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                 btrfs_put_bbio(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                         if (locked_ref->is_data &&
2542                             locked_ref->total_ref_mod < 0) {
2543                                 spin_lock(&delayed_refs->lock);
2544                                 delayed_refs->pending_csums -= ref->num_bytes;
2545                                 spin_unlock(&delayed_refs->lock);
2546                         }
2547                         btrfs_delayed_ref_unlock(locked_ref);
2548                         locked_ref = NULL;
2549                 }
2550                 btrfs_put_delayed_ref(ref);
2551                 count++;
2552                 cond_resched();
2553         }
2554
2555         /*
2556          * We don't want to include ref heads since we can have empty ref heads
2557          * and those will drastically skew our runtime down since we just do
2558          * accounting, no actual extent tree updates.
2559          */
2560         if (actual_count > 0) {
2561                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2562                 u64 avg;
2563
2564                 /*
2565                  * We weigh the current average higher than our current runtime
2566                  * to avoid large swings in the average.
2567                  */
2568                 spin_lock(&delayed_refs->lock);
2569                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2570                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2571                 spin_unlock(&delayed_refs->lock);
2572         }
2573         return 0;
2574 }
2575
2576 #ifdef SCRAMBLE_DELAYED_REFS
2577 /*
2578  * Normally delayed refs get processed in ascending bytenr order. This
2579  * correlates in most cases to the order added. To expose dependencies on this
2580  * order, we start to process the tree in the middle instead of the beginning
2581  */
2582 static u64 find_middle(struct rb_root *root)
2583 {
2584         struct rb_node *n = root->rb_node;
2585         struct btrfs_delayed_ref_node *entry;
2586         int alt = 1;
2587         u64 middle;
2588         u64 first = 0, last = 0;
2589
2590         n = rb_first(root);
2591         if (n) {
2592                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2593                 first = entry->bytenr;
2594         }
2595         n = rb_last(root);
2596         if (n) {
2597                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2598                 last = entry->bytenr;
2599         }
2600         n = root->rb_node;
2601
2602         while (n) {
2603                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2604                 WARN_ON(!entry->in_tree);
2605
2606                 middle = entry->bytenr;
2607
2608                 if (alt)
2609                         n = n->rb_left;
2610                 else
2611                         n = n->rb_right;
2612
2613                 alt = 1 - alt;
2614         }
2615         return middle;
2616 }
2617 #endif
2618
2619 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2620 {
2621         u64 num_bytes;
2622
2623         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2624                              sizeof(struct btrfs_extent_inline_ref));
2625         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2626                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2627
2628         /*
2629          * We don't ever fill up leaves all the way so multiply by 2 just to be
2630          * closer to what we're really going to want to ouse.
2631          */
2632         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2633 }
2634
2635 /*
2636  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2637  * would require to store the csums for that many bytes.
2638  */
2639 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2640 {
2641         u64 csum_size;
2642         u64 num_csums_per_leaf;
2643         u64 num_csums;
2644
2645         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
2646         num_csums_per_leaf = div64_u64(csum_size,
2647                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2648         num_csums = div64_u64(csum_bytes, root->sectorsize);
2649         num_csums += num_csums_per_leaf - 1;
2650         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2651         return num_csums;
2652 }
2653
2654 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2655                                        struct btrfs_root *root)
2656 {
2657         struct btrfs_block_rsv *global_rsv;
2658         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2659         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2660         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2661         u64 num_bytes, num_dirty_bgs_bytes;
2662         int ret = 0;
2663
2664         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2665         num_heads = heads_to_leaves(root, num_heads);
2666         if (num_heads > 1)
2667                 num_bytes += (num_heads - 1) * root->nodesize;
2668         num_bytes <<= 1;
2669         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2670         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2671                                                              num_dirty_bgs);
2672         global_rsv = &root->fs_info->global_block_rsv;
2673
2674         /*
2675          * If we can't allocate any more chunks lets make sure we have _lots_ of
2676          * wiggle room since running delayed refs can create more delayed refs.
2677          */
2678         if (global_rsv->space_info->full) {
2679                 num_dirty_bgs_bytes <<= 1;
2680                 num_bytes <<= 1;
2681         }
2682
2683         spin_lock(&global_rsv->lock);
2684         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2685                 ret = 1;
2686         spin_unlock(&global_rsv->lock);
2687         return ret;
2688 }
2689
2690 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2691                                        struct btrfs_root *root)
2692 {
2693         struct btrfs_fs_info *fs_info = root->fs_info;
2694         u64 num_entries =
2695                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2696         u64 avg_runtime;
2697         u64 val;
2698
2699         smp_mb();
2700         avg_runtime = fs_info->avg_delayed_ref_runtime;
2701         val = num_entries * avg_runtime;
2702         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2703                 return 1;
2704         if (val >= NSEC_PER_SEC / 2)
2705                 return 2;
2706
2707         return btrfs_check_space_for_delayed_refs(trans, root);
2708 }
2709
2710 struct async_delayed_refs {
2711         struct btrfs_root *root;
2712         int count;
2713         int error;
2714         int sync;
2715         struct completion wait;
2716         struct btrfs_work work;
2717 };
2718
2719 static void delayed_ref_async_start(struct btrfs_work *work)
2720 {
2721         struct async_delayed_refs *async;
2722         struct btrfs_trans_handle *trans;
2723         int ret;
2724
2725         async = container_of(work, struct async_delayed_refs, work);
2726
2727         trans = btrfs_join_transaction(async->root);
2728         if (IS_ERR(trans)) {
2729                 async->error = PTR_ERR(trans);
2730                 goto done;
2731         }
2732
2733         /*
2734          * trans->sync means that when we call end_transaciton, we won't
2735          * wait on delayed refs
2736          */
2737         trans->sync = true;
2738         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2739         if (ret)
2740                 async->error = ret;
2741
2742         ret = btrfs_end_transaction(trans, async->root);
2743         if (ret && !async->error)
2744                 async->error = ret;
2745 done:
2746         if (async->sync)
2747                 complete(&async->wait);
2748         else
2749                 kfree(async);
2750 }
2751
2752 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2753                                  unsigned long count, int wait)
2754 {
2755         struct async_delayed_refs *async;
2756         int ret;
2757
2758         async = kmalloc(sizeof(*async), GFP_NOFS);
2759         if (!async)
2760                 return -ENOMEM;
2761
2762         async->root = root->fs_info->tree_root;
2763         async->count = count;
2764         async->error = 0;
2765         if (wait)
2766                 async->sync = 1;
2767         else
2768                 async->sync = 0;
2769         init_completion(&async->wait);
2770
2771         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2772                         delayed_ref_async_start, NULL, NULL);
2773
2774         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2775
2776         if (wait) {
2777                 wait_for_completion(&async->wait);
2778                 ret = async->error;
2779                 kfree(async);
2780                 return ret;
2781         }
2782         return 0;
2783 }
2784
2785 /*
2786  * this starts processing the delayed reference count updates and
2787  * extent insertions we have queued up so far.  count can be
2788  * 0, which means to process everything in the tree at the start
2789  * of the run (but not newly added entries), or it can be some target
2790  * number you'd like to process.
2791  *
2792  * Returns 0 on success or if called with an aborted transaction
2793  * Returns <0 on error and aborts the transaction
2794  */
2795 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2796                            struct btrfs_root *root, unsigned long count)
2797 {
2798         struct rb_node *node;
2799         struct btrfs_delayed_ref_root *delayed_refs;
2800         struct btrfs_delayed_ref_head *head;
2801         int ret;
2802         int run_all = count == (unsigned long)-1;
2803
2804         /* We'll clean this up in btrfs_cleanup_transaction */
2805         if (trans->aborted)
2806                 return 0;
2807
2808         if (root == root->fs_info->extent_root)
2809                 root = root->fs_info->tree_root;
2810
2811         delayed_refs = &trans->transaction->delayed_refs;
2812         if (count == 0)
2813                 count = atomic_read(&delayed_refs->num_entries) * 2;
2814
2815 again:
2816 #ifdef SCRAMBLE_DELAYED_REFS
2817         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2818 #endif
2819         ret = __btrfs_run_delayed_refs(trans, root, count);
2820         if (ret < 0) {
2821                 btrfs_abort_transaction(trans, root, ret);
2822                 return ret;
2823         }
2824
2825         if (run_all) {
2826                 if (!list_empty(&trans->new_bgs))
2827                         btrfs_create_pending_block_groups(trans, root);
2828
2829                 spin_lock(&delayed_refs->lock);
2830                 node = rb_first(&delayed_refs->href_root);
2831                 if (!node) {
2832                         spin_unlock(&delayed_refs->lock);
2833                         goto out;
2834                 }
2835                 count = (unsigned long)-1;
2836
2837                 while (node) {
2838                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2839                                         href_node);
2840                         if (btrfs_delayed_ref_is_head(&head->node)) {
2841                                 struct btrfs_delayed_ref_node *ref;
2842
2843                                 ref = &head->node;
2844                                 atomic_inc(&ref->refs);
2845
2846                                 spin_unlock(&delayed_refs->lock);
2847                                 /*
2848                                  * Mutex was contended, block until it's
2849                                  * released and try again
2850                                  */
2851                                 mutex_lock(&head->mutex);
2852                                 mutex_unlock(&head->mutex);
2853
2854                                 btrfs_put_delayed_ref(ref);
2855                                 cond_resched();
2856                                 goto again;
2857                         } else {
2858                                 WARN_ON(1);
2859                         }
2860                         node = rb_next(node);
2861                 }
2862                 spin_unlock(&delayed_refs->lock);
2863                 cond_resched();
2864                 goto again;
2865         }
2866 out:
2867         ret = btrfs_delayed_qgroup_accounting(trans, root->fs_info);
2868         if (ret)
2869                 return ret;
2870         assert_qgroups_uptodate(trans);
2871         return 0;
2872 }
2873
2874 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2875                                 struct btrfs_root *root,
2876                                 u64 bytenr, u64 num_bytes, u64 flags,
2877                                 int level, int is_data)
2878 {
2879         struct btrfs_delayed_extent_op *extent_op;
2880         int ret;
2881
2882         extent_op = btrfs_alloc_delayed_extent_op();
2883         if (!extent_op)
2884                 return -ENOMEM;
2885
2886         extent_op->flags_to_set = flags;
2887         extent_op->update_flags = 1;
2888         extent_op->update_key = 0;
2889         extent_op->is_data = is_data ? 1 : 0;
2890         extent_op->level = level;
2891
2892         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2893                                           num_bytes, extent_op);
2894         if (ret)
2895                 btrfs_free_delayed_extent_op(extent_op);
2896         return ret;
2897 }
2898
2899 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2900                                       struct btrfs_root *root,
2901                                       struct btrfs_path *path,
2902                                       u64 objectid, u64 offset, u64 bytenr)
2903 {
2904         struct btrfs_delayed_ref_head *head;
2905         struct btrfs_delayed_ref_node *ref;
2906         struct btrfs_delayed_data_ref *data_ref;
2907         struct btrfs_delayed_ref_root *delayed_refs;
2908         struct rb_node *node;
2909         int ret = 0;
2910
2911         delayed_refs = &trans->transaction->delayed_refs;
2912         spin_lock(&delayed_refs->lock);
2913         head = btrfs_find_delayed_ref_head(trans, bytenr);
2914         if (!head) {
2915                 spin_unlock(&delayed_refs->lock);
2916                 return 0;
2917         }
2918
2919         if (!mutex_trylock(&head->mutex)) {
2920                 atomic_inc(&head->node.refs);
2921                 spin_unlock(&delayed_refs->lock);
2922
2923                 btrfs_release_path(path);
2924
2925                 /*
2926                  * Mutex was contended, block until it's released and let
2927                  * caller try again
2928                  */
2929                 mutex_lock(&head->mutex);
2930                 mutex_unlock(&head->mutex);
2931                 btrfs_put_delayed_ref(&head->node);
2932                 return -EAGAIN;
2933         }
2934         spin_unlock(&delayed_refs->lock);
2935
2936         spin_lock(&head->lock);
2937         node = rb_first(&head->ref_root);
2938         while (node) {
2939                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2940                 node = rb_next(node);
2941
2942                 /* If it's a shared ref we know a cross reference exists */
2943                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2944                         ret = 1;
2945                         break;
2946                 }
2947
2948                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2949
2950                 /*
2951                  * If our ref doesn't match the one we're currently looking at
2952                  * then we have a cross reference.
2953                  */
2954                 if (data_ref->root != root->root_key.objectid ||
2955                     data_ref->objectid != objectid ||
2956                     data_ref->offset != offset) {
2957                         ret = 1;
2958                         break;
2959                 }
2960         }
2961         spin_unlock(&head->lock);
2962         mutex_unlock(&head->mutex);
2963         return ret;
2964 }
2965
2966 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2967                                         struct btrfs_root *root,
2968                                         struct btrfs_path *path,
2969                                         u64 objectid, u64 offset, u64 bytenr)
2970 {
2971         struct btrfs_root *extent_root = root->fs_info->extent_root;
2972         struct extent_buffer *leaf;
2973         struct btrfs_extent_data_ref *ref;
2974         struct btrfs_extent_inline_ref *iref;
2975         struct btrfs_extent_item *ei;
2976         struct btrfs_key key;
2977         u32 item_size;
2978         int ret;
2979
2980         key.objectid = bytenr;
2981         key.offset = (u64)-1;
2982         key.type = BTRFS_EXTENT_ITEM_KEY;
2983
2984         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2985         if (ret < 0)
2986                 goto out;
2987         BUG_ON(ret == 0); /* Corruption */
2988
2989         ret = -ENOENT;
2990         if (path->slots[0] == 0)
2991                 goto out;
2992
2993         path->slots[0]--;
2994         leaf = path->nodes[0];
2995         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2996
2997         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2998                 goto out;
2999
3000         ret = 1;
3001         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3002 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3003         if (item_size < sizeof(*ei)) {
3004                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3005                 goto out;
3006         }
3007 #endif
3008         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3009
3010         if (item_size != sizeof(*ei) +
3011             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3012                 goto out;
3013
3014         if (btrfs_extent_generation(leaf, ei) <=
3015             btrfs_root_last_snapshot(&root->root_item))
3016                 goto out;
3017
3018         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3019         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3020             BTRFS_EXTENT_DATA_REF_KEY)
3021                 goto out;
3022
3023         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3024         if (btrfs_extent_refs(leaf, ei) !=
3025             btrfs_extent_data_ref_count(leaf, ref) ||
3026             btrfs_extent_data_ref_root(leaf, ref) !=
3027             root->root_key.objectid ||
3028             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3029             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3030                 goto out;
3031
3032         ret = 0;
3033 out:
3034         return ret;
3035 }
3036
3037 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3038                           struct btrfs_root *root,
3039                           u64 objectid, u64 offset, u64 bytenr)
3040 {
3041         struct btrfs_path *path;
3042         int ret;
3043         int ret2;
3044
3045         path = btrfs_alloc_path();
3046         if (!path)
3047                 return -ENOENT;
3048
3049         do {
3050                 ret = check_committed_ref(trans, root, path, objectid,
3051                                           offset, bytenr);
3052                 if (ret && ret != -ENOENT)
3053                         goto out;
3054
3055                 ret2 = check_delayed_ref(trans, root, path, objectid,
3056                                          offset, bytenr);
3057         } while (ret2 == -EAGAIN);
3058
3059         if (ret2 && ret2 != -ENOENT) {
3060                 ret = ret2;
3061                 goto out;
3062         }
3063
3064         if (ret != -ENOENT || ret2 != -ENOENT)
3065                 ret = 0;
3066 out:
3067         btrfs_free_path(path);
3068         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3069                 WARN_ON(ret > 0);
3070         return ret;
3071 }
3072
3073 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3074                            struct btrfs_root *root,
3075                            struct extent_buffer *buf,
3076                            int full_backref, int inc)
3077 {
3078         u64 bytenr;
3079         u64 num_bytes;
3080         u64 parent;
3081         u64 ref_root;
3082         u32 nritems;
3083         struct btrfs_key key;
3084         struct btrfs_file_extent_item *fi;
3085         int i;
3086         int level;
3087         int ret = 0;
3088         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3089                             u64, u64, u64, u64, u64, u64, int);
3090
3091
3092         if (btrfs_test_is_dummy_root(root))
3093                 return 0;
3094
3095         ref_root = btrfs_header_owner(buf);
3096         nritems = btrfs_header_nritems(buf);
3097         level = btrfs_header_level(buf);
3098
3099         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3100                 return 0;
3101
3102         if (inc)
3103                 process_func = btrfs_inc_extent_ref;
3104         else
3105                 process_func = btrfs_free_extent;
3106
3107         if (full_backref)
3108                 parent = buf->start;
3109         else
3110                 parent = 0;
3111
3112         for (i = 0; i < nritems; i++) {
3113                 if (level == 0) {
3114                         btrfs_item_key_to_cpu(buf, &key, i);
3115                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3116                                 continue;
3117                         fi = btrfs_item_ptr(buf, i,
3118                                             struct btrfs_file_extent_item);
3119                         if (btrfs_file_extent_type(buf, fi) ==
3120                             BTRFS_FILE_EXTENT_INLINE)
3121                                 continue;
3122                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3123                         if (bytenr == 0)
3124                                 continue;
3125
3126                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3127                         key.offset -= btrfs_file_extent_offset(buf, fi);
3128                         ret = process_func(trans, root, bytenr, num_bytes,
3129                                            parent, ref_root, key.objectid,
3130                                            key.offset, 1);
3131                         if (ret)
3132                                 goto fail;
3133                 } else {
3134                         bytenr = btrfs_node_blockptr(buf, i);
3135                         num_bytes = root->nodesize;
3136                         ret = process_func(trans, root, bytenr, num_bytes,
3137                                            parent, ref_root, level - 1, 0,
3138                                            1);
3139                         if (ret)
3140                                 goto fail;
3141                 }
3142         }
3143         return 0;
3144 fail:
3145         return ret;
3146 }
3147
3148 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3149                   struct extent_buffer *buf, int full_backref)
3150 {
3151         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3152 }
3153
3154 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3155                   struct extent_buffer *buf, int full_backref)
3156 {
3157         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3158 }
3159
3160 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3161                                  struct btrfs_root *root,
3162                                  struct btrfs_path *path,
3163                                  struct btrfs_block_group_cache *cache)
3164 {
3165         int ret;
3166         struct btrfs_root *extent_root = root->fs_info->extent_root;
3167         unsigned long bi;
3168         struct extent_buffer *leaf;
3169
3170         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3171         if (ret) {
3172                 if (ret > 0)
3173                         ret = -ENOENT;
3174                 goto fail;
3175         }
3176
3177         leaf = path->nodes[0];
3178         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3179         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3180         btrfs_mark_buffer_dirty(leaf);
3181 fail:
3182         btrfs_release_path(path);
3183         if (ret)
3184                 btrfs_abort_transaction(trans, root, ret);
3185         return ret;
3186
3187 }
3188
3189 static struct btrfs_block_group_cache *
3190 next_block_group(struct btrfs_root *root,
3191                  struct btrfs_block_group_cache *cache)
3192 {
3193         struct rb_node *node;
3194
3195         spin_lock(&root->fs_info->block_group_cache_lock);
3196
3197         /* If our block group was removed, we need a full search. */
3198         if (RB_EMPTY_NODE(&cache->cache_node)) {
3199                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3200
3201                 spin_unlock(&root->fs_info->block_group_cache_lock);
3202                 btrfs_put_block_group(cache);
3203                 cache = btrfs_lookup_first_block_group(root->fs_info,
3204                                                        next_bytenr);
3205                 return cache;
3206         }
3207         node = rb_next(&cache->cache_node);
3208         btrfs_put_block_group(cache);
3209         if (node) {
3210                 cache = rb_entry(node, struct btrfs_block_group_cache,
3211                                  cache_node);
3212                 btrfs_get_block_group(cache);
3213         } else
3214                 cache = NULL;
3215         spin_unlock(&root->fs_info->block_group_cache_lock);
3216         return cache;
3217 }
3218
3219 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3220                             struct btrfs_trans_handle *trans,
3221                             struct btrfs_path *path)
3222 {
3223         struct btrfs_root *root = block_group->fs_info->tree_root;
3224         struct inode *inode = NULL;
3225         u64 alloc_hint = 0;
3226         int dcs = BTRFS_DC_ERROR;
3227         u64 num_pages = 0;
3228         int retries = 0;
3229         int ret = 0;
3230
3231         /*
3232          * If this block group is smaller than 100 megs don't bother caching the
3233          * block group.
3234          */
3235         if (block_group->key.offset < (100 * 1024 * 1024)) {
3236                 spin_lock(&block_group->lock);
3237                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3238                 spin_unlock(&block_group->lock);
3239                 return 0;
3240         }
3241
3242         if (trans->aborted)
3243                 return 0;
3244 again:
3245         inode = lookup_free_space_inode(root, block_group, path);
3246         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3247                 ret = PTR_ERR(inode);
3248                 btrfs_release_path(path);
3249                 goto out;
3250         }
3251
3252         if (IS_ERR(inode)) {
3253                 BUG_ON(retries);
3254                 retries++;
3255
3256                 if (block_group->ro)
3257                         goto out_free;
3258
3259                 ret = create_free_space_inode(root, trans, block_group, path);
3260                 if (ret)
3261                         goto out_free;
3262                 goto again;
3263         }
3264
3265         /* We've already setup this transaction, go ahead and exit */
3266         if (block_group->cache_generation == trans->transid &&
3267             i_size_read(inode)) {
3268                 dcs = BTRFS_DC_SETUP;
3269                 goto out_put;
3270         }
3271
3272         /*
3273          * We want to set the generation to 0, that way if anything goes wrong
3274          * from here on out we know not to trust this cache when we load up next
3275          * time.
3276          */
3277         BTRFS_I(inode)->generation = 0;
3278         ret = btrfs_update_inode(trans, root, inode);
3279         if (ret) {
3280                 /*
3281                  * So theoretically we could recover from this, simply set the
3282                  * super cache generation to 0 so we know to invalidate the
3283                  * cache, but then we'd have to keep track of the block groups
3284                  * that fail this way so we know we _have_ to reset this cache
3285                  * before the next commit or risk reading stale cache.  So to
3286                  * limit our exposure to horrible edge cases lets just abort the
3287                  * transaction, this only happens in really bad situations
3288                  * anyway.
3289                  */
3290                 btrfs_abort_transaction(trans, root, ret);
3291                 goto out_put;
3292         }
3293         WARN_ON(ret);
3294
3295         if (i_size_read(inode) > 0) {
3296                 ret = btrfs_check_trunc_cache_free_space(root,
3297                                         &root->fs_info->global_block_rsv);
3298                 if (ret)
3299                         goto out_put;
3300
3301                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3302                 if (ret)
3303                         goto out_put;
3304         }
3305
3306         spin_lock(&block_group->lock);
3307         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3308             !btrfs_test_opt(root, SPACE_CACHE)) {
3309                 /*
3310                  * don't bother trying to write stuff out _if_
3311                  * a) we're not cached,
3312                  * b) we're with nospace_cache mount option.
3313                  */
3314                 dcs = BTRFS_DC_WRITTEN;
3315                 spin_unlock(&block_group->lock);
3316                 goto out_put;
3317         }
3318         spin_unlock(&block_group->lock);
3319
3320         /*
3321          * Try to preallocate enough space based on how big the block group is.
3322          * Keep in mind this has to include any pinned space which could end up
3323          * taking up quite a bit since it's not folded into the other space
3324          * cache.
3325          */
3326         num_pages = div_u64(block_group->key.offset, 256 * 1024 * 1024);
3327         if (!num_pages)
3328                 num_pages = 1;
3329
3330         num_pages *= 16;
3331         num_pages *= PAGE_CACHE_SIZE;
3332
3333         ret = btrfs_check_data_free_space(inode, num_pages, num_pages);
3334         if (ret)
3335                 goto out_put;
3336
3337         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3338                                               num_pages, num_pages,
3339                                               &alloc_hint);
3340         if (!ret)
3341                 dcs = BTRFS_DC_SETUP;
3342         btrfs_free_reserved_data_space(inode, num_pages);
3343
3344 out_put:
3345         iput(inode);
3346 out_free:
3347         btrfs_release_path(path);
3348 out:
3349         spin_lock(&block_group->lock);
3350         if (!ret && dcs == BTRFS_DC_SETUP)
3351                 block_group->cache_generation = trans->transid;
3352         block_group->disk_cache_state = dcs;
3353         spin_unlock(&block_group->lock);
3354
3355         return ret;
3356 }
3357
3358 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3359                             struct btrfs_root *root)
3360 {
3361         struct btrfs_block_group_cache *cache, *tmp;
3362         struct btrfs_transaction *cur_trans = trans->transaction;
3363         struct btrfs_path *path;
3364
3365         if (list_empty(&cur_trans->dirty_bgs) ||
3366             !btrfs_test_opt(root, SPACE_CACHE))
3367                 return 0;
3368
3369         path = btrfs_alloc_path();
3370         if (!path)
3371                 return -ENOMEM;
3372
3373         /* Could add new block groups, use _safe just in case */
3374         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3375                                  dirty_list) {
3376                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3377                         cache_save_setup(cache, trans, path);
3378         }
3379
3380         btrfs_free_path(path);
3381         return 0;
3382 }
3383
3384 /*
3385  * transaction commit does final block group cache writeback during a
3386  * critical section where nothing is allowed to change the FS.  This is
3387  * required in order for the cache to actually match the block group,
3388  * but can introduce a lot of latency into the commit.
3389  *
3390  * So, btrfs_start_dirty_block_groups is here to kick off block group
3391  * cache IO.  There's a chance we'll have to redo some of it if the
3392  * block group changes again during the commit, but it greatly reduces
3393  * the commit latency by getting rid of the easy block groups while
3394  * we're still allowing others to join the commit.
3395  */
3396 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3397                                    struct btrfs_root *root)
3398 {
3399         struct btrfs_block_group_cache *cache;
3400         struct btrfs_transaction *cur_trans = trans->transaction;
3401         int ret = 0;
3402         int should_put;
3403         struct btrfs_path *path = NULL;
3404         LIST_HEAD(dirty);
3405         struct list_head *io = &cur_trans->io_bgs;
3406         int num_started = 0;
3407         int loops = 0;
3408
3409         spin_lock(&cur_trans->dirty_bgs_lock);
3410         if (list_empty(&cur_trans->dirty_bgs)) {
3411                 spin_unlock(&cur_trans->dirty_bgs_lock);
3412                 return 0;
3413         }
3414         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3415         spin_unlock(&cur_trans->dirty_bgs_lock);
3416
3417 again:
3418         /*
3419          * make sure all the block groups on our dirty list actually
3420          * exist
3421          */
3422         btrfs_create_pending_block_groups(trans, root);
3423
3424         if (!path) {
3425                 path = btrfs_alloc_path();
3426                 if (!path)
3427                         return -ENOMEM;
3428         }
3429
3430         /*
3431          * cache_write_mutex is here only to save us from balance or automatic
3432          * removal of empty block groups deleting this block group while we are
3433          * writing out the cache
3434          */
3435         mutex_lock(&trans->transaction->cache_write_mutex);
3436         while (!list_empty(&dirty)) {
3437                 cache = list_first_entry(&dirty,
3438                                          struct btrfs_block_group_cache,
3439                                          dirty_list);
3440                 /*
3441                  * this can happen if something re-dirties a block
3442                  * group that is already under IO.  Just wait for it to
3443                  * finish and then do it all again
3444                  */
3445                 if (!list_empty(&cache->io_list)) {
3446                         list_del_init(&cache->io_list);
3447                         btrfs_wait_cache_io(root, trans, cache,
3448                                             &cache->io_ctl, path,
3449                                             cache->key.objectid);
3450                         btrfs_put_block_group(cache);
3451                 }
3452
3453
3454                 /*
3455                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3456                  * if it should update the cache_state.  Don't delete
3457                  * until after we wait.
3458                  *
3459                  * Since we're not running in the commit critical section
3460                  * we need the dirty_bgs_lock to protect from update_block_group
3461                  */
3462                 spin_lock(&cur_trans->dirty_bgs_lock);
3463                 list_del_init(&cache->dirty_list);
3464                 spin_unlock(&cur_trans->dirty_bgs_lock);
3465
3466                 should_put = 1;
3467
3468                 cache_save_setup(cache, trans, path);
3469
3470                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3471                         cache->io_ctl.inode = NULL;
3472                         ret = btrfs_write_out_cache(root, trans, cache, path);
3473                         if (ret == 0 && cache->io_ctl.inode) {
3474                                 num_started++;
3475                                 should_put = 0;
3476
3477                                 /*
3478                                  * the cache_write_mutex is protecting
3479                                  * the io_list
3480                                  */
3481                                 list_add_tail(&cache->io_list, io);
3482                         } else {
3483                                 /*
3484                                  * if we failed to write the cache, the
3485                                  * generation will be bad and life goes on
3486                                  */
3487                                 ret = 0;
3488                         }
3489                 }
3490                 if (!ret)
3491                         ret = write_one_cache_group(trans, root, path, cache);
3492
3493                 /* if its not on the io list, we need to put the block group */
3494                 if (should_put)
3495                         btrfs_put_block_group(cache);
3496
3497                 if (ret)
3498                         break;
3499
3500                 /*
3501                  * Avoid blocking other tasks for too long. It might even save
3502                  * us from writing caches for block groups that are going to be
3503                  * removed.
3504                  */
3505                 mutex_unlock(&trans->transaction->cache_write_mutex);
3506                 mutex_lock(&trans->transaction->cache_write_mutex);
3507         }
3508         mutex_unlock(&trans->transaction->cache_write_mutex);
3509
3510         /*
3511          * go through delayed refs for all the stuff we've just kicked off
3512          * and then loop back (just once)
3513          */
3514         ret = btrfs_run_delayed_refs(trans, root, 0);
3515         if (!ret && loops == 0) {
3516                 loops++;
3517                 spin_lock(&cur_trans->dirty_bgs_lock);
3518                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3519                 /*
3520                  * dirty_bgs_lock protects us from concurrent block group
3521                  * deletes too (not just cache_write_mutex).
3522                  */
3523                 if (!list_empty(&dirty)) {
3524                         spin_unlock(&cur_trans->dirty_bgs_lock);
3525                         goto again;
3526                 }
3527                 spin_unlock(&cur_trans->dirty_bgs_lock);
3528         }
3529
3530         btrfs_free_path(path);
3531         return ret;
3532 }
3533
3534 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3535                                    struct btrfs_root *root)
3536 {
3537         struct btrfs_block_group_cache *cache;
3538         struct btrfs_transaction *cur_trans = trans->transaction;
3539         int ret = 0;
3540         int should_put;
3541         struct btrfs_path *path;
3542         struct list_head *io = &cur_trans->io_bgs;
3543         int num_started = 0;
3544
3545         path = btrfs_alloc_path();
3546         if (!path)
3547                 return -ENOMEM;
3548
3549         /*
3550          * We don't need the lock here since we are protected by the transaction
3551          * commit.  We want to do the cache_save_setup first and then run the
3552          * delayed refs to make sure we have the best chance at doing this all
3553          * in one shot.
3554          */
3555         while (!list_empty(&cur_trans->dirty_bgs)) {
3556                 cache = list_first_entry(&cur_trans->dirty_bgs,
3557                                          struct btrfs_block_group_cache,
3558                                          dirty_list);
3559
3560                 /*
3561                  * this can happen if cache_save_setup re-dirties a block
3562                  * group that is already under IO.  Just wait for it to
3563                  * finish and then do it all again
3564                  */
3565                 if (!list_empty(&cache->io_list)) {
3566                         list_del_init(&cache->io_list);
3567                         btrfs_wait_cache_io(root, trans, cache,
3568                                             &cache->io_ctl, path,
3569                                             cache->key.objectid);
3570                         btrfs_put_block_group(cache);
3571                 }
3572
3573                 /*
3574                  * don't remove from the dirty list until after we've waited
3575                  * on any pending IO
3576                  */
3577                 list_del_init(&cache->dirty_list);
3578                 should_put = 1;
3579
3580                 cache_save_setup(cache, trans, path);
3581
3582                 if (!ret)
3583                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3584
3585                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3586                         cache->io_ctl.inode = NULL;
3587                         ret = btrfs_write_out_cache(root, trans, cache, path);
3588                         if (ret == 0 && cache->io_ctl.inode) {
3589                                 num_started++;
3590                                 should_put = 0;
3591                                 list_add_tail(&cache->io_list, io);
3592                         } else {
3593                                 /*
3594                                  * if we failed to write the cache, the
3595                                  * generation will be bad and life goes on
3596                                  */
3597                                 ret = 0;
3598                         }
3599                 }
3600                 if (!ret)
3601                         ret = write_one_cache_group(trans, root, path, cache);
3602
3603                 /* if its not on the io list, we need to put the block group */
3604                 if (should_put)
3605                         btrfs_put_block_group(cache);
3606         }
3607
3608         while (!list_empty(io)) {
3609                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3610                                          io_list);
3611                 list_del_init(&cache->io_list);
3612                 btrfs_wait_cache_io(root, trans, cache,
3613                                     &cache->io_ctl, path, cache->key.objectid);
3614                 btrfs_put_block_group(cache);
3615         }
3616
3617         btrfs_free_path(path);
3618         return ret;
3619 }
3620
3621 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3622 {
3623         struct btrfs_block_group_cache *block_group;
3624         int readonly = 0;
3625
3626         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3627         if (!block_group || block_group->ro)
3628                 readonly = 1;
3629         if (block_group)
3630                 btrfs_put_block_group(block_group);
3631         return readonly;
3632 }
3633
3634 static const char *alloc_name(u64 flags)
3635 {
3636         switch (flags) {
3637         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3638                 return "mixed";
3639         case BTRFS_BLOCK_GROUP_METADATA:
3640                 return "metadata";
3641         case BTRFS_BLOCK_GROUP_DATA:
3642                 return "data";
3643         case BTRFS_BLOCK_GROUP_SYSTEM:
3644                 return "system";
3645         default:
3646                 WARN_ON(1);
3647                 return "invalid-combination";
3648         };
3649 }
3650
3651 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3652                              u64 total_bytes, u64 bytes_used,
3653                              struct btrfs_space_info **space_info)
3654 {
3655         struct btrfs_space_info *found;
3656         int i;
3657         int factor;
3658         int ret;
3659
3660         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3661                      BTRFS_BLOCK_GROUP_RAID10))
3662                 factor = 2;
3663         else
3664                 factor = 1;
3665
3666         found = __find_space_info(info, flags);
3667         if (found) {
3668                 spin_lock(&found->lock);
3669                 found->total_bytes += total_bytes;
3670                 found->disk_total += total_bytes * factor;
3671                 found->bytes_used += bytes_used;
3672                 found->disk_used += bytes_used * factor;
3673                 found->full = 0;
3674                 spin_unlock(&found->lock);
3675                 *space_info = found;
3676                 return 0;
3677         }
3678         found = kzalloc(sizeof(*found), GFP_NOFS);
3679         if (!found)
3680                 return -ENOMEM;
3681
3682         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3683         if (ret) {
3684                 kfree(found);
3685                 return ret;
3686         }
3687
3688         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3689                 INIT_LIST_HEAD(&found->block_groups[i]);
3690         init_rwsem(&found->groups_sem);
3691         spin_lock_init(&found->lock);
3692         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3693         found->total_bytes = total_bytes;
3694         found->disk_total = total_bytes * factor;
3695         found->bytes_used = bytes_used;
3696         found->disk_used = bytes_used * factor;
3697         found->bytes_pinned = 0;
3698         found->bytes_reserved = 0;
3699         found->bytes_readonly = 0;
3700         found->bytes_may_use = 0;
3701         found->full = 0;
3702         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3703         found->chunk_alloc = 0;
3704         found->flush = 0;
3705         init_waitqueue_head(&found->wait);
3706         INIT_LIST_HEAD(&found->ro_bgs);
3707
3708         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3709                                     info->space_info_kobj, "%s",
3710                                     alloc_name(found->flags));
3711         if (ret) {
3712                 kfree(found);
3713                 return ret;
3714         }
3715
3716         *space_info = found;
3717         list_add_rcu(&found->list, &info->space_info);
3718         if (flags & BTRFS_BLOCK_GROUP_DATA)
3719                 info->data_sinfo = found;
3720
3721         return ret;
3722 }
3723
3724 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3725 {
3726         u64 extra_flags = chunk_to_extended(flags) &
3727                                 BTRFS_EXTENDED_PROFILE_MASK;
3728
3729         write_seqlock(&fs_info->profiles_lock);
3730         if (flags & BTRFS_BLOCK_GROUP_DATA)
3731                 fs_info->avail_data_alloc_bits |= extra_flags;
3732         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3733                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3734         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3735                 fs_info->avail_system_alloc_bits |= extra_flags;
3736         write_sequnlock(&fs_info->profiles_lock);
3737 }
3738
3739 /*
3740  * returns target flags in extended format or 0 if restripe for this
3741  * chunk_type is not in progress
3742  *
3743  * should be called with either volume_mutex or balance_lock held
3744  */
3745 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3746 {
3747         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3748         u64 target = 0;
3749
3750         if (!bctl)
3751                 return 0;
3752
3753         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3754             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3755                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3756         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3757                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3758                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3759         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3760                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3761                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3762         }
3763
3764         return target;
3765 }
3766
3767 /*
3768  * @flags: available profiles in extended format (see ctree.h)
3769  *
3770  * Returns reduced profile in chunk format.  If profile changing is in
3771  * progress (either running or paused) picks the target profile (if it's
3772  * already available), otherwise falls back to plain reducing.
3773  */
3774 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3775 {
3776         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3777         u64 target;
3778         u64 tmp;
3779
3780         /*
3781          * see if restripe for this chunk_type is in progress, if so
3782          * try to reduce to the target profile
3783          */
3784         spin_lock(&root->fs_info->balance_lock);
3785         target = get_restripe_target(root->fs_info, flags);
3786         if (target) {
3787                 /* pick target profile only if it's already available */
3788                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3789                         spin_unlock(&root->fs_info->balance_lock);
3790                         return extended_to_chunk(target);
3791                 }
3792         }
3793         spin_unlock(&root->fs_info->balance_lock);
3794
3795         /* First, mask out the RAID levels which aren't possible */
3796         if (num_devices == 1)
3797                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3798                            BTRFS_BLOCK_GROUP_RAID5);
3799         if (num_devices < 3)
3800                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3801         if (num_devices < 4)
3802                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3803
3804         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3805                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3806                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3807         flags &= ~tmp;
3808
3809         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3810                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3811         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3812                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3813         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3814                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3815         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3816                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3817         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3818                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3819
3820         return extended_to_chunk(flags | tmp);
3821 }
3822
3823 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3824 {
3825         unsigned seq;
3826         u64 flags;
3827
3828         do {
3829                 flags = orig_flags;
3830                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3831
3832                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3833                         flags |= root->fs_info->avail_data_alloc_bits;
3834                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3835                         flags |= root->fs_info->avail_system_alloc_bits;
3836                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3837                         flags |= root->fs_info->avail_metadata_alloc_bits;
3838         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3839
3840         return btrfs_reduce_alloc_profile(root, flags);
3841 }
3842
3843 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3844 {
3845         u64 flags;
3846         u64 ret;
3847
3848         if (data)
3849                 flags = BTRFS_BLOCK_GROUP_DATA;
3850         else if (root == root->fs_info->chunk_root)
3851                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3852         else
3853                 flags = BTRFS_BLOCK_GROUP_METADATA;
3854
3855         ret = get_alloc_profile(root, flags);
3856         return ret;
3857 }
3858
3859 /*
3860  * This will check the space that the inode allocates from to make sure we have
3861  * enough space for bytes.
3862  */
3863 int btrfs_check_data_free_space(struct inode *inode, u64 bytes, u64 write_bytes)
3864 {
3865         struct btrfs_space_info *data_sinfo;
3866         struct btrfs_root *root = BTRFS_I(inode)->root;
3867         struct btrfs_fs_info *fs_info = root->fs_info;
3868         u64 used;
3869         int ret = 0;
3870         int need_commit = 2;
3871         int have_pinned_space;
3872
3873         /* make sure bytes are sectorsize aligned */
3874         bytes = ALIGN(bytes, root->sectorsize);
3875
3876         if (btrfs_is_free_space_inode(inode)) {
3877                 need_commit = 0;
3878                 ASSERT(current->journal_info);
3879         }
3880
3881         data_sinfo = fs_info->data_sinfo;
3882         if (!data_sinfo)
3883                 goto alloc;
3884
3885 again:
3886         /* make sure we have enough space to handle the data first */
3887         spin_lock(&data_sinfo->lock);
3888         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3889                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3890                 data_sinfo->bytes_may_use;
3891
3892         if (used + bytes > data_sinfo->total_bytes) {
3893                 struct btrfs_trans_handle *trans;
3894
3895                 /*
3896                  * if we don't have enough free bytes in this space then we need
3897                  * to alloc a new chunk.
3898                  */
3899                 if (!data_sinfo->full) {
3900                         u64 alloc_target;
3901
3902                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3903                         spin_unlock(&data_sinfo->lock);
3904 alloc:
3905                         alloc_target = btrfs_get_alloc_profile(root, 1);
3906                         /*
3907                          * It is ugly that we don't call nolock join
3908                          * transaction for the free space inode case here.
3909                          * But it is safe because we only do the data space
3910                          * reservation for the free space cache in the
3911                          * transaction context, the common join transaction
3912                          * just increase the counter of the current transaction
3913                          * handler, doesn't try to acquire the trans_lock of
3914                          * the fs.
3915                          */
3916                         trans = btrfs_join_transaction(root);
3917                         if (IS_ERR(trans))
3918                                 return PTR_ERR(trans);
3919
3920                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3921                                              alloc_target,
3922                                              CHUNK_ALLOC_NO_FORCE);
3923                         btrfs_end_transaction(trans, root);
3924                         if (ret < 0) {
3925                                 if (ret != -ENOSPC)
3926                                         return ret;
3927                                 else {
3928                                         have_pinned_space = 1;
3929                                         goto commit_trans;
3930                                 }
3931                         }
3932
3933                         if (!data_sinfo)
3934                                 data_sinfo = fs_info->data_sinfo;
3935
3936                         goto again;
3937                 }
3938
3939                 /*
3940                  * If we don't have enough pinned space to deal with this
3941                  * allocation, and no removed chunk in current transaction,
3942                  * don't bother committing the transaction.
3943                  */
3944                 have_pinned_space = percpu_counter_compare(
3945                         &data_sinfo->total_bytes_pinned,
3946                         used + bytes - data_sinfo->total_bytes);
3947                 spin_unlock(&data_sinfo->lock);
3948
3949                 /* commit the current transaction and try again */
3950 commit_trans:
3951                 if (need_commit &&
3952                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3953                         need_commit--;
3954
3955                         trans = btrfs_join_transaction(root);
3956                         if (IS_ERR(trans))
3957                                 return PTR_ERR(trans);
3958                         if (have_pinned_space >= 0 ||
3959                             trans->transaction->have_free_bgs ||
3960                             need_commit > 0) {
3961                                 ret = btrfs_commit_transaction(trans, root);
3962                                 if (ret)
3963                                         return ret;
3964                                 /*
3965                                  * make sure that all running delayed iput are
3966                                  * done
3967                                  */
3968                                 down_write(&root->fs_info->delayed_iput_sem);
3969                                 up_write(&root->fs_info->delayed_iput_sem);
3970                                 goto again;
3971                         } else {
3972                                 btrfs_end_transaction(trans, root);
3973                         }
3974                 }
3975
3976                 trace_btrfs_space_reservation(root->fs_info,
3977                                               "space_info:enospc",
3978                                               data_sinfo->flags, bytes, 1);
3979                 return -ENOSPC;
3980         }
3981         ret = btrfs_qgroup_reserve(root, write_bytes);
3982         if (ret)
3983                 goto out;
3984         data_sinfo->bytes_may_use += bytes;
3985         trace_btrfs_space_reservation(root->fs_info, "space_info",
3986                                       data_sinfo->flags, bytes, 1);
3987 out:
3988         spin_unlock(&data_sinfo->lock);
3989
3990         return ret;
3991 }
3992
3993 /*
3994  * Called if we need to clear a data reservation for this inode.
3995  */
3996 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3997 {
3998         struct btrfs_root *root = BTRFS_I(inode)->root;
3999         struct btrfs_space_info *data_sinfo;
4000
4001         /* make sure bytes are sectorsize aligned */
4002         bytes = ALIGN(bytes, root->sectorsize);
4003
4004         data_sinfo = root->fs_info->data_sinfo;
4005         spin_lock(&data_sinfo->lock);
4006         WARN_ON(data_sinfo->bytes_may_use < bytes);
4007         data_sinfo->bytes_may_use -= bytes;
4008         trace_btrfs_space_reservation(root->fs_info, "space_info",
4009                                       data_sinfo->flags, bytes, 0);
4010         spin_unlock(&data_sinfo->lock);
4011 }
4012
4013 static void force_metadata_allocation(struct btrfs_fs_info *info)
4014 {
4015         struct list_head *head = &info->space_info;
4016         struct btrfs_space_info *found;
4017
4018         rcu_read_lock();
4019         list_for_each_entry_rcu(found, head, list) {
4020                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4021                         found->force_alloc = CHUNK_ALLOC_FORCE;
4022         }
4023         rcu_read_unlock();
4024 }
4025
4026 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4027 {
4028         return (global->size << 1);
4029 }
4030
4031 static int should_alloc_chunk(struct btrfs_root *root,
4032                               struct btrfs_space_info *sinfo, int force)
4033 {
4034         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4035         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4036         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4037         u64 thresh;
4038
4039         if (force == CHUNK_ALLOC_FORCE)
4040                 return 1;
4041
4042         /*
4043          * We need to take into account the global rsv because for all intents
4044          * and purposes it's used space.  Don't worry about locking the
4045          * global_rsv, it doesn't change except when the transaction commits.
4046          */
4047         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4048                 num_allocated += calc_global_rsv_need_space(global_rsv);
4049
4050         /*
4051          * in limited mode, we want to have some free space up to
4052          * about 1% of the FS size.
4053          */
4054         if (force == CHUNK_ALLOC_LIMITED) {
4055                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4056                 thresh = max_t(u64, 64 * 1024 * 1024,
4057                                div_factor_fine(thresh, 1));
4058
4059                 if (num_bytes - num_allocated < thresh)
4060                         return 1;
4061         }
4062
4063         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
4064                 return 0;
4065         return 1;
4066 }
4067
4068 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
4069 {
4070         u64 num_dev;
4071
4072         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4073                     BTRFS_BLOCK_GROUP_RAID0 |
4074                     BTRFS_BLOCK_GROUP_RAID5 |
4075                     BTRFS_BLOCK_GROUP_RAID6))
4076                 num_dev = root->fs_info->fs_devices->rw_devices;
4077         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4078                 num_dev = 2;
4079         else
4080                 num_dev = 1;    /* DUP or single */
4081
4082         /* metadata for updaing devices and chunk tree */
4083         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
4084 }
4085
4086 static void check_system_chunk(struct btrfs_trans_handle *trans,
4087                                struct btrfs_root *root, u64 type)
4088 {
4089         struct btrfs_space_info *info;
4090         u64 left;
4091         u64 thresh;
4092
4093         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4094         spin_lock(&info->lock);
4095         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4096                 info->bytes_reserved - info->bytes_readonly;
4097         spin_unlock(&info->lock);
4098
4099         thresh = get_system_chunk_thresh(root, type);
4100         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4101                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4102                         left, thresh, type);
4103                 dump_space_info(info, 0, 0);
4104         }
4105
4106         if (left < thresh) {
4107                 u64 flags;
4108
4109                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4110                 btrfs_alloc_chunk(trans, root, flags);
4111         }
4112 }
4113
4114 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4115                           struct btrfs_root *extent_root, u64 flags, int force)
4116 {
4117         struct btrfs_space_info *space_info;
4118         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4119         int wait_for_alloc = 0;
4120         int ret = 0;
4121
4122         /* Don't re-enter if we're already allocating a chunk */
4123         if (trans->allocating_chunk)
4124                 return -ENOSPC;
4125
4126         space_info = __find_space_info(extent_root->fs_info, flags);
4127         if (!space_info) {
4128                 ret = update_space_info(extent_root->fs_info, flags,
4129                                         0, 0, &space_info);
4130                 BUG_ON(ret); /* -ENOMEM */
4131         }
4132         BUG_ON(!space_info); /* Logic error */
4133
4134 again:
4135         spin_lock(&space_info->lock);
4136         if (force < space_info->force_alloc)
4137                 force = space_info->force_alloc;
4138         if (space_info->full) {
4139                 if (should_alloc_chunk(extent_root, space_info, force))
4140                         ret = -ENOSPC;
4141                 else
4142                         ret = 0;
4143                 spin_unlock(&space_info->lock);
4144                 return ret;
4145         }
4146
4147         if (!should_alloc_chunk(extent_root, space_info, force)) {
4148                 spin_unlock(&space_info->lock);
4149                 return 0;
4150         } else if (space_info->chunk_alloc) {
4151                 wait_for_alloc = 1;
4152         } else {
4153                 space_info->chunk_alloc = 1;
4154         }
4155
4156         spin_unlock(&space_info->lock);
4157
4158         mutex_lock(&fs_info->chunk_mutex);
4159
4160         /*
4161          * The chunk_mutex is held throughout the entirety of a chunk
4162          * allocation, so once we've acquired the chunk_mutex we know that the
4163          * other guy is done and we need to recheck and see if we should
4164          * allocate.
4165          */
4166         if (wait_for_alloc) {
4167                 mutex_unlock(&fs_info->chunk_mutex);
4168                 wait_for_alloc = 0;
4169                 goto again;
4170         }
4171
4172         trans->allocating_chunk = true;
4173
4174         /*
4175          * If we have mixed data/metadata chunks we want to make sure we keep
4176          * allocating mixed chunks instead of individual chunks.
4177          */
4178         if (btrfs_mixed_space_info(space_info))
4179                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4180
4181         /*
4182          * if we're doing a data chunk, go ahead and make sure that
4183          * we keep a reasonable number of metadata chunks allocated in the
4184          * FS as well.
4185          */
4186         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4187                 fs_info->data_chunk_allocations++;
4188                 if (!(fs_info->data_chunk_allocations %
4189                       fs_info->metadata_ratio))
4190                         force_metadata_allocation(fs_info);
4191         }
4192
4193         /*
4194          * Check if we have enough space in SYSTEM chunk because we may need
4195          * to update devices.
4196          */
4197         check_system_chunk(trans, extent_root, flags);
4198
4199         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4200         trans->allocating_chunk = false;
4201
4202         spin_lock(&space_info->lock);
4203         if (ret < 0 && ret != -ENOSPC)
4204                 goto out;
4205         if (ret)
4206                 space_info->full = 1;
4207         else
4208                 ret = 1;
4209
4210         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4211 out:
4212         space_info->chunk_alloc = 0;
4213         spin_unlock(&space_info->lock);
4214         mutex_unlock(&fs_info->chunk_mutex);
4215         return ret;
4216 }
4217
4218 static int can_overcommit(struct btrfs_root *root,
4219                           struct btrfs_space_info *space_info, u64 bytes,
4220                           enum btrfs_reserve_flush_enum flush)
4221 {
4222         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4223         u64 profile = btrfs_get_alloc_profile(root, 0);
4224         u64 space_size;
4225         u64 avail;
4226         u64 used;
4227
4228         used = space_info->bytes_used + space_info->bytes_reserved +
4229                 space_info->bytes_pinned + space_info->bytes_readonly;
4230
4231         /*
4232          * We only want to allow over committing if we have lots of actual space
4233          * free, but if we don't have enough space to handle the global reserve
4234          * space then we could end up having a real enospc problem when trying
4235          * to allocate a chunk or some other such important allocation.
4236          */
4237         spin_lock(&global_rsv->lock);
4238         space_size = calc_global_rsv_need_space(global_rsv);
4239         spin_unlock(&global_rsv->lock);
4240         if (used + space_size >= space_info->total_bytes)
4241                 return 0;
4242
4243         used += space_info->bytes_may_use;
4244
4245         spin_lock(&root->fs_info->free_chunk_lock);
4246         avail = root->fs_info->free_chunk_space;
4247         spin_unlock(&root->fs_info->free_chunk_lock);
4248
4249         /*
4250          * If we have dup, raid1 or raid10 then only half of the free
4251          * space is actually useable.  For raid56, the space info used
4252          * doesn't include the parity drive, so we don't have to
4253          * change the math
4254          */
4255         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4256                        BTRFS_BLOCK_GROUP_RAID1 |
4257                        BTRFS_BLOCK_GROUP_RAID10))
4258                 avail >>= 1;
4259
4260         /*
4261          * If we aren't flushing all things, let us overcommit up to
4262          * 1/2th of the space. If we can flush, don't let us overcommit
4263          * too much, let it overcommit up to 1/8 of the space.
4264          */
4265         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4266                 avail >>= 3;
4267         else
4268                 avail >>= 1;
4269
4270         if (used + bytes < space_info->total_bytes + avail)
4271                 return 1;
4272         return 0;
4273 }
4274
4275 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4276                                          unsigned long nr_pages, int nr_items)
4277 {
4278         struct super_block *sb = root->fs_info->sb;
4279
4280         if (down_read_trylock(&sb->s_umount)) {
4281                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4282                 up_read(&sb->s_umount);
4283         } else {
4284                 /*
4285                  * We needn't worry the filesystem going from r/w to r/o though
4286                  * we don't acquire ->s_umount mutex, because the filesystem
4287                  * should guarantee the delalloc inodes list be empty after
4288                  * the filesystem is readonly(all dirty pages are written to
4289                  * the disk).
4290                  */
4291                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4292                 if (!current->journal_info)
4293                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4294         }
4295 }
4296
4297 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4298 {
4299         u64 bytes;
4300         int nr;
4301
4302         bytes = btrfs_calc_trans_metadata_size(root, 1);
4303         nr = (int)div64_u64(to_reclaim, bytes);
4304         if (!nr)
4305                 nr = 1;
4306         return nr;
4307 }
4308
4309 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4310
4311 /*
4312  * shrink metadata reservation for delalloc
4313  */
4314 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4315                             bool wait_ordered)
4316 {
4317         struct btrfs_block_rsv *block_rsv;
4318         struct btrfs_space_info *space_info;
4319         struct btrfs_trans_handle *trans;
4320         u64 delalloc_bytes;
4321         u64 max_reclaim;
4322         long time_left;
4323         unsigned long nr_pages;
4324         int loops;
4325         int items;
4326         enum btrfs_reserve_flush_enum flush;
4327
4328         /* Calc the number of the pages we need flush for space reservation */
4329         items = calc_reclaim_items_nr(root, to_reclaim);
4330         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4331
4332         trans = (struct btrfs_trans_handle *)current->journal_info;
4333         block_rsv = &root->fs_info->delalloc_block_rsv;
4334         space_info = block_rsv->space_info;
4335
4336         delalloc_bytes = percpu_counter_sum_positive(
4337                                                 &root->fs_info->delalloc_bytes);
4338         if (delalloc_bytes == 0) {
4339                 if (trans)
4340                         return;
4341                 if (wait_ordered)
4342                         btrfs_wait_ordered_roots(root->fs_info, items);
4343                 return;
4344         }
4345
4346         loops = 0;
4347         while (delalloc_bytes && loops < 3) {
4348                 max_reclaim = min(delalloc_bytes, to_reclaim);
4349                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4350                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4351                 /*
4352                  * We need to wait for the async pages to actually start before
4353                  * we do anything.
4354                  */
4355                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4356                 if (!max_reclaim)
4357                         goto skip_async;
4358
4359                 if (max_reclaim <= nr_pages)
4360                         max_reclaim = 0;
4361                 else
4362                         max_reclaim -= nr_pages;
4363
4364                 wait_event(root->fs_info->async_submit_wait,
4365                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4366                            (int)max_reclaim);
4367 skip_async:
4368                 if (!trans)
4369                         flush = BTRFS_RESERVE_FLUSH_ALL;
4370                 else
4371                         flush = BTRFS_RESERVE_NO_FLUSH;
4372                 spin_lock(&space_info->lock);
4373                 if (can_overcommit(root, space_info, orig, flush)) {
4374                         spin_unlock(&space_info->lock);
4375                         break;
4376                 }
4377                 spin_unlock(&space_info->lock);
4378
4379                 loops++;
4380                 if (wait_ordered && !trans) {
4381                         btrfs_wait_ordered_roots(root->fs_info, items);
4382                 } else {
4383                         time_left = schedule_timeout_killable(1);
4384                         if (time_left)
4385                                 break;
4386                 }
4387                 delalloc_bytes = percpu_counter_sum_positive(
4388                                                 &root->fs_info->delalloc_bytes);
4389         }
4390 }
4391
4392 /**
4393  * maybe_commit_transaction - possibly commit the transaction if its ok to
4394  * @root - the root we're allocating for
4395  * @bytes - the number of bytes we want to reserve
4396  * @force - force the commit
4397  *
4398  * This will check to make sure that committing the transaction will actually
4399  * get us somewhere and then commit the transaction if it does.  Otherwise it
4400  * will return -ENOSPC.
4401  */
4402 static int may_commit_transaction(struct btrfs_root *root,
4403                                   struct btrfs_space_info *space_info,
4404                                   u64 bytes, int force)
4405 {
4406         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4407         struct btrfs_trans_handle *trans;
4408
4409         trans = (struct btrfs_trans_handle *)current->journal_info;
4410         if (trans)
4411                 return -EAGAIN;
4412
4413         if (force)
4414                 goto commit;
4415
4416         /* See if there is enough pinned space to make this reservation */
4417         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4418                                    bytes) >= 0)
4419                 goto commit;
4420
4421         /*
4422          * See if there is some space in the delayed insertion reservation for
4423          * this reservation.
4424          */
4425         if (space_info != delayed_rsv->space_info)
4426                 return -ENOSPC;
4427
4428         spin_lock(&delayed_rsv->lock);
4429         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4430                                    bytes - delayed_rsv->size) >= 0) {
4431                 spin_unlock(&delayed_rsv->lock);
4432                 return -ENOSPC;
4433         }
4434         spin_unlock(&delayed_rsv->lock);
4435
4436 commit:
4437         trans = btrfs_join_transaction(root);
4438         if (IS_ERR(trans))
4439                 return -ENOSPC;
4440
4441         return btrfs_commit_transaction(trans, root);
4442 }
4443
4444 enum flush_state {
4445         FLUSH_DELAYED_ITEMS_NR  =       1,
4446         FLUSH_DELAYED_ITEMS     =       2,
4447         FLUSH_DELALLOC          =       3,
4448         FLUSH_DELALLOC_WAIT     =       4,
4449         ALLOC_CHUNK             =       5,
4450         COMMIT_TRANS            =       6,
4451 };
4452
4453 static int flush_space(struct btrfs_root *root,
4454                        struct btrfs_space_info *space_info, u64 num_bytes,
4455                        u64 orig_bytes, int state)
4456 {
4457         struct btrfs_trans_handle *trans;
4458         int nr;
4459         int ret = 0;
4460
4461         switch (state) {
4462         case FLUSH_DELAYED_ITEMS_NR:
4463         case FLUSH_DELAYED_ITEMS:
4464                 if (state == FLUSH_DELAYED_ITEMS_NR)
4465                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4466                 else
4467                         nr = -1;
4468
4469                 trans = btrfs_join_transaction(root);
4470                 if (IS_ERR(trans)) {
4471                         ret = PTR_ERR(trans);
4472                         break;
4473                 }
4474                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4475                 btrfs_end_transaction(trans, root);
4476                 break;
4477         case FLUSH_DELALLOC:
4478         case FLUSH_DELALLOC_WAIT:
4479                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4480                                 state == FLUSH_DELALLOC_WAIT);
4481                 break;
4482         case ALLOC_CHUNK:
4483                 trans = btrfs_join_transaction(root);
4484                 if (IS_ERR(trans)) {
4485                         ret = PTR_ERR(trans);
4486                         break;
4487                 }
4488                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4489                                      btrfs_get_alloc_profile(root, 0),
4490                                      CHUNK_ALLOC_NO_FORCE);
4491                 btrfs_end_transaction(trans, root);
4492                 if (ret == -ENOSPC)
4493                         ret = 0;
4494                 break;
4495         case COMMIT_TRANS:
4496                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4497                 break;
4498         default:
4499                 ret = -ENOSPC;
4500                 break;
4501         }
4502
4503         return ret;
4504 }
4505
4506 static inline u64
4507 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4508                                  struct btrfs_space_info *space_info)
4509 {
4510         u64 used;
4511         u64 expected;
4512         u64 to_reclaim;
4513
4514         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4515                                 16 * 1024 * 1024);
4516         spin_lock(&space_info->lock);
4517         if (can_overcommit(root, space_info, to_reclaim,
4518                            BTRFS_RESERVE_FLUSH_ALL)) {
4519                 to_reclaim = 0;
4520                 goto out;
4521         }
4522
4523         used = space_info->bytes_used + space_info->bytes_reserved +
4524                space_info->bytes_pinned + space_info->bytes_readonly +
4525                space_info->bytes_may_use;
4526         if (can_overcommit(root, space_info, 1024 * 1024,
4527                            BTRFS_RESERVE_FLUSH_ALL))
4528                 expected = div_factor_fine(space_info->total_bytes, 95);
4529         else
4530                 expected = div_factor_fine(space_info->total_bytes, 90);
4531
4532         if (used > expected)
4533                 to_reclaim = used - expected;
4534         else
4535                 to_reclaim = 0;
4536         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4537                                      space_info->bytes_reserved);
4538 out:
4539         spin_unlock(&space_info->lock);
4540
4541         return to_reclaim;
4542 }
4543
4544 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4545                                         struct btrfs_fs_info *fs_info, u64 used)
4546 {
4547         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4548
4549         /* If we're just plain full then async reclaim just slows us down. */
4550         if (space_info->bytes_used >= thresh)
4551                 return 0;
4552
4553         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4554                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4555 }
4556
4557 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4558                                        struct btrfs_fs_info *fs_info,
4559                                        int flush_state)
4560 {
4561         u64 used;
4562
4563         spin_lock(&space_info->lock);
4564         /*
4565          * We run out of space and have not got any free space via flush_space,
4566          * so don't bother doing async reclaim.
4567          */
4568         if (flush_state > COMMIT_TRANS && space_info->full) {
4569                 spin_unlock(&space_info->lock);
4570                 return 0;
4571         }
4572
4573         used = space_info->bytes_used + space_info->bytes_reserved +
4574                space_info->bytes_pinned + space_info->bytes_readonly +
4575                space_info->bytes_may_use;
4576         if (need_do_async_reclaim(space_info, fs_info, used)) {
4577                 spin_unlock(&space_info->lock);
4578                 return 1;
4579         }
4580         spin_unlock(&space_info->lock);
4581
4582         return 0;
4583 }
4584
4585 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4586 {
4587         struct btrfs_fs_info *fs_info;
4588         struct btrfs_space_info *space_info;
4589         u64 to_reclaim;
4590         int flush_state;
4591
4592         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4593         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4594
4595         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4596                                                       space_info);
4597         if (!to_reclaim)
4598                 return;
4599
4600         flush_state = FLUSH_DELAYED_ITEMS_NR;
4601         do {
4602                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4603                             to_reclaim, flush_state);
4604                 flush_state++;
4605                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4606                                                  flush_state))
4607                         return;
4608         } while (flush_state < COMMIT_TRANS);
4609 }
4610
4611 void btrfs_init_async_reclaim_work(struct work_struct *work)
4612 {
4613         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4614 }
4615
4616 /**
4617  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4618  * @root - the root we're allocating for
4619  * @block_rsv - the block_rsv we're allocating for
4620  * @orig_bytes - the number of bytes we want
4621  * @flush - whether or not we can flush to make our reservation
4622  *
4623  * This will reserve orgi_bytes number of bytes from the space info associated
4624  * with the block_rsv.  If there is not enough space it will make an attempt to
4625  * flush out space to make room.  It will do this by flushing delalloc if
4626  * possible or committing the transaction.  If flush is 0 then no attempts to
4627  * regain reservations will be made and this will fail if there is not enough
4628  * space already.
4629  */
4630 static int reserve_metadata_bytes(struct btrfs_root *root,
4631                                   struct btrfs_block_rsv *block_rsv,
4632                                   u64 orig_bytes,
4633                                   enum btrfs_reserve_flush_enum flush)
4634 {
4635         struct btrfs_space_info *space_info = block_rsv->space_info;
4636         u64 used;
4637         u64 num_bytes = orig_bytes;
4638         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4639         int ret = 0;
4640         bool flushing = false;
4641
4642 again:
4643         ret = 0;
4644         spin_lock(&space_info->lock);
4645         /*
4646          * We only want to wait if somebody other than us is flushing and we
4647          * are actually allowed to flush all things.
4648          */
4649         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4650                space_info->flush) {
4651                 spin_unlock(&space_info->lock);
4652                 /*
4653                  * If we have a trans handle we can't wait because the flusher
4654                  * may have to commit the transaction, which would mean we would
4655                  * deadlock since we are waiting for the flusher to finish, but
4656                  * hold the current transaction open.
4657                  */
4658                 if (current->journal_info)
4659                         return -EAGAIN;
4660                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4661                 /* Must have been killed, return */
4662                 if (ret)
4663                         return -EINTR;
4664
4665                 spin_lock(&space_info->lock);
4666         }
4667
4668         ret = -ENOSPC;
4669         used = space_info->bytes_used + space_info->bytes_reserved +
4670                 space_info->bytes_pinned + space_info->bytes_readonly +
4671                 space_info->bytes_may_use;
4672
4673         /*
4674          * The idea here is that we've not already over-reserved the block group
4675          * then we can go ahead and save our reservation first and then start
4676          * flushing if we need to.  Otherwise if we've already overcommitted
4677          * lets start flushing stuff first and then come back and try to make
4678          * our reservation.
4679          */
4680         if (used <= space_info->total_bytes) {
4681                 if (used + orig_bytes <= space_info->total_bytes) {
4682                         space_info->bytes_may_use += orig_bytes;
4683                         trace_btrfs_space_reservation(root->fs_info,
4684                                 "space_info", space_info->flags, orig_bytes, 1);
4685                         ret = 0;
4686                 } else {
4687                         /*
4688                          * Ok set num_bytes to orig_bytes since we aren't
4689                          * overocmmitted, this way we only try and reclaim what
4690                          * we need.
4691                          */
4692                         num_bytes = orig_bytes;
4693                 }
4694         } else {
4695                 /*
4696                  * Ok we're over committed, set num_bytes to the overcommitted
4697                  * amount plus the amount of bytes that we need for this
4698                  * reservation.
4699                  */
4700                 num_bytes = used - space_info->total_bytes +
4701                         (orig_bytes * 2);
4702         }
4703
4704         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4705                 space_info->bytes_may_use += orig_bytes;
4706                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4707                                               space_info->flags, orig_bytes,
4708                                               1);
4709                 ret = 0;
4710         }
4711
4712         /*
4713          * Couldn't make our reservation, save our place so while we're trying
4714          * to reclaim space we can actually use it instead of somebody else
4715          * stealing it from us.
4716          *
4717          * We make the other tasks wait for the flush only when we can flush
4718          * all things.
4719          */
4720         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4721                 flushing = true;
4722                 space_info->flush = 1;
4723         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4724                 used += orig_bytes;
4725                 /*
4726                  * We will do the space reservation dance during log replay,
4727                  * which means we won't have fs_info->fs_root set, so don't do
4728                  * the async reclaim as we will panic.
4729                  */
4730                 if (!root->fs_info->log_root_recovering &&
4731                     need_do_async_reclaim(space_info, root->fs_info, used) &&
4732                     !work_busy(&root->fs_info->async_reclaim_work))
4733                         queue_work(system_unbound_wq,
4734                                    &root->fs_info->async_reclaim_work);
4735         }
4736         spin_unlock(&space_info->lock);
4737
4738         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4739                 goto out;
4740
4741         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4742                           flush_state);
4743         flush_state++;
4744
4745         /*
4746          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4747          * would happen. So skip delalloc flush.
4748          */
4749         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4750             (flush_state == FLUSH_DELALLOC ||
4751              flush_state == FLUSH_DELALLOC_WAIT))
4752                 flush_state = ALLOC_CHUNK;
4753
4754         if (!ret)
4755                 goto again;
4756         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4757                  flush_state < COMMIT_TRANS)
4758                 goto again;
4759         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4760                  flush_state <= COMMIT_TRANS)
4761                 goto again;
4762
4763 out:
4764         if (ret == -ENOSPC &&
4765             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4766                 struct btrfs_block_rsv *global_rsv =
4767                         &root->fs_info->global_block_rsv;
4768
4769                 if (block_rsv != global_rsv &&
4770                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4771                         ret = 0;
4772         }
4773         if (ret == -ENOSPC)
4774                 trace_btrfs_space_reservation(root->fs_info,
4775                                               "space_info:enospc",
4776                                               space_info->flags, orig_bytes, 1);
4777         if (flushing) {
4778                 spin_lock(&space_info->lock);
4779                 space_info->flush = 0;
4780                 wake_up_all(&space_info->wait);
4781                 spin_unlock(&space_info->lock);
4782         }
4783         return ret;
4784 }
4785
4786 static struct btrfs_block_rsv *get_block_rsv(
4787                                         const struct btrfs_trans_handle *trans,
4788                                         const struct btrfs_root *root)
4789 {
4790         struct btrfs_block_rsv *block_rsv = NULL;
4791
4792         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4793                 block_rsv = trans->block_rsv;
4794
4795         if (root == root->fs_info->csum_root && trans->adding_csums)
4796                 block_rsv = trans->block_rsv;
4797
4798         if (root == root->fs_info->uuid_root)
4799                 block_rsv = trans->block_rsv;
4800
4801         if (!block_rsv)
4802                 block_rsv = root->block_rsv;
4803
4804         if (!block_rsv)
4805                 block_rsv = &root->fs_info->empty_block_rsv;
4806
4807         return block_rsv;
4808 }
4809
4810 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4811                                u64 num_bytes)
4812 {
4813         int ret = -ENOSPC;
4814         spin_lock(&block_rsv->lock);
4815         if (block_rsv->reserved >= num_bytes) {
4816                 block_rsv->reserved -= num_bytes;
4817                 if (block_rsv->reserved < block_rsv->size)
4818                         block_rsv->full = 0;
4819                 ret = 0;
4820         }
4821         spin_unlock(&block_rsv->lock);
4822         return ret;
4823 }
4824
4825 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4826                                 u64 num_bytes, int update_size)
4827 {
4828         spin_lock(&block_rsv->lock);
4829         block_rsv->reserved += num_bytes;
4830         if (update_size)
4831                 block_rsv->size += num_bytes;
4832         else if (block_rsv->reserved >= block_rsv->size)
4833                 block_rsv->full = 1;
4834         spin_unlock(&block_rsv->lock);
4835 }
4836
4837 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4838                              struct btrfs_block_rsv *dest, u64 num_bytes,
4839                              int min_factor)
4840 {
4841         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4842         u64 min_bytes;
4843
4844         if (global_rsv->space_info != dest->space_info)
4845                 return -ENOSPC;
4846
4847         spin_lock(&global_rsv->lock);
4848         min_bytes = div_factor(global_rsv->size, min_factor);
4849         if (global_rsv->reserved < min_bytes + num_bytes) {
4850                 spin_unlock(&global_rsv->lock);
4851                 return -ENOSPC;
4852         }
4853         global_rsv->reserved -= num_bytes;
4854         if (global_rsv->reserved < global_rsv->size)
4855                 global_rsv->full = 0;
4856         spin_unlock(&global_rsv->lock);
4857
4858         block_rsv_add_bytes(dest, num_bytes, 1);
4859         return 0;
4860 }
4861
4862 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4863                                     struct btrfs_block_rsv *block_rsv,
4864                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4865 {
4866         struct btrfs_space_info *space_info = block_rsv->space_info;
4867
4868         spin_lock(&block_rsv->lock);
4869         if (num_bytes == (u64)-1)
4870                 num_bytes = block_rsv->size;
4871         block_rsv->size -= num_bytes;
4872         if (block_rsv->reserved >= block_rsv->size) {
4873                 num_bytes = block_rsv->reserved - block_rsv->size;
4874                 block_rsv->reserved = block_rsv->size;
4875                 block_rsv->full = 1;
4876         } else {
4877                 num_bytes = 0;
4878         }
4879         spin_unlock(&block_rsv->lock);
4880
4881         if (num_bytes > 0) {
4882                 if (dest) {
4883                         spin_lock(&dest->lock);
4884                         if (!dest->full) {
4885                                 u64 bytes_to_add;
4886
4887                                 bytes_to_add = dest->size - dest->reserved;
4888                                 bytes_to_add = min(num_bytes, bytes_to_add);
4889                                 dest->reserved += bytes_to_add;
4890                                 if (dest->reserved >= dest->size)
4891                                         dest->full = 1;
4892                                 num_bytes -= bytes_to_add;
4893                         }
4894                         spin_unlock(&dest->lock);
4895                 }
4896                 if (num_bytes) {
4897                         spin_lock(&space_info->lock);
4898                         space_info->bytes_may_use -= num_bytes;
4899                         trace_btrfs_space_reservation(fs_info, "space_info",
4900                                         space_info->flags, num_bytes, 0);
4901                         spin_unlock(&space_info->lock);
4902                 }
4903         }
4904 }
4905
4906 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4907                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4908 {
4909         int ret;
4910
4911         ret = block_rsv_use_bytes(src, num_bytes);
4912         if (ret)
4913                 return ret;
4914
4915         block_rsv_add_bytes(dst, num_bytes, 1);
4916         return 0;
4917 }
4918
4919 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4920 {
4921         memset(rsv, 0, sizeof(*rsv));
4922         spin_lock_init(&rsv->lock);
4923         rsv->type = type;
4924 }
4925
4926 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4927                                               unsigned short type)
4928 {
4929         struct btrfs_block_rsv *block_rsv;
4930         struct btrfs_fs_info *fs_info = root->fs_info;
4931
4932         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4933         if (!block_rsv)
4934                 return NULL;
4935
4936         btrfs_init_block_rsv(block_rsv, type);
4937         block_rsv->space_info = __find_space_info(fs_info,
4938                                                   BTRFS_BLOCK_GROUP_METADATA);
4939         return block_rsv;
4940 }
4941
4942 void btrfs_free_block_rsv(struct btrfs_root *root,
4943                           struct btrfs_block_rsv *rsv)
4944 {
4945         if (!rsv)
4946                 return;
4947         btrfs_block_rsv_release(root, rsv, (u64)-1);
4948         kfree(rsv);
4949 }
4950
4951 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
4952 {
4953         kfree(rsv);
4954 }
4955
4956 int btrfs_block_rsv_add(struct btrfs_root *root,
4957                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4958                         enum btrfs_reserve_flush_enum flush)
4959 {
4960         int ret;
4961
4962         if (num_bytes == 0)
4963                 return 0;
4964
4965         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4966         if (!ret) {
4967                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4968                 return 0;
4969         }
4970
4971         return ret;
4972 }
4973
4974 int btrfs_block_rsv_check(struct btrfs_root *root,
4975                           struct btrfs_block_rsv *block_rsv, int min_factor)
4976 {
4977         u64 num_bytes = 0;
4978         int ret = -ENOSPC;
4979
4980         if (!block_rsv)
4981                 return 0;
4982
4983         spin_lock(&block_rsv->lock);
4984         num_bytes = div_factor(block_rsv->size, min_factor);
4985         if (block_rsv->reserved >= num_bytes)
4986                 ret = 0;
4987         spin_unlock(&block_rsv->lock);
4988
4989         return ret;
4990 }
4991
4992 int btrfs_block_rsv_refill(struct btrfs_root *root,
4993                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4994                            enum btrfs_reserve_flush_enum flush)
4995 {
4996         u64 num_bytes = 0;
4997         int ret = -ENOSPC;
4998
4999         if (!block_rsv)
5000                 return 0;
5001
5002         spin_lock(&block_rsv->lock);
5003         num_bytes = min_reserved;
5004         if (block_rsv->reserved >= num_bytes)
5005                 ret = 0;
5006         else
5007                 num_bytes -= block_rsv->reserved;
5008         spin_unlock(&block_rsv->lock);
5009
5010         if (!ret)
5011                 return 0;
5012
5013         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5014         if (!ret) {
5015                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5016                 return 0;
5017         }
5018
5019         return ret;
5020 }
5021
5022 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5023                             struct btrfs_block_rsv *dst_rsv,
5024                             u64 num_bytes)
5025 {
5026         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5027 }
5028
5029 void btrfs_block_rsv_release(struct btrfs_root *root,
5030                              struct btrfs_block_rsv *block_rsv,
5031                              u64 num_bytes)
5032 {
5033         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5034         if (global_rsv == block_rsv ||
5035             block_rsv->space_info != global_rsv->space_info)
5036                 global_rsv = NULL;
5037         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5038                                 num_bytes);
5039 }
5040
5041 /*
5042  * helper to calculate size of global block reservation.
5043  * the desired value is sum of space used by extent tree,
5044  * checksum tree and root tree
5045  */
5046 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5047 {
5048         struct btrfs_space_info *sinfo;
5049         u64 num_bytes;
5050         u64 meta_used;
5051         u64 data_used;
5052         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5053
5054         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5055         spin_lock(&sinfo->lock);
5056         data_used = sinfo->bytes_used;
5057         spin_unlock(&sinfo->lock);
5058
5059         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5060         spin_lock(&sinfo->lock);
5061         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5062                 data_used = 0;
5063         meta_used = sinfo->bytes_used;
5064         spin_unlock(&sinfo->lock);
5065
5066         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5067                     csum_size * 2;
5068         num_bytes += div_u64(data_used + meta_used, 50);
5069
5070         if (num_bytes * 3 > meta_used)
5071                 num_bytes = div_u64(meta_used, 3);
5072
5073         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5074 }
5075
5076 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5077 {
5078         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5079         struct btrfs_space_info *sinfo = block_rsv->space_info;
5080         u64 num_bytes;
5081
5082         num_bytes = calc_global_metadata_size(fs_info);
5083
5084         spin_lock(&sinfo->lock);
5085         spin_lock(&block_rsv->lock);
5086
5087         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
5088
5089         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5090                     sinfo->bytes_reserved + sinfo->bytes_readonly +
5091                     sinfo->bytes_may_use;
5092
5093         if (sinfo->total_bytes > num_bytes) {
5094                 num_bytes = sinfo->total_bytes - num_bytes;
5095                 block_rsv->reserved += num_bytes;
5096                 sinfo->bytes_may_use += num_bytes;
5097                 trace_btrfs_space_reservation(fs_info, "space_info",
5098                                       sinfo->flags, num_bytes, 1);
5099         }
5100
5101         if (block_rsv->reserved >= block_rsv->size) {
5102                 num_bytes = block_rsv->reserved - block_rsv->size;
5103                 sinfo->bytes_may_use -= num_bytes;
5104                 trace_btrfs_space_reservation(fs_info, "space_info",
5105                                       sinfo->flags, num_bytes, 0);
5106                 block_rsv->reserved = block_rsv->size;
5107                 block_rsv->full = 1;
5108         }
5109
5110         spin_unlock(&block_rsv->lock);
5111         spin_unlock(&sinfo->lock);
5112 }
5113
5114 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5115 {
5116         struct btrfs_space_info *space_info;
5117
5118         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5119         fs_info->chunk_block_rsv.space_info = space_info;
5120
5121         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5122         fs_info->global_block_rsv.space_info = space_info;
5123         fs_info->delalloc_block_rsv.space_info = space_info;
5124         fs_info->trans_block_rsv.space_info = space_info;
5125         fs_info->empty_block_rsv.space_info = space_info;
5126         fs_info->delayed_block_rsv.space_info = space_info;
5127
5128         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5129         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5130         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5131         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5132         if (fs_info->quota_root)
5133                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5134         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5135
5136         update_global_block_rsv(fs_info);
5137 }
5138
5139 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5140 {
5141         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5142                                 (u64)-1);
5143         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5144         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5145         WARN_ON(fs_info->trans_block_rsv.size > 0);
5146         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5147         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5148         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5149         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5150         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5151 }
5152
5153 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5154                                   struct btrfs_root *root)
5155 {
5156         if (!trans->block_rsv)
5157                 return;
5158
5159         if (!trans->bytes_reserved)
5160                 return;
5161
5162         trace_btrfs_space_reservation(root->fs_info, "transaction",
5163                                       trans->transid, trans->bytes_reserved, 0);
5164         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5165         trans->bytes_reserved = 0;
5166 }
5167
5168 /* Can only return 0 or -ENOSPC */
5169 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5170                                   struct inode *inode)
5171 {
5172         struct btrfs_root *root = BTRFS_I(inode)->root;
5173         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5174         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5175
5176         /*
5177          * We need to hold space in order to delete our orphan item once we've
5178          * added it, so this takes the reservation so we can release it later
5179          * when we are truly done with the orphan item.
5180          */
5181         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5182         trace_btrfs_space_reservation(root->fs_info, "orphan",
5183                                       btrfs_ino(inode), num_bytes, 1);
5184         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5185 }
5186
5187 void btrfs_orphan_release_metadata(struct inode *inode)
5188 {
5189         struct btrfs_root *root = BTRFS_I(inode)->root;
5190         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5191         trace_btrfs_space_reservation(root->fs_info, "orphan",
5192                                       btrfs_ino(inode), num_bytes, 0);
5193         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5194 }
5195
5196 /*
5197  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5198  * root: the root of the parent directory
5199  * rsv: block reservation
5200  * items: the number of items that we need do reservation
5201  * qgroup_reserved: used to return the reserved size in qgroup
5202  *
5203  * This function is used to reserve the space for snapshot/subvolume
5204  * creation and deletion. Those operations are different with the
5205  * common file/directory operations, they change two fs/file trees
5206  * and root tree, the number of items that the qgroup reserves is
5207  * different with the free space reservation. So we can not use
5208  * the space reseravtion mechanism in start_transaction().
5209  */
5210 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5211                                      struct btrfs_block_rsv *rsv,
5212                                      int items,
5213                                      u64 *qgroup_reserved,
5214                                      bool use_global_rsv)
5215 {
5216         u64 num_bytes;
5217         int ret;
5218         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5219
5220         if (root->fs_info->quota_enabled) {
5221                 /* One for parent inode, two for dir entries */
5222                 num_bytes = 3 * root->nodesize;
5223                 ret = btrfs_qgroup_reserve(root, num_bytes);
5224                 if (ret)
5225                         return ret;
5226         } else {
5227                 num_bytes = 0;
5228         }
5229
5230         *qgroup_reserved = num_bytes;
5231
5232         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5233         rsv->space_info = __find_space_info(root->fs_info,
5234                                             BTRFS_BLOCK_GROUP_METADATA);
5235         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5236                                   BTRFS_RESERVE_FLUSH_ALL);
5237
5238         if (ret == -ENOSPC && use_global_rsv)
5239                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5240
5241         if (ret) {
5242                 if (*qgroup_reserved)
5243                         btrfs_qgroup_free(root, *qgroup_reserved);
5244         }
5245
5246         return ret;
5247 }
5248
5249 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5250                                       struct btrfs_block_rsv *rsv,
5251                                       u64 qgroup_reserved)
5252 {
5253         btrfs_block_rsv_release(root, rsv, (u64)-1);
5254 }
5255
5256 /**
5257  * drop_outstanding_extent - drop an outstanding extent
5258  * @inode: the inode we're dropping the extent for
5259  * @num_bytes: the number of bytes we're relaseing.
5260  *
5261  * This is called when we are freeing up an outstanding extent, either called
5262  * after an error or after an extent is written.  This will return the number of
5263  * reserved extents that need to be freed.  This must be called with
5264  * BTRFS_I(inode)->lock held.
5265  */
5266 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5267 {
5268         unsigned drop_inode_space = 0;
5269         unsigned dropped_extents = 0;
5270         unsigned num_extents = 0;
5271
5272         num_extents = (unsigned)div64_u64(num_bytes +
5273                                           BTRFS_MAX_EXTENT_SIZE - 1,
5274                                           BTRFS_MAX_EXTENT_SIZE);
5275         ASSERT(num_extents);
5276         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5277         BTRFS_I(inode)->outstanding_extents -= num_extents;
5278
5279         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5280             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5281                                &BTRFS_I(inode)->runtime_flags))
5282                 drop_inode_space = 1;
5283
5284         /*
5285          * If we have more or the same amount of outsanding extents than we have
5286          * reserved then we need to leave the reserved extents count alone.
5287          */
5288         if (BTRFS_I(inode)->outstanding_extents >=
5289             BTRFS_I(inode)->reserved_extents)
5290                 return drop_inode_space;
5291
5292         dropped_extents = BTRFS_I(inode)->reserved_extents -
5293                 BTRFS_I(inode)->outstanding_extents;
5294         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5295         return dropped_extents + drop_inode_space;
5296 }
5297
5298 /**
5299  * calc_csum_metadata_size - return the amount of metada space that must be
5300  *      reserved/free'd for the given bytes.
5301  * @inode: the inode we're manipulating
5302  * @num_bytes: the number of bytes in question
5303  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5304  *
5305  * This adjusts the number of csum_bytes in the inode and then returns the
5306  * correct amount of metadata that must either be reserved or freed.  We
5307  * calculate how many checksums we can fit into one leaf and then divide the
5308  * number of bytes that will need to be checksumed by this value to figure out
5309  * how many checksums will be required.  If we are adding bytes then the number
5310  * may go up and we will return the number of additional bytes that must be
5311  * reserved.  If it is going down we will return the number of bytes that must
5312  * be freed.
5313  *
5314  * This must be called with BTRFS_I(inode)->lock held.
5315  */
5316 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5317                                    int reserve)
5318 {
5319         struct btrfs_root *root = BTRFS_I(inode)->root;
5320         u64 old_csums, num_csums;
5321
5322         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5323             BTRFS_I(inode)->csum_bytes == 0)
5324                 return 0;
5325
5326         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5327         if (reserve)
5328                 BTRFS_I(inode)->csum_bytes += num_bytes;
5329         else
5330                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5331         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5332
5333         /* No change, no need to reserve more */
5334         if (old_csums == num_csums)
5335                 return 0;
5336
5337         if (reserve)
5338                 return btrfs_calc_trans_metadata_size(root,
5339                                                       num_csums - old_csums);
5340
5341         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5342 }
5343
5344 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5345 {
5346         struct btrfs_root *root = BTRFS_I(inode)->root;
5347         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5348         u64 to_reserve = 0;
5349         u64 csum_bytes;
5350         unsigned nr_extents = 0;
5351         int extra_reserve = 0;
5352         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5353         int ret = 0;
5354         bool delalloc_lock = true;
5355         u64 to_free = 0;
5356         unsigned dropped;
5357
5358         /* If we are a free space inode we need to not flush since we will be in
5359          * the middle of a transaction commit.  We also don't need the delalloc
5360          * mutex since we won't race with anybody.  We need this mostly to make
5361          * lockdep shut its filthy mouth.
5362          */
5363         if (btrfs_is_free_space_inode(inode)) {
5364                 flush = BTRFS_RESERVE_NO_FLUSH;
5365                 delalloc_lock = false;
5366         }
5367
5368         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5369             btrfs_transaction_in_commit(root->fs_info))
5370                 schedule_timeout(1);
5371
5372         if (delalloc_lock)
5373                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5374
5375         num_bytes = ALIGN(num_bytes, root->sectorsize);
5376
5377         spin_lock(&BTRFS_I(inode)->lock);
5378         nr_extents = (unsigned)div64_u64(num_bytes +
5379                                          BTRFS_MAX_EXTENT_SIZE - 1,
5380                                          BTRFS_MAX_EXTENT_SIZE);
5381         BTRFS_I(inode)->outstanding_extents += nr_extents;
5382         nr_extents = 0;
5383
5384         if (BTRFS_I(inode)->outstanding_extents >
5385             BTRFS_I(inode)->reserved_extents)
5386                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5387                         BTRFS_I(inode)->reserved_extents;
5388
5389         /*
5390          * Add an item to reserve for updating the inode when we complete the
5391          * delalloc io.
5392          */
5393         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5394                       &BTRFS_I(inode)->runtime_flags)) {
5395                 nr_extents++;
5396                 extra_reserve = 1;
5397         }
5398
5399         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5400         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5401         csum_bytes = BTRFS_I(inode)->csum_bytes;
5402         spin_unlock(&BTRFS_I(inode)->lock);
5403
5404         if (root->fs_info->quota_enabled) {
5405                 ret = btrfs_qgroup_reserve(root, nr_extents * root->nodesize);
5406                 if (ret)
5407                         goto out_fail;
5408         }
5409
5410         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5411         if (unlikely(ret)) {
5412                 if (root->fs_info->quota_enabled)
5413                         btrfs_qgroup_free(root, nr_extents * root->nodesize);
5414                 goto out_fail;
5415         }
5416
5417         spin_lock(&BTRFS_I(inode)->lock);
5418         if (extra_reserve) {
5419                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5420                         &BTRFS_I(inode)->runtime_flags);
5421                 nr_extents--;
5422         }
5423         BTRFS_I(inode)->reserved_extents += nr_extents;
5424         spin_unlock(&BTRFS_I(inode)->lock);
5425
5426         if (delalloc_lock)
5427                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5428
5429         if (to_reserve)
5430                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5431                                               btrfs_ino(inode), to_reserve, 1);
5432         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5433
5434         return 0;
5435
5436 out_fail:
5437         spin_lock(&BTRFS_I(inode)->lock);
5438         dropped = drop_outstanding_extent(inode, num_bytes);
5439         /*
5440          * If the inodes csum_bytes is the same as the original
5441          * csum_bytes then we know we haven't raced with any free()ers
5442          * so we can just reduce our inodes csum bytes and carry on.
5443          */
5444         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5445                 calc_csum_metadata_size(inode, num_bytes, 0);
5446         } else {
5447                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5448                 u64 bytes;
5449
5450                 /*
5451                  * This is tricky, but first we need to figure out how much we
5452                  * free'd from any free-ers that occured during this
5453                  * reservation, so we reset ->csum_bytes to the csum_bytes
5454                  * before we dropped our lock, and then call the free for the
5455                  * number of bytes that were freed while we were trying our
5456                  * reservation.
5457                  */
5458                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5459                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5460                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5461
5462
5463                 /*
5464                  * Now we need to see how much we would have freed had we not
5465                  * been making this reservation and our ->csum_bytes were not
5466                  * artificially inflated.
5467                  */
5468                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5469                 bytes = csum_bytes - orig_csum_bytes;
5470                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5471
5472                 /*
5473                  * Now reset ->csum_bytes to what it should be.  If bytes is
5474                  * more than to_free then we would have free'd more space had we
5475                  * not had an artificially high ->csum_bytes, so we need to free
5476                  * the remainder.  If bytes is the same or less then we don't
5477                  * need to do anything, the other free-ers did the correct
5478                  * thing.
5479                  */
5480                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5481                 if (bytes > to_free)
5482                         to_free = bytes - to_free;
5483                 else
5484                         to_free = 0;
5485         }
5486         spin_unlock(&BTRFS_I(inode)->lock);
5487         if (dropped)
5488                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5489
5490         if (to_free) {
5491                 btrfs_block_rsv_release(root, block_rsv, to_free);
5492                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5493                                               btrfs_ino(inode), to_free, 0);
5494         }
5495         if (delalloc_lock)
5496                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5497         return ret;
5498 }
5499
5500 /**
5501  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5502  * @inode: the inode to release the reservation for
5503  * @num_bytes: the number of bytes we're releasing
5504  *
5505  * This will release the metadata reservation for an inode.  This can be called
5506  * once we complete IO for a given set of bytes to release their metadata
5507  * reservations.
5508  */
5509 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5510 {
5511         struct btrfs_root *root = BTRFS_I(inode)->root;
5512         u64 to_free = 0;
5513         unsigned dropped;
5514
5515         num_bytes = ALIGN(num_bytes, root->sectorsize);
5516         spin_lock(&BTRFS_I(inode)->lock);
5517         dropped = drop_outstanding_extent(inode, num_bytes);
5518
5519         if (num_bytes)
5520                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5521         spin_unlock(&BTRFS_I(inode)->lock);
5522         if (dropped > 0)
5523                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5524
5525         if (btrfs_test_is_dummy_root(root))
5526                 return;
5527
5528         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5529                                       btrfs_ino(inode), to_free, 0);
5530
5531         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5532                                 to_free);
5533 }
5534
5535 /**
5536  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5537  * @inode: inode we're writing to
5538  * @num_bytes: the number of bytes we want to allocate
5539  *
5540  * This will do the following things
5541  *
5542  * o reserve space in the data space info for num_bytes
5543  * o reserve space in the metadata space info based on number of outstanding
5544  *   extents and how much csums will be needed
5545  * o add to the inodes ->delalloc_bytes
5546  * o add it to the fs_info's delalloc inodes list.
5547  *
5548  * This will return 0 for success and -ENOSPC if there is no space left.
5549  */
5550 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5551 {
5552         int ret;
5553
5554         ret = btrfs_check_data_free_space(inode, num_bytes, num_bytes);
5555         if (ret)
5556                 return ret;
5557
5558         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5559         if (ret) {
5560                 btrfs_free_reserved_data_space(inode, num_bytes);
5561                 return ret;
5562         }
5563
5564         return 0;
5565 }
5566
5567 /**
5568  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5569  * @inode: inode we're releasing space for
5570  * @num_bytes: the number of bytes we want to free up
5571  *
5572  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5573  * called in the case that we don't need the metadata AND data reservations
5574  * anymore.  So if there is an error or we insert an inline extent.
5575  *
5576  * This function will release the metadata space that was not used and will
5577  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5578  * list if there are no delalloc bytes left.
5579  */
5580 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5581 {
5582         btrfs_delalloc_release_metadata(inode, num_bytes);
5583         btrfs_free_reserved_data_space(inode, num_bytes);
5584 }
5585
5586 static int update_block_group(struct btrfs_trans_handle *trans,
5587                               struct btrfs_root *root, u64 bytenr,
5588                               u64 num_bytes, int alloc)
5589 {
5590         struct btrfs_block_group_cache *cache = NULL;
5591         struct btrfs_fs_info *info = root->fs_info;
5592         u64 total = num_bytes;
5593         u64 old_val;
5594         u64 byte_in_group;
5595         int factor;
5596
5597         /* block accounting for super block */
5598         spin_lock(&info->delalloc_root_lock);
5599         old_val = btrfs_super_bytes_used(info->super_copy);
5600         if (alloc)
5601                 old_val += num_bytes;
5602         else
5603                 old_val -= num_bytes;
5604         btrfs_set_super_bytes_used(info->super_copy, old_val);
5605         spin_unlock(&info->delalloc_root_lock);
5606
5607         while (total) {
5608                 cache = btrfs_lookup_block_group(info, bytenr);
5609                 if (!cache)
5610                         return -ENOENT;
5611                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5612                                     BTRFS_BLOCK_GROUP_RAID1 |
5613                                     BTRFS_BLOCK_GROUP_RAID10))
5614                         factor = 2;
5615                 else
5616                         factor = 1;
5617                 /*
5618                  * If this block group has free space cache written out, we
5619                  * need to make sure to load it if we are removing space.  This
5620                  * is because we need the unpinning stage to actually add the
5621                  * space back to the block group, otherwise we will leak space.
5622                  */
5623                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5624                         cache_block_group(cache, 1);
5625
5626                 byte_in_group = bytenr - cache->key.objectid;
5627                 WARN_ON(byte_in_group > cache->key.offset);
5628
5629                 spin_lock(&cache->space_info->lock);
5630                 spin_lock(&cache->lock);
5631
5632                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5633                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5634                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5635
5636                 old_val = btrfs_block_group_used(&cache->item);
5637                 num_bytes = min(total, cache->key.offset - byte_in_group);
5638                 if (alloc) {
5639                         old_val += num_bytes;
5640                         btrfs_set_block_group_used(&cache->item, old_val);
5641                         cache->reserved -= num_bytes;
5642                         cache->space_info->bytes_reserved -= num_bytes;
5643                         cache->space_info->bytes_used += num_bytes;
5644                         cache->space_info->disk_used += num_bytes * factor;
5645                         spin_unlock(&cache->lock);
5646                         spin_unlock(&cache->space_info->lock);
5647                 } else {
5648                         old_val -= num_bytes;
5649                         btrfs_set_block_group_used(&cache->item, old_val);
5650                         cache->pinned += num_bytes;
5651                         cache->space_info->bytes_pinned += num_bytes;
5652                         cache->space_info->bytes_used -= num_bytes;
5653                         cache->space_info->disk_used -= num_bytes * factor;
5654                         spin_unlock(&cache->lock);
5655                         spin_unlock(&cache->space_info->lock);
5656
5657                         set_extent_dirty(info->pinned_extents,
5658                                          bytenr, bytenr + num_bytes - 1,
5659                                          GFP_NOFS | __GFP_NOFAIL);
5660                         /*
5661                          * No longer have used bytes in this block group, queue
5662                          * it for deletion.
5663                          */
5664                         if (old_val == 0) {
5665                                 spin_lock(&info->unused_bgs_lock);
5666                                 if (list_empty(&cache->bg_list)) {
5667                                         btrfs_get_block_group(cache);
5668                                         list_add_tail(&cache->bg_list,
5669                                                       &info->unused_bgs);
5670                                 }
5671                                 spin_unlock(&info->unused_bgs_lock);
5672                         }
5673                 }
5674
5675                 spin_lock(&trans->transaction->dirty_bgs_lock);
5676                 if (list_empty(&cache->dirty_list)) {
5677                         list_add_tail(&cache->dirty_list,
5678                                       &trans->transaction->dirty_bgs);
5679                                 trans->transaction->num_dirty_bgs++;
5680                         btrfs_get_block_group(cache);
5681                 }
5682                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5683
5684                 btrfs_put_block_group(cache);
5685                 total -= num_bytes;
5686                 bytenr += num_bytes;
5687         }
5688         return 0;
5689 }
5690
5691 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5692 {
5693         struct btrfs_block_group_cache *cache;
5694         u64 bytenr;
5695
5696         spin_lock(&root->fs_info->block_group_cache_lock);
5697         bytenr = root->fs_info->first_logical_byte;
5698         spin_unlock(&root->fs_info->block_group_cache_lock);
5699
5700         if (bytenr < (u64)-1)
5701                 return bytenr;
5702
5703         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5704         if (!cache)
5705                 return 0;
5706
5707         bytenr = cache->key.objectid;
5708         btrfs_put_block_group(cache);
5709
5710         return bytenr;
5711 }
5712
5713 static int pin_down_extent(struct btrfs_root *root,
5714                            struct btrfs_block_group_cache *cache,
5715                            u64 bytenr, u64 num_bytes, int reserved)
5716 {
5717         spin_lock(&cache->space_info->lock);
5718         spin_lock(&cache->lock);
5719         cache->pinned += num_bytes;
5720         cache->space_info->bytes_pinned += num_bytes;
5721         if (reserved) {
5722                 cache->reserved -= num_bytes;
5723                 cache->space_info->bytes_reserved -= num_bytes;
5724         }
5725         spin_unlock(&cache->lock);
5726         spin_unlock(&cache->space_info->lock);
5727
5728         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5729                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5730         if (reserved)
5731                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5732         return 0;
5733 }
5734
5735 /*
5736  * this function must be called within transaction
5737  */
5738 int btrfs_pin_extent(struct btrfs_root *root,
5739                      u64 bytenr, u64 num_bytes, int reserved)
5740 {
5741         struct btrfs_block_group_cache *cache;
5742
5743         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5744         BUG_ON(!cache); /* Logic error */
5745
5746         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5747
5748         btrfs_put_block_group(cache);
5749         return 0;
5750 }
5751
5752 /*
5753  * this function must be called within transaction
5754  */
5755 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5756                                     u64 bytenr, u64 num_bytes)
5757 {
5758         struct btrfs_block_group_cache *cache;
5759         int ret;
5760
5761         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5762         if (!cache)
5763                 return -EINVAL;
5764
5765         /*
5766          * pull in the free space cache (if any) so that our pin
5767          * removes the free space from the cache.  We have load_only set
5768          * to one because the slow code to read in the free extents does check
5769          * the pinned extents.
5770          */
5771         cache_block_group(cache, 1);
5772
5773         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5774
5775         /* remove us from the free space cache (if we're there at all) */
5776         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5777         btrfs_put_block_group(cache);
5778         return ret;
5779 }
5780
5781 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5782 {
5783         int ret;
5784         struct btrfs_block_group_cache *block_group;
5785         struct btrfs_caching_control *caching_ctl;
5786
5787         block_group = btrfs_lookup_block_group(root->fs_info, start);
5788         if (!block_group)
5789                 return -EINVAL;
5790
5791         cache_block_group(block_group, 0);
5792         caching_ctl = get_caching_control(block_group);
5793
5794         if (!caching_ctl) {
5795                 /* Logic error */
5796                 BUG_ON(!block_group_cache_done(block_group));
5797                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5798         } else {
5799                 mutex_lock(&caching_ctl->mutex);
5800
5801                 if (start >= caching_ctl->progress) {
5802                         ret = add_excluded_extent(root, start, num_bytes);
5803                 } else if (start + num_bytes <= caching_ctl->progress) {
5804                         ret = btrfs_remove_free_space(block_group,
5805                                                       start, num_bytes);
5806                 } else {
5807                         num_bytes = caching_ctl->progress - start;
5808                         ret = btrfs_remove_free_space(block_group,
5809                                                       start, num_bytes);
5810                         if (ret)
5811                                 goto out_lock;
5812
5813                         num_bytes = (start + num_bytes) -
5814                                 caching_ctl->progress;
5815                         start = caching_ctl->progress;
5816                         ret = add_excluded_extent(root, start, num_bytes);
5817                 }
5818 out_lock:
5819                 mutex_unlock(&caching_ctl->mutex);
5820                 put_caching_control(caching_ctl);
5821         }
5822         btrfs_put_block_group(block_group);
5823         return ret;
5824 }
5825
5826 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5827                                  struct extent_buffer *eb)
5828 {
5829         struct btrfs_file_extent_item *item;
5830         struct btrfs_key key;
5831         int found_type;
5832         int i;
5833
5834         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5835                 return 0;
5836
5837         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5838                 btrfs_item_key_to_cpu(eb, &key, i);
5839                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5840                         continue;
5841                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5842                 found_type = btrfs_file_extent_type(eb, item);
5843                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5844                         continue;
5845                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5846                         continue;
5847                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5848                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5849                 __exclude_logged_extent(log, key.objectid, key.offset);
5850         }
5851
5852         return 0;
5853 }
5854
5855 /**
5856  * btrfs_update_reserved_bytes - update the block_group and space info counters
5857  * @cache:      The cache we are manipulating
5858  * @num_bytes:  The number of bytes in question
5859  * @reserve:    One of the reservation enums
5860  * @delalloc:   The blocks are allocated for the delalloc write
5861  *
5862  * This is called by the allocator when it reserves space, or by somebody who is
5863  * freeing space that was never actually used on disk.  For example if you
5864  * reserve some space for a new leaf in transaction A and before transaction A
5865  * commits you free that leaf, you call this with reserve set to 0 in order to
5866  * clear the reservation.
5867  *
5868  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5869  * ENOSPC accounting.  For data we handle the reservation through clearing the
5870  * delalloc bits in the io_tree.  We have to do this since we could end up
5871  * allocating less disk space for the amount of data we have reserved in the
5872  * case of compression.
5873  *
5874  * If this is a reservation and the block group has become read only we cannot
5875  * make the reservation and return -EAGAIN, otherwise this function always
5876  * succeeds.
5877  */
5878 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5879                                        u64 num_bytes, int reserve, int delalloc)
5880 {
5881         struct btrfs_space_info *space_info = cache->space_info;
5882         int ret = 0;
5883
5884         spin_lock(&space_info->lock);
5885         spin_lock(&cache->lock);
5886         if (reserve != RESERVE_FREE) {
5887                 if (cache->ro) {
5888                         ret = -EAGAIN;
5889                 } else {
5890                         cache->reserved += num_bytes;
5891                         space_info->bytes_reserved += num_bytes;
5892                         if (reserve == RESERVE_ALLOC) {
5893                                 trace_btrfs_space_reservation(cache->fs_info,
5894                                                 "space_info", space_info->flags,
5895                                                 num_bytes, 0);
5896                                 space_info->bytes_may_use -= num_bytes;
5897                         }
5898
5899                         if (delalloc)
5900                                 cache->delalloc_bytes += num_bytes;
5901                 }
5902         } else {
5903                 if (cache->ro)
5904                         space_info->bytes_readonly += num_bytes;
5905                 cache->reserved -= num_bytes;
5906                 space_info->bytes_reserved -= num_bytes;
5907
5908                 if (delalloc)
5909                         cache->delalloc_bytes -= num_bytes;
5910         }
5911         spin_unlock(&cache->lock);
5912         spin_unlock(&space_info->lock);
5913         return ret;
5914 }
5915
5916 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5917                                 struct btrfs_root *root)
5918 {
5919         struct btrfs_fs_info *fs_info = root->fs_info;
5920         struct btrfs_caching_control *next;
5921         struct btrfs_caching_control *caching_ctl;
5922         struct btrfs_block_group_cache *cache;
5923
5924         down_write(&fs_info->commit_root_sem);
5925
5926         list_for_each_entry_safe(caching_ctl, next,
5927                                  &fs_info->caching_block_groups, list) {
5928                 cache = caching_ctl->block_group;
5929                 if (block_group_cache_done(cache)) {
5930                         cache->last_byte_to_unpin = (u64)-1;
5931                         list_del_init(&caching_ctl->list);
5932                         put_caching_control(caching_ctl);
5933                 } else {
5934                         cache->last_byte_to_unpin = caching_ctl->progress;
5935                 }
5936         }
5937
5938         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5939                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5940         else
5941                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5942
5943         up_write(&fs_info->commit_root_sem);
5944
5945         update_global_block_rsv(fs_info);
5946 }
5947
5948 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
5949                               const bool return_free_space)
5950 {
5951         struct btrfs_fs_info *fs_info = root->fs_info;
5952         struct btrfs_block_group_cache *cache = NULL;
5953         struct btrfs_space_info *space_info;
5954         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5955         u64 len;
5956         bool readonly;
5957
5958         while (start <= end) {
5959                 readonly = false;
5960                 if (!cache ||
5961                     start >= cache->key.objectid + cache->key.offset) {
5962                         if (cache)
5963                                 btrfs_put_block_group(cache);
5964                         cache = btrfs_lookup_block_group(fs_info, start);
5965                         BUG_ON(!cache); /* Logic error */
5966                 }
5967
5968                 len = cache->key.objectid + cache->key.offset - start;
5969                 len = min(len, end + 1 - start);
5970
5971                 if (start < cache->last_byte_to_unpin) {
5972                         len = min(len, cache->last_byte_to_unpin - start);
5973                         if (return_free_space)
5974                                 btrfs_add_free_space(cache, start, len);
5975                 }
5976
5977                 start += len;
5978                 space_info = cache->space_info;
5979
5980                 spin_lock(&space_info->lock);
5981                 spin_lock(&cache->lock);
5982                 cache->pinned -= len;
5983                 space_info->bytes_pinned -= len;
5984                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
5985                 if (cache->ro) {
5986                         space_info->bytes_readonly += len;
5987                         readonly = true;
5988                 }
5989                 spin_unlock(&cache->lock);
5990                 if (!readonly && global_rsv->space_info == space_info) {
5991                         spin_lock(&global_rsv->lock);
5992                         if (!global_rsv->full) {
5993                                 len = min(len, global_rsv->size -
5994                                           global_rsv->reserved);
5995                                 global_rsv->reserved += len;
5996                                 space_info->bytes_may_use += len;
5997                                 if (global_rsv->reserved >= global_rsv->size)
5998                                         global_rsv->full = 1;
5999                         }
6000                         spin_unlock(&global_rsv->lock);
6001                 }
6002                 spin_unlock(&space_info->lock);
6003         }
6004
6005         if (cache)
6006                 btrfs_put_block_group(cache);
6007         return 0;
6008 }
6009
6010 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6011                                struct btrfs_root *root)
6012 {
6013         struct btrfs_fs_info *fs_info = root->fs_info;
6014         struct extent_io_tree *unpin;
6015         u64 start;
6016         u64 end;
6017         int ret;
6018
6019         if (trans->aborted)
6020                 return 0;
6021
6022         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6023                 unpin = &fs_info->freed_extents[1];
6024         else
6025                 unpin = &fs_info->freed_extents[0];
6026
6027         while (1) {
6028                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6029                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6030                                             EXTENT_DIRTY, NULL);
6031                 if (ret) {
6032                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6033                         break;
6034                 }
6035
6036                 if (btrfs_test_opt(root, DISCARD))
6037                         ret = btrfs_discard_extent(root, start,
6038                                                    end + 1 - start, NULL);
6039
6040                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
6041                 unpin_extent_range(root, start, end, true);
6042                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6043                 cond_resched();
6044         }
6045
6046         return 0;
6047 }
6048
6049 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6050                              u64 owner, u64 root_objectid)
6051 {
6052         struct btrfs_space_info *space_info;
6053         u64 flags;
6054
6055         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6056                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6057                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6058                 else
6059                         flags = BTRFS_BLOCK_GROUP_METADATA;
6060         } else {
6061                 flags = BTRFS_BLOCK_GROUP_DATA;
6062         }
6063
6064         space_info = __find_space_info(fs_info, flags);
6065         BUG_ON(!space_info); /* Logic bug */
6066         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6067 }
6068
6069
6070 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6071                                 struct btrfs_root *root,
6072                                 u64 bytenr, u64 num_bytes, u64 parent,
6073                                 u64 root_objectid, u64 owner_objectid,
6074                                 u64 owner_offset, int refs_to_drop,
6075                                 struct btrfs_delayed_extent_op *extent_op,
6076                                 int no_quota)
6077 {
6078         struct btrfs_key key;
6079         struct btrfs_path *path;
6080         struct btrfs_fs_info *info = root->fs_info;
6081         struct btrfs_root *extent_root = info->extent_root;
6082         struct extent_buffer *leaf;
6083         struct btrfs_extent_item *ei;
6084         struct btrfs_extent_inline_ref *iref;
6085         int ret;
6086         int is_data;
6087         int extent_slot = 0;
6088         int found_extent = 0;
6089         int num_to_del = 1;
6090         u32 item_size;
6091         u64 refs;
6092         int last_ref = 0;
6093         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_SUB_EXCL;
6094         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6095                                                  SKINNY_METADATA);
6096
6097         if (!info->quota_enabled || !is_fstree(root_objectid))
6098                 no_quota = 1;
6099
6100         path = btrfs_alloc_path();
6101         if (!path)
6102                 return -ENOMEM;
6103
6104         path->reada = 1;
6105         path->leave_spinning = 1;
6106
6107         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6108         BUG_ON(!is_data && refs_to_drop != 1);
6109
6110         if (is_data)
6111                 skinny_metadata = 0;
6112
6113         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6114                                     bytenr, num_bytes, parent,
6115                                     root_objectid, owner_objectid,
6116                                     owner_offset);
6117         if (ret == 0) {
6118                 extent_slot = path->slots[0];
6119                 while (extent_slot >= 0) {
6120                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6121                                               extent_slot);
6122                         if (key.objectid != bytenr)
6123                                 break;
6124                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6125                             key.offset == num_bytes) {
6126                                 found_extent = 1;
6127                                 break;
6128                         }
6129                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6130                             key.offset == owner_objectid) {
6131                                 found_extent = 1;
6132                                 break;
6133                         }
6134                         if (path->slots[0] - extent_slot > 5)
6135                                 break;
6136                         extent_slot--;
6137                 }
6138 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6139                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6140                 if (found_extent && item_size < sizeof(*ei))
6141                         found_extent = 0;
6142 #endif
6143                 if (!found_extent) {
6144                         BUG_ON(iref);
6145                         ret = remove_extent_backref(trans, extent_root, path,
6146                                                     NULL, refs_to_drop,
6147                                                     is_data, &last_ref);
6148                         if (ret) {
6149                                 btrfs_abort_transaction(trans, extent_root, ret);
6150                                 goto out;
6151                         }
6152                         btrfs_release_path(path);
6153                         path->leave_spinning = 1;
6154
6155                         key.objectid = bytenr;
6156                         key.type = BTRFS_EXTENT_ITEM_KEY;
6157                         key.offset = num_bytes;
6158
6159                         if (!is_data && skinny_metadata) {
6160                                 key.type = BTRFS_METADATA_ITEM_KEY;
6161                                 key.offset = owner_objectid;
6162                         }
6163
6164                         ret = btrfs_search_slot(trans, extent_root,
6165                                                 &key, path, -1, 1);
6166                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6167                                 /*
6168                                  * Couldn't find our skinny metadata item,
6169                                  * see if we have ye olde extent item.
6170                                  */
6171                                 path->slots[0]--;
6172                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6173                                                       path->slots[0]);
6174                                 if (key.objectid == bytenr &&
6175                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6176                                     key.offset == num_bytes)
6177                                         ret = 0;
6178                         }
6179
6180                         if (ret > 0 && skinny_metadata) {
6181                                 skinny_metadata = false;
6182                                 key.objectid = bytenr;
6183                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6184                                 key.offset = num_bytes;
6185                                 btrfs_release_path(path);
6186                                 ret = btrfs_search_slot(trans, extent_root,
6187                                                         &key, path, -1, 1);
6188                         }
6189
6190                         if (ret) {
6191                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6192                                         ret, bytenr);
6193                                 if (ret > 0)
6194                                         btrfs_print_leaf(extent_root,
6195                                                          path->nodes[0]);
6196                         }
6197                         if (ret < 0) {
6198                                 btrfs_abort_transaction(trans, extent_root, ret);
6199                                 goto out;
6200                         }
6201                         extent_slot = path->slots[0];
6202                 }
6203         } else if (WARN_ON(ret == -ENOENT)) {
6204                 btrfs_print_leaf(extent_root, path->nodes[0]);
6205                 btrfs_err(info,
6206                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6207                         bytenr, parent, root_objectid, owner_objectid,
6208                         owner_offset);
6209                 btrfs_abort_transaction(trans, extent_root, ret);
6210                 goto out;
6211         } else {
6212                 btrfs_abort_transaction(trans, extent_root, ret);
6213                 goto out;
6214         }
6215
6216         leaf = path->nodes[0];
6217         item_size = btrfs_item_size_nr(leaf, extent_slot);
6218 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6219         if (item_size < sizeof(*ei)) {
6220                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6221                 ret = convert_extent_item_v0(trans, extent_root, path,
6222                                              owner_objectid, 0);
6223                 if (ret < 0) {
6224                         btrfs_abort_transaction(trans, extent_root, ret);
6225                         goto out;
6226                 }
6227
6228                 btrfs_release_path(path);
6229                 path->leave_spinning = 1;
6230
6231                 key.objectid = bytenr;
6232                 key.type = BTRFS_EXTENT_ITEM_KEY;
6233                 key.offset = num_bytes;
6234
6235                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6236                                         -1, 1);
6237                 if (ret) {
6238                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6239                                 ret, bytenr);
6240                         btrfs_print_leaf(extent_root, path->nodes[0]);
6241                 }
6242                 if (ret < 0) {
6243                         btrfs_abort_transaction(trans, extent_root, ret);
6244                         goto out;
6245                 }
6246
6247                 extent_slot = path->slots[0];
6248                 leaf = path->nodes[0];
6249                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6250         }
6251 #endif
6252         BUG_ON(item_size < sizeof(*ei));
6253         ei = btrfs_item_ptr(leaf, extent_slot,
6254                             struct btrfs_extent_item);
6255         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6256             key.type == BTRFS_EXTENT_ITEM_KEY) {
6257                 struct btrfs_tree_block_info *bi;
6258                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6259                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6260                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6261         }
6262
6263         refs = btrfs_extent_refs(leaf, ei);
6264         if (refs < refs_to_drop) {
6265                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6266                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6267                 ret = -EINVAL;
6268                 btrfs_abort_transaction(trans, extent_root, ret);
6269                 goto out;
6270         }
6271         refs -= refs_to_drop;
6272
6273         if (refs > 0) {
6274                 type = BTRFS_QGROUP_OPER_SUB_SHARED;
6275                 if (extent_op)
6276                         __run_delayed_extent_op(extent_op, leaf, ei);
6277                 /*
6278                  * In the case of inline back ref, reference count will
6279                  * be updated by remove_extent_backref
6280                  */
6281                 if (iref) {
6282                         BUG_ON(!found_extent);
6283                 } else {
6284                         btrfs_set_extent_refs(leaf, ei, refs);
6285                         btrfs_mark_buffer_dirty(leaf);
6286                 }
6287                 if (found_extent) {
6288                         ret = remove_extent_backref(trans, extent_root, path,
6289                                                     iref, refs_to_drop,
6290                                                     is_data, &last_ref);
6291                         if (ret) {
6292                                 btrfs_abort_transaction(trans, extent_root, ret);
6293                                 goto out;
6294                         }
6295                 }
6296                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6297                                  root_objectid);
6298         } else {
6299                 if (found_extent) {
6300                         BUG_ON(is_data && refs_to_drop !=
6301                                extent_data_ref_count(root, path, iref));
6302                         if (iref) {
6303                                 BUG_ON(path->slots[0] != extent_slot);
6304                         } else {
6305                                 BUG_ON(path->slots[0] != extent_slot + 1);
6306                                 path->slots[0] = extent_slot;
6307                                 num_to_del = 2;
6308                         }
6309                 }
6310
6311                 last_ref = 1;
6312                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6313                                       num_to_del);
6314                 if (ret) {
6315                         btrfs_abort_transaction(trans, extent_root, ret);
6316                         goto out;
6317                 }
6318                 btrfs_release_path(path);
6319
6320                 if (is_data) {
6321                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6322                         if (ret) {
6323                                 btrfs_abort_transaction(trans, extent_root, ret);
6324                                 goto out;
6325                         }
6326                 }
6327
6328                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6329                 if (ret) {
6330                         btrfs_abort_transaction(trans, extent_root, ret);
6331                         goto out;
6332                 }
6333         }
6334         btrfs_release_path(path);
6335
6336         /* Deal with the quota accounting */
6337         if (!ret && last_ref && !no_quota) {
6338                 int mod_seq = 0;
6339
6340                 if (owner_objectid >= BTRFS_FIRST_FREE_OBJECTID &&
6341                     type == BTRFS_QGROUP_OPER_SUB_SHARED)
6342                         mod_seq = 1;
6343
6344                 ret = btrfs_qgroup_record_ref(trans, info, root_objectid,
6345                                               bytenr, num_bytes, type,
6346                                               mod_seq);
6347         }
6348 out:
6349         btrfs_free_path(path);
6350         return ret;
6351 }
6352
6353 /*
6354  * when we free an block, it is possible (and likely) that we free the last
6355  * delayed ref for that extent as well.  This searches the delayed ref tree for
6356  * a given extent, and if there are no other delayed refs to be processed, it
6357  * removes it from the tree.
6358  */
6359 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6360                                       struct btrfs_root *root, u64 bytenr)
6361 {
6362         struct btrfs_delayed_ref_head *head;
6363         struct btrfs_delayed_ref_root *delayed_refs;
6364         int ret = 0;
6365
6366         delayed_refs = &trans->transaction->delayed_refs;
6367         spin_lock(&delayed_refs->lock);
6368         head = btrfs_find_delayed_ref_head(trans, bytenr);
6369         if (!head)
6370                 goto out_delayed_unlock;
6371
6372         spin_lock(&head->lock);
6373         if (rb_first(&head->ref_root))
6374                 goto out;
6375
6376         if (head->extent_op) {
6377                 if (!head->must_insert_reserved)
6378                         goto out;
6379                 btrfs_free_delayed_extent_op(head->extent_op);
6380                 head->extent_op = NULL;
6381         }
6382
6383         /*
6384          * waiting for the lock here would deadlock.  If someone else has it
6385          * locked they are already in the process of dropping it anyway
6386          */
6387         if (!mutex_trylock(&head->mutex))
6388                 goto out;
6389
6390         /*
6391          * at this point we have a head with no other entries.  Go
6392          * ahead and process it.
6393          */
6394         head->node.in_tree = 0;
6395         rb_erase(&head->href_node, &delayed_refs->href_root);
6396
6397         atomic_dec(&delayed_refs->num_entries);
6398
6399         /*
6400          * we don't take a ref on the node because we're removing it from the
6401          * tree, so we just steal the ref the tree was holding.
6402          */
6403         delayed_refs->num_heads--;
6404         if (head->processing == 0)
6405                 delayed_refs->num_heads_ready--;
6406         head->processing = 0;
6407         spin_unlock(&head->lock);
6408         spin_unlock(&delayed_refs->lock);
6409
6410         BUG_ON(head->extent_op);
6411         if (head->must_insert_reserved)
6412                 ret = 1;
6413
6414         mutex_unlock(&head->mutex);
6415         btrfs_put_delayed_ref(&head->node);
6416         return ret;
6417 out:
6418         spin_unlock(&head->lock);
6419
6420 out_delayed_unlock:
6421         spin_unlock(&delayed_refs->lock);
6422         return 0;
6423 }
6424
6425 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6426                            struct btrfs_root *root,
6427                            struct extent_buffer *buf,
6428                            u64 parent, int last_ref)
6429 {
6430         int pin = 1;
6431         int ret;
6432
6433         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6434                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6435                                         buf->start, buf->len,
6436                                         parent, root->root_key.objectid,
6437                                         btrfs_header_level(buf),
6438                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6439                 BUG_ON(ret); /* -ENOMEM */
6440         }
6441
6442         if (!last_ref)
6443                 return;
6444
6445         if (btrfs_header_generation(buf) == trans->transid) {
6446                 struct btrfs_block_group_cache *cache;
6447
6448                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6449                         ret = check_ref_cleanup(trans, root, buf->start);
6450                         if (!ret)
6451                                 goto out;
6452                 }
6453
6454                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6455
6456                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6457                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6458                         btrfs_put_block_group(cache);
6459                         goto out;
6460                 }
6461
6462                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6463
6464                 btrfs_add_free_space(cache, buf->start, buf->len);
6465                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6466                 btrfs_put_block_group(cache);
6467                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6468                 pin = 0;
6469         }
6470 out:
6471         if (pin)
6472                 add_pinned_bytes(root->fs_info, buf->len,
6473                                  btrfs_header_level(buf),
6474                                  root->root_key.objectid);
6475
6476         /*
6477          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6478          * anymore.
6479          */
6480         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6481 }
6482
6483 /* Can return -ENOMEM */
6484 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6485                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6486                       u64 owner, u64 offset, int no_quota)
6487 {
6488         int ret;
6489         struct btrfs_fs_info *fs_info = root->fs_info;
6490
6491         if (btrfs_test_is_dummy_root(root))
6492                 return 0;
6493
6494         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6495
6496         /*
6497          * tree log blocks never actually go into the extent allocation
6498          * tree, just update pinning info and exit early.
6499          */
6500         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6501                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6502                 /* unlocks the pinned mutex */
6503                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6504                 ret = 0;
6505         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6506                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6507                                         num_bytes,
6508                                         parent, root_objectid, (int)owner,
6509                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6510         } else {
6511                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6512                                                 num_bytes,
6513                                                 parent, root_objectid, owner,
6514                                                 offset, BTRFS_DROP_DELAYED_REF,
6515                                                 NULL, no_quota);
6516         }
6517         return ret;
6518 }
6519
6520 /*
6521  * when we wait for progress in the block group caching, its because
6522  * our allocation attempt failed at least once.  So, we must sleep
6523  * and let some progress happen before we try again.
6524  *
6525  * This function will sleep at least once waiting for new free space to
6526  * show up, and then it will check the block group free space numbers
6527  * for our min num_bytes.  Another option is to have it go ahead
6528  * and look in the rbtree for a free extent of a given size, but this
6529  * is a good start.
6530  *
6531  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6532  * any of the information in this block group.
6533  */
6534 static noinline void
6535 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6536                                 u64 num_bytes)
6537 {
6538         struct btrfs_caching_control *caching_ctl;
6539
6540         caching_ctl = get_caching_control(cache);
6541         if (!caching_ctl)
6542                 return;
6543
6544         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6545                    (cache->free_space_ctl->free_space >= num_bytes));
6546
6547         put_caching_control(caching_ctl);
6548 }
6549
6550 static noinline int
6551 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6552 {
6553         struct btrfs_caching_control *caching_ctl;
6554         int ret = 0;
6555
6556         caching_ctl = get_caching_control(cache);
6557         if (!caching_ctl)
6558                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6559
6560         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6561         if (cache->cached == BTRFS_CACHE_ERROR)
6562                 ret = -EIO;
6563         put_caching_control(caching_ctl);
6564         return ret;
6565 }
6566
6567 int __get_raid_index(u64 flags)
6568 {
6569         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6570                 return BTRFS_RAID_RAID10;
6571         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6572                 return BTRFS_RAID_RAID1;
6573         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6574                 return BTRFS_RAID_DUP;
6575         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6576                 return BTRFS_RAID_RAID0;
6577         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6578                 return BTRFS_RAID_RAID5;
6579         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6580                 return BTRFS_RAID_RAID6;
6581
6582         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6583 }
6584
6585 int get_block_group_index(struct btrfs_block_group_cache *cache)
6586 {
6587         return __get_raid_index(cache->flags);
6588 }
6589
6590 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6591         [BTRFS_RAID_RAID10]     = "raid10",
6592         [BTRFS_RAID_RAID1]      = "raid1",
6593         [BTRFS_RAID_DUP]        = "dup",
6594         [BTRFS_RAID_RAID0]      = "raid0",
6595         [BTRFS_RAID_SINGLE]     = "single",
6596         [BTRFS_RAID_RAID5]      = "raid5",
6597         [BTRFS_RAID_RAID6]      = "raid6",
6598 };
6599
6600 static const char *get_raid_name(enum btrfs_raid_types type)
6601 {
6602         if (type >= BTRFS_NR_RAID_TYPES)
6603                 return NULL;
6604
6605         return btrfs_raid_type_names[type];
6606 }
6607
6608 enum btrfs_loop_type {
6609         LOOP_CACHING_NOWAIT = 0,
6610         LOOP_CACHING_WAIT = 1,
6611         LOOP_ALLOC_CHUNK = 2,
6612         LOOP_NO_EMPTY_SIZE = 3,
6613 };
6614
6615 static inline void
6616 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6617                        int delalloc)
6618 {
6619         if (delalloc)
6620                 down_read(&cache->data_rwsem);
6621 }
6622
6623 static inline void
6624 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6625                        int delalloc)
6626 {
6627         btrfs_get_block_group(cache);
6628         if (delalloc)
6629                 down_read(&cache->data_rwsem);
6630 }
6631
6632 static struct btrfs_block_group_cache *
6633 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6634                    struct btrfs_free_cluster *cluster,
6635                    int delalloc)
6636 {
6637         struct btrfs_block_group_cache *used_bg;
6638         bool locked = false;
6639 again:
6640         spin_lock(&cluster->refill_lock);
6641         if (locked) {
6642                 if (used_bg == cluster->block_group)
6643                         return used_bg;
6644
6645                 up_read(&used_bg->data_rwsem);
6646                 btrfs_put_block_group(used_bg);
6647         }
6648
6649         used_bg = cluster->block_group;
6650         if (!used_bg)
6651                 return NULL;
6652
6653         if (used_bg == block_group)
6654                 return used_bg;
6655
6656         btrfs_get_block_group(used_bg);
6657
6658         if (!delalloc)
6659                 return used_bg;
6660
6661         if (down_read_trylock(&used_bg->data_rwsem))
6662                 return used_bg;
6663
6664         spin_unlock(&cluster->refill_lock);
6665         down_read(&used_bg->data_rwsem);
6666         locked = true;
6667         goto again;
6668 }
6669
6670 static inline void
6671 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6672                          int delalloc)
6673 {
6674         if (delalloc)
6675                 up_read(&cache->data_rwsem);
6676         btrfs_put_block_group(cache);
6677 }
6678
6679 /*
6680  * walks the btree of allocated extents and find a hole of a given size.
6681  * The key ins is changed to record the hole:
6682  * ins->objectid == start position
6683  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6684  * ins->offset == the size of the hole.
6685  * Any available blocks before search_start are skipped.
6686  *
6687  * If there is no suitable free space, we will record the max size of
6688  * the free space extent currently.
6689  */
6690 static noinline int find_free_extent(struct btrfs_root *orig_root,
6691                                      u64 num_bytes, u64 empty_size,
6692                                      u64 hint_byte, struct btrfs_key *ins,
6693                                      u64 flags, int delalloc)
6694 {
6695         int ret = 0;
6696         struct btrfs_root *root = orig_root->fs_info->extent_root;
6697         struct btrfs_free_cluster *last_ptr = NULL;
6698         struct btrfs_block_group_cache *block_group = NULL;
6699         u64 search_start = 0;
6700         u64 max_extent_size = 0;
6701         int empty_cluster = 2 * 1024 * 1024;
6702         struct btrfs_space_info *space_info;
6703         int loop = 0;
6704         int index = __get_raid_index(flags);
6705         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6706                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6707         bool failed_cluster_refill = false;
6708         bool failed_alloc = false;
6709         bool use_cluster = true;
6710         bool have_caching_bg = false;
6711
6712         WARN_ON(num_bytes < root->sectorsize);
6713         ins->type = BTRFS_EXTENT_ITEM_KEY;
6714         ins->objectid = 0;
6715         ins->offset = 0;
6716
6717         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6718
6719         space_info = __find_space_info(root->fs_info, flags);
6720         if (!space_info) {
6721                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6722                 return -ENOSPC;
6723         }
6724
6725         /*
6726          * If the space info is for both data and metadata it means we have a
6727          * small filesystem and we can't use the clustering stuff.
6728          */
6729         if (btrfs_mixed_space_info(space_info))
6730                 use_cluster = false;
6731
6732         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6733                 last_ptr = &root->fs_info->meta_alloc_cluster;
6734                 if (!btrfs_test_opt(root, SSD))
6735                         empty_cluster = 64 * 1024;
6736         }
6737
6738         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6739             btrfs_test_opt(root, SSD)) {
6740                 last_ptr = &root->fs_info->data_alloc_cluster;
6741         }
6742
6743         if (last_ptr) {
6744                 spin_lock(&last_ptr->lock);
6745                 if (last_ptr->block_group)
6746                         hint_byte = last_ptr->window_start;
6747                 spin_unlock(&last_ptr->lock);
6748         }
6749
6750         search_start = max(search_start, first_logical_byte(root, 0));
6751         search_start = max(search_start, hint_byte);
6752
6753         if (!last_ptr)
6754                 empty_cluster = 0;
6755
6756         if (search_start == hint_byte) {
6757                 block_group = btrfs_lookup_block_group(root->fs_info,
6758                                                        search_start);
6759                 /*
6760                  * we don't want to use the block group if it doesn't match our
6761                  * allocation bits, or if its not cached.
6762                  *
6763                  * However if we are re-searching with an ideal block group
6764                  * picked out then we don't care that the block group is cached.
6765                  */
6766                 if (block_group && block_group_bits(block_group, flags) &&
6767                     block_group->cached != BTRFS_CACHE_NO) {
6768                         down_read(&space_info->groups_sem);
6769                         if (list_empty(&block_group->list) ||
6770                             block_group->ro) {
6771                                 /*
6772                                  * someone is removing this block group,
6773                                  * we can't jump into the have_block_group
6774                                  * target because our list pointers are not
6775                                  * valid
6776                                  */
6777                                 btrfs_put_block_group(block_group);
6778                                 up_read(&space_info->groups_sem);
6779                         } else {
6780                                 index = get_block_group_index(block_group);
6781                                 btrfs_lock_block_group(block_group, delalloc);
6782                                 goto have_block_group;
6783                         }
6784                 } else if (block_group) {
6785                         btrfs_put_block_group(block_group);
6786                 }
6787         }
6788 search:
6789         have_caching_bg = false;
6790         down_read(&space_info->groups_sem);
6791         list_for_each_entry(block_group, &space_info->block_groups[index],
6792                             list) {
6793                 u64 offset;
6794                 int cached;
6795
6796                 btrfs_grab_block_group(block_group, delalloc);
6797                 search_start = block_group->key.objectid;
6798
6799                 /*
6800                  * this can happen if we end up cycling through all the
6801                  * raid types, but we want to make sure we only allocate
6802                  * for the proper type.
6803                  */
6804                 if (!block_group_bits(block_group, flags)) {
6805                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6806                                 BTRFS_BLOCK_GROUP_RAID1 |
6807                                 BTRFS_BLOCK_GROUP_RAID5 |
6808                                 BTRFS_BLOCK_GROUP_RAID6 |
6809                                 BTRFS_BLOCK_GROUP_RAID10;
6810
6811                         /*
6812                          * if they asked for extra copies and this block group
6813                          * doesn't provide them, bail.  This does allow us to
6814                          * fill raid0 from raid1.
6815                          */
6816                         if ((flags & extra) && !(block_group->flags & extra))
6817                                 goto loop;
6818                 }
6819
6820 have_block_group:
6821                 cached = block_group_cache_done(block_group);
6822                 if (unlikely(!cached)) {
6823                         ret = cache_block_group(block_group, 0);
6824                         BUG_ON(ret < 0);
6825                         ret = 0;
6826                 }
6827
6828                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6829                         goto loop;
6830                 if (unlikely(block_group->ro))
6831                         goto loop;
6832
6833                 /*
6834                  * Ok we want to try and use the cluster allocator, so
6835                  * lets look there
6836                  */
6837                 if (last_ptr) {
6838                         struct btrfs_block_group_cache *used_block_group;
6839                         unsigned long aligned_cluster;
6840                         /*
6841                          * the refill lock keeps out other
6842                          * people trying to start a new cluster
6843                          */
6844                         used_block_group = btrfs_lock_cluster(block_group,
6845                                                               last_ptr,
6846                                                               delalloc);
6847                         if (!used_block_group)
6848                                 goto refill_cluster;
6849
6850                         if (used_block_group != block_group &&
6851                             (used_block_group->ro ||
6852                              !block_group_bits(used_block_group, flags)))
6853                                 goto release_cluster;
6854
6855                         offset = btrfs_alloc_from_cluster(used_block_group,
6856                                                 last_ptr,
6857                                                 num_bytes,
6858                                                 used_block_group->key.objectid,
6859                                                 &max_extent_size);
6860                         if (offset) {
6861                                 /* we have a block, we're done */
6862                                 spin_unlock(&last_ptr->refill_lock);
6863                                 trace_btrfs_reserve_extent_cluster(root,
6864                                                 used_block_group,
6865                                                 search_start, num_bytes);
6866                                 if (used_block_group != block_group) {
6867                                         btrfs_release_block_group(block_group,
6868                                                                   delalloc);
6869                                         block_group = used_block_group;
6870                                 }
6871                                 goto checks;
6872                         }
6873
6874                         WARN_ON(last_ptr->block_group != used_block_group);
6875 release_cluster:
6876                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6877                          * set up a new clusters, so lets just skip it
6878                          * and let the allocator find whatever block
6879                          * it can find.  If we reach this point, we
6880                          * will have tried the cluster allocator
6881                          * plenty of times and not have found
6882                          * anything, so we are likely way too
6883                          * fragmented for the clustering stuff to find
6884                          * anything.
6885                          *
6886                          * However, if the cluster is taken from the
6887                          * current block group, release the cluster
6888                          * first, so that we stand a better chance of
6889                          * succeeding in the unclustered
6890                          * allocation.  */
6891                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6892                             used_block_group != block_group) {
6893                                 spin_unlock(&last_ptr->refill_lock);
6894                                 btrfs_release_block_group(used_block_group,
6895                                                           delalloc);
6896                                 goto unclustered_alloc;
6897                         }
6898
6899                         /*
6900                          * this cluster didn't work out, free it and
6901                          * start over
6902                          */
6903                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6904
6905                         if (used_block_group != block_group)
6906                                 btrfs_release_block_group(used_block_group,
6907                                                           delalloc);
6908 refill_cluster:
6909                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6910                                 spin_unlock(&last_ptr->refill_lock);
6911                                 goto unclustered_alloc;
6912                         }
6913
6914                         aligned_cluster = max_t(unsigned long,
6915                                                 empty_cluster + empty_size,
6916                                               block_group->full_stripe_len);
6917
6918                         /* allocate a cluster in this block group */
6919                         ret = btrfs_find_space_cluster(root, block_group,
6920                                                        last_ptr, search_start,
6921                                                        num_bytes,
6922                                                        aligned_cluster);
6923                         if (ret == 0) {
6924                                 /*
6925                                  * now pull our allocation out of this
6926                                  * cluster
6927                                  */
6928                                 offset = btrfs_alloc_from_cluster(block_group,
6929                                                         last_ptr,
6930                                                         num_bytes,
6931                                                         search_start,
6932                                                         &max_extent_size);
6933                                 if (offset) {
6934                                         /* we found one, proceed */
6935                                         spin_unlock(&last_ptr->refill_lock);
6936                                         trace_btrfs_reserve_extent_cluster(root,
6937                                                 block_group, search_start,
6938                                                 num_bytes);
6939                                         goto checks;
6940                                 }
6941                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6942                                    && !failed_cluster_refill) {
6943                                 spin_unlock(&last_ptr->refill_lock);
6944
6945                                 failed_cluster_refill = true;
6946                                 wait_block_group_cache_progress(block_group,
6947                                        num_bytes + empty_cluster + empty_size);
6948                                 goto have_block_group;
6949                         }
6950
6951                         /*
6952                          * at this point we either didn't find a cluster
6953                          * or we weren't able to allocate a block from our
6954                          * cluster.  Free the cluster we've been trying
6955                          * to use, and go to the next block group
6956                          */
6957                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6958                         spin_unlock(&last_ptr->refill_lock);
6959                         goto loop;
6960                 }
6961
6962 unclustered_alloc:
6963                 spin_lock(&block_group->free_space_ctl->tree_lock);
6964                 if (cached &&
6965                     block_group->free_space_ctl->free_space <
6966                     num_bytes + empty_cluster + empty_size) {
6967                         if (block_group->free_space_ctl->free_space >
6968                             max_extent_size)
6969                                 max_extent_size =
6970                                         block_group->free_space_ctl->free_space;
6971                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6972                         goto loop;
6973                 }
6974                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6975
6976                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6977                                                     num_bytes, empty_size,
6978                                                     &max_extent_size);
6979                 /*
6980                  * If we didn't find a chunk, and we haven't failed on this
6981                  * block group before, and this block group is in the middle of
6982                  * caching and we are ok with waiting, then go ahead and wait
6983                  * for progress to be made, and set failed_alloc to true.
6984                  *
6985                  * If failed_alloc is true then we've already waited on this
6986                  * block group once and should move on to the next block group.
6987                  */
6988                 if (!offset && !failed_alloc && !cached &&
6989                     loop > LOOP_CACHING_NOWAIT) {
6990                         wait_block_group_cache_progress(block_group,
6991                                                 num_bytes + empty_size);
6992                         failed_alloc = true;
6993                         goto have_block_group;
6994                 } else if (!offset) {
6995                         if (!cached)
6996                                 have_caching_bg = true;
6997                         goto loop;
6998                 }
6999 checks:
7000                 search_start = ALIGN(offset, root->stripesize);
7001
7002                 /* move on to the next group */
7003                 if (search_start + num_bytes >
7004                     block_group->key.objectid + block_group->key.offset) {
7005                         btrfs_add_free_space(block_group, offset, num_bytes);
7006                         goto loop;
7007                 }
7008
7009                 if (offset < search_start)
7010                         btrfs_add_free_space(block_group, offset,
7011                                              search_start - offset);
7012                 BUG_ON(offset > search_start);
7013
7014                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7015                                                   alloc_type, delalloc);
7016                 if (ret == -EAGAIN) {
7017                         btrfs_add_free_space(block_group, offset, num_bytes);
7018                         goto loop;
7019                 }
7020
7021                 /* we are all good, lets return */
7022                 ins->objectid = search_start;
7023                 ins->offset = num_bytes;
7024
7025                 trace_btrfs_reserve_extent(orig_root, block_group,
7026                                            search_start, num_bytes);
7027                 btrfs_release_block_group(block_group, delalloc);
7028                 break;
7029 loop:
7030                 failed_cluster_refill = false;
7031                 failed_alloc = false;
7032                 BUG_ON(index != get_block_group_index(block_group));
7033                 btrfs_release_block_group(block_group, delalloc);
7034         }
7035         up_read(&space_info->groups_sem);
7036
7037         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7038                 goto search;
7039
7040         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7041                 goto search;
7042
7043         /*
7044          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7045          *                      caching kthreads as we move along
7046          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7047          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7048          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7049          *                      again
7050          */
7051         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7052                 index = 0;
7053                 loop++;
7054                 if (loop == LOOP_ALLOC_CHUNK) {
7055                         struct btrfs_trans_handle *trans;
7056                         int exist = 0;
7057
7058                         trans = current->journal_info;
7059                         if (trans)
7060                                 exist = 1;
7061                         else
7062                                 trans = btrfs_join_transaction(root);
7063
7064                         if (IS_ERR(trans)) {
7065                                 ret = PTR_ERR(trans);
7066                                 goto out;
7067                         }
7068
7069                         ret = do_chunk_alloc(trans, root, flags,
7070                                              CHUNK_ALLOC_FORCE);
7071                         /*
7072                          * Do not bail out on ENOSPC since we
7073                          * can do more things.
7074                          */
7075                         if (ret < 0 && ret != -ENOSPC)
7076                                 btrfs_abort_transaction(trans,
7077                                                         root, ret);
7078                         else
7079                                 ret = 0;
7080                         if (!exist)
7081                                 btrfs_end_transaction(trans, root);
7082                         if (ret)
7083                                 goto out;
7084                 }
7085
7086                 if (loop == LOOP_NO_EMPTY_SIZE) {
7087                         empty_size = 0;
7088                         empty_cluster = 0;
7089                 }
7090
7091                 goto search;
7092         } else if (!ins->objectid) {
7093                 ret = -ENOSPC;
7094         } else if (ins->objectid) {
7095                 ret = 0;
7096         }
7097 out:
7098         if (ret == -ENOSPC)
7099                 ins->offset = max_extent_size;
7100         return ret;
7101 }
7102
7103 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7104                             int dump_block_groups)
7105 {
7106         struct btrfs_block_group_cache *cache;
7107         int index = 0;
7108
7109         spin_lock(&info->lock);
7110         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7111                info->flags,
7112                info->total_bytes - info->bytes_used - info->bytes_pinned -
7113                info->bytes_reserved - info->bytes_readonly,
7114                (info->full) ? "" : "not ");
7115         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7116                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7117                info->total_bytes, info->bytes_used, info->bytes_pinned,
7118                info->bytes_reserved, info->bytes_may_use,
7119                info->bytes_readonly);
7120         spin_unlock(&info->lock);
7121
7122         if (!dump_block_groups)
7123                 return;
7124
7125         down_read(&info->groups_sem);
7126 again:
7127         list_for_each_entry(cache, &info->block_groups[index], list) {
7128                 spin_lock(&cache->lock);
7129                 printk(KERN_INFO "BTRFS: "
7130                            "block group %llu has %llu bytes, "
7131                            "%llu used %llu pinned %llu reserved %s\n",
7132                        cache->key.objectid, cache->key.offset,
7133                        btrfs_block_group_used(&cache->item), cache->pinned,
7134                        cache->reserved, cache->ro ? "[readonly]" : "");
7135                 btrfs_dump_free_space(cache, bytes);
7136                 spin_unlock(&cache->lock);
7137         }
7138         if (++index < BTRFS_NR_RAID_TYPES)
7139                 goto again;
7140         up_read(&info->groups_sem);
7141 }
7142
7143 int btrfs_reserve_extent(struct btrfs_root *root,
7144                          u64 num_bytes, u64 min_alloc_size,
7145                          u64 empty_size, u64 hint_byte,
7146                          struct btrfs_key *ins, int is_data, int delalloc)
7147 {
7148         bool final_tried = false;
7149         u64 flags;
7150         int ret;
7151
7152         flags = btrfs_get_alloc_profile(root, is_data);
7153 again:
7154         WARN_ON(num_bytes < root->sectorsize);
7155         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7156                                flags, delalloc);
7157
7158         if (ret == -ENOSPC) {
7159                 if (!final_tried && ins->offset) {
7160                         num_bytes = min(num_bytes >> 1, ins->offset);
7161                         num_bytes = round_down(num_bytes, root->sectorsize);
7162                         num_bytes = max(num_bytes, min_alloc_size);
7163                         if (num_bytes == min_alloc_size)
7164                                 final_tried = true;
7165                         goto again;
7166                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7167                         struct btrfs_space_info *sinfo;
7168
7169                         sinfo = __find_space_info(root->fs_info, flags);
7170                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7171                                 flags, num_bytes);
7172                         if (sinfo)
7173                                 dump_space_info(sinfo, num_bytes, 1);
7174                 }
7175         }
7176
7177         return ret;
7178 }
7179
7180 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7181                                         u64 start, u64 len,
7182                                         int pin, int delalloc)
7183 {
7184         struct btrfs_block_group_cache *cache;
7185         int ret = 0;
7186
7187         cache = btrfs_lookup_block_group(root->fs_info, start);
7188         if (!cache) {
7189                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7190                         start);
7191                 return -ENOSPC;
7192         }
7193
7194         if (pin)
7195                 pin_down_extent(root, cache, start, len, 1);
7196         else {
7197                 if (btrfs_test_opt(root, DISCARD))
7198                         ret = btrfs_discard_extent(root, start, len, NULL);
7199                 btrfs_add_free_space(cache, start, len);
7200                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7201         }
7202
7203         btrfs_put_block_group(cache);
7204
7205         trace_btrfs_reserved_extent_free(root, start, len);
7206
7207         return ret;
7208 }
7209
7210 int btrfs_free_reserved_extent(struct btrfs_root *root,
7211                                u64 start, u64 len, int delalloc)
7212 {
7213         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7214 }
7215
7216 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7217                                        u64 start, u64 len)
7218 {
7219         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7220 }
7221
7222 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7223                                       struct btrfs_root *root,
7224                                       u64 parent, u64 root_objectid,
7225                                       u64 flags, u64 owner, u64 offset,
7226                                       struct btrfs_key *ins, int ref_mod)
7227 {
7228         int ret;
7229         struct btrfs_fs_info *fs_info = root->fs_info;
7230         struct btrfs_extent_item *extent_item;
7231         struct btrfs_extent_inline_ref *iref;
7232         struct btrfs_path *path;
7233         struct extent_buffer *leaf;
7234         int type;
7235         u32 size;
7236
7237         if (parent > 0)
7238                 type = BTRFS_SHARED_DATA_REF_KEY;
7239         else
7240                 type = BTRFS_EXTENT_DATA_REF_KEY;
7241
7242         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7243
7244         path = btrfs_alloc_path();
7245         if (!path)
7246                 return -ENOMEM;
7247
7248         path->leave_spinning = 1;
7249         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7250                                       ins, size);
7251         if (ret) {
7252                 btrfs_free_path(path);
7253                 return ret;
7254         }
7255
7256         leaf = path->nodes[0];
7257         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7258                                      struct btrfs_extent_item);
7259         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7260         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7261         btrfs_set_extent_flags(leaf, extent_item,
7262                                flags | BTRFS_EXTENT_FLAG_DATA);
7263
7264         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7265         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7266         if (parent > 0) {
7267                 struct btrfs_shared_data_ref *ref;
7268                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7269                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7270                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7271         } else {
7272                 struct btrfs_extent_data_ref *ref;
7273                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7274                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7275                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7276                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7277                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7278         }
7279
7280         btrfs_mark_buffer_dirty(path->nodes[0]);
7281         btrfs_free_path(path);
7282
7283         /* Always set parent to 0 here since its exclusive anyway. */
7284         ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7285                                       ins->objectid, ins->offset,
7286                                       BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7287         if (ret)
7288                 return ret;
7289
7290         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7291         if (ret) { /* -ENOENT, logic error */
7292                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7293                         ins->objectid, ins->offset);
7294                 BUG();
7295         }
7296         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7297         return ret;
7298 }
7299
7300 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7301                                      struct btrfs_root *root,
7302                                      u64 parent, u64 root_objectid,
7303                                      u64 flags, struct btrfs_disk_key *key,
7304                                      int level, struct btrfs_key *ins,
7305                                      int no_quota)
7306 {
7307         int ret;
7308         struct btrfs_fs_info *fs_info = root->fs_info;
7309         struct btrfs_extent_item *extent_item;
7310         struct btrfs_tree_block_info *block_info;
7311         struct btrfs_extent_inline_ref *iref;
7312         struct btrfs_path *path;
7313         struct extent_buffer *leaf;
7314         u32 size = sizeof(*extent_item) + sizeof(*iref);
7315         u64 num_bytes = ins->offset;
7316         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7317                                                  SKINNY_METADATA);
7318
7319         if (!skinny_metadata)
7320                 size += sizeof(*block_info);
7321
7322         path = btrfs_alloc_path();
7323         if (!path) {
7324                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7325                                                    root->nodesize);
7326                 return -ENOMEM;
7327         }
7328
7329         path->leave_spinning = 1;
7330         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7331                                       ins, size);
7332         if (ret) {
7333                 btrfs_free_path(path);
7334                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7335                                                    root->nodesize);
7336                 return ret;
7337         }
7338
7339         leaf = path->nodes[0];
7340         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7341                                      struct btrfs_extent_item);
7342         btrfs_set_extent_refs(leaf, extent_item, 1);
7343         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7344         btrfs_set_extent_flags(leaf, extent_item,
7345                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7346
7347         if (skinny_metadata) {
7348                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7349                 num_bytes = root->nodesize;
7350         } else {
7351                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7352                 btrfs_set_tree_block_key(leaf, block_info, key);
7353                 btrfs_set_tree_block_level(leaf, block_info, level);
7354                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7355         }
7356
7357         if (parent > 0) {
7358                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7359                 btrfs_set_extent_inline_ref_type(leaf, iref,
7360                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7361                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7362         } else {
7363                 btrfs_set_extent_inline_ref_type(leaf, iref,
7364                                                  BTRFS_TREE_BLOCK_REF_KEY);
7365                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7366         }
7367
7368         btrfs_mark_buffer_dirty(leaf);
7369         btrfs_free_path(path);
7370
7371         if (!no_quota) {
7372                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7373                                               ins->objectid, num_bytes,
7374                                               BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7375                 if (ret)
7376                         return ret;
7377         }
7378
7379         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7380                                  1);
7381         if (ret) { /* -ENOENT, logic error */
7382                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7383                         ins->objectid, ins->offset);
7384                 BUG();
7385         }
7386
7387         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7388         return ret;
7389 }
7390
7391 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7392                                      struct btrfs_root *root,
7393                                      u64 root_objectid, u64 owner,
7394                                      u64 offset, struct btrfs_key *ins)
7395 {
7396         int ret;
7397
7398         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7399
7400         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7401                                          ins->offset, 0,
7402                                          root_objectid, owner, offset,
7403                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7404         return ret;
7405 }
7406
7407 /*
7408  * this is used by the tree logging recovery code.  It records that
7409  * an extent has been allocated and makes sure to clear the free
7410  * space cache bits as well
7411  */
7412 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7413                                    struct btrfs_root *root,
7414                                    u64 root_objectid, u64 owner, u64 offset,
7415                                    struct btrfs_key *ins)
7416 {
7417         int ret;
7418         struct btrfs_block_group_cache *block_group;
7419
7420         /*
7421          * Mixed block groups will exclude before processing the log so we only
7422          * need to do the exlude dance if this fs isn't mixed.
7423          */
7424         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7425                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7426                 if (ret)
7427                         return ret;
7428         }
7429
7430         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7431         if (!block_group)
7432                 return -EINVAL;
7433
7434         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7435                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7436         BUG_ON(ret); /* logic error */
7437         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7438                                          0, owner, offset, ins, 1);
7439         btrfs_put_block_group(block_group);
7440         return ret;
7441 }
7442
7443 static struct extent_buffer *
7444 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7445                       u64 bytenr, int level)
7446 {
7447         struct extent_buffer *buf;
7448
7449         buf = btrfs_find_create_tree_block(root, bytenr);
7450         if (!buf)
7451                 return ERR_PTR(-ENOMEM);
7452         btrfs_set_header_generation(buf, trans->transid);
7453         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7454         btrfs_tree_lock(buf);
7455         clean_tree_block(trans, root->fs_info, buf);
7456         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7457
7458         btrfs_set_lock_blocking(buf);
7459         btrfs_set_buffer_uptodate(buf);
7460
7461         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7462                 buf->log_index = root->log_transid % 2;
7463                 /*
7464                  * we allow two log transactions at a time, use different
7465                  * EXENT bit to differentiate dirty pages.
7466                  */
7467                 if (buf->log_index == 0)
7468                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7469                                         buf->start + buf->len - 1, GFP_NOFS);
7470                 else
7471                         set_extent_new(&root->dirty_log_pages, buf->start,
7472                                         buf->start + buf->len - 1, GFP_NOFS);
7473         } else {
7474                 buf->log_index = -1;
7475                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7476                          buf->start + buf->len - 1, GFP_NOFS);
7477         }
7478         trans->blocks_used++;
7479         /* this returns a buffer locked for blocking */
7480         return buf;
7481 }
7482
7483 static struct btrfs_block_rsv *
7484 use_block_rsv(struct btrfs_trans_handle *trans,
7485               struct btrfs_root *root, u32 blocksize)
7486 {
7487         struct btrfs_block_rsv *block_rsv;
7488         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7489         int ret;
7490         bool global_updated = false;
7491
7492         block_rsv = get_block_rsv(trans, root);
7493
7494         if (unlikely(block_rsv->size == 0))
7495                 goto try_reserve;
7496 again:
7497         ret = block_rsv_use_bytes(block_rsv, blocksize);
7498         if (!ret)
7499                 return block_rsv;
7500
7501         if (block_rsv->failfast)
7502                 return ERR_PTR(ret);
7503
7504         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7505                 global_updated = true;
7506                 update_global_block_rsv(root->fs_info);
7507                 goto again;
7508         }
7509
7510         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7511                 static DEFINE_RATELIMIT_STATE(_rs,
7512                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7513                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7514                 if (__ratelimit(&_rs))
7515                         WARN(1, KERN_DEBUG
7516                                 "BTRFS: block rsv returned %d\n", ret);
7517         }
7518 try_reserve:
7519         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7520                                      BTRFS_RESERVE_NO_FLUSH);
7521         if (!ret)
7522                 return block_rsv;
7523         /*
7524          * If we couldn't reserve metadata bytes try and use some from
7525          * the global reserve if its space type is the same as the global
7526          * reservation.
7527          */
7528         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7529             block_rsv->space_info == global_rsv->space_info) {
7530                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7531                 if (!ret)
7532                         return global_rsv;
7533         }
7534         return ERR_PTR(ret);
7535 }
7536
7537 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7538                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7539 {
7540         block_rsv_add_bytes(block_rsv, blocksize, 0);
7541         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7542 }
7543
7544 /*
7545  * finds a free extent and does all the dirty work required for allocation
7546  * returns the key for the extent through ins, and a tree buffer for
7547  * the first block of the extent through buf.
7548  *
7549  * returns the tree buffer or an ERR_PTR on error.
7550  */
7551 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7552                                         struct btrfs_root *root,
7553                                         u64 parent, u64 root_objectid,
7554                                         struct btrfs_disk_key *key, int level,
7555                                         u64 hint, u64 empty_size)
7556 {
7557         struct btrfs_key ins;
7558         struct btrfs_block_rsv *block_rsv;
7559         struct extent_buffer *buf;
7560         struct btrfs_delayed_extent_op *extent_op;
7561         u64 flags = 0;
7562         int ret;
7563         u32 blocksize = root->nodesize;
7564         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7565                                                  SKINNY_METADATA);
7566
7567         if (btrfs_test_is_dummy_root(root)) {
7568                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7569                                             level);
7570                 if (!IS_ERR(buf))
7571                         root->alloc_bytenr += blocksize;
7572                 return buf;
7573         }
7574
7575         block_rsv = use_block_rsv(trans, root, blocksize);
7576         if (IS_ERR(block_rsv))
7577                 return ERR_CAST(block_rsv);
7578
7579         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7580                                    empty_size, hint, &ins, 0, 0);
7581         if (ret)
7582                 goto out_unuse;
7583
7584         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
7585         if (IS_ERR(buf)) {
7586                 ret = PTR_ERR(buf);
7587                 goto out_free_reserved;
7588         }
7589
7590         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7591                 if (parent == 0)
7592                         parent = ins.objectid;
7593                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7594         } else
7595                 BUG_ON(parent > 0);
7596
7597         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7598                 extent_op = btrfs_alloc_delayed_extent_op();
7599                 if (!extent_op) {
7600                         ret = -ENOMEM;
7601                         goto out_free_buf;
7602                 }
7603                 if (key)
7604                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7605                 else
7606                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7607                 extent_op->flags_to_set = flags;
7608                 if (skinny_metadata)
7609                         extent_op->update_key = 0;
7610                 else
7611                         extent_op->update_key = 1;
7612                 extent_op->update_flags = 1;
7613                 extent_op->is_data = 0;
7614                 extent_op->level = level;
7615
7616                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7617                                                  ins.objectid, ins.offset,
7618                                                  parent, root_objectid, level,
7619                                                  BTRFS_ADD_DELAYED_EXTENT,
7620                                                  extent_op, 0);
7621                 if (ret)
7622                         goto out_free_delayed;
7623         }
7624         return buf;
7625
7626 out_free_delayed:
7627         btrfs_free_delayed_extent_op(extent_op);
7628 out_free_buf:
7629         free_extent_buffer(buf);
7630 out_free_reserved:
7631         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
7632 out_unuse:
7633         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7634         return ERR_PTR(ret);
7635 }
7636
7637 struct walk_control {
7638         u64 refs[BTRFS_MAX_LEVEL];
7639         u64 flags[BTRFS_MAX_LEVEL];
7640         struct btrfs_key update_progress;
7641         int stage;
7642         int level;
7643         int shared_level;
7644         int update_ref;
7645         int keep_locks;
7646         int reada_slot;
7647         int reada_count;
7648         int for_reloc;
7649 };
7650
7651 #define DROP_REFERENCE  1
7652 #define UPDATE_BACKREF  2
7653
7654 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7655                                      struct btrfs_root *root,
7656                                      struct walk_control *wc,
7657                                      struct btrfs_path *path)
7658 {
7659         u64 bytenr;
7660         u64 generation;
7661         u64 refs;
7662         u64 flags;
7663         u32 nritems;
7664         u32 blocksize;
7665         struct btrfs_key key;
7666         struct extent_buffer *eb;
7667         int ret;
7668         int slot;
7669         int nread = 0;
7670
7671         if (path->slots[wc->level] < wc->reada_slot) {
7672                 wc->reada_count = wc->reada_count * 2 / 3;
7673                 wc->reada_count = max(wc->reada_count, 2);
7674         } else {
7675                 wc->reada_count = wc->reada_count * 3 / 2;
7676                 wc->reada_count = min_t(int, wc->reada_count,
7677                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7678         }
7679
7680         eb = path->nodes[wc->level];
7681         nritems = btrfs_header_nritems(eb);
7682         blocksize = root->nodesize;
7683
7684         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7685                 if (nread >= wc->reada_count)
7686                         break;
7687
7688                 cond_resched();
7689                 bytenr = btrfs_node_blockptr(eb, slot);
7690                 generation = btrfs_node_ptr_generation(eb, slot);
7691
7692                 if (slot == path->slots[wc->level])
7693                         goto reada;
7694
7695                 if (wc->stage == UPDATE_BACKREF &&
7696                     generation <= root->root_key.offset)
7697                         continue;
7698
7699                 /* We don't lock the tree block, it's OK to be racy here */
7700                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7701                                                wc->level - 1, 1, &refs,
7702                                                &flags);
7703                 /* We don't care about errors in readahead. */
7704                 if (ret < 0)
7705                         continue;
7706                 BUG_ON(refs == 0);
7707
7708                 if (wc->stage == DROP_REFERENCE) {
7709                         if (refs == 1)
7710                                 goto reada;
7711
7712                         if (wc->level == 1 &&
7713                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7714                                 continue;
7715                         if (!wc->update_ref ||
7716                             generation <= root->root_key.offset)
7717                                 continue;
7718                         btrfs_node_key_to_cpu(eb, &key, slot);
7719                         ret = btrfs_comp_cpu_keys(&key,
7720                                                   &wc->update_progress);
7721                         if (ret < 0)
7722                                 continue;
7723                 } else {
7724                         if (wc->level == 1 &&
7725                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7726                                 continue;
7727                 }
7728 reada:
7729                 readahead_tree_block(root, bytenr);
7730                 nread++;
7731         }
7732         wc->reada_slot = slot;
7733 }
7734
7735 static int account_leaf_items(struct btrfs_trans_handle *trans,
7736                               struct btrfs_root *root,
7737                               struct extent_buffer *eb)
7738 {
7739         int nr = btrfs_header_nritems(eb);
7740         int i, extent_type, ret;
7741         struct btrfs_key key;
7742         struct btrfs_file_extent_item *fi;
7743         u64 bytenr, num_bytes;
7744
7745         for (i = 0; i < nr; i++) {
7746                 btrfs_item_key_to_cpu(eb, &key, i);
7747
7748                 if (key.type != BTRFS_EXTENT_DATA_KEY)
7749                         continue;
7750
7751                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
7752                 /* filter out non qgroup-accountable extents  */
7753                 extent_type = btrfs_file_extent_type(eb, fi);
7754
7755                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
7756                         continue;
7757
7758                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
7759                 if (!bytenr)
7760                         continue;
7761
7762                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
7763
7764                 ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7765                                               root->objectid,
7766                                               bytenr, num_bytes,
7767                                               BTRFS_QGROUP_OPER_SUB_SUBTREE, 0);
7768                 if (ret)
7769                         return ret;
7770         }
7771         return 0;
7772 }
7773
7774 /*
7775  * Walk up the tree from the bottom, freeing leaves and any interior
7776  * nodes which have had all slots visited. If a node (leaf or
7777  * interior) is freed, the node above it will have it's slot
7778  * incremented. The root node will never be freed.
7779  *
7780  * At the end of this function, we should have a path which has all
7781  * slots incremented to the next position for a search. If we need to
7782  * read a new node it will be NULL and the node above it will have the
7783  * correct slot selected for a later read.
7784  *
7785  * If we increment the root nodes slot counter past the number of
7786  * elements, 1 is returned to signal completion of the search.
7787  */
7788 static int adjust_slots_upwards(struct btrfs_root *root,
7789                                 struct btrfs_path *path, int root_level)
7790 {
7791         int level = 0;
7792         int nr, slot;
7793         struct extent_buffer *eb;
7794
7795         if (root_level == 0)
7796                 return 1;
7797
7798         while (level <= root_level) {
7799                 eb = path->nodes[level];
7800                 nr = btrfs_header_nritems(eb);
7801                 path->slots[level]++;
7802                 slot = path->slots[level];
7803                 if (slot >= nr || level == 0) {
7804                         /*
7805                          * Don't free the root -  we will detect this
7806                          * condition after our loop and return a
7807                          * positive value for caller to stop walking the tree.
7808                          */
7809                         if (level != root_level) {
7810                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7811                                 path->locks[level] = 0;
7812
7813                                 free_extent_buffer(eb);
7814                                 path->nodes[level] = NULL;
7815                                 path->slots[level] = 0;
7816                         }
7817                 } else {
7818                         /*
7819                          * We have a valid slot to walk back down
7820                          * from. Stop here so caller can process these
7821                          * new nodes.
7822                          */
7823                         break;
7824                 }
7825
7826                 level++;
7827         }
7828
7829         eb = path->nodes[root_level];
7830         if (path->slots[root_level] >= btrfs_header_nritems(eb))
7831                 return 1;
7832
7833         return 0;
7834 }
7835
7836 /*
7837  * root_eb is the subtree root and is locked before this function is called.
7838  */
7839 static int account_shared_subtree(struct btrfs_trans_handle *trans,
7840                                   struct btrfs_root *root,
7841                                   struct extent_buffer *root_eb,
7842                                   u64 root_gen,
7843                                   int root_level)
7844 {
7845         int ret = 0;
7846         int level;
7847         struct extent_buffer *eb = root_eb;
7848         struct btrfs_path *path = NULL;
7849
7850         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
7851         BUG_ON(root_eb == NULL);
7852
7853         if (!root->fs_info->quota_enabled)
7854                 return 0;
7855
7856         if (!extent_buffer_uptodate(root_eb)) {
7857                 ret = btrfs_read_buffer(root_eb, root_gen);
7858                 if (ret)
7859                         goto out;
7860         }
7861
7862         if (root_level == 0) {
7863                 ret = account_leaf_items(trans, root, root_eb);
7864                 goto out;
7865         }
7866
7867         path = btrfs_alloc_path();
7868         if (!path)
7869                 return -ENOMEM;
7870
7871         /*
7872          * Walk down the tree.  Missing extent blocks are filled in as
7873          * we go. Metadata is accounted every time we read a new
7874          * extent block.
7875          *
7876          * When we reach a leaf, we account for file extent items in it,
7877          * walk back up the tree (adjusting slot pointers as we go)
7878          * and restart the search process.
7879          */
7880         extent_buffer_get(root_eb); /* For path */
7881         path->nodes[root_level] = root_eb;
7882         path->slots[root_level] = 0;
7883         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
7884 walk_down:
7885         level = root_level;
7886         while (level >= 0) {
7887                 if (path->nodes[level] == NULL) {
7888                         int parent_slot;
7889                         u64 child_gen;
7890                         u64 child_bytenr;
7891
7892                         /* We need to get child blockptr/gen from
7893                          * parent before we can read it. */
7894                         eb = path->nodes[level + 1];
7895                         parent_slot = path->slots[level + 1];
7896                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
7897                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
7898
7899                         eb = read_tree_block(root, child_bytenr, child_gen);
7900                         if (!eb || !extent_buffer_uptodate(eb)) {
7901                                 ret = -EIO;
7902                                 goto out;
7903                         }
7904
7905                         path->nodes[level] = eb;
7906                         path->slots[level] = 0;
7907
7908                         btrfs_tree_read_lock(eb);
7909                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
7910                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
7911
7912                         ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7913                                                 root->objectid,
7914                                                 child_bytenr,
7915                                                 root->nodesize,
7916                                                 BTRFS_QGROUP_OPER_SUB_SUBTREE,
7917                                                 0);
7918                         if (ret)
7919                                 goto out;
7920
7921                 }
7922
7923                 if (level == 0) {
7924                         ret = account_leaf_items(trans, root, path->nodes[level]);
7925                         if (ret)
7926                                 goto out;
7927
7928                         /* Nonzero return here means we completed our search */
7929                         ret = adjust_slots_upwards(root, path, root_level);
7930                         if (ret)
7931                                 break;
7932
7933                         /* Restart search with new slots */
7934                         goto walk_down;
7935                 }
7936
7937                 level--;
7938         }
7939
7940         ret = 0;
7941 out:
7942         btrfs_free_path(path);
7943
7944         return ret;
7945 }
7946
7947 /*
7948  * helper to process tree block while walking down the tree.
7949  *
7950  * when wc->stage == UPDATE_BACKREF, this function updates
7951  * back refs for pointers in the block.
7952  *
7953  * NOTE: return value 1 means we should stop walking down.
7954  */
7955 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7956                                    struct btrfs_root *root,
7957                                    struct btrfs_path *path,
7958                                    struct walk_control *wc, int lookup_info)
7959 {
7960         int level = wc->level;
7961         struct extent_buffer *eb = path->nodes[level];
7962         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7963         int ret;
7964
7965         if (wc->stage == UPDATE_BACKREF &&
7966             btrfs_header_owner(eb) != root->root_key.objectid)
7967                 return 1;
7968
7969         /*
7970          * when reference count of tree block is 1, it won't increase
7971          * again. once full backref flag is set, we never clear it.
7972          */
7973         if (lookup_info &&
7974             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7975              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7976                 BUG_ON(!path->locks[level]);
7977                 ret = btrfs_lookup_extent_info(trans, root,
7978                                                eb->start, level, 1,
7979                                                &wc->refs[level],
7980                                                &wc->flags[level]);
7981                 BUG_ON(ret == -ENOMEM);
7982                 if (ret)
7983                         return ret;
7984                 BUG_ON(wc->refs[level] == 0);
7985         }
7986
7987         if (wc->stage == DROP_REFERENCE) {
7988                 if (wc->refs[level] > 1)
7989                         return 1;
7990
7991                 if (path->locks[level] && !wc->keep_locks) {
7992                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7993                         path->locks[level] = 0;
7994                 }
7995                 return 0;
7996         }
7997
7998         /* wc->stage == UPDATE_BACKREF */
7999         if (!(wc->flags[level] & flag)) {
8000                 BUG_ON(!path->locks[level]);
8001                 ret = btrfs_inc_ref(trans, root, eb, 1);
8002                 BUG_ON(ret); /* -ENOMEM */
8003                 ret = btrfs_dec_ref(trans, root, eb, 0);
8004                 BUG_ON(ret); /* -ENOMEM */
8005                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8006                                                   eb->len, flag,
8007                                                   btrfs_header_level(eb), 0);
8008                 BUG_ON(ret); /* -ENOMEM */
8009                 wc->flags[level] |= flag;
8010         }
8011
8012         /*
8013          * the block is shared by multiple trees, so it's not good to
8014          * keep the tree lock
8015          */
8016         if (path->locks[level] && level > 0) {
8017                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8018                 path->locks[level] = 0;
8019         }
8020         return 0;
8021 }
8022
8023 /*
8024  * helper to process tree block pointer.
8025  *
8026  * when wc->stage == DROP_REFERENCE, this function checks
8027  * reference count of the block pointed to. if the block
8028  * is shared and we need update back refs for the subtree
8029  * rooted at the block, this function changes wc->stage to
8030  * UPDATE_BACKREF. if the block is shared and there is no
8031  * need to update back, this function drops the reference
8032  * to the block.
8033  *
8034  * NOTE: return value 1 means we should stop walking down.
8035  */
8036 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8037                                  struct btrfs_root *root,
8038                                  struct btrfs_path *path,
8039                                  struct walk_control *wc, int *lookup_info)
8040 {
8041         u64 bytenr;
8042         u64 generation;
8043         u64 parent;
8044         u32 blocksize;
8045         struct btrfs_key key;
8046         struct extent_buffer *next;
8047         int level = wc->level;
8048         int reada = 0;
8049         int ret = 0;
8050         bool need_account = false;
8051
8052         generation = btrfs_node_ptr_generation(path->nodes[level],
8053                                                path->slots[level]);
8054         /*
8055          * if the lower level block was created before the snapshot
8056          * was created, we know there is no need to update back refs
8057          * for the subtree
8058          */
8059         if (wc->stage == UPDATE_BACKREF &&
8060             generation <= root->root_key.offset) {
8061                 *lookup_info = 1;
8062                 return 1;
8063         }
8064
8065         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8066         blocksize = root->nodesize;
8067
8068         next = btrfs_find_tree_block(root->fs_info, bytenr);
8069         if (!next) {
8070                 next = btrfs_find_create_tree_block(root, bytenr);
8071                 if (!next)
8072                         return -ENOMEM;
8073                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8074                                                level - 1);
8075                 reada = 1;
8076         }
8077         btrfs_tree_lock(next);
8078         btrfs_set_lock_blocking(next);
8079
8080         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8081                                        &wc->refs[level - 1],
8082                                        &wc->flags[level - 1]);
8083         if (ret < 0) {
8084                 btrfs_tree_unlock(next);
8085                 return ret;
8086         }
8087
8088         if (unlikely(wc->refs[level - 1] == 0)) {
8089                 btrfs_err(root->fs_info, "Missing references.");
8090                 BUG();
8091         }
8092         *lookup_info = 0;
8093
8094         if (wc->stage == DROP_REFERENCE) {
8095                 if (wc->refs[level - 1] > 1) {
8096                         need_account = true;
8097                         if (level == 1 &&
8098                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8099                                 goto skip;
8100
8101                         if (!wc->update_ref ||
8102                             generation <= root->root_key.offset)
8103                                 goto skip;
8104
8105                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8106                                               path->slots[level]);
8107                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8108                         if (ret < 0)
8109                                 goto skip;
8110
8111                         wc->stage = UPDATE_BACKREF;
8112                         wc->shared_level = level - 1;
8113                 }
8114         } else {
8115                 if (level == 1 &&
8116                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8117                         goto skip;
8118         }
8119
8120         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8121                 btrfs_tree_unlock(next);
8122                 free_extent_buffer(next);
8123                 next = NULL;
8124                 *lookup_info = 1;
8125         }
8126
8127         if (!next) {
8128                 if (reada && level == 1)
8129                         reada_walk_down(trans, root, wc, path);
8130                 next = read_tree_block(root, bytenr, generation);
8131                 if (!next || !extent_buffer_uptodate(next)) {
8132                         free_extent_buffer(next);
8133                         return -EIO;
8134                 }
8135                 btrfs_tree_lock(next);
8136                 btrfs_set_lock_blocking(next);
8137         }
8138
8139         level--;
8140         BUG_ON(level != btrfs_header_level(next));
8141         path->nodes[level] = next;
8142         path->slots[level] = 0;
8143         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8144         wc->level = level;
8145         if (wc->level == 1)
8146                 wc->reada_slot = 0;
8147         return 0;
8148 skip:
8149         wc->refs[level - 1] = 0;
8150         wc->flags[level - 1] = 0;
8151         if (wc->stage == DROP_REFERENCE) {
8152                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8153                         parent = path->nodes[level]->start;
8154                 } else {
8155                         BUG_ON(root->root_key.objectid !=
8156                                btrfs_header_owner(path->nodes[level]));
8157                         parent = 0;
8158                 }
8159
8160                 if (need_account) {
8161                         ret = account_shared_subtree(trans, root, next,
8162                                                      generation, level - 1);
8163                         if (ret) {
8164                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8165                                         "%d accounting shared subtree. Quota "
8166                                         "is out of sync, rescan required.\n",
8167                                         root->fs_info->sb->s_id, ret);
8168                         }
8169                 }
8170                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8171                                 root->root_key.objectid, level - 1, 0, 0);
8172                 BUG_ON(ret); /* -ENOMEM */
8173         }
8174         btrfs_tree_unlock(next);
8175         free_extent_buffer(next);
8176         *lookup_info = 1;
8177         return 1;
8178 }
8179
8180 /*
8181  * helper to process tree block while walking up the tree.
8182  *
8183  * when wc->stage == DROP_REFERENCE, this function drops
8184  * reference count on the block.
8185  *
8186  * when wc->stage == UPDATE_BACKREF, this function changes
8187  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8188  * to UPDATE_BACKREF previously while processing the block.
8189  *
8190  * NOTE: return value 1 means we should stop walking up.
8191  */
8192 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8193                                  struct btrfs_root *root,
8194                                  struct btrfs_path *path,
8195                                  struct walk_control *wc)
8196 {
8197         int ret;
8198         int level = wc->level;
8199         struct extent_buffer *eb = path->nodes[level];
8200         u64 parent = 0;
8201
8202         if (wc->stage == UPDATE_BACKREF) {
8203                 BUG_ON(wc->shared_level < level);
8204                 if (level < wc->shared_level)
8205                         goto out;
8206
8207                 ret = find_next_key(path, level + 1, &wc->update_progress);
8208                 if (ret > 0)
8209                         wc->update_ref = 0;
8210
8211                 wc->stage = DROP_REFERENCE;
8212                 wc->shared_level = -1;
8213                 path->slots[level] = 0;
8214
8215                 /*
8216                  * check reference count again if the block isn't locked.
8217                  * we should start walking down the tree again if reference
8218                  * count is one.
8219                  */
8220                 if (!path->locks[level]) {
8221                         BUG_ON(level == 0);
8222                         btrfs_tree_lock(eb);
8223                         btrfs_set_lock_blocking(eb);
8224                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8225
8226                         ret = btrfs_lookup_extent_info(trans, root,
8227                                                        eb->start, level, 1,
8228                                                        &wc->refs[level],
8229                                                        &wc->flags[level]);
8230                         if (ret < 0) {
8231                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8232                                 path->locks[level] = 0;
8233                                 return ret;
8234                         }
8235                         BUG_ON(wc->refs[level] == 0);
8236                         if (wc->refs[level] == 1) {
8237                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8238                                 path->locks[level] = 0;
8239                                 return 1;
8240                         }
8241                 }
8242         }
8243
8244         /* wc->stage == DROP_REFERENCE */
8245         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8246
8247         if (wc->refs[level] == 1) {
8248                 if (level == 0) {
8249                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8250                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8251                         else
8252                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8253                         BUG_ON(ret); /* -ENOMEM */
8254                         ret = account_leaf_items(trans, root, eb);
8255                         if (ret) {
8256                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8257                                         "%d accounting leaf items. Quota "
8258                                         "is out of sync, rescan required.\n",
8259                                         root->fs_info->sb->s_id, ret);
8260                         }
8261                 }
8262                 /* make block locked assertion in clean_tree_block happy */
8263                 if (!path->locks[level] &&
8264                     btrfs_header_generation(eb) == trans->transid) {
8265                         btrfs_tree_lock(eb);
8266                         btrfs_set_lock_blocking(eb);
8267                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8268                 }
8269                 clean_tree_block(trans, root->fs_info, eb);
8270         }
8271
8272         if (eb == root->node) {
8273                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8274                         parent = eb->start;
8275                 else
8276                         BUG_ON(root->root_key.objectid !=
8277                                btrfs_header_owner(eb));
8278         } else {
8279                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8280                         parent = path->nodes[level + 1]->start;
8281                 else
8282                         BUG_ON(root->root_key.objectid !=
8283                                btrfs_header_owner(path->nodes[level + 1]));
8284         }
8285
8286         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8287 out:
8288         wc->refs[level] = 0;
8289         wc->flags[level] = 0;
8290         return 0;
8291 }
8292
8293 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8294                                    struct btrfs_root *root,
8295                                    struct btrfs_path *path,
8296                                    struct walk_control *wc)
8297 {
8298         int level = wc->level;
8299         int lookup_info = 1;
8300         int ret;
8301
8302         while (level >= 0) {
8303                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8304                 if (ret > 0)
8305                         break;
8306
8307                 if (level == 0)
8308                         break;
8309
8310                 if (path->slots[level] >=
8311                     btrfs_header_nritems(path->nodes[level]))
8312                         break;
8313
8314                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8315                 if (ret > 0) {
8316                         path->slots[level]++;
8317                         continue;
8318                 } else if (ret < 0)
8319                         return ret;
8320                 level = wc->level;
8321         }
8322         return 0;
8323 }
8324
8325 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8326                                  struct btrfs_root *root,
8327                                  struct btrfs_path *path,
8328                                  struct walk_control *wc, int max_level)
8329 {
8330         int level = wc->level;
8331         int ret;
8332
8333         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8334         while (level < max_level && path->nodes[level]) {
8335                 wc->level = level;
8336                 if (path->slots[level] + 1 <
8337                     btrfs_header_nritems(path->nodes[level])) {
8338                         path->slots[level]++;
8339                         return 0;
8340                 } else {
8341                         ret = walk_up_proc(trans, root, path, wc);
8342                         if (ret > 0)
8343                                 return 0;
8344
8345                         if (path->locks[level]) {
8346                                 btrfs_tree_unlock_rw(path->nodes[level],
8347                                                      path->locks[level]);
8348                                 path->locks[level] = 0;
8349                         }
8350                         free_extent_buffer(path->nodes[level]);
8351                         path->nodes[level] = NULL;
8352                         level++;
8353                 }
8354         }
8355         return 1;
8356 }
8357
8358 /*
8359  * drop a subvolume tree.
8360  *
8361  * this function traverses the tree freeing any blocks that only
8362  * referenced by the tree.
8363  *
8364  * when a shared tree block is found. this function decreases its
8365  * reference count by one. if update_ref is true, this function
8366  * also make sure backrefs for the shared block and all lower level
8367  * blocks are properly updated.
8368  *
8369  * If called with for_reloc == 0, may exit early with -EAGAIN
8370  */
8371 int btrfs_drop_snapshot(struct btrfs_root *root,
8372                          struct btrfs_block_rsv *block_rsv, int update_ref,
8373                          int for_reloc)
8374 {
8375         struct btrfs_path *path;
8376         struct btrfs_trans_handle *trans;
8377         struct btrfs_root *tree_root = root->fs_info->tree_root;
8378         struct btrfs_root_item *root_item = &root->root_item;
8379         struct walk_control *wc;
8380         struct btrfs_key key;
8381         int err = 0;
8382         int ret;
8383         int level;
8384         bool root_dropped = false;
8385
8386         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8387
8388         path = btrfs_alloc_path();
8389         if (!path) {
8390                 err = -ENOMEM;
8391                 goto out;
8392         }
8393
8394         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8395         if (!wc) {
8396                 btrfs_free_path(path);
8397                 err = -ENOMEM;
8398                 goto out;
8399         }
8400
8401         trans = btrfs_start_transaction(tree_root, 0);
8402         if (IS_ERR(trans)) {
8403                 err = PTR_ERR(trans);
8404                 goto out_free;
8405         }
8406
8407         if (block_rsv)
8408                 trans->block_rsv = block_rsv;
8409
8410         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8411                 level = btrfs_header_level(root->node);
8412                 path->nodes[level] = btrfs_lock_root_node(root);
8413                 btrfs_set_lock_blocking(path->nodes[level]);
8414                 path->slots[level] = 0;
8415                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8416                 memset(&wc->update_progress, 0,
8417                        sizeof(wc->update_progress));
8418         } else {
8419                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8420                 memcpy(&wc->update_progress, &key,
8421                        sizeof(wc->update_progress));
8422
8423                 level = root_item->drop_level;
8424                 BUG_ON(level == 0);
8425                 path->lowest_level = level;
8426                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8427                 path->lowest_level = 0;
8428                 if (ret < 0) {
8429                         err = ret;
8430                         goto out_end_trans;
8431                 }
8432                 WARN_ON(ret > 0);
8433
8434                 /*
8435                  * unlock our path, this is safe because only this
8436                  * function is allowed to delete this snapshot
8437                  */
8438                 btrfs_unlock_up_safe(path, 0);
8439
8440                 level = btrfs_header_level(root->node);
8441                 while (1) {
8442                         btrfs_tree_lock(path->nodes[level]);
8443                         btrfs_set_lock_blocking(path->nodes[level]);
8444                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8445
8446                         ret = btrfs_lookup_extent_info(trans, root,
8447                                                 path->nodes[level]->start,
8448                                                 level, 1, &wc->refs[level],
8449                                                 &wc->flags[level]);
8450                         if (ret < 0) {
8451                                 err = ret;
8452                                 goto out_end_trans;
8453                         }
8454                         BUG_ON(wc->refs[level] == 0);
8455
8456                         if (level == root_item->drop_level)
8457                                 break;
8458
8459                         btrfs_tree_unlock(path->nodes[level]);
8460                         path->locks[level] = 0;
8461                         WARN_ON(wc->refs[level] != 1);
8462                         level--;
8463                 }
8464         }
8465
8466         wc->level = level;
8467         wc->shared_level = -1;
8468         wc->stage = DROP_REFERENCE;
8469         wc->update_ref = update_ref;
8470         wc->keep_locks = 0;
8471         wc->for_reloc = for_reloc;
8472         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8473
8474         while (1) {
8475
8476                 ret = walk_down_tree(trans, root, path, wc);
8477                 if (ret < 0) {
8478                         err = ret;
8479                         break;
8480                 }
8481
8482                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8483                 if (ret < 0) {
8484                         err = ret;
8485                         break;
8486                 }
8487
8488                 if (ret > 0) {
8489                         BUG_ON(wc->stage != DROP_REFERENCE);
8490                         break;
8491                 }
8492
8493                 if (wc->stage == DROP_REFERENCE) {
8494                         level = wc->level;
8495                         btrfs_node_key(path->nodes[level],
8496                                        &root_item->drop_progress,
8497                                        path->slots[level]);
8498                         root_item->drop_level = level;
8499                 }
8500
8501                 BUG_ON(wc->level == 0);
8502                 if (btrfs_should_end_transaction(trans, tree_root) ||
8503                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8504                         ret = btrfs_update_root(trans, tree_root,
8505                                                 &root->root_key,
8506                                                 root_item);
8507                         if (ret) {
8508                                 btrfs_abort_transaction(trans, tree_root, ret);
8509                                 err = ret;
8510                                 goto out_end_trans;
8511                         }
8512
8513                         /*
8514                          * Qgroup update accounting is run from
8515                          * delayed ref handling. This usually works
8516                          * out because delayed refs are normally the
8517                          * only way qgroup updates are added. However,
8518                          * we may have added updates during our tree
8519                          * walk so run qgroups here to make sure we
8520                          * don't lose any updates.
8521                          */
8522                         ret = btrfs_delayed_qgroup_accounting(trans,
8523                                                               root->fs_info);
8524                         if (ret)
8525                                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8526                                                    "running qgroup updates "
8527                                                    "during snapshot delete. "
8528                                                    "Quota is out of sync, "
8529                                                    "rescan required.\n", ret);
8530
8531                         btrfs_end_transaction_throttle(trans, tree_root);
8532                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8533                                 pr_debug("BTRFS: drop snapshot early exit\n");
8534                                 err = -EAGAIN;
8535                                 goto out_free;
8536                         }
8537
8538                         trans = btrfs_start_transaction(tree_root, 0);
8539                         if (IS_ERR(trans)) {
8540                                 err = PTR_ERR(trans);
8541                                 goto out_free;
8542                         }
8543                         if (block_rsv)
8544                                 trans->block_rsv = block_rsv;
8545                 }
8546         }
8547         btrfs_release_path(path);
8548         if (err)
8549                 goto out_end_trans;
8550
8551         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8552         if (ret) {
8553                 btrfs_abort_transaction(trans, tree_root, ret);
8554                 goto out_end_trans;
8555         }
8556
8557         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8558                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8559                                       NULL, NULL);
8560                 if (ret < 0) {
8561                         btrfs_abort_transaction(trans, tree_root, ret);
8562                         err = ret;
8563                         goto out_end_trans;
8564                 } else if (ret > 0) {
8565                         /* if we fail to delete the orphan item this time
8566                          * around, it'll get picked up the next time.
8567                          *
8568                          * The most common failure here is just -ENOENT.
8569                          */
8570                         btrfs_del_orphan_item(trans, tree_root,
8571                                               root->root_key.objectid);
8572                 }
8573         }
8574
8575         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8576                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
8577         } else {
8578                 free_extent_buffer(root->node);
8579                 free_extent_buffer(root->commit_root);
8580                 btrfs_put_fs_root(root);
8581         }
8582         root_dropped = true;
8583 out_end_trans:
8584         ret = btrfs_delayed_qgroup_accounting(trans, tree_root->fs_info);
8585         if (ret)
8586                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8587                                    "running qgroup updates "
8588                                    "during snapshot delete. "
8589                                    "Quota is out of sync, "
8590                                    "rescan required.\n", ret);
8591
8592         btrfs_end_transaction_throttle(trans, tree_root);
8593 out_free:
8594         kfree(wc);
8595         btrfs_free_path(path);
8596 out:
8597         /*
8598          * So if we need to stop dropping the snapshot for whatever reason we
8599          * need to make sure to add it back to the dead root list so that we
8600          * keep trying to do the work later.  This also cleans up roots if we
8601          * don't have it in the radix (like when we recover after a power fail
8602          * or unmount) so we don't leak memory.
8603          */
8604         if (!for_reloc && root_dropped == false)
8605                 btrfs_add_dead_root(root);
8606         if (err && err != -EAGAIN)
8607                 btrfs_std_error(root->fs_info, err);
8608         return err;
8609 }
8610
8611 /*
8612  * drop subtree rooted at tree block 'node'.
8613  *
8614  * NOTE: this function will unlock and release tree block 'node'
8615  * only used by relocation code
8616  */
8617 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8618                         struct btrfs_root *root,
8619                         struct extent_buffer *node,
8620                         struct extent_buffer *parent)
8621 {
8622         struct btrfs_path *path;
8623         struct walk_control *wc;
8624         int level;
8625         int parent_level;
8626         int ret = 0;
8627         int wret;
8628
8629         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8630
8631         path = btrfs_alloc_path();
8632         if (!path)
8633                 return -ENOMEM;
8634
8635         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8636         if (!wc) {
8637                 btrfs_free_path(path);
8638                 return -ENOMEM;
8639         }
8640
8641         btrfs_assert_tree_locked(parent);
8642         parent_level = btrfs_header_level(parent);
8643         extent_buffer_get(parent);
8644         path->nodes[parent_level] = parent;
8645         path->slots[parent_level] = btrfs_header_nritems(parent);
8646
8647         btrfs_assert_tree_locked(node);
8648         level = btrfs_header_level(node);
8649         path->nodes[level] = node;
8650         path->slots[level] = 0;
8651         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8652
8653         wc->refs[parent_level] = 1;
8654         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8655         wc->level = level;
8656         wc->shared_level = -1;
8657         wc->stage = DROP_REFERENCE;
8658         wc->update_ref = 0;
8659         wc->keep_locks = 1;
8660         wc->for_reloc = 1;
8661         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8662
8663         while (1) {
8664                 wret = walk_down_tree(trans, root, path, wc);
8665                 if (wret < 0) {
8666                         ret = wret;
8667                         break;
8668                 }
8669
8670                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8671                 if (wret < 0)
8672                         ret = wret;
8673                 if (wret != 0)
8674                         break;
8675         }
8676
8677         kfree(wc);
8678         btrfs_free_path(path);
8679         return ret;
8680 }
8681
8682 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8683 {
8684         u64 num_devices;
8685         u64 stripped;
8686
8687         /*
8688          * if restripe for this chunk_type is on pick target profile and
8689          * return, otherwise do the usual balance
8690          */
8691         stripped = get_restripe_target(root->fs_info, flags);
8692         if (stripped)
8693                 return extended_to_chunk(stripped);
8694
8695         num_devices = root->fs_info->fs_devices->rw_devices;
8696
8697         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8698                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8699                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8700
8701         if (num_devices == 1) {
8702                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8703                 stripped = flags & ~stripped;
8704
8705                 /* turn raid0 into single device chunks */
8706                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8707                         return stripped;
8708
8709                 /* turn mirroring into duplication */
8710                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8711                              BTRFS_BLOCK_GROUP_RAID10))
8712                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8713         } else {
8714                 /* they already had raid on here, just return */
8715                 if (flags & stripped)
8716                         return flags;
8717
8718                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8719                 stripped = flags & ~stripped;
8720
8721                 /* switch duplicated blocks with raid1 */
8722                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8723                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8724
8725                 /* this is drive concat, leave it alone */
8726         }
8727
8728         return flags;
8729 }
8730
8731 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8732 {
8733         struct btrfs_space_info *sinfo = cache->space_info;
8734         u64 num_bytes;
8735         u64 min_allocable_bytes;
8736         int ret = -ENOSPC;
8737
8738
8739         /*
8740          * We need some metadata space and system metadata space for
8741          * allocating chunks in some corner cases until we force to set
8742          * it to be readonly.
8743          */
8744         if ((sinfo->flags &
8745              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8746             !force)
8747                 min_allocable_bytes = 1 * 1024 * 1024;
8748         else
8749                 min_allocable_bytes = 0;
8750
8751         spin_lock(&sinfo->lock);
8752         spin_lock(&cache->lock);
8753
8754         if (cache->ro) {
8755                 ret = 0;
8756                 goto out;
8757         }
8758
8759         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8760                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8761
8762         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8763             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8764             min_allocable_bytes <= sinfo->total_bytes) {
8765                 sinfo->bytes_readonly += num_bytes;
8766                 cache->ro = 1;
8767                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
8768                 ret = 0;
8769         }
8770 out:
8771         spin_unlock(&cache->lock);
8772         spin_unlock(&sinfo->lock);
8773         return ret;
8774 }
8775
8776 int btrfs_set_block_group_ro(struct btrfs_root *root,
8777                              struct btrfs_block_group_cache *cache)
8778
8779 {
8780         struct btrfs_trans_handle *trans;
8781         u64 alloc_flags;
8782         int ret;
8783
8784         BUG_ON(cache->ro);
8785
8786 again:
8787         trans = btrfs_join_transaction(root);
8788         if (IS_ERR(trans))
8789                 return PTR_ERR(trans);
8790
8791         /*
8792          * we're not allowed to set block groups readonly after the dirty
8793          * block groups cache has started writing.  If it already started,
8794          * back off and let this transaction commit
8795          */
8796         mutex_lock(&root->fs_info->ro_block_group_mutex);
8797         if (trans->transaction->dirty_bg_run) {
8798                 u64 transid = trans->transid;
8799
8800                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
8801                 btrfs_end_transaction(trans, root);
8802
8803                 ret = btrfs_wait_for_commit(root, transid);
8804                 if (ret)
8805                         return ret;
8806                 goto again;
8807         }
8808
8809
8810         ret = set_block_group_ro(cache, 0);
8811         if (!ret)
8812                 goto out;
8813         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8814         ret = do_chunk_alloc(trans, root, alloc_flags,
8815                              CHUNK_ALLOC_FORCE);
8816         if (ret < 0)
8817                 goto out;
8818         ret = set_block_group_ro(cache, 0);
8819 out:
8820         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
8821                 alloc_flags = update_block_group_flags(root, cache->flags);
8822                 check_system_chunk(trans, root, alloc_flags);
8823         }
8824         mutex_unlock(&root->fs_info->ro_block_group_mutex);
8825
8826         btrfs_end_transaction(trans, root);
8827         return ret;
8828 }
8829
8830 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8831                             struct btrfs_root *root, u64 type)
8832 {
8833         u64 alloc_flags = get_alloc_profile(root, type);
8834         return do_chunk_alloc(trans, root, alloc_flags,
8835                               CHUNK_ALLOC_FORCE);
8836 }
8837
8838 /*
8839  * helper to account the unused space of all the readonly block group in the
8840  * space_info. takes mirrors into account.
8841  */
8842 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8843 {
8844         struct btrfs_block_group_cache *block_group;
8845         u64 free_bytes = 0;
8846         int factor;
8847
8848         /* It's df, we don't care if it's racey */
8849         if (list_empty(&sinfo->ro_bgs))
8850                 return 0;
8851
8852         spin_lock(&sinfo->lock);
8853         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
8854                 spin_lock(&block_group->lock);
8855
8856                 if (!block_group->ro) {
8857                         spin_unlock(&block_group->lock);
8858                         continue;
8859                 }
8860
8861                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8862                                           BTRFS_BLOCK_GROUP_RAID10 |
8863                                           BTRFS_BLOCK_GROUP_DUP))
8864                         factor = 2;
8865                 else
8866                         factor = 1;
8867
8868                 free_bytes += (block_group->key.offset -
8869                                btrfs_block_group_used(&block_group->item)) *
8870                                factor;
8871
8872                 spin_unlock(&block_group->lock);
8873         }
8874         spin_unlock(&sinfo->lock);
8875
8876         return free_bytes;
8877 }
8878
8879 void btrfs_set_block_group_rw(struct btrfs_root *root,
8880                               struct btrfs_block_group_cache *cache)
8881 {
8882         struct btrfs_space_info *sinfo = cache->space_info;
8883         u64 num_bytes;
8884
8885         BUG_ON(!cache->ro);
8886
8887         spin_lock(&sinfo->lock);
8888         spin_lock(&cache->lock);
8889         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8890                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8891         sinfo->bytes_readonly -= num_bytes;
8892         cache->ro = 0;
8893         list_del_init(&cache->ro_list);
8894         spin_unlock(&cache->lock);
8895         spin_unlock(&sinfo->lock);
8896 }
8897
8898 /*
8899  * checks to see if its even possible to relocate this block group.
8900  *
8901  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8902  * ok to go ahead and try.
8903  */
8904 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8905 {
8906         struct btrfs_block_group_cache *block_group;
8907         struct btrfs_space_info *space_info;
8908         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8909         struct btrfs_device *device;
8910         struct btrfs_trans_handle *trans;
8911         u64 min_free;
8912         u64 dev_min = 1;
8913         u64 dev_nr = 0;
8914         u64 target;
8915         int index;
8916         int full = 0;
8917         int ret = 0;
8918
8919         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8920
8921         /* odd, couldn't find the block group, leave it alone */
8922         if (!block_group)
8923                 return -1;
8924
8925         min_free = btrfs_block_group_used(&block_group->item);
8926
8927         /* no bytes used, we're good */
8928         if (!min_free)
8929                 goto out;
8930
8931         space_info = block_group->space_info;
8932         spin_lock(&space_info->lock);
8933
8934         full = space_info->full;
8935
8936         /*
8937          * if this is the last block group we have in this space, we can't
8938          * relocate it unless we're able to allocate a new chunk below.
8939          *
8940          * Otherwise, we need to make sure we have room in the space to handle
8941          * all of the extents from this block group.  If we can, we're good
8942          */
8943         if ((space_info->total_bytes != block_group->key.offset) &&
8944             (space_info->bytes_used + space_info->bytes_reserved +
8945              space_info->bytes_pinned + space_info->bytes_readonly +
8946              min_free < space_info->total_bytes)) {
8947                 spin_unlock(&space_info->lock);
8948                 goto out;
8949         }
8950         spin_unlock(&space_info->lock);
8951
8952         /*
8953          * ok we don't have enough space, but maybe we have free space on our
8954          * devices to allocate new chunks for relocation, so loop through our
8955          * alloc devices and guess if we have enough space.  if this block
8956          * group is going to be restriped, run checks against the target
8957          * profile instead of the current one.
8958          */
8959         ret = -1;
8960
8961         /*
8962          * index:
8963          *      0: raid10
8964          *      1: raid1
8965          *      2: dup
8966          *      3: raid0
8967          *      4: single
8968          */
8969         target = get_restripe_target(root->fs_info, block_group->flags);
8970         if (target) {
8971                 index = __get_raid_index(extended_to_chunk(target));
8972         } else {
8973                 /*
8974                  * this is just a balance, so if we were marked as full
8975                  * we know there is no space for a new chunk
8976                  */
8977                 if (full)
8978                         goto out;
8979
8980                 index = get_block_group_index(block_group);
8981         }
8982
8983         if (index == BTRFS_RAID_RAID10) {
8984                 dev_min = 4;
8985                 /* Divide by 2 */
8986                 min_free >>= 1;
8987         } else if (index == BTRFS_RAID_RAID1) {
8988                 dev_min = 2;
8989         } else if (index == BTRFS_RAID_DUP) {
8990                 /* Multiply by 2 */
8991                 min_free <<= 1;
8992         } else if (index == BTRFS_RAID_RAID0) {
8993                 dev_min = fs_devices->rw_devices;
8994                 min_free = div64_u64(min_free, dev_min);
8995         }
8996
8997         /* We need to do this so that we can look at pending chunks */
8998         trans = btrfs_join_transaction(root);
8999         if (IS_ERR(trans)) {
9000                 ret = PTR_ERR(trans);
9001                 goto out;
9002         }
9003
9004         mutex_lock(&root->fs_info->chunk_mutex);
9005         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9006                 u64 dev_offset;
9007
9008                 /*
9009                  * check to make sure we can actually find a chunk with enough
9010                  * space to fit our block group in.
9011                  */
9012                 if (device->total_bytes > device->bytes_used + min_free &&
9013                     !device->is_tgtdev_for_dev_replace) {
9014                         ret = find_free_dev_extent(trans, device, min_free,
9015                                                    &dev_offset, NULL);
9016                         if (!ret)
9017                                 dev_nr++;
9018
9019                         if (dev_nr >= dev_min)
9020                                 break;
9021
9022                         ret = -1;
9023                 }
9024         }
9025         mutex_unlock(&root->fs_info->chunk_mutex);
9026         btrfs_end_transaction(trans, root);
9027 out:
9028         btrfs_put_block_group(block_group);
9029         return ret;
9030 }
9031
9032 static int find_first_block_group(struct btrfs_root *root,
9033                 struct btrfs_path *path, struct btrfs_key *key)
9034 {
9035         int ret = 0;
9036         struct btrfs_key found_key;
9037         struct extent_buffer *leaf;
9038         int slot;
9039
9040         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9041         if (ret < 0)
9042                 goto out;
9043
9044         while (1) {
9045                 slot = path->slots[0];
9046                 leaf = path->nodes[0];
9047                 if (slot >= btrfs_header_nritems(leaf)) {
9048                         ret = btrfs_next_leaf(root, path);
9049                         if (ret == 0)
9050                                 continue;
9051                         if (ret < 0)
9052                                 goto out;
9053                         break;
9054                 }
9055                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9056
9057                 if (found_key.objectid >= key->objectid &&
9058                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9059                         ret = 0;
9060                         goto out;
9061                 }
9062                 path->slots[0]++;
9063         }
9064 out:
9065         return ret;
9066 }
9067
9068 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9069 {
9070         struct btrfs_block_group_cache *block_group;
9071         u64 last = 0;
9072
9073         while (1) {
9074                 struct inode *inode;
9075
9076                 block_group = btrfs_lookup_first_block_group(info, last);
9077                 while (block_group) {
9078                         spin_lock(&block_group->lock);
9079                         if (block_group->iref)
9080                                 break;
9081                         spin_unlock(&block_group->lock);
9082                         block_group = next_block_group(info->tree_root,
9083                                                        block_group);
9084                 }
9085                 if (!block_group) {
9086                         if (last == 0)
9087                                 break;
9088                         last = 0;
9089                         continue;
9090                 }
9091
9092                 inode = block_group->inode;
9093                 block_group->iref = 0;
9094                 block_group->inode = NULL;
9095                 spin_unlock(&block_group->lock);
9096                 iput(inode);
9097                 last = block_group->key.objectid + block_group->key.offset;
9098                 btrfs_put_block_group(block_group);
9099         }
9100 }
9101
9102 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9103 {
9104         struct btrfs_block_group_cache *block_group;
9105         struct btrfs_space_info *space_info;
9106         struct btrfs_caching_control *caching_ctl;
9107         struct rb_node *n;
9108
9109         down_write(&info->commit_root_sem);
9110         while (!list_empty(&info->caching_block_groups)) {
9111                 caching_ctl = list_entry(info->caching_block_groups.next,
9112                                          struct btrfs_caching_control, list);
9113                 list_del(&caching_ctl->list);
9114                 put_caching_control(caching_ctl);
9115         }
9116         up_write(&info->commit_root_sem);
9117
9118         spin_lock(&info->unused_bgs_lock);
9119         while (!list_empty(&info->unused_bgs)) {
9120                 block_group = list_first_entry(&info->unused_bgs,
9121                                                struct btrfs_block_group_cache,
9122                                                bg_list);
9123                 list_del_init(&block_group->bg_list);
9124                 btrfs_put_block_group(block_group);
9125         }
9126         spin_unlock(&info->unused_bgs_lock);
9127
9128         spin_lock(&info->block_group_cache_lock);
9129         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9130                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9131                                        cache_node);
9132                 rb_erase(&block_group->cache_node,
9133                          &info->block_group_cache_tree);
9134                 RB_CLEAR_NODE(&block_group->cache_node);
9135                 spin_unlock(&info->block_group_cache_lock);
9136
9137                 down_write(&block_group->space_info->groups_sem);
9138                 list_del(&block_group->list);
9139                 up_write(&block_group->space_info->groups_sem);
9140
9141                 if (block_group->cached == BTRFS_CACHE_STARTED)
9142                         wait_block_group_cache_done(block_group);
9143
9144                 /*
9145                  * We haven't cached this block group, which means we could
9146                  * possibly have excluded extents on this block group.
9147                  */
9148                 if (block_group->cached == BTRFS_CACHE_NO ||
9149                     block_group->cached == BTRFS_CACHE_ERROR)
9150                         free_excluded_extents(info->extent_root, block_group);
9151
9152                 btrfs_remove_free_space_cache(block_group);
9153                 btrfs_put_block_group(block_group);
9154
9155                 spin_lock(&info->block_group_cache_lock);
9156         }
9157         spin_unlock(&info->block_group_cache_lock);
9158
9159         /* now that all the block groups are freed, go through and
9160          * free all the space_info structs.  This is only called during
9161          * the final stages of unmount, and so we know nobody is
9162          * using them.  We call synchronize_rcu() once before we start,
9163          * just to be on the safe side.
9164          */
9165         synchronize_rcu();
9166
9167         release_global_block_rsv(info);
9168
9169         while (!list_empty(&info->space_info)) {
9170                 int i;
9171
9172                 space_info = list_entry(info->space_info.next,
9173                                         struct btrfs_space_info,
9174                                         list);
9175                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9176                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9177                             space_info->bytes_reserved > 0 ||
9178                             space_info->bytes_may_use > 0)) {
9179                                 dump_space_info(space_info, 0, 0);
9180                         }
9181                 }
9182                 list_del(&space_info->list);
9183                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9184                         struct kobject *kobj;
9185                         kobj = space_info->block_group_kobjs[i];
9186                         space_info->block_group_kobjs[i] = NULL;
9187                         if (kobj) {
9188                                 kobject_del(kobj);
9189                                 kobject_put(kobj);
9190                         }
9191                 }
9192                 kobject_del(&space_info->kobj);
9193                 kobject_put(&space_info->kobj);
9194         }
9195         return 0;
9196 }
9197
9198 static void __link_block_group(struct btrfs_space_info *space_info,
9199                                struct btrfs_block_group_cache *cache)
9200 {
9201         int index = get_block_group_index(cache);
9202         bool first = false;
9203
9204         down_write(&space_info->groups_sem);
9205         if (list_empty(&space_info->block_groups[index]))
9206                 first = true;
9207         list_add_tail(&cache->list, &space_info->block_groups[index]);
9208         up_write(&space_info->groups_sem);
9209
9210         if (first) {
9211                 struct raid_kobject *rkobj;
9212                 int ret;
9213
9214                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9215                 if (!rkobj)
9216                         goto out_err;
9217                 rkobj->raid_type = index;
9218                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9219                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9220                                   "%s", get_raid_name(index));
9221                 if (ret) {
9222                         kobject_put(&rkobj->kobj);
9223                         goto out_err;
9224                 }
9225                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9226         }
9227
9228         return;
9229 out_err:
9230         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9231 }
9232
9233 static struct btrfs_block_group_cache *
9234 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9235 {
9236         struct btrfs_block_group_cache *cache;
9237
9238         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9239         if (!cache)
9240                 return NULL;
9241
9242         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9243                                         GFP_NOFS);
9244         if (!cache->free_space_ctl) {
9245                 kfree(cache);
9246                 return NULL;
9247         }
9248
9249         cache->key.objectid = start;
9250         cache->key.offset = size;
9251         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9252
9253         cache->sectorsize = root->sectorsize;
9254         cache->fs_info = root->fs_info;
9255         cache->full_stripe_len = btrfs_full_stripe_len(root,
9256                                                &root->fs_info->mapping_tree,
9257                                                start);
9258         atomic_set(&cache->count, 1);
9259         spin_lock_init(&cache->lock);
9260         init_rwsem(&cache->data_rwsem);
9261         INIT_LIST_HEAD(&cache->list);
9262         INIT_LIST_HEAD(&cache->cluster_list);
9263         INIT_LIST_HEAD(&cache->bg_list);
9264         INIT_LIST_HEAD(&cache->ro_list);
9265         INIT_LIST_HEAD(&cache->dirty_list);
9266         INIT_LIST_HEAD(&cache->io_list);
9267         btrfs_init_free_space_ctl(cache);
9268         atomic_set(&cache->trimming, 0);
9269
9270         return cache;
9271 }
9272
9273 int btrfs_read_block_groups(struct btrfs_root *root)
9274 {
9275         struct btrfs_path *path;
9276         int ret;
9277         struct btrfs_block_group_cache *cache;
9278         struct btrfs_fs_info *info = root->fs_info;
9279         struct btrfs_space_info *space_info;
9280         struct btrfs_key key;
9281         struct btrfs_key found_key;
9282         struct extent_buffer *leaf;
9283         int need_clear = 0;
9284         u64 cache_gen;
9285
9286         root = info->extent_root;
9287         key.objectid = 0;
9288         key.offset = 0;
9289         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9290         path = btrfs_alloc_path();
9291         if (!path)
9292                 return -ENOMEM;
9293         path->reada = 1;
9294
9295         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9296         if (btrfs_test_opt(root, SPACE_CACHE) &&
9297             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9298                 need_clear = 1;
9299         if (btrfs_test_opt(root, CLEAR_CACHE))
9300                 need_clear = 1;
9301
9302         while (1) {
9303                 ret = find_first_block_group(root, path, &key);
9304                 if (ret > 0)
9305                         break;
9306                 if (ret != 0)
9307                         goto error;
9308
9309                 leaf = path->nodes[0];
9310                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9311
9312                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9313                                                        found_key.offset);
9314                 if (!cache) {
9315                         ret = -ENOMEM;
9316                         goto error;
9317                 }
9318
9319                 if (need_clear) {
9320                         /*
9321                          * When we mount with old space cache, we need to
9322                          * set BTRFS_DC_CLEAR and set dirty flag.
9323                          *
9324                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9325                          *    truncate the old free space cache inode and
9326                          *    setup a new one.
9327                          * b) Setting 'dirty flag' makes sure that we flush
9328                          *    the new space cache info onto disk.
9329                          */
9330                         if (btrfs_test_opt(root, SPACE_CACHE))
9331                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9332                 }
9333
9334                 read_extent_buffer(leaf, &cache->item,
9335                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9336                                    sizeof(cache->item));
9337                 cache->flags = btrfs_block_group_flags(&cache->item);
9338
9339                 key.objectid = found_key.objectid + found_key.offset;
9340                 btrfs_release_path(path);
9341
9342                 /*
9343                  * We need to exclude the super stripes now so that the space
9344                  * info has super bytes accounted for, otherwise we'll think
9345                  * we have more space than we actually do.
9346                  */
9347                 ret = exclude_super_stripes(root, cache);
9348                 if (ret) {
9349                         /*
9350                          * We may have excluded something, so call this just in
9351                          * case.
9352                          */
9353                         free_excluded_extents(root, cache);
9354                         btrfs_put_block_group(cache);
9355                         goto error;
9356                 }
9357
9358                 /*
9359                  * check for two cases, either we are full, and therefore
9360                  * don't need to bother with the caching work since we won't
9361                  * find any space, or we are empty, and we can just add all
9362                  * the space in and be done with it.  This saves us _alot_ of
9363                  * time, particularly in the full case.
9364                  */
9365                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9366                         cache->last_byte_to_unpin = (u64)-1;
9367                         cache->cached = BTRFS_CACHE_FINISHED;
9368                         free_excluded_extents(root, cache);
9369                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9370                         cache->last_byte_to_unpin = (u64)-1;
9371                         cache->cached = BTRFS_CACHE_FINISHED;
9372                         add_new_free_space(cache, root->fs_info,
9373                                            found_key.objectid,
9374                                            found_key.objectid +
9375                                            found_key.offset);
9376                         free_excluded_extents(root, cache);
9377                 }
9378
9379                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9380                 if (ret) {
9381                         btrfs_remove_free_space_cache(cache);
9382                         btrfs_put_block_group(cache);
9383                         goto error;
9384                 }
9385
9386                 ret = update_space_info(info, cache->flags, found_key.offset,
9387                                         btrfs_block_group_used(&cache->item),
9388                                         &space_info);
9389                 if (ret) {
9390                         btrfs_remove_free_space_cache(cache);
9391                         spin_lock(&info->block_group_cache_lock);
9392                         rb_erase(&cache->cache_node,
9393                                  &info->block_group_cache_tree);
9394                         RB_CLEAR_NODE(&cache->cache_node);
9395                         spin_unlock(&info->block_group_cache_lock);
9396                         btrfs_put_block_group(cache);
9397                         goto error;
9398                 }
9399
9400                 cache->space_info = space_info;
9401                 spin_lock(&cache->space_info->lock);
9402                 cache->space_info->bytes_readonly += cache->bytes_super;
9403                 spin_unlock(&cache->space_info->lock);
9404
9405                 __link_block_group(space_info, cache);
9406
9407                 set_avail_alloc_bits(root->fs_info, cache->flags);
9408                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9409                         set_block_group_ro(cache, 1);
9410                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9411                         spin_lock(&info->unused_bgs_lock);
9412                         /* Should always be true but just in case. */
9413                         if (list_empty(&cache->bg_list)) {
9414                                 btrfs_get_block_group(cache);
9415                                 list_add_tail(&cache->bg_list,
9416                                               &info->unused_bgs);
9417                         }
9418                         spin_unlock(&info->unused_bgs_lock);
9419                 }
9420         }
9421
9422         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9423                 if (!(get_alloc_profile(root, space_info->flags) &
9424                       (BTRFS_BLOCK_GROUP_RAID10 |
9425                        BTRFS_BLOCK_GROUP_RAID1 |
9426                        BTRFS_BLOCK_GROUP_RAID5 |
9427                        BTRFS_BLOCK_GROUP_RAID6 |
9428                        BTRFS_BLOCK_GROUP_DUP)))
9429                         continue;
9430                 /*
9431                  * avoid allocating from un-mirrored block group if there are
9432                  * mirrored block groups.
9433                  */
9434                 list_for_each_entry(cache,
9435                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9436                                 list)
9437                         set_block_group_ro(cache, 1);
9438                 list_for_each_entry(cache,
9439                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9440                                 list)
9441                         set_block_group_ro(cache, 1);
9442         }
9443
9444         init_global_block_rsv(info);
9445         ret = 0;
9446 error:
9447         btrfs_free_path(path);
9448         return ret;
9449 }
9450
9451 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9452                                        struct btrfs_root *root)
9453 {
9454         struct btrfs_block_group_cache *block_group, *tmp;
9455         struct btrfs_root *extent_root = root->fs_info->extent_root;
9456         struct btrfs_block_group_item item;
9457         struct btrfs_key key;
9458         int ret = 0;
9459
9460         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9461                 if (ret)
9462                         goto next;
9463
9464                 spin_lock(&block_group->lock);
9465                 memcpy(&item, &block_group->item, sizeof(item));
9466                 memcpy(&key, &block_group->key, sizeof(key));
9467                 spin_unlock(&block_group->lock);
9468
9469                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9470                                         sizeof(item));
9471                 if (ret)
9472                         btrfs_abort_transaction(trans, extent_root, ret);
9473                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9474                                                key.objectid, key.offset);
9475                 if (ret)
9476                         btrfs_abort_transaction(trans, extent_root, ret);
9477 next:
9478                 list_del_init(&block_group->bg_list);
9479         }
9480 }
9481
9482 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9483                            struct btrfs_root *root, u64 bytes_used,
9484                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9485                            u64 size)
9486 {
9487         int ret;
9488         struct btrfs_root *extent_root;
9489         struct btrfs_block_group_cache *cache;
9490
9491         extent_root = root->fs_info->extent_root;
9492
9493         btrfs_set_log_full_commit(root->fs_info, trans);
9494
9495         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9496         if (!cache)
9497                 return -ENOMEM;
9498
9499         btrfs_set_block_group_used(&cache->item, bytes_used);
9500         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9501         btrfs_set_block_group_flags(&cache->item, type);
9502
9503         cache->flags = type;
9504         cache->last_byte_to_unpin = (u64)-1;
9505         cache->cached = BTRFS_CACHE_FINISHED;
9506         ret = exclude_super_stripes(root, cache);
9507         if (ret) {
9508                 /*
9509                  * We may have excluded something, so call this just in
9510                  * case.
9511                  */
9512                 free_excluded_extents(root, cache);
9513                 btrfs_put_block_group(cache);
9514                 return ret;
9515         }
9516
9517         add_new_free_space(cache, root->fs_info, chunk_offset,
9518                            chunk_offset + size);
9519
9520         free_excluded_extents(root, cache);
9521
9522         ret = btrfs_add_block_group_cache(root->fs_info, cache);
9523         if (ret) {
9524                 btrfs_remove_free_space_cache(cache);
9525                 btrfs_put_block_group(cache);
9526                 return ret;
9527         }
9528
9529         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
9530                                 &cache->space_info);
9531         if (ret) {
9532                 btrfs_remove_free_space_cache(cache);
9533                 spin_lock(&root->fs_info->block_group_cache_lock);
9534                 rb_erase(&cache->cache_node,
9535                          &root->fs_info->block_group_cache_tree);
9536                 RB_CLEAR_NODE(&cache->cache_node);
9537                 spin_unlock(&root->fs_info->block_group_cache_lock);
9538                 btrfs_put_block_group(cache);
9539                 return ret;
9540         }
9541         update_global_block_rsv(root->fs_info);
9542
9543         spin_lock(&cache->space_info->lock);
9544         cache->space_info->bytes_readonly += cache->bytes_super;
9545         spin_unlock(&cache->space_info->lock);
9546
9547         __link_block_group(cache->space_info, cache);
9548
9549         list_add_tail(&cache->bg_list, &trans->new_bgs);
9550
9551         set_avail_alloc_bits(extent_root->fs_info, type);
9552
9553         return 0;
9554 }
9555
9556 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9557 {
9558         u64 extra_flags = chunk_to_extended(flags) &
9559                                 BTRFS_EXTENDED_PROFILE_MASK;
9560
9561         write_seqlock(&fs_info->profiles_lock);
9562         if (flags & BTRFS_BLOCK_GROUP_DATA)
9563                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9564         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9565                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9566         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9567                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9568         write_sequnlock(&fs_info->profiles_lock);
9569 }
9570
9571 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9572                              struct btrfs_root *root, u64 group_start,
9573                              struct extent_map *em)
9574 {
9575         struct btrfs_path *path;
9576         struct btrfs_block_group_cache *block_group;
9577         struct btrfs_free_cluster *cluster;
9578         struct btrfs_root *tree_root = root->fs_info->tree_root;
9579         struct btrfs_key key;
9580         struct inode *inode;
9581         struct kobject *kobj = NULL;
9582         int ret;
9583         int index;
9584         int factor;
9585         struct btrfs_caching_control *caching_ctl = NULL;
9586         bool remove_em;
9587
9588         root = root->fs_info->extent_root;
9589
9590         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9591         BUG_ON(!block_group);
9592         BUG_ON(!block_group->ro);
9593
9594         /*
9595          * Free the reserved super bytes from this block group before
9596          * remove it.
9597          */
9598         free_excluded_extents(root, block_group);
9599
9600         memcpy(&key, &block_group->key, sizeof(key));
9601         index = get_block_group_index(block_group);
9602         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9603                                   BTRFS_BLOCK_GROUP_RAID1 |
9604                                   BTRFS_BLOCK_GROUP_RAID10))
9605                 factor = 2;
9606         else
9607                 factor = 1;
9608
9609         /* make sure this block group isn't part of an allocation cluster */
9610         cluster = &root->fs_info->data_alloc_cluster;
9611         spin_lock(&cluster->refill_lock);
9612         btrfs_return_cluster_to_free_space(block_group, cluster);
9613         spin_unlock(&cluster->refill_lock);
9614
9615         /*
9616          * make sure this block group isn't part of a metadata
9617          * allocation cluster
9618          */
9619         cluster = &root->fs_info->meta_alloc_cluster;
9620         spin_lock(&cluster->refill_lock);
9621         btrfs_return_cluster_to_free_space(block_group, cluster);
9622         spin_unlock(&cluster->refill_lock);
9623
9624         path = btrfs_alloc_path();
9625         if (!path) {
9626                 ret = -ENOMEM;
9627                 goto out;
9628         }
9629
9630         /*
9631          * get the inode first so any iput calls done for the io_list
9632          * aren't the final iput (no unlinks allowed now)
9633          */
9634         inode = lookup_free_space_inode(tree_root, block_group, path);
9635
9636         mutex_lock(&trans->transaction->cache_write_mutex);
9637         /*
9638          * make sure our free spache cache IO is done before remove the
9639          * free space inode
9640          */
9641         spin_lock(&trans->transaction->dirty_bgs_lock);
9642         if (!list_empty(&block_group->io_list)) {
9643                 list_del_init(&block_group->io_list);
9644
9645                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
9646
9647                 spin_unlock(&trans->transaction->dirty_bgs_lock);
9648                 btrfs_wait_cache_io(root, trans, block_group,
9649                                     &block_group->io_ctl, path,
9650                                     block_group->key.objectid);
9651                 btrfs_put_block_group(block_group);
9652                 spin_lock(&trans->transaction->dirty_bgs_lock);
9653         }
9654
9655         if (!list_empty(&block_group->dirty_list)) {
9656                 list_del_init(&block_group->dirty_list);
9657                 btrfs_put_block_group(block_group);
9658         }
9659         spin_unlock(&trans->transaction->dirty_bgs_lock);
9660         mutex_unlock(&trans->transaction->cache_write_mutex);
9661
9662         if (!IS_ERR(inode)) {
9663                 ret = btrfs_orphan_add(trans, inode);
9664                 if (ret) {
9665                         btrfs_add_delayed_iput(inode);
9666                         goto out;
9667                 }
9668                 clear_nlink(inode);
9669                 /* One for the block groups ref */
9670                 spin_lock(&block_group->lock);
9671                 if (block_group->iref) {
9672                         block_group->iref = 0;
9673                         block_group->inode = NULL;
9674                         spin_unlock(&block_group->lock);
9675                         iput(inode);
9676                 } else {
9677                         spin_unlock(&block_group->lock);
9678                 }
9679                 /* One for our lookup ref */
9680                 btrfs_add_delayed_iput(inode);
9681         }
9682
9683         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9684         key.offset = block_group->key.objectid;
9685         key.type = 0;
9686
9687         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9688         if (ret < 0)
9689                 goto out;
9690         if (ret > 0)
9691                 btrfs_release_path(path);
9692         if (ret == 0) {
9693                 ret = btrfs_del_item(trans, tree_root, path);
9694                 if (ret)
9695                         goto out;
9696                 btrfs_release_path(path);
9697         }
9698
9699         spin_lock(&root->fs_info->block_group_cache_lock);
9700         rb_erase(&block_group->cache_node,
9701                  &root->fs_info->block_group_cache_tree);
9702         RB_CLEAR_NODE(&block_group->cache_node);
9703
9704         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9705                 root->fs_info->first_logical_byte = (u64)-1;
9706         spin_unlock(&root->fs_info->block_group_cache_lock);
9707
9708         down_write(&block_group->space_info->groups_sem);
9709         /*
9710          * we must use list_del_init so people can check to see if they
9711          * are still on the list after taking the semaphore
9712          */
9713         list_del_init(&block_group->list);
9714         if (list_empty(&block_group->space_info->block_groups[index])) {
9715                 kobj = block_group->space_info->block_group_kobjs[index];
9716                 block_group->space_info->block_group_kobjs[index] = NULL;
9717                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9718         }
9719         up_write(&block_group->space_info->groups_sem);
9720         if (kobj) {
9721                 kobject_del(kobj);
9722                 kobject_put(kobj);
9723         }
9724
9725         if (block_group->has_caching_ctl)
9726                 caching_ctl = get_caching_control(block_group);
9727         if (block_group->cached == BTRFS_CACHE_STARTED)
9728                 wait_block_group_cache_done(block_group);
9729         if (block_group->has_caching_ctl) {
9730                 down_write(&root->fs_info->commit_root_sem);
9731                 if (!caching_ctl) {
9732                         struct btrfs_caching_control *ctl;
9733
9734                         list_for_each_entry(ctl,
9735                                     &root->fs_info->caching_block_groups, list)
9736                                 if (ctl->block_group == block_group) {
9737                                         caching_ctl = ctl;
9738                                         atomic_inc(&caching_ctl->count);
9739                                         break;
9740                                 }
9741                 }
9742                 if (caching_ctl)
9743                         list_del_init(&caching_ctl->list);
9744                 up_write(&root->fs_info->commit_root_sem);
9745                 if (caching_ctl) {
9746                         /* Once for the caching bgs list and once for us. */
9747                         put_caching_control(caching_ctl);
9748                         put_caching_control(caching_ctl);
9749                 }
9750         }
9751
9752         spin_lock(&trans->transaction->dirty_bgs_lock);
9753         if (!list_empty(&block_group->dirty_list)) {
9754                 WARN_ON(1);
9755         }
9756         if (!list_empty(&block_group->io_list)) {
9757                 WARN_ON(1);
9758         }
9759         spin_unlock(&trans->transaction->dirty_bgs_lock);
9760         btrfs_remove_free_space_cache(block_group);
9761
9762         spin_lock(&block_group->space_info->lock);
9763         list_del_init(&block_group->ro_list);
9764
9765         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
9766                 WARN_ON(block_group->space_info->total_bytes
9767                         < block_group->key.offset);
9768                 WARN_ON(block_group->space_info->bytes_readonly
9769                         < block_group->key.offset);
9770                 WARN_ON(block_group->space_info->disk_total
9771                         < block_group->key.offset * factor);
9772         }
9773         block_group->space_info->total_bytes -= block_group->key.offset;
9774         block_group->space_info->bytes_readonly -= block_group->key.offset;
9775         block_group->space_info->disk_total -= block_group->key.offset * factor;
9776
9777         spin_unlock(&block_group->space_info->lock);
9778
9779         memcpy(&key, &block_group->key, sizeof(key));
9780
9781         lock_chunks(root);
9782         if (!list_empty(&em->list)) {
9783                 /* We're in the transaction->pending_chunks list. */
9784                 free_extent_map(em);
9785         }
9786         spin_lock(&block_group->lock);
9787         block_group->removed = 1;
9788         /*
9789          * At this point trimming can't start on this block group, because we
9790          * removed the block group from the tree fs_info->block_group_cache_tree
9791          * so no one can't find it anymore and even if someone already got this
9792          * block group before we removed it from the rbtree, they have already
9793          * incremented block_group->trimming - if they didn't, they won't find
9794          * any free space entries because we already removed them all when we
9795          * called btrfs_remove_free_space_cache().
9796          *
9797          * And we must not remove the extent map from the fs_info->mapping_tree
9798          * to prevent the same logical address range and physical device space
9799          * ranges from being reused for a new block group. This is because our
9800          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
9801          * completely transactionless, so while it is trimming a range the
9802          * currently running transaction might finish and a new one start,
9803          * allowing for new block groups to be created that can reuse the same
9804          * physical device locations unless we take this special care.
9805          */
9806         remove_em = (atomic_read(&block_group->trimming) == 0);
9807         /*
9808          * Make sure a trimmer task always sees the em in the pinned_chunks list
9809          * if it sees block_group->removed == 1 (needs to lock block_group->lock
9810          * before checking block_group->removed).
9811          */
9812         if (!remove_em) {
9813                 /*
9814                  * Our em might be in trans->transaction->pending_chunks which
9815                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
9816                  * and so is the fs_info->pinned_chunks list.
9817                  *
9818                  * So at this point we must be holding the chunk_mutex to avoid
9819                  * any races with chunk allocation (more specifically at
9820                  * volumes.c:contains_pending_extent()), to ensure it always
9821                  * sees the em, either in the pending_chunks list or in the
9822                  * pinned_chunks list.
9823                  */
9824                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
9825         }
9826         spin_unlock(&block_group->lock);
9827
9828         if (remove_em) {
9829                 struct extent_map_tree *em_tree;
9830
9831                 em_tree = &root->fs_info->mapping_tree.map_tree;
9832                 write_lock(&em_tree->lock);
9833                 /*
9834                  * The em might be in the pending_chunks list, so make sure the
9835                  * chunk mutex is locked, since remove_extent_mapping() will
9836                  * delete us from that list.
9837                  */
9838                 remove_extent_mapping(em_tree, em);
9839                 write_unlock(&em_tree->lock);
9840                 /* once for the tree */
9841                 free_extent_map(em);
9842         }
9843
9844         unlock_chunks(root);
9845
9846         btrfs_put_block_group(block_group);
9847         btrfs_put_block_group(block_group);
9848
9849         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9850         if (ret > 0)
9851                 ret = -EIO;
9852         if (ret < 0)
9853                 goto out;
9854
9855         ret = btrfs_del_item(trans, root, path);
9856 out:
9857         btrfs_free_path(path);
9858         return ret;
9859 }
9860
9861 /*
9862  * Process the unused_bgs list and remove any that don't have any allocated
9863  * space inside of them.
9864  */
9865 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
9866 {
9867         struct btrfs_block_group_cache *block_group;
9868         struct btrfs_space_info *space_info;
9869         struct btrfs_root *root = fs_info->extent_root;
9870         struct btrfs_trans_handle *trans;
9871         int ret = 0;
9872
9873         if (!fs_info->open)
9874                 return;
9875
9876         spin_lock(&fs_info->unused_bgs_lock);
9877         while (!list_empty(&fs_info->unused_bgs)) {
9878                 u64 start, end;
9879
9880                 block_group = list_first_entry(&fs_info->unused_bgs,
9881                                                struct btrfs_block_group_cache,
9882                                                bg_list);
9883                 space_info = block_group->space_info;
9884                 list_del_init(&block_group->bg_list);
9885                 if (ret || btrfs_mixed_space_info(space_info)) {
9886                         btrfs_put_block_group(block_group);
9887                         continue;
9888                 }
9889                 spin_unlock(&fs_info->unused_bgs_lock);
9890
9891                 /* Don't want to race with allocators so take the groups_sem */
9892                 down_write(&space_info->groups_sem);
9893                 spin_lock(&block_group->lock);
9894                 if (block_group->reserved ||
9895                     btrfs_block_group_used(&block_group->item) ||
9896                     block_group->ro) {
9897                         /*
9898                          * We want to bail if we made new allocations or have
9899                          * outstanding allocations in this block group.  We do
9900                          * the ro check in case balance is currently acting on
9901                          * this block group.
9902                          */
9903                         spin_unlock(&block_group->lock);
9904                         up_write(&space_info->groups_sem);
9905                         goto next;
9906                 }
9907                 spin_unlock(&block_group->lock);
9908
9909                 /* We don't want to force the issue, only flip if it's ok. */
9910                 ret = set_block_group_ro(block_group, 0);
9911                 up_write(&space_info->groups_sem);
9912                 if (ret < 0) {
9913                         ret = 0;
9914                         goto next;
9915                 }
9916
9917                 /*
9918                  * Want to do this before we do anything else so we can recover
9919                  * properly if we fail to join the transaction.
9920                  */
9921                 /* 1 for btrfs_orphan_reserve_metadata() */
9922                 trans = btrfs_start_transaction(root, 1);
9923                 if (IS_ERR(trans)) {
9924                         btrfs_set_block_group_rw(root, block_group);
9925                         ret = PTR_ERR(trans);
9926                         goto next;
9927                 }
9928
9929                 /*
9930                  * We could have pending pinned extents for this block group,
9931                  * just delete them, we don't care about them anymore.
9932                  */
9933                 start = block_group->key.objectid;
9934                 end = start + block_group->key.offset - 1;
9935                 /*
9936                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
9937                  * btrfs_finish_extent_commit(). If we are at transaction N,
9938                  * another task might be running finish_extent_commit() for the
9939                  * previous transaction N - 1, and have seen a range belonging
9940                  * to the block group in freed_extents[] before we were able to
9941                  * clear the whole block group range from freed_extents[]. This
9942                  * means that task can lookup for the block group after we
9943                  * unpinned it from freed_extents[] and removed it, leading to
9944                  * a BUG_ON() at btrfs_unpin_extent_range().
9945                  */
9946                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
9947                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
9948                                   EXTENT_DIRTY, GFP_NOFS);
9949                 if (ret) {
9950                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9951                         btrfs_set_block_group_rw(root, block_group);
9952                         goto end_trans;
9953                 }
9954                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
9955                                   EXTENT_DIRTY, GFP_NOFS);
9956                 if (ret) {
9957                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9958                         btrfs_set_block_group_rw(root, block_group);
9959                         goto end_trans;
9960                 }
9961                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9962
9963                 /* Reset pinned so btrfs_put_block_group doesn't complain */
9964                 spin_lock(&space_info->lock);
9965                 spin_lock(&block_group->lock);
9966
9967                 space_info->bytes_pinned -= block_group->pinned;
9968                 space_info->bytes_readonly += block_group->pinned;
9969                 percpu_counter_add(&space_info->total_bytes_pinned,
9970                                    -block_group->pinned);
9971                 block_group->pinned = 0;
9972
9973                 spin_unlock(&block_group->lock);
9974                 spin_unlock(&space_info->lock);
9975
9976                 /*
9977                  * Btrfs_remove_chunk will abort the transaction if things go
9978                  * horribly wrong.
9979                  */
9980                 ret = btrfs_remove_chunk(trans, root,
9981                                          block_group->key.objectid);
9982 end_trans:
9983                 btrfs_end_transaction(trans, root);
9984 next:
9985                 btrfs_put_block_group(block_group);
9986                 spin_lock(&fs_info->unused_bgs_lock);
9987         }
9988         spin_unlock(&fs_info->unused_bgs_lock);
9989 }
9990
9991 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
9992 {
9993         struct btrfs_space_info *space_info;
9994         struct btrfs_super_block *disk_super;
9995         u64 features;
9996         u64 flags;
9997         int mixed = 0;
9998         int ret;
9999
10000         disk_super = fs_info->super_copy;
10001         if (!btrfs_super_root(disk_super))
10002                 return 1;
10003
10004         features = btrfs_super_incompat_flags(disk_super);
10005         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10006                 mixed = 1;
10007
10008         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10009         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10010         if (ret)
10011                 goto out;
10012
10013         if (mixed) {
10014                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10015                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10016         } else {
10017                 flags = BTRFS_BLOCK_GROUP_METADATA;
10018                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10019                 if (ret)
10020                         goto out;
10021
10022                 flags = BTRFS_BLOCK_GROUP_DATA;
10023                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10024         }
10025 out:
10026         return ret;
10027 }
10028
10029 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10030 {
10031         return unpin_extent_range(root, start, end, false);
10032 }
10033
10034 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10035 {
10036         struct btrfs_fs_info *fs_info = root->fs_info;
10037         struct btrfs_block_group_cache *cache = NULL;
10038         u64 group_trimmed;
10039         u64 start;
10040         u64 end;
10041         u64 trimmed = 0;
10042         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10043         int ret = 0;
10044
10045         /*
10046          * try to trim all FS space, our block group may start from non-zero.
10047          */
10048         if (range->len == total_bytes)
10049                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10050         else
10051                 cache = btrfs_lookup_block_group(fs_info, range->start);
10052
10053         while (cache) {
10054                 if (cache->key.objectid >= (range->start + range->len)) {
10055                         btrfs_put_block_group(cache);
10056                         break;
10057                 }
10058
10059                 start = max(range->start, cache->key.objectid);
10060                 end = min(range->start + range->len,
10061                                 cache->key.objectid + cache->key.offset);
10062
10063                 if (end - start >= range->minlen) {
10064                         if (!block_group_cache_done(cache)) {
10065                                 ret = cache_block_group(cache, 0);
10066                                 if (ret) {
10067                                         btrfs_put_block_group(cache);
10068                                         break;
10069                                 }
10070                                 ret = wait_block_group_cache_done(cache);
10071                                 if (ret) {
10072                                         btrfs_put_block_group(cache);
10073                                         break;
10074                                 }
10075                         }
10076                         ret = btrfs_trim_block_group(cache,
10077                                                      &group_trimmed,
10078                                                      start,
10079                                                      end,
10080                                                      range->minlen);
10081
10082                         trimmed += group_trimmed;
10083                         if (ret) {
10084                                 btrfs_put_block_group(cache);
10085                                 break;
10086                         }
10087                 }
10088
10089                 cache = next_block_group(fs_info->tree_root, cache);
10090         }
10091
10092         range->len = trimmed;
10093         return ret;
10094 }
10095
10096 /*
10097  * btrfs_{start,end}_write_no_snapshoting() are similar to
10098  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10099  * data into the page cache through nocow before the subvolume is snapshoted,
10100  * but flush the data into disk after the snapshot creation, or to prevent
10101  * operations while snapshoting is ongoing and that cause the snapshot to be
10102  * inconsistent (writes followed by expanding truncates for example).
10103  */
10104 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10105 {
10106         percpu_counter_dec(&root->subv_writers->counter);
10107         /*
10108          * Make sure counter is updated before we wake up
10109          * waiters.
10110          */
10111         smp_mb();
10112         if (waitqueue_active(&root->subv_writers->wait))
10113                 wake_up(&root->subv_writers->wait);
10114 }
10115
10116 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10117 {
10118         if (atomic_read(&root->will_be_snapshoted))
10119                 return 0;
10120
10121         percpu_counter_inc(&root->subv_writers->counter);
10122         /*
10123          * Make sure counter is updated before we check for snapshot creation.
10124          */
10125         smp_mb();
10126         if (atomic_read(&root->will_be_snapshoted)) {
10127                 btrfs_end_write_no_snapshoting(root);
10128                 return 0;
10129         }
10130         return 1;
10131 }