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