u64 generation;
7835         u64 refs;
7836         u64 flags;
7837         u32 nritems;
7838         u32 blocksize;
7839         struct btrfs_key key;
7840         struct extent_buffer *eb;
7841         int ret;
7842         int slot;
7843         int nread = 0;
7844
7845         if (path->slots[wc->level] < wc->reada_slot) {
7846                 wc->reada_count = wc->reada_count * 2 / 3;
7847                 wc->reada_count = max(wc->reada_count, 2);
7848         } else {
7849                 wc->reada_count = wc->reada_count * 3 / 2;
7850                 wc->reada_count = min_t(int, wc->reada_count,
7851                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7852         }
7853
7854         eb = path->nodes[wc->level];
7855         nritems = btrfs_header_nritems(eb);
7856         blocksize = root->nodesize;
7857
7858         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7859                 if (nread >= wc->reada_count)
7860                         break;
7861
7862                 cond_resched();
7863                 bytenr = btrfs_node_blockptr(eb, slot);
7864                 generation = btrfs_node_ptr_generation(eb, slot);
7865
7866                 if (slot == path->slots[wc->level])
7867                         goto reada;
7868
7869                 if (wc->stage == UPDATE_BACKREF &&
7870                     generation <= root->root_key.offset)
7871                         continue;
7872
7873                 /* We don't lock the tree block, it's OK to be racy here */
7874                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7875                                                wc->level - 1, 1, &refs,
7876                                                &flags);
7877                 /* We don't care about errors in readahead. */
7878                 if (ret < 0)
7879                         continue;
7880                 BUG_ON(refs == 0);
7881
7882                 if (wc->stage == DROP_REFERENCE) {
7883                         if (refs == 1)
7884                                 goto reada;
7885
7886                         if (wc->level == 1 &&
7887                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7888                                 continue;
7889                         if (!wc->update_ref ||
7890                             generation <= root->root_key.offset)
7891                                 continue;
7892                         btrfs_node_key_to_cpu(eb, &key, slot);
7893                         ret = btrfs_comp_cpu_keys(&key,
7894                                                   &wc->update_progress);
7895                         if (ret < 0)
7896                                 continue;
7897                 } else {
7898                         if (wc->level == 1 &&
7899                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7900                                 continue;
7901                 }
7902 reada:
7903                 readahead_tree_block(root, bytenr);
7904                 nread++;
7905         }
7906         wc->reada_slot = slot;
7907 }
7908
7909 /*
7910  * TODO: Modify related function to add related node/leaf to dirty_extent_root,
7911  * for later qgroup accounting.
7912  *
7913  * Current, this function does nothing.
7914  */
7915 static int account_leaf_items(struct btrfs_trans_handle *trans,
7916                               struct btrfs_root *root,
7917                               struct extent_buffer *eb)
7918 {
7919         int nr = btrfs_header_nritems(eb);
7920         int i, extent_type;
7921         struct btrfs_key key;
7922         struct btrfs_file_extent_item *fi;
7923         u64 bytenr, num_bytes;
7924
7925         for (i = 0; i < nr; i++) {
7926                 btrfs_item_key_to_cpu(eb, &key, i);
7927
7928                 if (key.type != BTRFS_EXTENT_DATA_KEY)
7929                         continue;
7930
7931                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
7932                 /* filter out non qgroup-accountable extents  */
7933                 extent_type = btrfs_file_extent_type(eb, fi);
7934
7935                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
7936                         continue;
7937
7938                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
7939                 if (!bytenr)
7940                         continue;
7941
7942                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
7943         }
7944         return 0;
7945 }
7946
7947 /*
7948  * Walk up the tree from the bottom, freeing leaves and any interior
7949  * nodes which have had all slots visited. If a node (leaf or
7950  * interior) is freed, the node above it will have it's slot
7951  * incremented. The root node will never be freed.
7952  *
7953  * At the end of this function, we should have a path which has all
7954  * slots incremented to the next position for a search. If we need to
7955  * read a new node it will be NULL and the node above it will have the
7956  * correct slot selected for a later read.
7957  *
7958  * If we increment the root nodes slot counter past the number of
7959  * elements, 1 is returned to signal completion of the search.
7960  */
7961 static int adjust_slots_upwards(struct btrfs_root *root,
7962                                 struct btrfs_path *path, int root_level)
7963 {
7964         int level = 0;
7965         int nr, slot;
7966         struct extent_buffer *eb;
7967
7968         if (root_level == 0)
7969                 return 1;
7970
7971         while (level <= root_level) {
7972                 eb = path->nodes[level];
7973                 nr = btrfs_header_nritems(eb);
7974                 path->slots[level]++;
7975                 slot = path->slots[level];
7976                 if (slot >= nr || level == 0) {
7977                         /*
7978                          * Don't free the root -  we will detect this
7979                          * condition after our loop and return a
7980                          * positive value for caller to stop walking the tree.
7981                          */
7982                         if (level != root_level) {
7983                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7984                                 path->locks[level] = 0;
7985
7986                                 free_extent_buffer(eb);
7987                                 path->nodes[level] = NULL;
7988                                 path->slots[level] = 0;
7989                         }
7990                 } else {
7991                         /*
7992                          * We have a valid slot to walk back down
7993                          * from. Stop here so caller can process these
7994                          * new nodes.
7995                          */
7996                         break;
7997                 }
7998
7999                 level++;
8000         }
8001
8002         eb = path->nodes[root_level];
8003         if (path->slots[root_level] >= btrfs_header_nritems(eb))
8004                 return 1;
8005
8006         return 0;
8007 }
8008
8009 /*
8010  * root_eb is the subtree root and is locked before this function is called.
8011  * TODO: Modify this function to mark all (including complete shared node)
8012  * to dirty_extent_root to allow it get accounted in qgroup.
8013  */
8014 static int account_shared_subtree(struct btrfs_trans_handle *trans,
8015                                   struct btrfs_root *root,
8016                                   struct extent_buffer *root_eb,
8017                                   u64 root_gen,
8018                                   int root_level)
8019 {
8020         int ret = 0;
8021         int level;
8022         struct extent_buffer *eb = root_eb;
8023         struct btrfs_path *path = NULL;
8024
8025         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
8026         BUG_ON(root_eb == NULL);
8027
8028         if (!root->fs_info->quota_enabled)
8029                 return 0;
8030
8031         if (!extent_buffer_uptodate(root_eb)) {
8032                 ret = btrfs_read_buffer(root_eb, root_gen);
8033                 if (ret)
8034                         goto out;
8035         }
8036
8037         if (root_level == 0) {
8038                 ret = account_leaf_items(trans, root, root_eb);
8039                 goto out;
8040         }
8041
8042         path = btrfs_alloc_path();
8043         if (!path)
8044                 return -ENOMEM;
8045
8046         /*
8047          * Walk down the tree.  Missing extent blocks are filled in as
8048          * we go. Metadata is accounted every time we read a new
8049          * extent block.
8050          *
8051          * When we reach a leaf, we account for file extent items in it,
8052          * walk back up the tree (adjusting slot pointers as we go)
8053          * and restart the search process.
8054          */
8055         extent_buffer_get(root_eb); /* For path */
8056         path->nodes[root_level] = root_eb;
8057         path->slots[root_level] = 0;
8058         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
8059 walk_down:
8060         level = root_level;
8061         while (level >= 0) {
8062                 if (path->nodes[level] == NULL) {
8063                         int parent_slot;
8064                         u64 child_gen;
8065                         u64 child_bytenr;
8066
8067                         /* We need to get child blockptr/gen from
8068                          * parent before we can read it. */
8069                         eb = path->nodes[level + 1];
8070                         parent_slot = path->slots[level + 1];
8071                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8072                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8073
8074                         eb = read_tree_block(root, child_bytenr, child_gen);
8075                         if (IS_ERR(eb)) {
8076                                 ret = PTR_ERR(eb);
8077                                 goto out;
8078                         } else if (!extent_buffer_uptodate(eb)) {
8079                                 free_extent_buffer(eb);
8080                                 ret = -EIO;
8081                                 goto out;
8082                         }
8083
8084                         path->nodes[level] = eb;
8085                         path->slots[level] = 0;
8086
8087                         btrfs_tree_read_lock(eb);
8088                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8089                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8090                 }
8091
8092                 if (level == 0) {
8093                         ret = account_leaf_items(trans, root, path->nodes[level]);
8094                         if (ret)
8095                                 goto out;
8096
8097                         /* Nonzero return here means we completed our search */
8098                         ret = adjust_slots_upwards(root, path, root_level);
8099                         if (ret)
8100                                 break;
8101
8102                         /* Restart search with new slots */
8103                         goto walk_down;
8104                 }
8105
8106                 level--;
8107         }
8108
8109         ret = 0;
8110 out:
8111         btrfs_free_path(path);
8112
8113         return ret;
8114 }
8115
8116 /*
8117  * helper to process tree block while walking down the tree.
8118  *
8119  * when wc->stage == UPDATE_BACKREF, this function updates
8120  * back refs for pointers in the block.
8121  *
8122  * NOTE: return value 1 means we should stop walking down.
8123  */
8124 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8125                                    struct btrfs_root *root,
8126                                    struct btrfs_path *path,
8127                                    struct walk_control *wc, int lookup_info)
8128 {
8129         int level = wc->level;
8130         struct extent_buffer *eb = path->nodes[level];
8131         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8132         int ret;
8133
8134         if (wc->stage == UPDATE_BACKREF &&
8135             btrfs_header_owner(eb) != root->root_key.objectid)
8136                 return 1;
8137
8138         /*
8139          * when reference count of tree block is 1, it won't increase
8140          * again. once full backref flag is set, we never clear it.
8141          */
8142         if (lookup_info &&
8143             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8144              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8145                 BUG_ON(!path->locks[level]);
8146                 ret = btrfs_lookup_extent_info(trans, root,
8147                                                eb->start, level, 1,
8148                                                &wc->refs[level],
8149                                                &wc->flags[level]);
8150                 BUG_ON(ret == -ENOMEM);
8151                 if (ret)
8152                         return ret;
8153                 BUG_ON(wc->refs[level] == 0);
8154         }
8155
8156         if (wc->stage == DROP_REFERENCE) {
8157                 if (wc->refs[level] > 1)
8158                         return 1;
8159
8160                 if (path->locks[level] && !wc->keep_locks) {
8161                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8162                         path->locks[level] = 0;
8163                 }
8164                 return 0;
8165         }
8166
8167         /* wc->stage == UPDATE_BACKREF */
8168         if (!(wc->flags[level] & flag)) {
8169                 BUG_ON(!path->locks[level]);
8170                 ret = btrfs_inc_ref(trans, root, eb, 1);
8171                 BUG_ON(ret); /* -ENOMEM */
8172                 ret = btrfs_dec_ref(trans, root, eb, 0);
8173                 BUG_ON(ret); /* -ENOMEM */
8174                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8175                                                   eb->len, flag,
8176                                                   btrfs_header_level(eb), 0);
8177                 BUG_ON(ret); /* -ENOMEM */
8178                 wc->flags[level] |= flag;
8179         }
8180
8181         /*
8182          * the block is shared by multiple trees, so it's not good to
8183          * keep the tree lock
8184          */
8185         if (path->locks[level] && level > 0) {
8186                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8187                 path->locks[level] = 0;
8188         }
8189         return 0;
8190 }
8191
8192 /*
8193  * helper to process tree block pointer.
8194  *
8195  * when wc->stage == DROP_REFERENCE, this function checks
8196  * reference count of the block pointed to. if the block
8197  * is shared and we need update back refs for the subtree
8198  * rooted at the block, this function changes wc->stage to
8199  * UPDATE_BACKREF. if the block is shared and there is no
8200  * need to update back, this function drops the reference
8201  * to the block.
8202  *
8203  * NOTE: return value 1 means we should stop walking down.
8204  */
8205 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8206                                  struct btrfs_root *root,
8207                                  struct btrfs_path *path,
8208                                  struct walk_control *wc, int *lookup_info)
8209 {
8210         u64 bytenr;
8211         u64 generation;
8212         u64 parent;
8213         u32 blocksize;
8214         struct btrfs_key key;
8215         struct extent_buffer *next;
8216         int level = wc->level;
8217         int reada = 0;
8218         int ret = 0;
8219         bool need_account = false;
8220
8221         generation = btrfs_node_ptr_generation(path->nodes[level],
8222                                                path->slots[level]);
8223         /*
8224          * if the lower level block was created before the snapshot
8225          * was created, we know there is no need to update back refs
8226          * for the subtree
8227          */
8228         if (wc->stage == UPDATE_BACKREF &&
8229             generation <= root->root_key.offset) {
8230                 *lookup_info = 1;
8231                 return 1;
8232         }
8233
8234         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8235         blocksize = root->nodesize;
8236
8237         next = btrfs_find_tree_block(root->fs_info, bytenr);
8238         if (!next) {
8239                 next = btrfs_find_create_tree_block(root, bytenr);
8240                 if (!next)
8241                         return -ENOMEM;
8242                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8243                                                level - 1);
8244                 reada = 1;
8245         }
8246         btrfs_tree_lock(next);
8247         btrfs_set_lock_blocking(next);
8248
8249         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8250                                        &wc->refs[level - 1],
8251                                        &wc->flags[level - 1]);
8252         if (ret < 0) {
8253                 btrfs_tree_unlock(next);
8254                 return ret;
8255         }
8256
8257         if (unlikely(wc->refs[level - 1] == 0)) {
8258                 btrfs_err(root->fs_info, "Missing references.");
8259                 BUG();
8260         }
8261         *lookup_info = 0;
8262
8263         if (wc->stage == DROP_REFERENCE) {
8264                 if (wc->refs[level - 1] > 1) {
8265                         need_account = true;
8266                         if (level == 1 &&
8267                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8268                                 goto skip;
8269
8270                         if (!wc->update_ref ||
8271                             generation <= root->root_key.offset)
8272                                 goto skip;
8273
8274                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8275                                               path->slots[level]);
8276                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8277                         if (ret < 0)
8278                                 goto skip;
8279
8280                         wc->stage = UPDATE_BACKREF;
8281                         wc->shared_level = level - 1;
8282                 }
8283         } else {
8284                 if (level == 1 &&
8285                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8286                         goto skip;
8287         }
8288
8289         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8290                 btrfs_tree_unlock(next);
8291                 free_extent_buffer(next);
8292                 next = NULL;
8293                 *lookup_info = 1;
8294         }
8295
8296         if (!next) {
8297                 if (reada && level == 1)
8298                         reada_walk_down(trans, root, wc, path);
8299                 next = read_tree_block(root, bytenr, generation);
8300                 if (IS_ERR(next)) {
8301                         return PTR_ERR(next);
8302                 } else if (!extent_buffer_uptodate(next)) {
8303                         free_extent_buffer(next);
8304                         return -EIO;
8305                 }
8306                 btrfs_tree_lock(next);
8307                 btrfs_set_lock_blocking(next);
8308         }
8309
8310         level--;
8311         BUG_ON(level != btrfs_header_level(next));
8312         path->nodes[level] = next;
8313         path->slots[level] = 0;
8314         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8315         wc->level = level;
8316         if (wc->level == 1)
8317                 wc->reada_slot = 0;
8318         return 0;
8319 skip:
8320         wc->refs[level - 1] = 0;
8321         wc->flags[level - 1] = 0;
8322         if (wc->stage == DROP_REFERENCE) {
8323                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8324                         parent = path->nodes[level]->start;
8325                 } else {
8326                         BUG_ON(root->root_key.objectid !=
8327                                btrfs_header_owner(path->nodes[level]));
8328                         parent = 0;
8329                 }
8330
8331                 if (need_account) {
8332                         ret = account_shared_subtree(trans, root, next,
8333                                                      generation, level - 1);
8334                         if (ret) {
8335                                 btrfs_err_rl(root->fs_info,
8336                                         "Error "
8337                                         "%d accounting shared subtree. Quota "
8338                                         "is out of sync, rescan required.",
8339                                         ret);
8340                         }
8341                 }
8342                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8343                                 root->root_key.objectid, level - 1, 0, 0);
8344                 BUG_ON(ret); /* -ENOMEM */
8345         }
8346         btrfs_tree_unlock(next);
8347         free_extent_buffer(next);
8348         *lookup_info = 1;
8349         return 1;
8350 }
8351
8352 /*
8353  * helper to process tree block while walking up the tree.
8354  *
8355  * when wc->stage == DROP_REFERENCE, this function drops
8356  * reference count on the block.
8357  *
8358  * when wc->stage == UPDATE_BACKREF, this function changes
8359  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8360  * to UPDATE_BACKREF previously while processing the block.
8361  *
8362  * NOTE: return value 1 means we should stop walking up.
8363  */
8364 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8365                                  struct btrfs_root *root,
8366                                  struct btrfs_path *path,
8367                                  struct walk_control *wc)
8368 {
8369         int ret;
8370         int level = wc->level;
8371         struct extent_buffer *eb = path->nodes[level];
8372         u64 parent = 0;
8373
8374         if (wc->stage == UPDATE_BACKREF) {
8375                 BUG_ON(wc->shared_level < level);
8376                 if (level < wc->shared_level)
8377                         goto out;
8378
8379                 ret = find_next_key(path, level + 1, &wc->update_progress);
8380                 if (ret > 0)
8381                         wc->update_ref = 0;
8382
8383                 wc->stage = DROP_REFERENCE;
8384                 wc->shared_level = -1;
8385                 path->slots[level] = 0;
8386
8387                 /*
8388                  * check reference count again if the block isn't locked.
8389                  * we should start walking down the tree again if reference
8390                  * count is one.
8391                  */
8392                 if (!path->locks[level]) {
8393                         BUG_ON(level == 0);
8394                         btrfs_tree_lock(eb);
8395                         btrfs_set_lock_blocking(eb);
8396                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8397
8398                         ret = btrfs_lookup_extent_info(trans, root,
8399                                                        eb->start, level, 1,
8400                                                        &wc->refs[level],
8401                                                        &wc->flags[level]);
8402                         if (ret < 0) {
8403                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8404                                 path->locks[level] = 0;
8405                                 return ret;
8406                         }
8407                         BUG_ON(wc->refs[level] == 0);
8408                         if (wc->refs[level] == 1) {
8409                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8410                                 path->locks[level] = 0;
8411                                 return 1;
8412                         }
8413                 }
8414         }
8415
8416         /* wc->stage == DROP_REFERENCE */
8417         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8418
8419         if (wc->refs[level] == 1) {
8420                 if (level == 0) {
8421                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8422                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8423                         else
8424                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8425                         BUG_ON(ret); /* -ENOMEM */
8426                         ret = account_leaf_items(trans, root, eb);
8427                         if (ret) {
8428                                 btrfs_err_rl(root->fs_info,
8429                                         "error "
8430                                         "%d accounting leaf items. Quota "
8431                                         "is out of sync, rescan required.",
8432                                         ret);
8433                         }
8434                 }
8435                 /* make block locked assertion in clean_tree_block happy */
8436                 if (!path->locks[level] &&
8437                     btrfs_header_generation(eb) == trans->transid) {
8438                         btrfs_tree_lock(eb);
8439                         btrfs_set_lock_blocking(eb);
8440                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8441                 }
8442                 clean_tree_block(trans, root->fs_info, eb);
8443         }
8444
8445         if (eb == root->node) {
8446                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8447                         parent = eb->start;
8448                 else
8449                         BUG_ON(root->root_key.objectid !=
8450                                btrfs_header_owner(eb));
8451         } else {
8452                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8453                         parent = path->nodes[level + 1]->start;
8454                 else
8455                         BUG_ON(root->root_key.objectid !=
8456                                btrfs_header_owner(path->nodes[level + 1]));
8457         }
8458
8459         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8460 out:
8461         wc->refs[level] = 0;
8462         wc->flags[level] = 0;
8463         return 0;
8464 }
8465
8466 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8467                                    struct btrfs_root *root,
8468                                    struct btrfs_path *path,
8469                                    struct walk_control *wc)
8470 {
8471         int level = wc->level;
8472         int lookup_info = 1;
8473         int ret;
8474
8475         while (level >= 0) {
8476                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8477                 if (ret > 0)
8478                         break;
8479
8480                 if (level == 0)
8481                         break;
8482
8483                 if (path->slots[level] >=
8484                     btrfs_header_nritems(path->nodes[level]))
8485                         break;
8486
8487                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8488                 if (ret > 0) {
8489                         path->slots[level]++;
8490                         continue;
8491                 } else if (ret < 0)
8492                         return ret;
8493                 level = wc->level;
8494         }
8495         return 0;
8496 }
8497
8498 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8499                                  struct btrfs_root *root,
8500                                  struct btrfs_path *path,
8501                                  struct walk_control *wc, int max_level)
8502 {
8503         int level = wc->level;
8504         int ret;
8505
8506         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8507         while (level < max_level && path->nodes[level]) {
8508                 wc->level = level;
8509                 if (path->slots[level] + 1 <
8510                     btrfs_header_nritems(path->nodes[level])) {
8511                         path->slots[level]++;
8512                         return 0;
8513                 } else {
8514                         ret = walk_up_proc(trans, root, path, wc);
8515                         if (ret > 0)
8516                                 return 0;
8517
8518                         if (path->locks[level]) {
8519                                 btrfs_tree_unlock_rw(path->nodes[level],
8520                                                      path->locks[level]);
8521                                 path->locks[level] = 0;
8522                         }
8523                         free_extent_buffer(path->nodes[level]);
8524                         path->nodes[level] = NULL;
8525                         level++;
8526                 }
8527         }
8528         return 1;
8529 }
8530
8531 /*
8532  * drop a subvolume tree.
8533  *
8534  * this function traverses the tree freeing any blocks that only
8535  * referenced by the tree.
8536  *
8537  * when a shared tree block is found. this function decreases its
8538  * reference count by one. if update_ref is true, this function
8539  * also make sure backrefs for the shared block and all lower level
8540  * blocks are properly updated.
8541  *
8542  * If called with for_reloc == 0, may exit early with -EAGAIN
8543  */
8544 int btrfs_drop_snapshot(struct btrfs_root *root,
8545                          struct btrfs_block_rsv *block_rsv, int update_ref,
8546                          int for_reloc)
8547 {
8548         struct btrfs_path *path;
8549         struct btrfs_trans_handle *trans;
8550         struct btrfs_root *tree_root = root->fs_info->tree_root;
8551         struct btrfs_root_item *root_item = &root->root_item;
8552         struct walk_control *wc;
8553         struct btrfs_key key;
8554         int err = 0;
8555         int ret;
8556         int level;
8557         bool root_dropped = false;
8558
8559         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8560
8561         path = btrfs_alloc_path();
8562         if (!path) {
8563                 err = -ENOMEM;
8564                 goto out;
8565         }
8566
8567         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8568         if (!wc) {
8569                 btrfs_free_path(path);
8570                 err = -ENOMEM;
8571                 goto out;
8572         }
8573
8574         trans = btrfs_start_transaction(tree_root, 0);
8575         if (IS_ERR(trans)) {
8576                 err = PTR_ERR(trans);
8577                 goto out_free;
8578         }
8579
8580         if (block_rsv)
8581                 trans->block_rsv = block_rsv;
8582
8583         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8584                 level = btrfs_header_level(root->node);
8585                 path->nodes[level] = btrfs_lock_root_node(root);
8586                 btrfs_set_lock_blocking(path->nodes[level]);
8587                 path->slots[level] = 0;
8588                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8589                 memset(&wc->update_progress, 0,
8590                        sizeof(wc->update_progress));
8591         } else {
8592                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8593                 memcpy(&wc->update_progress, &key,
8594                        sizeof(wc->update_progress));
8595
8596                 level = root_item->drop_level;
8597                 BUG_ON(level == 0);
8598                 path->lowest_level = level;
8599                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8600                 path->lowest_level = 0;
8601                 if (ret < 0) {
8602                         err = ret;
8603                         goto out_end_trans;
8604                 }
8605                 WARN_ON(ret > 0);
8606
8607                 /*
8608                  * unlock our path, this is safe because only this
8609                  * function is allowed to delete this snapshot
8610                  */
8611                 btrfs_unlock_up_safe(path, 0);
8612
8613                 level = btrfs_header_level(root->node);
8614                 while (1) {
8615                         btrfs_tree_lock(path->nodes[level]);
8616                         btrfs_set_lock_blocking(path->nodes[level]);
8617                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8618
8619                         ret = btrfs_lookup_extent_info(trans, root,
8620                                                 path->nodes[level]->start,
8621                                                 level, 1, &wc->refs[level],
8622                                                 &wc->flags[level]);
8623                         if (ret < 0) {
8624                                 err = ret;
8625                                 goto out_end_trans;
8626                         }
8627                         BUG_ON(wc->refs[level] == 0);
8628
8629                         if (level == root_item->drop_level)
8630                                 break;
8631
8632                         btrfs_tree_unlock(path->nodes[level]);
8633                         path->locks[level] = 0;
8634                         WARN_ON(wc->refs[level] != 1);
8635                         level--;
8636                 }
8637         }
8638
8639         wc->level = level;
8640         wc->shared_level = -1;
8641         wc->stage = DROP_REFERENCE;
8642         wc->update_ref = update_ref;
8643         wc->keep_locks = 0;
8644         wc->for_reloc = for_reloc;
8645         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8646
8647         while (1) {
8648
8649                 ret = walk_down_tree(trans, root, path, wc);
8650                 if (ret < 0) {
8651                         err = ret;
8652                         break;
8653                 }
8654
8655                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8656                 if (ret < 0) {
8657                         err = ret;
8658                         break;
8659                 }
8660
8661                 if (ret > 0) {
8662                         BUG_ON(wc->stage != DROP_REFERENCE);
8663                         break;
8664                 }
8665
8666                 if (wc->stage == DROP_REFERENCE) {
8667                         level = wc->level;
8668                         btrfs_node_key(path->nodes[level],
8669                                        &root_item->drop_progress,
8670                                        path->slots[level]);
8671                         root_item->drop_level = level;
8672                 }
8673
8674                 BUG_ON(wc->level == 0);
8675                 if (btrfs_should_end_transaction(trans, tree_root) ||
8676                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8677                         ret = btrfs_update_root(trans, tree_root,
8678                                                 &root->root_key,
8679                                                 root_item);
8680                         if (ret) {
8681                                 btrfs_abort_transaction(trans, tree_root, ret);
8682                                 err = ret;
8683                                 goto out_end_trans;
8684                         }
8685
8686                         btrfs_end_transaction_throttle(trans, tree_root);
8687                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8688                                 pr_debug("BTRFS: drop snapshot early exit\n");
8689                                 err = -EAGAIN;
8690                                 goto out_free;
8691                         }
8692
8693                         trans = btrfs_start_transaction(tree_root, 0);
8694                         if (IS_ERR(trans)) {
8695                                 err = PTR_ERR(trans);
8696                                 goto out_free;
8697                         }
8698                         if (block_rsv)
8699                                 trans->block_rsv = block_rsv;
8700                 }
8701         }
8702         btrfs_release_path(path);
8703         if (err)
8704                 goto out_end_trans;
8705
8706         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8707         if (ret) {
8708                 btrfs_abort_transaction(trans, tree_root, ret);
8709                 goto out_end_trans;
8710         }
8711
8712         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8713                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8714                                       NULL, NULL);
8715                 if (ret < 0) {
8716                         btrfs_abort_transaction(trans, tree_root, ret);
8717                         err = ret;
8718                         goto out_end_trans;
8719                 } else if (ret > 0) {
8720                         /* if we fail to delete the orphan item this time
8721                          * around, it'll get picked up the next time.
8722                          *
8723                          * The most common failure here is just -ENOENT.
8724                          */
8725                         btrfs_del_orphan_item(trans, tree_root,
8726                                               root->root_key.objectid);
8727                 }
8728         }
8729
8730         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8731                 btrfs_add_dropped_root(trans, root);
8732         } else {
8733                 free_extent_buffer(root->node);
8734                 free_extent_buffer(root->commit_root);
8735                 btrfs_put_fs_root(root);
8736         }
8737         root_dropped = true;
8738 out_end_trans:
8739         btrfs_end_transaction_throttle(trans, tree_root);
8740 out_free:
8741         kfree(wc);
8742         btrfs_free_path(path);
8743 out:
8744         /*
8745          * So if we need to stop dropping the snapshot for whatever reason we
8746          * need to make sure to add it back to the dead root list so that we
8747          * keep trying to do the work later.  This also cleans up roots if we
8748          * don't have it in the radix (like when we recover after a power fail
8749          * or unmount) so we don't leak memory.
8750          */
8751         if (!for_reloc && root_dropped == false)
8752                 btrfs_add_dead_root(root);
8753         if (err && err != -EAGAIN)
8754                 btrfs_std_error(root->fs_info, err, NULL);
8755         return err;
8756 }
8757
8758 /*
8759  * drop subtree rooted at tree block 'node'.
8760  *
8761  * NOTE: this function will unlock and release tree block 'node'
8762  * only used by relocation code
8763  */
8764 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8765                         struct btrfs_root *root,
8766                         struct extent_buffer *node,
8767                         struct extent_buffer *parent)
8768 {
8769         struct btrfs_path *path;
8770         struct walk_control *wc;
8771         int level;
8772         int parent_level;
8773         int ret = 0;
8774         int wret;
8775
8776         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8777
8778         path = btrfs_alloc_path();
8779         if (!path)
8780                 return -ENOMEM;
8781
8782         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8783         if (!wc) {
8784                 btrfs_free_path(path);
8785                 return -ENOMEM;
8786         }
8787
8788         btrfs_assert_tree_locked(parent);
8789         parent_level = btrfs_header_level(parent);
8790         extent_buffer_get(parent);
8791         path->nodes[parent_level] = parent;
8792         path->slots[parent_level] = btrfs_header_nritems(parent);
8793
8794         btrfs_assert_tree_locked(node);
8795         level = btrfs_header_level(node);
8796         path->nodes[level] = node;
8797         path->slots[level] = 0;
8798         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8799
8800         wc->refs[parent_level] = 1;
8801         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8802         wc->level = level;
8803         wc->shared_level = -1;
8804         wc->stage = DROP_REFERENCE;
8805         wc->update_ref = 0;
8806         wc->keep_locks = 1;
8807         wc->for_reloc = 1;
8808         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8809
8810         while (1) {
8811                 wret = walk_down_tree(trans, root, path, wc);
8812                 if (wret < 0) {
8813                         ret = wret;
8814                         break;
8815                 }
8816
8817                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8818                 if (wret < 0)
8819                         ret = wret;
8820                 if (wret != 0)
8821                         break;
8822         }
8823
8824         kfree(wc);
8825         btrfs_free_path(path);
8826         return ret;
8827 }
8828
8829 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8830 {
8831         u64 num_devices;
8832         u64 stripped;
8833
8834         /*
8835          * if restripe for this chunk_type is on pick target profile and
8836          * return, otherwise do the usual balance
8837          */
8838         stripped = get_restripe_target(root->fs_info, flags);
8839         if (stripped)
8840                 return extended_to_chunk(stripped);
8841
8842         num_devices = root->fs_info->fs_devices->rw_devices;
8843
8844         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8845                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8846                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8847
8848         if (num_devices == 1) {
8849                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8850                 stripped = flags & ~stripped;
8851
8852                 /* turn raid0 into single device chunks */
8853                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8854                         return stripped;
8855
8856                 /* turn mirroring into duplication */
8857                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8858                              BTRFS_BLOCK_GROUP_RAID10))
8859                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8860         } else {
8861                 /* they already had raid on here, just return */
8862                 if (flags & stripped)
8863                         return flags;
8864
8865                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8866                 stripped = flags & ~stripped;
8867
8868                 /* switch duplicated blocks with raid1 */
8869                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8870                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8871
8872                 /* this is drive concat, leave it alone */
8873         }
8874
8875         return flags;
8876 }
8877
8878 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8879 {
8880         struct btrfs_space_info *sinfo = cache->space_info;
8881         u64 num_bytes;
8882         u64 min_allocable_bytes;
8883         int ret = -ENOSPC;
8884
8885         /*
8886          * We need some metadata space and system metadata space for
8887          * allocating chunks in some corner cases until we force to set
8888          * it to be readonly.
8889          */
8890         if ((sinfo->flags &
8891              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8892             !force)
8893                 min_allocable_bytes = 1 * 1024 * 1024;
8894         else
8895                 min_allocable_bytes = 0;
8896
8897         spin_lock(&sinfo->lock);
8898         spin_lock(&cache->lock);
8899
8900         if (cache->ro) {
8901                 cache->ro++;
8902                 ret = 0;
8903                 goto out;
8904         }
8905
8906         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8907                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8908
8909         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8910             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8911             min_allocable_bytes <= sinfo->total_bytes) {
8912                 sinfo->bytes_readonly += num_bytes;
8913                 cache->ro++;
8914                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
8915                 ret = 0;
8916         }
8917 out:
8918         spin_unlock(&cache->lock);
8919         spin_unlock(&sinfo->lock);
8920         return ret;
8921 }
8922
8923 int btrfs_inc_block_group_ro(struct btrfs_root *root,
8924                              struct btrfs_block_group_cache *cache)
8925
8926 {
8927         struct btrfs_trans_handle *trans;
8928         u64 alloc_flags;
8929         int ret;
8930
8931 again:
8932         trans = btrfs_join_transaction(root);
8933         if (IS_ERR(trans))
8934                 return PTR_ERR(trans);
8935
8936         /*
8937          * we're not allowed to set block groups readonly after the dirty
8938          * block groups cache has started writing.  If it already started,
8939          * back off and let this transaction commit
8940          */
8941         mutex_lock(&root->fs_info->ro_block_group_mutex);
8942         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
8943                 u64 transid = trans->transid;
8944
8945                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
8946                 btrfs_end_transaction(trans, root);
8947
8948                 ret = btrfs_wait_for_commit(root, transid);
8949                 if (ret)
8950                         return ret;
8951                 goto again;
8952         }
8953
8954         /*
8955          * if we are changing raid levels, try to allocate a corresponding
8956          * block group with the new raid level.
8957          */
8958         alloc_flags = update_block_group_flags(root, cache->flags);
8959         if (alloc_flags != cache->flags) {
8960                 ret = do_chunk_alloc(trans, root, alloc_flags,
8961                                      CHUNK_ALLOC_FORCE);
8962                 /*
8963                  * ENOSPC is allowed here, we may have enough space
8964                  * already allocated at the new raid level to
8965                  * carry on
8966                  */
8967                 if (ret == -ENOSPC)
8968                         ret = 0;
8969                 if (ret < 0)
8970                         goto out;
8971         }
8972
8973         ret = inc_block_group_ro(cache, 0);
8974         if (!ret)
8975                 goto out;
8976         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8977         ret = do_chunk_alloc(trans, root, alloc_flags,
8978                              CHUNK_ALLOC_FORCE);
8979         if (ret < 0)
8980                 goto out;
8981         ret = inc_block_group_ro(cache, 0);
8982 out:
8983         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
8984                 alloc_flags = update_block_group_flags(root, cache->flags);
8985                 lock_chunks(root->fs_info->chunk_root);
8986                 check_system_chunk(trans, root, alloc_flags);
8987                 unlock_chunks(root->fs_info->chunk_root);
8988         }
8989         mutex_unlock(&root->fs_info->ro_block_group_mutex);
8990
8991         btrfs_end_transaction(trans, root);
8992         return ret;
8993 }
8994
8995 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8996                             struct btrfs_root *root, u64 type)
8997 {
8998         u64 alloc_flags = get_alloc_profile(root, type);
8999         return do_chunk_alloc(trans, root, alloc_flags,
9000                               CHUNK_ALLOC_FORCE);
9001 }
9002
9003 /*
9004  * helper to account the unused space of all the readonly block group in the
9005  * space_info. takes mirrors into account.
9006  */
9007 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9008 {
9009         struct btrfs_block_group_cache *block_group;
9010         u64 free_bytes = 0;
9011         int factor;
9012
9013         /* It's df, we don't care if it's racey */
9014         if (list_empty(&sinfo->ro_bgs))
9015                 return 0;
9016
9017         spin_lock(&sinfo->lock);
9018         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9019                 spin_lock(&block_group->lock);
9020
9021                 if (!block_group->ro) {
9022                         spin_unlock(&block_group->lock);
9023                         continue;
9024                 }
9025
9026                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9027                                           BTRFS_BLOCK_GROUP_RAID10 |
9028                                           BTRFS_BLOCK_GROUP_DUP))
9029                         factor = 2;
9030                 else
9031                         factor = 1;
9032
9033                 free_bytes += (block_group->key.offset -
9034                                btrfs_block_group_used(&block_group->item)) *
9035                                factor;
9036
9037                 spin_unlock(&block_group->lock);
9038         }
9039         spin_unlock(&sinfo->lock);
9040
9041         return free_bytes;
9042 }
9043
9044 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9045                               struct btrfs_block_group_cache *cache)
9046 {
9047         struct btrfs_space_info *sinfo = cache->space_info;
9048         u64 num_bytes;
9049
9050         BUG_ON(!cache->ro);
9051
9052         spin_lock(&sinfo->lock);
9053         spin_lock(&cache->lock);
9054         if (!--cache->ro) {
9055                 num_bytes = cache->key.offset - cache->reserved -
9056                             cache->pinned - cache->bytes_super -
9057                             btrfs_block_group_used(&cache->item);
9058                 sinfo->bytes_readonly -= num_bytes;
9059                 list_del_init(&cache->ro_list);
9060         }
9061         spin_unlock(&cache->lock);
9062         spin_unlock(&sinfo->lock);
9063 }
9064
9065 /*
9066  * checks to see if its even possible to relocate this block group.
9067  *
9068  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9069  * ok to go ahead and try.
9070  */
9071 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9072 {
9073         struct btrfs_block_group_cache *block_group;
9074         struct btrfs_space_info *space_info;
9075         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9076         struct btrfs_device *device;
9077         struct btrfs_trans_handle *trans;
9078         u64 min_free;
9079         u64 dev_min = 1;
9080         u64 dev_nr = 0;
9081         u64 target;
9082         int index;
9083         int full = 0;
9084         int ret = 0;
9085
9086         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9087
9088         /* odd, couldn't find the block group, leave it alone */
9089         if (!block_group)
9090                 return -1;
9091
9092         min_free = btrfs_block_group_used(&block_group->item);
9093
9094         /* no bytes used, we're good */
9095         if (!min_free)
9096                 goto out;
9097
9098         space_info = block_group->space_info;
9099         spin_lock(&space_info->lock);
9100
9101         full = space_info->full;
9102
9103         /*
9104          * if this is the last block group we have in this space, we can't
9105          * relocate it unless we're able to allocate a new chunk below.
9106          *
9107          * Otherwise, we need to make sure we have room in the space to handle
9108          * all of the extents from this block group.  If we can, we're good
9109          */
9110         if ((space_info->total_bytes != block_group->key.offset) &&
9111             (space_info->bytes_used + space_info->bytes_reserved +
9112              space_info->bytes_pinned + space_info->bytes_readonly +
9113              min_free < space_info->total_bytes)) {
9114                 spin_unlock(&space_info->lock);
9115                 goto out;
9116         }
9117         spin_unlock(&space_info->lock);
9118
9119         /*
9120          * ok we don't have enough space, but maybe we have free space on our
9121          * devices to allocate new chunks for relocation, so loop through our
9122          * alloc devices and guess if we have enough space.  if this block
9123          * group is going to be restriped, run checks against the target
9124          * profile instead of the current one.
9125          */
9126         ret = -1;
9127
9128         /*
9129          * index:
9130          *      0: raid10
9131          *      1: raid1
9132          *      2: dup
9133          *      3: raid0
9134          *      4: single
9135          */
9136         target = get_restripe_target(root->fs_info, block_group->flags);
9137         if (target) {
9138                 index = __get_raid_index(extended_to_chunk(target));
9139         } else {
9140                 /*
9141                  * this is just a balance, so if we were marked as full
9142                  * we know there is no space for a new chunk
9143                  */
9144                 if (full)
9145                         goto out;
9146
9147                 index = get_block_group_index(block_group);
9148         }
9149
9150         if (index == BTRFS_RAID_RAID10) {
9151                 dev_min = 4;
9152                 /* Divide by 2 */
9153                 min_free >>= 1;
9154         } else if (index == BTRFS_RAID_RAID1) {
9155                 dev_min = 2;
9156         } else if (index == BTRFS_RAID_DUP) {
9157                 /* Multiply by 2 */
9158                 min_free <<= 1;
9159         } else if (index == BTRFS_RAID_RAID0) {
9160                 dev_min = fs_devices->rw_devices;
9161                 min_free = div64_u64(min_free, dev_min);
9162         }
9163
9164         /* We need to do this so that we can look at pending chunks */
9165         trans = btrfs_join_transaction(root);
9166         if (IS_ERR(trans)) {
9167                 ret = PTR_ERR(trans);
9168                 goto out;
9169         }
9170
9171         mutex_lock(&root->fs_info->chunk_mutex);
9172         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9173                 u64 dev_offset;
9174
9175                 /*
9176                  * check to make sure we can actually find a chunk with enough
9177                  * space to fit our block group in.
9178                  */
9179                 if (device->total_bytes > device->bytes_used + min_free &&
9180                     !device->is_tgtdev_for_dev_replace) {
9181                         ret = find_free_dev_extent(trans, device, min_free,
9182                                                    &dev_offset, NULL);
9183                         if (!ret)
9184                                 dev_nr++;
9185
9186                         if (dev_nr >= dev_min)
9187                                 break;
9188
9189                         ret = -1;
9190                 }
9191         }
9192         mutex_unlock(&root->fs_info->chunk_mutex);
9193         btrfs_end_transaction(trans, root);
9194 out:
9195         btrfs_put_block_group(block_group);
9196         return ret;
9197 }
9198
9199 static int find_first_block_group(struct btrfs_root *root,
9200                 struct btrfs_path *path, struct btrfs_key *key)
9201 {
9202         int ret = 0;
9203         struct btrfs_key found_key;
9204         struct extent_buffer *leaf;
9205         int slot;
9206
9207         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9208         if (ret < 0)
9209                 goto out;
9210
9211         while (1) {
9212                 slot = path->slots[0];
9213                 leaf = path->nodes[0];
9214                 if (slot >= btrfs_header_nritems(leaf)) {
9215                         ret = btrfs_next_leaf(root, path);
9216                         if (ret == 0)
9217                                 continue;
9218                         if (ret < 0)
9219                                 goto out;
9220                         break;
9221                 }
9222                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9223
9224                 if (found_key.objectid >= key->objectid &&
9225                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9226                         ret = 0;
9227                         goto out;
9228                 }
9229                 path->slots[0]++;
9230         }
9231 out:
9232         return ret;
9233 }
9234
9235 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9236 {
9237         struct btrfs_block_group_cache *block_group;
9238         u64 last = 0;
9239
9240         while (1) {
9241                 struct inode *inode;
9242
9243                 block_group = btrfs_lookup_first_block_group(info, last);
9244                 while (block_group) {
9245                         spin_lock(&block_group->lock);
9246                         if (block_group->iref)
9247                                 break;
9248                         spin_unlock(&block_group->lock);
9249                         block_group = next_block_group(info->tree_root,
9250                                                        block_group);
9251                 }
9252                 if (!block_group) {
9253                         if (last == 0)
9254                                 break;
9255                         last = 0;
9256                         continue;
9257                 }
9258
9259                 inode = block_group->inode;
9260                 block_group->iref = 0;
9261                 block_group->inode = NULL;
9262                 spin_unlock(&block_group->lock);
9263                 iput(inode);
9264                 last = block_group->key.objectid + block_group->key.offset;
9265                 btrfs_put_block_group(block_group);
9266         }
9267 }
9268
9269 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9270 {
9271         struct btrfs_block_group_cache *block_group;
9272         struct btrfs_space_info *space_info;
9273         struct btrfs_caching_control *caching_ctl;
9274         struct rb_node *n;
9275
9276         down_write(&info->commit_root_sem);
9277         while (!list_empty(&info->caching_block_groups)) {
9278                 caching_ctl = list_entry(info->caching_block_groups.next,
9279                                          struct btrfs_caching_control, list);
9280                 list_del(&caching_ctl->list);
9281                 put_caching_control(caching_ctl);
9282         }
9283         up_write(&info->commit_root_sem);
9284
9285         spin_lock(&info->unused_bgs_lock);
9286         while (!list_empty(&info->unused_bgs)) {
9287                 block_group = list_first_entry(&info->unused_bgs,
9288                                                struct btrfs_block_group_cache,
9289                                                bg_list);
9290                 list_del_init(&block_group->bg_list);
9291                 btrfs_put_block_group(block_group);
9292         }
9293         spin_unlock(&info->unused_bgs_lock);
9294
9295         spin_lock(&info->block_group_cache_lock);
9296         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9297                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9298                                        cache_node);
9299                 rb_erase(&block_group->cache_node,
9300                          &info->block_group_cache_tree);
9301                 RB_CLEAR_NODE(&block_group->cache_node);
9302                 spin_unlock(&info->block_group_cache_lock);
9303
9304                 down_write(&block_group->space_info->groups_sem);
9305                 list_del(&block_group->list);
9306                 up_write(&block_group->space_info->groups_sem);
9307
9308                 if (block_group->cached == BTRFS_CACHE_STARTED)
9309                         wait_block_group_cache_done(block_group);
9310
9311                 /*
9312                  * We haven't cached this block group, which means we could
9313                  * possibly have excluded extents on this block group.
9314                  */
9315                 if (block_group->cached == BTRFS_CACHE_NO ||
9316                     block_group->cached == BTRFS_CACHE_ERROR)
9317                         free_excluded_extents(info->extent_root, block_group);
9318
9319                 btrfs_remove_free_space_cache(block_group);
9320                 btrfs_put_block_group(block_group);
9321
9322                 spin_lock(&info->block_group_cache_lock);
9323         }
9324         spin_unlock(&info->block_group_cache_lock);
9325
9326         /* now that all the block groups are freed, go through and
9327          * free all the space_info structs.  This is only called during
9328          * the final stages of unmount, and so we know nobody is
9329          * using them.  We call synchronize_rcu() once before we start,
9330          * just to be on the safe side.
9331          */
9332         synchronize_rcu();
9333
9334         release_global_block_rsv(info);
9335
9336         while (!list_empty(&info->space_info)) {
9337                 int i;
9338
9339                 space_info = list_entry(info->space_info.next,
9340                                         struct btrfs_space_info,
9341                                         list);
9342                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9343                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9344                             space_info->bytes_reserved > 0 ||
9345                             space_info->bytes_may_use > 0)) {
9346                                 dump_space_info(space_info, 0, 0);
9347                         }
9348                 }
9349                 list_del(&space_info->list);
9350                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9351                         struct kobject *kobj;
9352                         kobj = space_info->block_group_kobjs[i];
9353                         space_info->block_group_kobjs[i] = NULL;
9354                         if (kobj) {
9355                                 kobject_del(kobj);
9356                                 kobject_put(kobj);
9357                         }
9358                 }
9359                 kobject_del(&space_info->kobj);
9360                 kobject_put(&space_info->kobj);
9361         }
9362         return 0;
9363 }
9364
9365 static void __link_block_group(struct btrfs_space_info *space_info,
9366                                struct btrfs_block_group_cache *cache)
9367 {
9368         int index = get_block_group_index(cache);
9369         bool first = false;
9370
9371         down_write(&space_info->groups_sem);
9372         if (list_empty(&space_info->block_groups[index]))
9373                 first = true;
9374         list_add_tail(&cache->list, &space_info->block_groups[index]);
9375         up_write(&space_info->groups_sem);
9376
9377         if (first) {
9378                 struct raid_kobject *rkobj;
9379                 int ret;
9380
9381                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9382                 if (!rkobj)
9383                         goto out_err;
9384                 rkobj->raid_type = index;
9385                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9386                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9387                                   "%s", get_raid_name(index));
9388                 if (ret) {
9389                         kobject_put(&rkobj->kobj);
9390                         goto out_err;
9391                 }
9392                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9393         }
9394
9395         return;
9396 out_err:
9397         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9398 }
9399
9400 static struct btrfs_block_group_cache *
9401 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9402 {
9403         struct btrfs_block_group_cache *cache;
9404
9405         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9406         if (!cache)
9407                 return NULL;
9408
9409         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9410                                         GFP_NOFS);
9411         if (!cache->free_space_ctl) {
9412                 kfree(cache);
9413                 return NULL;
9414         }
9415
9416         cache->key.objectid = start;
9417         cache->key.offset = size;
9418         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9419
9420         cache->sectorsize = root->sectorsize;
9421         cache->fs_info = root->fs_info;
9422         cache->full_stripe_len = btrfs_full_stripe_len(root,
9423                                                &root->fs_info->mapping_tree,
9424                                                start);
9425         atomic_set(&cache->count, 1);
9426         spin_lock_init(&cache->lock);
9427         init_rwsem(&cache->data_rwsem);
9428         INIT_LIST_HEAD(&cache->list);
9429         INIT_LIST_HEAD(&cache->cluster_list);
9430         INIT_LIST_HEAD(&cache->bg_list);
9431         INIT_LIST_HEAD(&cache->ro_list);
9432         INIT_LIST_HEAD(&cache->dirty_list);
9433         INIT_LIST_HEAD(&cache->io_list);
9434         btrfs_init_free_space_ctl(cache);
9435         atomic_set(&cache->trimming, 0);
9436
9437         return cache;
9438 }
9439
9440 int btrfs_read_block_groups(struct btrfs_root *root)
9441 {
9442         struct btrfs_path *path;
9443         int ret;
9444         struct btrfs_block_group_cache *cache;
9445         struct btrfs_fs_info *info = root->fs_info;
9446         struct btrfs_space_info *space_info;
9447         struct btrfs_key key;
9448         struct btrfs_key found_key;
9449         struct extent_buffer *leaf;
9450         int need_clear = 0;
9451         u64 cache_gen;
9452
9453         root = info->extent_root;
9454         key.objectid = 0;
9455         key.offset = 0;
9456         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9457         path = btrfs_alloc_path();
9458         if (!path)
9459                 return -ENOMEM;
9460         path->reada = 1;
9461
9462         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9463         if (btrfs_test_opt(root, SPACE_CACHE) &&
9464             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9465                 need_clear = 1;
9466         if (btrfs_test_opt(root, CLEAR_CACHE))
9467                 need_clear = 1;
9468
9469         while (1) {
9470                 ret = find_first_block_group(root, path, &key);
9471                 if (ret > 0)
9472                         break;
9473                 if (ret != 0)
9474                         goto error;
9475
9476                 leaf = path->nodes[0];
9477                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9478
9479                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9480                                                        found_key.offset);
9481                 if (!cache) {
9482                         ret = -ENOMEM;
9483                         goto error;
9484                 }
9485
9486                 if (need_clear) {
9487                         /*
9488                          * When we mount with old space cache, we need to
9489                          * set BTRFS_DC_CLEAR and set dirty flag.
9490                          *
9491                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9492                          *    truncate the old free space cache inode and
9493                          *    setup a new one.
9494                          * b) Setting 'dirty flag' makes sure that we flush
9495                          *    the new space cache info onto disk.
9496                          */
9497                         if (btrfs_test_opt(root, SPACE_CACHE))
9498                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9499                 }
9500
9501                 read_extent_buffer(leaf, &cache->item,
9502                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9503                                    sizeof(cache->item));
9504                 cache->flags = btrfs_block_group_flags(&cache->item);
9505
9506                 key.objectid = found_key.objectid + found_key.offset;
9507                 btrfs_release_path(path);
9508
9509                 /*
9510                  * We need to exclude the super stripes now so that the space
9511                  * info has super bytes accounted for, otherwise we'll think
9512                  * we have more space than we actually do.
9513                  */
9514                 ret = exclude_super_stripes(root, cache);
9515                 if (ret) {
9516                         /*
9517                          * We may have excluded something, so call this just in
9518                          * case.
9519                          */
9520                         free_excluded_extents(root, cache);
9521                         btrfs_put_block_group(cache);
9522                         goto error;
9523                 }
9524
9525                 /*
9526                  * check for two cases, either we are full, and therefore
9527                  * don't need to bother with the caching work since we won't
9528                  * find any space, or we are empty, and we can just add all
9529                  * the space in and be done with it.  This saves us _alot_ of
9530                  * time, particularly in the full case.
9531                  */
9532                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9533                         cache->last_byte_to_unpin = (u64)-1;
9534                         cache->cached = BTRFS_CACHE_FINISHED;
9535                         free_excluded_extents(root, cache);
9536                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9537                         cache->last_byte_to_unpin = (u64)-1;
9538                         cache->cached = BTRFS_CACHE_FINISHED;
9539                         add_new_free_space(cache, root->fs_info,
9540                                            found_key.objectid,
9541                                            found_key.objectid +
9542                                            found_key.offset);
9543                         free_excluded_extents(root, cache);
9544                 }
9545
9546                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9547                 if (ret) {
9548                         btrfs_remove_free_space_cache(cache);
9549                         btrfs_put_block_group(cache);
9550                         goto error;
9551                 }
9552
9553                 ret = update_space_info(info, cache->flags, found_key.offset,
9554                                         btrfs_block_group_used(&cache->item),
9555                                         &space_info);
9556                 if (ret) {
9557                         btrfs_remove_free_space_cache(cache);
9558                         spin_lock(&info->block_group_cache_lock);
9559                         rb_erase(&cache->cache_node,
9560                                  &info->block_group_cache_tree);
9561                         RB_CLEAR_NODE(&cache->cache_node);
9562                         spin_unlock(&info->block_group_cache_lock);
9563                         btrfs_put_block_group(cache);
9564                         goto error;
9565                 }
9566
9567                 cache->space_info = space_info;
9568                 spin_lock(&cache->space_info->lock);
9569                 cache->space_info->bytes_readonly += cache->bytes_super;
9570                 spin_unlock(&cache->space_info->lock);
9571
9572                 __link_block_group(space_info, cache);
9573
9574                 set_avail_alloc_bits(root->fs_info, cache->flags);
9575                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9576                         inc_block_group_ro(cache, 1);
9577                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9578                         spin_lock(&info->unused_bgs_lock);
9579                         /* Should always be true but just in case. */
9580                         if (list_empty(&cache->bg_list)) {
9581                                 btrfs_get_block_group(cache);
9582                                 list_add_tail(&cache->bg_list,
9583                                               &info->unused_bgs);
9584                         }
9585                         spin_unlock(&info->unused_bgs_lock);
9586                 }
9587         }
9588
9589         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9590                 if (!(get_alloc_profile(root, space_info->flags) &
9591                       (BTRFS_BLOCK_GROUP_RAID10 |
9592                        BTRFS_BLOCK_GROUP_RAID1 |
9593                        BTRFS_BLOCK_GROUP_RAID5 |
9594                        BTRFS_BLOCK_GROUP_RAID6 |
9595                        BTRFS_BLOCK_GROUP_DUP)))
9596                         continue;
9597                 /*
9598                  * avoid allocating from un-mirrored block group if there are
9599                  * mirrored block groups.
9600                  */
9601                 list_for_each_entry(cache,
9602                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9603                                 list)
9604                         inc_block_group_ro(cache, 1);
9605                 list_for_each_entry(cache,
9606                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9607                                 list)
9608                         inc_block_group_ro(cache, 1);
9609         }
9610
9611         init_global_block_rsv(info);
9612         ret = 0;
9613 error:
9614         btrfs_free_path(path);
9615         return ret;
9616 }
9617
9618 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9619                                        struct btrfs_root *root)
9620 {
9621         struct btrfs_block_group_cache *block_group, *tmp;
9622         struct btrfs_root *extent_root = root->fs_info->extent_root;
9623         struct btrfs_block_group_item item;
9624         struct btrfs_key key;
9625         int ret = 0;
9626         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
9627
9628         trans->can_flush_pending_bgs = false;
9629         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9630                 if (ret)
9631                         goto next;
9632
9633                 spin_lock(&block_group->lock);
9634                 memcpy(&item, &block_group->item, sizeof(item));
9635                 memcpy(&key, &block_group->key, sizeof(key));
9636                 spin_unlock(&block_group->lock);
9637
9638                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9639                                         sizeof(item));
9640                 if (ret)
9641                         btrfs_abort_transaction(trans, extent_root, ret);
9642                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9643                                                key.objectid, key.offset);
9644                 if (ret)
9645                         btrfs_abort_transaction(trans, extent_root, ret);
9646 next:
9647                 list_del_init(&block_group->bg_list);
9648         }
9649         trans->can_flush_pending_bgs = can_flush_pending_bgs;
9650 }
9651
9652 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9653                            struct btrfs_root *root, u64 bytes_used,
9654                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9655                            u64 size)
9656 {
9657         int ret;
9658         struct btrfs_root *extent_root;
9659         struct btrfs_block_group_cache *cache;
9660
9661         extent_root = root->fs_info->extent_root;
9662
9663         btrfs_set_log_full_commit(root->fs_info, trans);
9664
9665         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9666         if (!cache)
9667                 return -ENOMEM;
9668
9669         btrfs_set_block_group_used(&cache->item, bytes_used);
9670         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9671         btrfs_set_block_group_flags(&cache->item, type);
9672
9673         cache->flags = type;
9674         cache->last_byte_to_unpin = (u64)-1;
9675         cache->cached = BTRFS_CACHE_FINISHED;
9676         ret = exclude_super_stripes(root, cache);
9677         if (ret) {
9678                 /*
9679                  * We may have excluded something, so call this just in
9680                  * case.
9681                  */
9682                 free_excluded_extents(root, cache);
9683                 btrfs_put_block_group(cache);
9684                 return ret;
9685         }
9686
9687         add_new_free_space(cache, root->fs_info, chunk_offset,
9688                            chunk_offset + size);
9689
9690         free_excluded_extents(root, cache);
9691
9692 #ifdef CONFIG_BTRFS_DEBUG
9693         if (btrfs_should_fragment_free_space(root, cache)) {
9694                 u64 new_bytes_used = size - bytes_used;
9695
9696                 bytes_used += new_bytes_used >> 1;
9697                 fragment_free_space(root, cache);
9698         }
9699 #endif
9700         /*
9701          * Call to ensure the corresponding space_info object is created and
9702          * assigned to our block group, but don't update its counters just yet.
9703          * We want our bg to be added to the rbtree with its ->space_info set.
9704          */
9705         ret = update_space_info(root->fs_info, cache->flags, 0, 0,
9706                                 &cache->space_info);
9707         if (ret) {
9708                 btrfs_remove_free_space_cache(cache);
9709                 btrfs_put_block_group(cache);
9710                 return ret;
9711         }
9712
9713         ret = btrfs_add_block_group_cache(root->fs_info, cache);
9714         if (ret) {
9715                 btrfs_remove_free_space_cache(cache);
9716                 btrfs_put_block_group(cache);
9717                 return ret;
9718         }
9719
9720         /*
9721          * Now that our block group has its ->space_info set and is inserted in
9722          * the rbtree, update the space info's counters.
9723          */
9724         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
9725                                 &cache->space_info);
9726         if (ret) {
9727                 btrfs_remove_free_space_cache(cache);
9728                 spin_lock(&root->fs_info->block_group_cache_lock);
9729                 rb_erase(&cache->cache_node,
9730                          &root->fs_info->block_group_cache_tree);
9731                 RB_CLEAR_NODE(&cache->cache_node);
9732                 spin_unlock(&root->fs_info->block_group_cache_lock);
9733                 btrfs_put_block_group(cache);
9734                 return ret;
9735         }
9736         update_global_block_rsv(root->fs_info);
9737
9738         spin_lock(&cache->space_info->lock);
9739         cache->space_info->bytes_readonly += cache->bytes_super;
9740         spin_unlock(&cache->space_info->lock);
9741
9742         __link_block_group(cache->space_info, cache);
9743
9744         list_add_tail(&cache->bg_list, &trans->new_bgs);
9745
9746         set_avail_alloc_bits(extent_root->fs_info, type);
9747
9748         return 0;
9749 }
9750
9751 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9752 {
9753         u64 extra_flags = chunk_to_extended(flags) &
9754                                 BTRFS_EXTENDED_PROFILE_MASK;
9755
9756         write_seqlock(&fs_info->profiles_lock);
9757         if (flags & BTRFS_BLOCK_GROUP_DATA)
9758                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9759         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9760                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9761         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9762                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9763         write_sequnlock(&fs_info->profiles_lock);
9764 }
9765
9766 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9767                              struct btrfs_root *root, u64 group_start,
9768                              struct extent_map *em)
9769 {
9770         struct btrfs_path *path;
9771         struct btrfs_block_group_cache *block_group;
9772         struct btrfs_free_cluster *cluster;
9773         struct btrfs_root *tree_root = root->fs_info->tree_root;
9774         struct btrfs_key key;
9775         struct inode *inode;
9776         struct kobject *kobj = NULL;
9777         int ret;
9778         int index;
9779         int factor;
9780         struct btrfs_caching_control *caching_ctl = NULL;
9781         bool remove_em;
9782
9783         root = root->fs_info->extent_root;
9784
9785         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9786         BUG_ON(!block_group);
9787         BUG_ON(!block_group->ro);
9788
9789         /*
9790          * Free the reserved super bytes from this block group before
9791          * remove it.
9792          */
9793         free_excluded_extents(root, block_group);
9794
9795         memcpy(&key, &block_group->key, sizeof(key));
9796         index = get_block_group_index(block_group);
9797         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9798                                   BTRFS_BLOCK_GROUP_RAID1 |
9799                                   BTRFS_BLOCK_GROUP_RAID10))
9800                 factor = 2;
9801         else
9802                 factor = 1;
9803
9804         /* make sure this block group isn't part of an allocation cluster */
9805         cluster = &root->fs_info->data_alloc_cluster;
9806         spin_lock(&cluster->refill_lock);
9807         btrfs_return_cluster_to_free_space(block_group, cluster);
9808         spin_unlock(&cluster->refill_lock);
9809
9810         /*
9811          * make sure this block group isn't part of a metadata
9812          * allocation cluster
9813          */
9814         cluster = &root->fs_info->meta_alloc_cluster;
9815         spin_lock(&cluster->refill_lock);
9816         btrfs_return_cluster_to_free_space(block_group, cluster);
9817         spin_unlock(&cluster->refill_lock);
9818
9819         path = btrfs_alloc_path();
9820         if (!path) {
9821                 ret = -ENOMEM;
9822                 goto out;
9823         }
9824
9825         /*
9826          * get the inode first so any iput calls done for the io_list
9827          * aren't the final iput (no unlinks allowed now)
9828          */
9829         inode = lookup_free_space_inode(tree_root, block_group, path);
9830
9831         mutex_lock(&trans->transaction->cache_write_mutex);
9832         /*
9833          * make sure our free spache cache IO is done before remove the
9834          * free space inode
9835          */
9836         spin_lock(&trans->transaction->dirty_bgs_lock);
9837         if (!list_empty(&block_group->io_list)) {
9838                 list_del_init(&block_group->io_list);
9839
9840                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
9841
9842                 spin_unlock(&trans->transaction->dirty_bgs_lock);
9843                 btrfs_wait_cache_io(root, trans, block_group,
9844                                     &block_group->io_ctl, path,
9845                                     block_group->key.objectid);
9846                 btrfs_put_block_group(block_group);
9847                 spin_lock(&trans->transaction->dirty_bgs_lock);
9848         }
9849
9850         if (!list_empty(&block_group->dirty_list)) {
9851                 list_del_init(&block_group->dirty_list);
9852                 btrfs_put_block_group(block_group);
9853         }
9854         spin_unlock(&trans->transaction->dirty_bgs_lock);
9855         mutex_unlock(&trans->transaction->cache_write_mutex);
9856
9857         if (!IS_ERR(inode)) {
9858                 ret = btrfs_orphan_add(trans, inode);
9859                 if (ret) {
9860                         btrfs_add_delayed_iput(inode);
9861                         goto out;
9862                 }
9863                 clear_nlink(inode);
9864                 /* One for the block groups ref */
9865                 spin_lock(&block_group->lock);
9866                 if (block_group->iref) {
9867                         block_group->iref = 0;
9868                         block_group->inode = NULL;
9869                         spin_unlock(&block_group->lock);
9870                         iput(inode);
9871                 } else {
9872                         spin_unlock(&block_group->lock);
9873                 }
9874                 /* One for our lookup ref */
9875                 btrfs_add_delayed_iput(inode);
9876         }
9877
9878         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9879         key.offset = block_group->key.objectid;
9880         key.type = 0;
9881
9882         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9883         if (ret < 0)
9884                 goto out;
9885         if (ret > 0)
9886                 btrfs_release_path(path);
9887         if (ret == 0) {
9888                 ret = btrfs_del_item(trans, tree_root, path);
9889                 if (ret)
9890                         goto out;
9891                 btrfs_release_path(path);
9892         }
9893
9894         spin_lock(&root->fs_info->block_group_cache_lock);
9895         rb_erase(&block_group->cache_node,
9896                  &root->fs_info->block_group_cache_tree);
9897         RB_CLEAR_NODE(&block_group->cache_node);
9898
9899         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9900                 root->fs_info->first_logical_byte = (u64)-1;
9901         spin_unlock(&root->fs_info->block_group_cache_lock);
9902
9903         down_write(&block_group->space_info->groups_sem);
9904         /*
9905          * we must use list_del_init so people can check to see if they
9906          * are still on the list after taking the semaphore
9907          */
9908         list_del_init(&block_group->list);
9909         if (list_empty(&block_group->space_info->block_groups[index])) {
9910                 kobj = block_group->space_info->block_group_kobjs[index];
9911                 block_group->space_info->block_group_kobjs[index] = NULL;
9912                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9913         }
9914         up_write(&block_group->space_info->groups_sem);
9915         if (kobj) {
9916                 kobject_del(kobj);
9917                 kobject_put(kobj);
9918         }
9919
9920         if (block_group->has_caching_ctl)
9921                 caching_ctl = get_caching_control(block_group);
9922         if (block_group->cached == BTRFS_CACHE_STARTED)
9923                 wait_block_group_cache_done(block_group);
9924         if (block_group->has_caching_ctl) {
9925                 down_write(&root->fs_info->commit_root_sem);
9926                 if (!caching_ctl) {
9927                         struct btrfs_caching_control *ctl;
9928
9929                         list_for_each_entry(ctl,
9930                                     &root->fs_info->caching_block_groups, list)
9931                                 if (ctl->block_group == block_group) {
9932                                         caching_ctl = ctl;
9933                                         atomic_inc(&caching_ctl->count);
9934                                         break;
9935                                 }
9936                 }
9937                 if (caching_ctl)
9938                         list_del_init(&caching_ctl->list);
9939                 up_write(&root->fs_info->commit_root_sem);
9940                 if (caching_ctl) {
9941                         /* Once for the caching bgs list and once for us. */
9942                         put_caching_control(caching_ctl);
9943                         put_caching_control(caching_ctl);
9944                 }
9945         }
9946
9947         spin_lock(&trans->transaction->dirty_bgs_lock);
9948         if (!list_empty(&block_group->dirty_list)) {
9949                 WARN_ON(1);
9950         }
9951         if (!list_empty(&block_group->io_list)) {
9952                 WARN_ON(1);
9953         }
9954         spin_unlock(&trans->transaction->dirty_bgs_lock);
9955         btrfs_remove_free_space_cache(block_group);
9956
9957         spin_lock(&block_group->space_info->lock);
9958         list_del_init(&block_group->ro_list);
9959
9960         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
9961                 WARN_ON(block_group->space_info->total_bytes
9962                         < block_group->key.offset);
9963                 WARN_ON(block_group->space_info->bytes_readonly
9964                         < block_group->key.offset);
9965                 WARN_ON(block_group->space_info->disk_total
9966                         < block_group->key.offset * factor);
9967         }
9968         block_group->space_info->total_bytes -= block_group->key.offset;
9969         block_group->space_info->bytes_readonly -= block_group->key.offset;
9970         block_group->space_info->disk_total -= block_group->key.offset * factor;
9971
9972         spin_unlock(&block_group->space_info->lock);
9973
9974         memcpy(&key, &block_group->key, sizeof(key));
9975
9976         lock_chunks(root);
9977         if (!list_empty(&em->list)) {
9978                 /* We're in the transaction->pending_chunks list. */
9979                 free_extent_map(em);
9980         }
9981         spin_lock(&block_group->lock);
9982         block_group->removed = 1;
9983         /*
9984          * At this point trimming can't start on this block group, because we
9985          * removed the block group from the tree fs_info->block_group_cache_tree
9986          * so no one can't find it anymore and even if someone already got this
9987          * block group before we removed it from the rbtree, they have already
9988          * incremented block_group->trimming - if they didn't, they won't find
9989          * any free space entries because we already removed them all when we
9990          * called btrfs_remove_free_space_cache().
9991          *
9992          * And we must not remove the extent map from the fs_info->mapping_tree
9993          * to prevent the same logical address range and physical device space
9994          * ranges from being reused for a new block group. This is because our
9995          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
9996          * completely transactionless, so while it is trimming a range the
9997          * currently running transaction might finish and a new one start,
9998          * allowing for new block groups to be created that can reuse the same
9999          * physical device locations unless we take this special care.
10000          *
10001          * There may also be an implicit trim operation if the file system
10002          * is mounted with -odiscard. The same protections must remain
10003          * in place until the extents have been discarded completely when
10004          * the transaction commit has completed.
10005          */
10006         remove_em = (atomic_read(&block_group->trimming) == 0);
10007         /*
10008          * Make sure a trimmer task always sees the em in the pinned_chunks list
10009          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10010          * before checking block_group->removed).
10011          */
10012         if (!remove_em) {
10013                 /*
10014                  * Our em might be in trans->transaction->pending_chunks which
10015                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10016                  * and so is the fs_info->pinned_chunks list.
10017                  *
10018                  * So at this point we must be holding the chunk_mutex to avoid
10019                  * any races with chunk allocation (more specifically at
10020                  * volumes.c:contains_pending_extent()), to ensure it always
10021                  * sees the em, either in the pending_chunks list or in the
10022                  * pinned_chunks list.
10023                  */
10024                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10025         }
10026         spin_unlock(&block_group->lock);
10027
10028         if (remove_em) {
10029                 struct extent_map_tree *em_tree;
10030
10031                 em_tree = &root->fs_info->mapping_tree.map_tree;
10032                 write_lock(&em_tree->lock);
10033                 /*
10034                  * The em might be in the pending_chunks list, so make sure the
10035                  * chunk mutex is locked, since remove_extent_mapping() will
10036                  * delete us from that list.
10037                  */
10038                 remove_extent_mapping(em_tree, em);
10039                 write_unlock(&em_tree->lock);
10040                 /* once for the tree */
10041                 free_extent_map(em);
10042         }
10043
10044         unlock_chunks(root);
10045
10046         btrfs_put_block_group(block_group);
10047         btrfs_put_block_group(block_group);
10048
10049         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10050         if (ret > 0)
10051                 ret = -EIO;
10052         if (ret < 0)
10053                 goto out;
10054
10055         ret = btrfs_del_item(trans, root, path);
10056 out:
10057         btrfs_free_path(path);
10058         return ret;
10059 }
10060
10061 /*
10062  * Process the unused_bgs list and remove any that don't have any allocated
10063  * space inside of them.
10064  */
10065 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10066 {
10067         struct btrfs_block_group_cache *block_group;
10068         struct btrfs_space_info *space_info;
10069         struct btrfs_root *root = fs_info->extent_root;
10070         struct btrfs_trans_handle *trans;
10071         int ret = 0;
10072
10073         if (!fs_info->open)
10074                 return;
10075
10076         spin_lock(&fs_info->unused_bgs_lock);
10077         while (!list_empty(&fs_info->unused_bgs)) {
10078                 u64 start, end;
10079                 int trimming;
10080
10081                 block_group = list_first_entry(&fs_info->unused_bgs,
10082                                                struct btrfs_block_group_cache,
10083                                                bg_list);
10084                 space_info = block_group->space_info;
10085                 list_del_init(&block_group->bg_list);
10086                 if (ret || btrfs_mixed_space_info(space_info)) {
10087                         btrfs_put_block_group(block_group);
10088                         continue;
10089                 }
10090                 spin_unlock(&fs_info->unused_bgs_lock);
10091
10092                 mutex_lock(&root->fs_info->delete_unused_bgs_mutex);
10093
10094                 /* Don't want to race with allocators so take the groups_sem */
10095                 down_write(&space_info->groups_sem);
10096                 spin_lock(&block_group->lock);
10097                 if (block_group->reserved ||
10098                     btrfs_block_group_used(&block_group->item) ||
10099                     block_group->ro) {
10100                         /*
10101                          * We want to bail if we made new allocations or have
10102                          * outstanding allocations in this block group.  We do
10103                          * the ro check in case balance is currently acting on
10104                          * this block group.
10105                          */
10106                         spin_unlock(&block_group->lock);
10107                         up_write(&space_info->groups_sem);
10108                         goto next;
10109                 }
10110                 spin_unlock(&block_group->lock);
10111
10112                 /* We don't want to force the issue, only flip if it's ok. */
10113                 ret = inc_block_group_ro(block_group, 0);
10114                 up_write(&space_info->groups_sem);
10115                 if (ret < 0) {
10116                         ret = 0;
10117                         goto next;
10118                 }
10119
10120                 /*
10121                  * Want to do this before we do anything else so we can recover
10122                  * properly if we fail to join the transaction.
10123                  */
10124                 /* 1 for btrfs_orphan_reserve_metadata() */
10125                 trans = btrfs_start_transaction(root, 1);
10126                 if (IS_ERR(trans)) {
10127                         btrfs_dec_block_group_ro(root, block_group);
10128                         ret = PTR_ERR(trans);
10129                         goto next;
10130                 }
10131
10132                 /*
10133                  * We could have pending pinned extents for this block group,
10134                  * just delete them, we don't care about them anymore.
10135                  */
10136                 start = block_group->key.objectid;
10137                 end = start + block_group->key.offset - 1;
10138                 /*
10139                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10140                  * btrfs_finish_extent_commit(). If we are at transaction N,
10141                  * another task might be running finish_extent_commit() for the
10142                  * previous transaction N - 1, and have seen a range belonging
10143                  * to the block group in freed_extents[] before we were able to
10144                  * clear the whole block group range from freed_extents[]. This
10145                  * means that task can lookup for the block group after we
10146                  * unpinned it from freed_extents[] and removed it, leading to
10147                  * a BUG_ON() at btrfs_unpin_extent_range().
10148                  */
10149                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10150                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10151                                   EXTENT_DIRTY, GFP_NOFS);
10152                 if (ret) {
10153                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10154                         btrfs_dec_block_group_ro(root, block_group);
10155                         goto end_trans;
10156                 }
10157                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10158                                   EXTENT_DIRTY, GFP_NOFS);
10159                 if (ret) {
10160                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10161                         btrfs_dec_block_group_ro(root, block_group);
10162                         goto end_trans;
10163                 }
10164                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10165
10166                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10167                 spin_lock(&space_info->lock);
10168                 spin_lock(&block_group->lock);
10169
10170                 space_info->bytes_pinned -= block_group->pinned;
10171                 space_info->bytes_readonly += block_group->pinned;
10172                 percpu_counter_add(&space_info->total_bytes_pinned,
10173                                    -block_group->pinned);
10174                 block_group->pinned = 0;
10175
10176                 spin_unlock(&block_group->lock);
10177                 spin_unlock(&space_info->lock);
10178
10179                 /* DISCARD can flip during remount */
10180                 trimming = btrfs_test_opt(root, DISCARD);
10181
10182                 /* Implicit trim during transaction commit. */
10183                 if (trimming)
10184                         btrfs_get_block_group_trimming(block_group);
10185
10186                 /*
10187                  * Btrfs_remove_chunk will abort the transaction if things go
10188                  * horribly wrong.
10189                  */
10190                 ret = btrfs_remove_chunk(trans, root,
10191                                          block_group->key.objectid);
10192
10193                 if (ret) {
10194                         if (trimming)
10195                                 btrfs_put_block_group_trimming(block_group);
10196                         goto end_trans;
10197                 }
10198
10199                 /*
10200                  * If we're not mounted with -odiscard, we can just forget
10201                  * about this block group. Otherwise we'll need to wait
10202                  * until transaction commit to do the actual discard.
10203                  */
10204                 if (trimming) {
10205                         WARN_ON(!list_empty(&block_group->bg_list));
10206                         spin_lock(&trans->transaction->deleted_bgs_lock);
10207                         list_move(&block_group->bg_list,
10208                                   &trans->transaction->deleted_bgs);
10209                         spin_unlock(&trans->transaction->deleted_bgs_lock);
10210                         btrfs_get_block_group(block_group);
10211                 }
10212 end_trans:
10213                 btrfs_end_transaction(trans, root);
10214 next:
10215                 mutex_unlock(&root->fs_info->delete_unused_bgs_mutex);
10216                 btrfs_put_block_group(block_group);
10217                 spin_lock(&fs_info->unused_bgs_lock);
10218         }
10219         spin_unlock(&fs_info->unused_bgs_lock);
10220 }
10221
10222 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10223 {
10224         struct btrfs_space_info *space_info;
10225         struct btrfs_super_block *disk_super;
10226         u64 features;
10227         u64 flags;
10228         int mixed = 0;
10229         int ret;
10230
10231         disk_super = fs_info->super_copy;
10232         if (!btrfs_super_root(disk_super))
10233                 return 1;
10234
10235         features = btrfs_super_incompat_flags(disk_super);
10236         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10237                 mixed = 1;
10238
10239         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10240         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10241         if (ret)
10242                 goto out;
10243
10244         if (mixed) {
10245                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10246                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10247         } else {
10248                 flags = BTRFS_BLOCK_GROUP_METADATA;
10249                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10250                 if (ret)
10251                         goto out;
10252
10253                 flags = BTRFS_BLOCK_GROUP_DATA;
10254                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10255         }
10256 out:
10257         return ret;
10258 }
10259
10260 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10261 {
10262         return unpin_extent_range(root, start, end, false);
10263 }
10264
10265 /*
10266  * It used to be that old block groups would be left around forever.
10267  * Iterating over them would be enough to trim unused space.  Since we
10268  * now automatically remove them, we also need to iterate over unallocated
10269  * space.
10270  *
10271  * We don't want a transaction for this since the discard may take a
10272  * substantial amount of time.  We don't require that a transaction be
10273  * running, but we do need to take a running transaction into account
10274  * to ensure that we're not discarding chunks that were released in
10275  * the current transaction.
10276  *
10277  * Holding the chunks lock will prevent other threads from allocating
10278  * or releasing chunks, but it won't prevent a running transaction
10279  * from committing and releasing the memory that the pending chunks
10280  * list head uses.  For that, we need to take a reference to the
10281  * transaction.
10282  */
10283 static int btrfs_trim_free_extents(struct btrfs_device *device,
10284                                    u64 minlen, u64 *trimmed)
10285 {
10286         u64 start = 0, len = 0;
10287         int ret;
10288
10289         *trimmed = 0;
10290
10291         /* Not writeable = nothing to do. */
10292         if (!device->writeable)
10293                 return 0;
10294
10295         /* No free space = nothing to do. */
10296         if (device->total_bytes <= device->bytes_used)
10297                 return 0;
10298
10299         ret = 0;
10300
10301         while (1) {
10302                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
10303                 struct btrfs_transaction *trans;
10304                 u64 bytes;
10305
10306                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10307                 if (ret)
10308                         return ret;
10309
10310                 down_read(&fs_info->commit_root_sem);
10311
10312                 spin_lock(&fs_info->trans_lock);
10313                 trans = fs_info->running_transaction;
10314                 if (trans)
10315                         atomic_inc(&trans->use_count);
10316                 spin_unlock(&fs_info->trans_lock);
10317
10318                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10319                                                  &start, &len);
10320                 if (trans)
10321                         btrfs_put_transaction(trans);
10322
10323                 if (ret) {
10324                         up_read(&fs_info->commit_root_sem);
10325                         mutex_unlock(&fs_info->chunk_mutex);
10326                         if (ret == -ENOSPC)
10327                                 ret = 0;
10328                         break;
10329                 }
10330
10331                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10332                 up_read(&fs_info->commit_root_sem);
10333                 mutex_unlock(&fs_info->chunk_mutex);
10334
10335                 if (ret)
10336                         break;
10337
10338                 start += len;
10339                 *trimmed += bytes;
10340
10341                 if (fatal_signal_pending(current)) {
10342                         ret = -ERESTARTSYS;
10343                         break;
10344                 }
10345
10346                 cond_resched();
10347         }
10348
10349         return ret;
10350 }
10351
10352 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10353 {
10354         struct btrfs_fs_info *fs_info = root->fs_info;
10355         struct btrfs_block_group_cache *cache = NULL;
10356         struct btrfs_device *device;
10357         struct list_head *devices;
10358         u64 group_trimmed;
10359         u64 start;
10360         u64 end;
10361         u64 trimmed = 0;
10362         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10363         int ret = 0;
10364
10365         /*
10366          * try to trim all FS space, our block group may start from non-zero.
10367          */
10368         if (range->len == total_bytes)
10369                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10370         else
10371                 cache = btrfs_lookup_block_group(fs_info, range->start);
10372
10373         while (cache) {
10374                 if (cache->key.objectid >= (range->start + range->len)) {
10375                         btrfs_put_block_group(cache);
10376                         break;
10377                 }
10378
10379                 start = max(range->start, cache->key.objectid);
10380                 end = min(range->start + range->len,
10381                                 cache->key.objectid + cache->key.offset);
10382
10383                 if (end - start >= range->minlen) {
10384                         if (!block_group_cache_done(cache)) {
10385                                 ret = cache_block_group(cache, 0);
10386                                 if (ret) {
10387                                         btrfs_put_block_group(cache);
10388                                         break;
10389                                 }
10390                                 ret = wait_block_group_cache_done(cache);
10391                                 if (ret) {
10392                                         btrfs_put_block_group(cache);
10393                                         break;
10394                                 }
10395                         }
10396                         ret = btrfs_trim_block_group(cache,
10397                                                      &group_trimmed,
10398                                                      start,
10399                                                      end,
10400                                                      range->minlen);
10401
10402                         trimmed += group_trimmed;
10403                         if (ret) {
10404                                 btrfs_put_block_group(cache);
10405                                 break;
10406                         }
10407                 }
10408
10409                 cache = next_block_group(fs_info->tree_root, cache);
10410         }
10411
10412         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
10413         devices = &root->fs_info->fs_devices->alloc_list;
10414         list_for_each_entry(device, devices, dev_alloc_list) {
10415                 ret = btrfs_trim_free_extents(device, range->minlen,
10416                                               &group_trimmed);
10417                 if (ret)
10418                         break;
10419
10420                 trimmed += group_trimmed;
10421         }
10422         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
10423
10424         range->len = trimmed;
10425         return ret;
10426 }
10427
10428 /*
10429  * btrfs_{start,end}_write_no_snapshoting() are similar to
10430  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10431  * data into the page cache through nocow before the subvolume is snapshoted,
10432  * but flush the data into disk after the snapshot creation, or to prevent
10433  * operations while snapshoting is ongoing and that cause the snapshot to be
10434  * inconsistent (writes followed by expanding truncates for example).
10435  */
10436 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10437 {
10438         percpu_counter_dec(&root->subv_writers->counter);
10439         /*
10440          * Make sure counter is updated before we wake up waiters.
10441          */
10442         smp_mb();
10443         if (waitqueue_active(&root->subv_writers->wait))
10444                 wake_up(&root->subv_writers->wait);
10445 }
10446
10447 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10448 {
10449         if (atomic_read(&root->will_be_snapshoted))
10450                 return 0;
10451
10452         percpu_counter_inc(&root->subv_writers->counter);
10453         /*
10454          * Make sure counter is updated before we check for snapshot creation.
10455          */
10456         smp_mb();
10457         if (atomic_read(&root->will_be_snapshoted)) {
10458                 btrfs_end_write_no_snapshoting(root);
10459                 return 0;
10460         }
10461         return 1;
10462 }