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         return ret;
3184
3185 }
3186
3187 static struct btrfs_block_group_cache *
3188 next_block_group(struct btrfs_root *root,
3189                  struct btrfs_block_group_cache *cache)
3190 {
3191         struct rb_node *node;
3192
3193         spin_lock(&root->fs_info->block_group_cache_lock);
3194
3195         /* If our block group was removed, we need a full search. */
3196         if (RB_EMPTY_NODE(&cache->cache_node)) {
3197                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3198
3199                 spin_unlock(&root->fs_info->block_group_cache_lock);
3200                 btrfs_put_block_group(cache);
3201                 cache = btrfs_lookup_first_block_group(root->fs_info,
3202                                                        next_bytenr);
3203                 return cache;
3204         }
3205         node = rb_next(&cache->cache_node);
3206         btrfs_put_block_group(cache);
3207         if (node) {
3208                 cache = rb_entry(node, struct btrfs_block_group_cache,
3209                                  cache_node);
3210                 btrfs_get_block_group(cache);
3211         } else
3212                 cache = NULL;
3213         spin_unlock(&root->fs_info->block_group_cache_lock);
3214         return cache;
3215 }
3216
3217 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3218                             struct btrfs_trans_handle *trans,
3219                             struct btrfs_path *path)
3220 {
3221         struct btrfs_root *root = block_group->fs_info->tree_root;
3222         struct inode *inode = NULL;
3223         u64 alloc_hint = 0;
3224         int dcs = BTRFS_DC_ERROR;
3225         u64 num_pages = 0;
3226         int retries = 0;
3227         int ret = 0;
3228
3229         /*
3230          * If this block group is smaller than 100 megs don't bother caching the
3231          * block group.
3232          */
3233         if (block_group->key.offset < (100 * 1024 * 1024)) {
3234                 spin_lock(&block_group->lock);
3235                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3236                 spin_unlock(&block_group->lock);
3237                 return 0;
3238         }
3239
3240         if (trans->aborted)
3241                 return 0;
3242 again:
3243         inode = lookup_free_space_inode(root, block_group, path);
3244         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3245                 ret = PTR_ERR(inode);
3246                 btrfs_release_path(path);
3247                 goto out;
3248         }
3249
3250         if (IS_ERR(inode)) {
3251                 BUG_ON(retries);
3252                 retries++;
3253
3254                 if (block_group->ro)
3255                         goto out_free;
3256
3257                 ret = create_free_space_inode(root, trans, block_group, path);
3258                 if (ret)
3259                         goto out_free;
3260                 goto again;
3261         }
3262
3263         /* We've already setup this transaction, go ahead and exit */
3264         if (block_group->cache_generation == trans->transid &&
3265             i_size_read(inode)) {
3266                 dcs = BTRFS_DC_SETUP;
3267                 goto out_put;
3268         }
3269
3270         /*
3271          * We want to set the generation to 0, that way if anything goes wrong
3272          * from here on out we know not to trust this cache when we load up next
3273          * time.
3274          */
3275         BTRFS_I(inode)->generation = 0;
3276         ret = btrfs_update_inode(trans, root, inode);
3277         if (ret) {
3278                 /*
3279                  * So theoretically we could recover from this, simply set the
3280                  * super cache generation to 0 so we know to invalidate the
3281                  * cache, but then we'd have to keep track of the block groups
3282                  * that fail this way so we know we _have_ to reset this cache
3283                  * before the next commit or risk reading stale cache.  So to
3284                  * limit our exposure to horrible edge cases lets just abort the
3285                  * transaction, this only happens in really bad situations
3286                  * anyway.
3287                  */
3288                 btrfs_abort_transaction(trans, root, ret);
3289                 goto out_put;
3290         }
3291         WARN_ON(ret);
3292
3293         if (i_size_read(inode) > 0) {
3294                 ret = btrfs_check_trunc_cache_free_space(root,
3295                                         &root->fs_info->global_block_rsv);
3296                 if (ret)
3297                         goto out_put;
3298
3299                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3300                 if (ret)
3301                         goto out_put;
3302         }
3303
3304         spin_lock(&block_group->lock);
3305         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3306             !btrfs_test_opt(root, SPACE_CACHE)) {
3307                 /*
3308                  * don't bother trying to write stuff out _if_
3309                  * a) we're not cached,
3310                  * b) we're with nospace_cache mount option.
3311                  */
3312                 dcs = BTRFS_DC_WRITTEN;
3313                 spin_unlock(&block_group->lock);
3314                 goto out_put;
3315         }
3316         spin_unlock(&block_group->lock);
3317
3318         /*
3319          * Try to preallocate enough space based on how big the block group is.
3320          * Keep in mind this has to include any pinned space which could end up
3321          * taking up quite a bit since it's not folded into the other space
3322          * cache.
3323          */
3324         num_pages = div_u64(block_group->key.offset, 256 * 1024 * 1024);
3325         if (!num_pages)
3326                 num_pages = 1;
3327
3328         num_pages *= 16;
3329         num_pages *= PAGE_CACHE_SIZE;
3330
3331         ret = btrfs_check_data_free_space(inode, num_pages, num_pages);
3332         if (ret)
3333                 goto out_put;
3334
3335         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3336                                               num_pages, num_pages,
3337                                               &alloc_hint);
3338         if (!ret)
3339                 dcs = BTRFS_DC_SETUP;
3340         btrfs_free_reserved_data_space(inode, num_pages);
3341
3342 out_put:
3343         iput(inode);
3344 out_free:
3345         btrfs_release_path(path);
3346 out:
3347         spin_lock(&block_group->lock);
3348         if (!ret && dcs == BTRFS_DC_SETUP)
3349                 block_group->cache_generation = trans->transid;
3350         block_group->disk_cache_state = dcs;
3351         spin_unlock(&block_group->lock);
3352
3353         return ret;
3354 }
3355
3356 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3357                             struct btrfs_root *root)
3358 {
3359         struct btrfs_block_group_cache *cache, *tmp;
3360         struct btrfs_transaction *cur_trans = trans->transaction;
3361         struct btrfs_path *path;
3362
3363         if (list_empty(&cur_trans->dirty_bgs) ||
3364             !btrfs_test_opt(root, SPACE_CACHE))
3365                 return 0;
3366
3367         path = btrfs_alloc_path();
3368         if (!path)
3369                 return -ENOMEM;
3370
3371         /* Could add new block groups, use _safe just in case */
3372         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3373                                  dirty_list) {
3374                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3375                         cache_save_setup(cache, trans, path);
3376         }
3377
3378         btrfs_free_path(path);
3379         return 0;
3380 }
3381
3382 /*
3383  * transaction commit does final block group cache writeback during a
3384  * critical section where nothing is allowed to change the FS.  This is
3385  * required in order for the cache to actually match the block group,
3386  * but can introduce a lot of latency into the commit.
3387  *
3388  * So, btrfs_start_dirty_block_groups is here to kick off block group
3389  * cache IO.  There's a chance we'll have to redo some of it if the
3390  * block group changes again during the commit, but it greatly reduces
3391  * the commit latency by getting rid of the easy block groups while
3392  * we're still allowing others to join the commit.
3393  */
3394 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3395                                    struct btrfs_root *root)
3396 {
3397         struct btrfs_block_group_cache *cache;
3398         struct btrfs_transaction *cur_trans = trans->transaction;
3399         int ret = 0;
3400         int should_put;
3401         struct btrfs_path *path = NULL;
3402         LIST_HEAD(dirty);
3403         struct list_head *io = &cur_trans->io_bgs;
3404         int num_started = 0;
3405         int loops = 0;
3406
3407         spin_lock(&cur_trans->dirty_bgs_lock);
3408         if (list_empty(&cur_trans->dirty_bgs)) {
3409                 spin_unlock(&cur_trans->dirty_bgs_lock);
3410                 return 0;
3411         }
3412         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3413         spin_unlock(&cur_trans->dirty_bgs_lock);
3414
3415 again:
3416         /*
3417          * make sure all the block groups on our dirty list actually
3418          * exist
3419          */
3420         btrfs_create_pending_block_groups(trans, root);
3421
3422         if (!path) {
3423                 path = btrfs_alloc_path();
3424                 if (!path)
3425                         return -ENOMEM;
3426         }
3427
3428         /*
3429          * cache_write_mutex is here only to save us from balance or automatic
3430          * removal of empty block groups deleting this block group while we are
3431          * writing out the cache
3432          */
3433         mutex_lock(&trans->transaction->cache_write_mutex);
3434         while (!list_empty(&dirty)) {
3435                 cache = list_first_entry(&dirty,
3436                                          struct btrfs_block_group_cache,
3437                                          dirty_list);
3438                 /*
3439                  * this can happen if something re-dirties a block
3440                  * group that is already under IO.  Just wait for it to
3441                  * finish and then do it all again
3442                  */
3443                 if (!list_empty(&cache->io_list)) {
3444                         list_del_init(&cache->io_list);
3445                         btrfs_wait_cache_io(root, trans, cache,
3446                                             &cache->io_ctl, path,
3447                                             cache->key.objectid);
3448                         btrfs_put_block_group(cache);
3449                 }
3450
3451
3452                 /*
3453                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3454                  * if it should update the cache_state.  Don't delete
3455                  * until after we wait.
3456                  *
3457                  * Since we're not running in the commit critical section
3458                  * we need the dirty_bgs_lock to protect from update_block_group
3459                  */
3460                 spin_lock(&cur_trans->dirty_bgs_lock);
3461                 list_del_init(&cache->dirty_list);
3462                 spin_unlock(&cur_trans->dirty_bgs_lock);
3463
3464                 should_put = 1;
3465
3466                 cache_save_setup(cache, trans, path);
3467
3468                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3469                         cache->io_ctl.inode = NULL;
3470                         ret = btrfs_write_out_cache(root, trans, cache, path);
3471                         if (ret == 0 && cache->io_ctl.inode) {
3472                                 num_started++;
3473                                 should_put = 0;
3474
3475                                 /*
3476                                  * the cache_write_mutex is protecting
3477                                  * the io_list
3478                                  */
3479                                 list_add_tail(&cache->io_list, io);
3480                         } else {
3481                                 /*
3482                                  * if we failed to write the cache, the
3483                                  * generation will be bad and life goes on
3484                                  */
3485                                 ret = 0;
3486                         }
3487                 }
3488                 if (!ret) {
3489                         ret = write_one_cache_group(trans, root, path, cache);
3490                         /*
3491                          * Our block group might still be attached to the list
3492                          * of new block groups in the transaction handle of some
3493                          * other task (struct btrfs_trans_handle->new_bgs). This
3494                          * means its block group item isn't yet in the extent
3495                          * tree. If this happens ignore the error, as we will
3496                          * try again later in the critical section of the
3497                          * transaction commit.
3498                          */
3499                         if (ret == -ENOENT) {
3500                                 ret = 0;
3501                                 spin_lock(&cur_trans->dirty_bgs_lock);
3502                                 if (list_empty(&cache->dirty_list)) {
3503                                         list_add_tail(&cache->dirty_list,
3504                                                       &cur_trans->dirty_bgs);
3505                                         btrfs_get_block_group(cache);
3506                                 }
3507                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3508                         } else if (ret) {
3509                                 btrfs_abort_transaction(trans, root, ret);
3510                         }
3511                 }
3512
3513                 /* if its not on the io list, we need to put the block group */
3514                 if (should_put)
3515                         btrfs_put_block_group(cache);
3516
3517                 if (ret)
3518                         break;
3519
3520                 /*
3521                  * Avoid blocking other tasks for too long. It might even save
3522                  * us from writing caches for block groups that are going to be
3523                  * removed.
3524                  */
3525                 mutex_unlock(&trans->transaction->cache_write_mutex);
3526                 mutex_lock(&trans->transaction->cache_write_mutex);
3527         }
3528         mutex_unlock(&trans->transaction->cache_write_mutex);
3529
3530         /*
3531          * go through delayed refs for all the stuff we've just kicked off
3532          * and then loop back (just once)
3533          */
3534         ret = btrfs_run_delayed_refs(trans, root, 0);
3535         if (!ret && loops == 0) {
3536                 loops++;
3537                 spin_lock(&cur_trans->dirty_bgs_lock);
3538                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3539                 /*
3540                  * dirty_bgs_lock protects us from concurrent block group
3541                  * deletes too (not just cache_write_mutex).
3542                  */
3543                 if (!list_empty(&dirty)) {
3544                         spin_unlock(&cur_trans->dirty_bgs_lock);
3545                         goto again;
3546                 }
3547                 spin_unlock(&cur_trans->dirty_bgs_lock);
3548         }
3549
3550         btrfs_free_path(path);
3551         return ret;
3552 }
3553
3554 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3555                                    struct btrfs_root *root)
3556 {
3557         struct btrfs_block_group_cache *cache;
3558         struct btrfs_transaction *cur_trans = trans->transaction;
3559         int ret = 0;
3560         int should_put;
3561         struct btrfs_path *path;
3562         struct list_head *io = &cur_trans->io_bgs;
3563         int num_started = 0;
3564
3565         path = btrfs_alloc_path();
3566         if (!path)
3567                 return -ENOMEM;
3568
3569         /*
3570          * We don't need the lock here since we are protected by the transaction
3571          * commit.  We want to do the cache_save_setup first and then run the
3572          * delayed refs to make sure we have the best chance at doing this all
3573          * in one shot.
3574          */
3575         while (!list_empty(&cur_trans->dirty_bgs)) {
3576                 cache = list_first_entry(&cur_trans->dirty_bgs,
3577                                          struct btrfs_block_group_cache,
3578                                          dirty_list);
3579
3580                 /*
3581                  * this can happen if cache_save_setup re-dirties a block
3582                  * group that is already under IO.  Just wait for it to
3583                  * finish and then do it all again
3584                  */
3585                 if (!list_empty(&cache->io_list)) {
3586                         list_del_init(&cache->io_list);
3587                         btrfs_wait_cache_io(root, trans, cache,
3588                                             &cache->io_ctl, path,
3589                                             cache->key.objectid);
3590                         btrfs_put_block_group(cache);
3591                 }
3592
3593                 /*
3594                  * don't remove from the dirty list until after we've waited
3595                  * on any pending IO
3596                  */
3597                 list_del_init(&cache->dirty_list);
3598                 should_put = 1;
3599
3600                 cache_save_setup(cache, trans, path);
3601
3602                 if (!ret)
3603                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3604
3605                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3606                         cache->io_ctl.inode = NULL;
3607                         ret = btrfs_write_out_cache(root, trans, cache, path);
3608                         if (ret == 0 && cache->io_ctl.inode) {
3609                                 num_started++;
3610                                 should_put = 0;
3611                                 list_add_tail(&cache->io_list, io);
3612                         } else {
3613                                 /*
3614                                  * if we failed to write the cache, the
3615                                  * generation will be bad and life goes on
3616                                  */
3617                                 ret = 0;
3618                         }
3619                 }
3620                 if (!ret) {
3621                         ret = write_one_cache_group(trans, root, path, cache);
3622                         if (ret)
3623                                 btrfs_abort_transaction(trans, root, ret);
3624                 }
3625
3626                 /* if its not on the io list, we need to put the block group */
3627                 if (should_put)
3628                         btrfs_put_block_group(cache);
3629         }
3630
3631         while (!list_empty(io)) {
3632                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3633                                          io_list);
3634                 list_del_init(&cache->io_list);
3635                 btrfs_wait_cache_io(root, trans, cache,
3636                                     &cache->io_ctl, path, cache->key.objectid);
3637                 btrfs_put_block_group(cache);
3638         }
3639
3640         btrfs_free_path(path);
3641         return ret;
3642 }
3643
3644 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3645 {
3646         struct btrfs_block_group_cache *block_group;
3647         int readonly = 0;
3648
3649         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3650         if (!block_group || block_group->ro)
3651                 readonly = 1;
3652         if (block_group)
3653                 btrfs_put_block_group(block_group);
3654         return readonly;
3655 }
3656
3657 static const char *alloc_name(u64 flags)
3658 {
3659         switch (flags) {
3660         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3661                 return "mixed";
3662         case BTRFS_BLOCK_GROUP_METADATA:
3663                 return "metadata";
3664         case BTRFS_BLOCK_GROUP_DATA:
3665                 return "data";
3666         case BTRFS_BLOCK_GROUP_SYSTEM:
3667                 return "system";
3668         default:
3669                 WARN_ON(1);
3670                 return "invalid-combination";
3671         };
3672 }
3673
3674 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3675                              u64 total_bytes, u64 bytes_used,
3676                              struct btrfs_space_info **space_info)
3677 {
3678         struct btrfs_space_info *found;
3679         int i;
3680         int factor;
3681         int ret;
3682
3683         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3684                      BTRFS_BLOCK_GROUP_RAID10))
3685                 factor = 2;
3686         else
3687                 factor = 1;
3688
3689         found = __find_space_info(info, flags);
3690         if (found) {
3691                 spin_lock(&found->lock);
3692                 found->total_bytes += total_bytes;
3693                 found->disk_total += total_bytes * factor;
3694                 found->bytes_used += bytes_used;
3695                 found->disk_used += bytes_used * factor;
3696                 found->full = 0;
3697                 spin_unlock(&found->lock);
3698                 *space_info = found;
3699                 return 0;
3700         }
3701         found = kzalloc(sizeof(*found), GFP_NOFS);
3702         if (!found)
3703                 return -ENOMEM;
3704
3705         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3706         if (ret) {
3707                 kfree(found);
3708                 return ret;
3709         }
3710
3711         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3712                 INIT_LIST_HEAD(&found->block_groups[i]);
3713         init_rwsem(&found->groups_sem);
3714         spin_lock_init(&found->lock);
3715         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3716         found->total_bytes = total_bytes;
3717         found->disk_total = total_bytes * factor;
3718         found->bytes_used = bytes_used;
3719         found->disk_used = bytes_used * factor;
3720         found->bytes_pinned = 0;
3721         found->bytes_reserved = 0;
3722         found->bytes_readonly = 0;
3723         found->bytes_may_use = 0;
3724         found->full = 0;
3725         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3726         found->chunk_alloc = 0;
3727         found->flush = 0;
3728         init_waitqueue_head(&found->wait);
3729         INIT_LIST_HEAD(&found->ro_bgs);
3730
3731         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3732                                     info->space_info_kobj, "%s",
3733                                     alloc_name(found->flags));
3734         if (ret) {
3735                 kfree(found);
3736                 return ret;
3737         }
3738
3739         *space_info = found;
3740         list_add_rcu(&found->list, &info->space_info);
3741         if (flags & BTRFS_BLOCK_GROUP_DATA)
3742                 info->data_sinfo = found;
3743
3744         return ret;
3745 }
3746
3747 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3748 {
3749         u64 extra_flags = chunk_to_extended(flags) &
3750                                 BTRFS_EXTENDED_PROFILE_MASK;
3751
3752         write_seqlock(&fs_info->profiles_lock);
3753         if (flags & BTRFS_BLOCK_GROUP_DATA)
3754                 fs_info->avail_data_alloc_bits |= extra_flags;
3755         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3756                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3757         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3758                 fs_info->avail_system_alloc_bits |= extra_flags;
3759         write_sequnlock(&fs_info->profiles_lock);
3760 }
3761
3762 /*
3763  * returns target flags in extended format or 0 if restripe for this
3764  * chunk_type is not in progress
3765  *
3766  * should be called with either volume_mutex or balance_lock held
3767  */
3768 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3769 {
3770         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3771         u64 target = 0;
3772
3773         if (!bctl)
3774                 return 0;
3775
3776         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3777             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3778                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3779         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3780                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3781                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3782         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3783                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3784                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3785         }
3786
3787         return target;
3788 }
3789
3790 /*
3791  * @flags: available profiles in extended format (see ctree.h)
3792  *
3793  * Returns reduced profile in chunk format.  If profile changing is in
3794  * progress (either running or paused) picks the target profile (if it's
3795  * already available), otherwise falls back to plain reducing.
3796  */
3797 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3798 {
3799         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3800         u64 target;
3801         u64 tmp;
3802
3803         /*
3804          * see if restripe for this chunk_type is in progress, if so
3805          * try to reduce to the target profile
3806          */
3807         spin_lock(&root->fs_info->balance_lock);
3808         target = get_restripe_target(root->fs_info, flags);
3809         if (target) {
3810                 /* pick target profile only if it's already available */
3811                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3812                         spin_unlock(&root->fs_info->balance_lock);
3813                         return extended_to_chunk(target);
3814                 }
3815         }
3816         spin_unlock(&root->fs_info->balance_lock);
3817
3818         /* First, mask out the RAID levels which aren't possible */
3819         if (num_devices == 1)
3820                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3821                            BTRFS_BLOCK_GROUP_RAID5);
3822         if (num_devices < 3)
3823                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3824         if (num_devices < 4)
3825                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3826
3827         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3828                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3829                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3830         flags &= ~tmp;
3831
3832         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3833                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3834         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3835                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3836         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3837                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3838         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3839                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3840         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3841                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3842
3843         return extended_to_chunk(flags | tmp);
3844 }
3845
3846 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3847 {
3848         unsigned seq;
3849         u64 flags;
3850
3851         do {
3852                 flags = orig_flags;
3853                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3854
3855                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3856                         flags |= root->fs_info->avail_data_alloc_bits;
3857                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3858                         flags |= root->fs_info->avail_system_alloc_bits;
3859                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3860                         flags |= root->fs_info->avail_metadata_alloc_bits;
3861         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3862
3863         return btrfs_reduce_alloc_profile(root, flags);
3864 }
3865
3866 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3867 {
3868         u64 flags;
3869         u64 ret;
3870
3871         if (data)
3872                 flags = BTRFS_BLOCK_GROUP_DATA;
3873         else if (root == root->fs_info->chunk_root)
3874                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3875         else
3876                 flags = BTRFS_BLOCK_GROUP_METADATA;
3877
3878         ret = get_alloc_profile(root, flags);
3879         return ret;
3880 }
3881
3882 /*
3883  * This will check the space that the inode allocates from to make sure we have
3884  * enough space for bytes.
3885  */
3886 int btrfs_check_data_free_space(struct inode *inode, u64 bytes, u64 write_bytes)
3887 {
3888         struct btrfs_space_info *data_sinfo;
3889         struct btrfs_root *root = BTRFS_I(inode)->root;
3890         struct btrfs_fs_info *fs_info = root->fs_info;
3891         u64 used;
3892         int ret = 0;
3893         int need_commit = 2;
3894         int have_pinned_space;
3895
3896         /* make sure bytes are sectorsize aligned */
3897         bytes = ALIGN(bytes, root->sectorsize);
3898
3899         if (btrfs_is_free_space_inode(inode)) {
3900                 need_commit = 0;
3901                 ASSERT(current->journal_info);
3902         }
3903
3904         data_sinfo = fs_info->data_sinfo;
3905         if (!data_sinfo)
3906                 goto alloc;
3907
3908 again:
3909         /* make sure we have enough space to handle the data first */
3910         spin_lock(&data_sinfo->lock);
3911         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3912                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3913                 data_sinfo->bytes_may_use;
3914
3915         if (used + bytes > data_sinfo->total_bytes) {
3916                 struct btrfs_trans_handle *trans;
3917
3918                 /*
3919                  * if we don't have enough free bytes in this space then we need
3920                  * to alloc a new chunk.
3921                  */
3922                 if (!data_sinfo->full) {
3923                         u64 alloc_target;
3924
3925                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3926                         spin_unlock(&data_sinfo->lock);
3927 alloc:
3928                         alloc_target = btrfs_get_alloc_profile(root, 1);
3929                         /*
3930                          * It is ugly that we don't call nolock join
3931                          * transaction for the free space inode case here.
3932                          * But it is safe because we only do the data space
3933                          * reservation for the free space cache in the
3934                          * transaction context, the common join transaction
3935                          * just increase the counter of the current transaction
3936                          * handler, doesn't try to acquire the trans_lock of
3937                          * the fs.
3938                          */
3939                         trans = btrfs_join_transaction(root);
3940                         if (IS_ERR(trans))
3941                                 return PTR_ERR(trans);
3942
3943                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3944                                              alloc_target,
3945                                              CHUNK_ALLOC_NO_FORCE);
3946                         btrfs_end_transaction(trans, root);
3947                         if (ret < 0) {
3948                                 if (ret != -ENOSPC)
3949                                         return ret;
3950                                 else {
3951                                         have_pinned_space = 1;
3952                                         goto commit_trans;
3953                                 }
3954                         }
3955
3956                         if (!data_sinfo)
3957                                 data_sinfo = fs_info->data_sinfo;
3958
3959                         goto again;
3960                 }
3961
3962                 /*
3963                  * If we don't have enough pinned space to deal with this
3964                  * allocation, and no removed chunk in current transaction,
3965                  * don't bother committing the transaction.
3966                  */
3967                 have_pinned_space = percpu_counter_compare(
3968                         &data_sinfo->total_bytes_pinned,
3969                         used + bytes - data_sinfo->total_bytes);
3970                 spin_unlock(&data_sinfo->lock);
3971
3972                 /* commit the current transaction and try again */
3973 commit_trans:
3974                 if (need_commit &&
3975                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3976                         need_commit--;
3977
3978                         trans = btrfs_join_transaction(root);
3979                         if (IS_ERR(trans))
3980                                 return PTR_ERR(trans);
3981                         if (have_pinned_space >= 0 ||
3982                             trans->transaction->have_free_bgs ||
3983                             need_commit > 0) {
3984                                 ret = btrfs_commit_transaction(trans, root);
3985                                 if (ret)
3986                                         return ret;
3987                                 /*
3988                                  * make sure that all running delayed iput are
3989                                  * done
3990                                  */
3991                                 down_write(&root->fs_info->delayed_iput_sem);
3992                                 up_write(&root->fs_info->delayed_iput_sem);
3993                                 goto again;
3994                         } else {
3995                                 btrfs_end_transaction(trans, root);
3996                         }
3997                 }
3998
3999                 trace_btrfs_space_reservation(root->fs_info,
4000                                               "space_info:enospc",
4001                                               data_sinfo->flags, bytes, 1);
4002                 return -ENOSPC;
4003         }
4004         ret = btrfs_qgroup_reserve(root, write_bytes);
4005         if (ret)
4006                 goto out;
4007         data_sinfo->bytes_may_use += bytes;
4008         trace_btrfs_space_reservation(root->fs_info, "space_info",
4009                                       data_sinfo->flags, bytes, 1);
4010 out:
4011         spin_unlock(&data_sinfo->lock);
4012
4013         return ret;
4014 }
4015
4016 /*
4017  * Called if we need to clear a data reservation for this inode.
4018  */
4019 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
4020 {
4021         struct btrfs_root *root = BTRFS_I(inode)->root;
4022         struct btrfs_space_info *data_sinfo;
4023
4024         /* make sure bytes are sectorsize aligned */
4025         bytes = ALIGN(bytes, root->sectorsize);
4026
4027         data_sinfo = root->fs_info->data_sinfo;
4028         spin_lock(&data_sinfo->lock);
4029         WARN_ON(data_sinfo->bytes_may_use < bytes);
4030         data_sinfo->bytes_may_use -= bytes;
4031         trace_btrfs_space_reservation(root->fs_info, "space_info",
4032                                       data_sinfo->flags, bytes, 0);
4033         spin_unlock(&data_sinfo->lock);
4034 }
4035
4036 static void force_metadata_allocation(struct btrfs_fs_info *info)
4037 {
4038         struct list_head *head = &info->space_info;
4039         struct btrfs_space_info *found;
4040
4041         rcu_read_lock();
4042         list_for_each_entry_rcu(found, head, list) {
4043                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4044                         found->force_alloc = CHUNK_ALLOC_FORCE;
4045         }
4046         rcu_read_unlock();
4047 }
4048
4049 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4050 {
4051         return (global->size << 1);
4052 }
4053
4054 static int should_alloc_chunk(struct btrfs_root *root,
4055                               struct btrfs_space_info *sinfo, int force)
4056 {
4057         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4058         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4059         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4060         u64 thresh;
4061
4062         if (force == CHUNK_ALLOC_FORCE)
4063                 return 1;
4064
4065         /*
4066          * We need to take into account the global rsv because for all intents
4067          * and purposes it's used space.  Don't worry about locking the
4068          * global_rsv, it doesn't change except when the transaction commits.
4069          */
4070         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4071                 num_allocated += calc_global_rsv_need_space(global_rsv);
4072
4073         /*
4074          * in limited mode, we want to have some free space up to
4075          * about 1% of the FS size.
4076          */
4077         if (force == CHUNK_ALLOC_LIMITED) {
4078                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4079                 thresh = max_t(u64, 64 * 1024 * 1024,
4080                                div_factor_fine(thresh, 1));
4081
4082                 if (num_bytes - num_allocated < thresh)
4083                         return 1;
4084         }
4085
4086         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
4087                 return 0;
4088         return 1;
4089 }
4090
4091 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
4092 {
4093         u64 num_dev;
4094
4095         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4096                     BTRFS_BLOCK_GROUP_RAID0 |
4097                     BTRFS_BLOCK_GROUP_RAID5 |
4098                     BTRFS_BLOCK_GROUP_RAID6))
4099                 num_dev = root->fs_info->fs_devices->rw_devices;
4100         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4101                 num_dev = 2;
4102         else
4103                 num_dev = 1;    /* DUP or single */
4104
4105         /* metadata for updaing devices and chunk tree */
4106         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
4107 }
4108
4109 static void check_system_chunk(struct btrfs_trans_handle *trans,
4110                                struct btrfs_root *root, u64 type)
4111 {
4112         struct btrfs_space_info *info;
4113         u64 left;
4114         u64 thresh;
4115
4116         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4117         spin_lock(&info->lock);
4118         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4119                 info->bytes_reserved - info->bytes_readonly;
4120         spin_unlock(&info->lock);
4121
4122         thresh = get_system_chunk_thresh(root, type);
4123         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4124                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4125                         left, thresh, type);
4126                 dump_space_info(info, 0, 0);
4127         }
4128
4129         if (left < thresh) {
4130                 u64 flags;
4131
4132                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4133                 btrfs_alloc_chunk(trans, root, flags);
4134         }
4135 }
4136
4137 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4138                           struct btrfs_root *extent_root, u64 flags, int force)
4139 {
4140         struct btrfs_space_info *space_info;
4141         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4142         int wait_for_alloc = 0;
4143         int ret = 0;
4144
4145         /* Don't re-enter if we're already allocating a chunk */
4146         if (trans->allocating_chunk)
4147                 return -ENOSPC;
4148
4149         space_info = __find_space_info(extent_root->fs_info, flags);
4150         if (!space_info) {
4151                 ret = update_space_info(extent_root->fs_info, flags,
4152                                         0, 0, &space_info);
4153                 BUG_ON(ret); /* -ENOMEM */
4154         }
4155         BUG_ON(!space_info); /* Logic error */
4156
4157 again:
4158         spin_lock(&space_info->lock);
4159         if (force < space_info->force_alloc)
4160                 force = space_info->force_alloc;
4161         if (space_info->full) {
4162                 if (should_alloc_chunk(extent_root, space_info, force))
4163                         ret = -ENOSPC;
4164                 else
4165                         ret = 0;
4166                 spin_unlock(&space_info->lock);
4167                 return ret;
4168         }
4169
4170         if (!should_alloc_chunk(extent_root, space_info, force)) {
4171                 spin_unlock(&space_info->lock);
4172                 return 0;
4173         } else if (space_info->chunk_alloc) {
4174                 wait_for_alloc = 1;
4175         } else {
4176                 space_info->chunk_alloc = 1;
4177         }
4178
4179         spin_unlock(&space_info->lock);
4180
4181         mutex_lock(&fs_info->chunk_mutex);
4182
4183         /*
4184          * The chunk_mutex is held throughout the entirety of a chunk
4185          * allocation, so once we've acquired the chunk_mutex we know that the
4186          * other guy is done and we need to recheck and see if we should
4187          * allocate.
4188          */
4189         if (wait_for_alloc) {
4190                 mutex_unlock(&fs_info->chunk_mutex);
4191                 wait_for_alloc = 0;
4192                 goto again;
4193         }
4194
4195         trans->allocating_chunk = true;
4196
4197         /*
4198          * If we have mixed data/metadata chunks we want to make sure we keep
4199          * allocating mixed chunks instead of individual chunks.
4200          */
4201         if (btrfs_mixed_space_info(space_info))
4202                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4203
4204         /*
4205          * if we're doing a data chunk, go ahead and make sure that
4206          * we keep a reasonable number of metadata chunks allocated in the
4207          * FS as well.
4208          */
4209         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4210                 fs_info->data_chunk_allocations++;
4211                 if (!(fs_info->data_chunk_allocations %
4212                       fs_info->metadata_ratio))
4213                         force_metadata_allocation(fs_info);
4214         }
4215
4216         /*
4217          * Check if we have enough space in SYSTEM chunk because we may need
4218          * to update devices.
4219          */
4220         check_system_chunk(trans, extent_root, flags);
4221
4222         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4223         trans->allocating_chunk = false;
4224
4225         spin_lock(&space_info->lock);
4226         if (ret < 0 && ret != -ENOSPC)
4227                 goto out;
4228         if (ret)
4229                 space_info->full = 1;
4230         else
4231                 ret = 1;
4232
4233         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4234 out:
4235         space_info->chunk_alloc = 0;
4236         spin_unlock(&space_info->lock);
4237         mutex_unlock(&fs_info->chunk_mutex);
4238         return ret;
4239 }
4240
4241 static int can_overcommit(struct btrfs_root *root,
4242                           struct btrfs_space_info *space_info, u64 bytes,
4243                           enum btrfs_reserve_flush_enum flush)
4244 {
4245         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4246         u64 profile = btrfs_get_alloc_profile(root, 0);
4247         u64 space_size;
4248         u64 avail;
4249         u64 used;
4250
4251         used = space_info->bytes_used + space_info->bytes_reserved +
4252                 space_info->bytes_pinned + space_info->bytes_readonly;
4253
4254         /*
4255          * We only want to allow over committing if we have lots of actual space
4256          * free, but if we don't have enough space to handle the global reserve
4257          * space then we could end up having a real enospc problem when trying
4258          * to allocate a chunk or some other such important allocation.
4259          */
4260         spin_lock(&global_rsv->lock);
4261         space_size = calc_global_rsv_need_space(global_rsv);
4262         spin_unlock(&global_rsv->lock);
4263         if (used + space_size >= space_info->total_bytes)
4264                 return 0;
4265
4266         used += space_info->bytes_may_use;
4267
4268         spin_lock(&root->fs_info->free_chunk_lock);
4269         avail = root->fs_info->free_chunk_space;
4270         spin_unlock(&root->fs_info->free_chunk_lock);
4271
4272         /*
4273          * If we have dup, raid1 or raid10 then only half of the free
4274          * space is actually useable.  For raid56, the space info used
4275          * doesn't include the parity drive, so we don't have to
4276          * change the math
4277          */
4278         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4279                        BTRFS_BLOCK_GROUP_RAID1 |
4280                        BTRFS_BLOCK_GROUP_RAID10))
4281                 avail >>= 1;
4282
4283         /*
4284          * If we aren't flushing all things, let us overcommit up to
4285          * 1/2th of the space. If we can flush, don't let us overcommit
4286          * too much, let it overcommit up to 1/8 of the space.
4287          */
4288         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4289                 avail >>= 3;
4290         else
4291                 avail >>= 1;
4292
4293         if (used + bytes < space_info->total_bytes + avail)
4294                 return 1;
4295         return 0;
4296 }
4297
4298 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4299                                          unsigned long nr_pages, int nr_items)
4300 {
4301         struct super_block *sb = root->fs_info->sb;
4302
4303         if (down_read_trylock(&sb->s_umount)) {
4304                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4305                 up_read(&sb->s_umount);
4306         } else {
4307                 /*
4308                  * We needn't worry the filesystem going from r/w to r/o though
4309                  * we don't acquire ->s_umount mutex, because the filesystem
4310                  * should guarantee the delalloc inodes list be empty after
4311                  * the filesystem is readonly(all dirty pages are written to
4312                  * the disk).
4313                  */
4314                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4315                 if (!current->journal_info)
4316                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4317         }
4318 }
4319
4320 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4321 {
4322         u64 bytes;
4323         int nr;
4324
4325         bytes = btrfs_calc_trans_metadata_size(root, 1);
4326         nr = (int)div64_u64(to_reclaim, bytes);
4327         if (!nr)
4328                 nr = 1;
4329         return nr;
4330 }
4331
4332 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4333
4334 /*
4335  * shrink metadata reservation for delalloc
4336  */
4337 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4338                             bool wait_ordered)
4339 {
4340         struct btrfs_block_rsv *block_rsv;
4341         struct btrfs_space_info *space_info;
4342         struct btrfs_trans_handle *trans;
4343         u64 delalloc_bytes;
4344         u64 max_reclaim;
4345         long time_left;
4346         unsigned long nr_pages;
4347         int loops;
4348         int items;
4349         enum btrfs_reserve_flush_enum flush;
4350
4351         /* Calc the number of the pages we need flush for space reservation */
4352         items = calc_reclaim_items_nr(root, to_reclaim);
4353         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4354
4355         trans = (struct btrfs_trans_handle *)current->journal_info;
4356         block_rsv = &root->fs_info->delalloc_block_rsv;
4357         space_info = block_rsv->space_info;
4358
4359         delalloc_bytes = percpu_counter_sum_positive(
4360                                                 &root->fs_info->delalloc_bytes);
4361         if (delalloc_bytes == 0) {
4362                 if (trans)
4363                         return;
4364                 if (wait_ordered)
4365                         btrfs_wait_ordered_roots(root->fs_info, items);
4366                 return;
4367         }
4368
4369         loops = 0;
4370         while (delalloc_bytes && loops < 3) {
4371                 max_reclaim = min(delalloc_bytes, to_reclaim);
4372                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4373                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4374                 /*
4375                  * We need to wait for the async pages to actually start before
4376                  * we do anything.
4377                  */
4378                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4379                 if (!max_reclaim)
4380                         goto skip_async;
4381
4382                 if (max_reclaim <= nr_pages)
4383                         max_reclaim = 0;
4384                 else
4385                         max_reclaim -= nr_pages;
4386
4387                 wait_event(root->fs_info->async_submit_wait,
4388                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4389                            (int)max_reclaim);
4390 skip_async:
4391                 if (!trans)
4392                         flush = BTRFS_RESERVE_FLUSH_ALL;
4393                 else
4394                         flush = BTRFS_RESERVE_NO_FLUSH;
4395                 spin_lock(&space_info->lock);
4396                 if (can_overcommit(root, space_info, orig, flush)) {
4397                         spin_unlock(&space_info->lock);
4398                         break;
4399                 }
4400                 spin_unlock(&space_info->lock);
4401
4402                 loops++;
4403                 if (wait_ordered && !trans) {
4404                         btrfs_wait_ordered_roots(root->fs_info, items);
4405                 } else {
4406                         time_left = schedule_timeout_killable(1);
4407                         if (time_left)
4408                                 break;
4409                 }
4410                 delalloc_bytes = percpu_counter_sum_positive(
4411                                                 &root->fs_info->delalloc_bytes);
4412         }
4413 }
4414
4415 /**
4416  * maybe_commit_transaction - possibly commit the transaction if its ok to
4417  * @root - the root we're allocating for
4418  * @bytes - the number of bytes we want to reserve
4419  * @force - force the commit
4420  *
4421  * This will check to make sure that committing the transaction will actually
4422  * get us somewhere and then commit the transaction if it does.  Otherwise it
4423  * will return -ENOSPC.
4424  */
4425 static int may_commit_transaction(struct btrfs_root *root,
4426                                   struct btrfs_space_info *space_info,
4427                                   u64 bytes, int force)
4428 {
4429         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4430         struct btrfs_trans_handle *trans;
4431
4432         trans = (struct btrfs_trans_handle *)current->journal_info;
4433         if (trans)
4434                 return -EAGAIN;
4435
4436         if (force)
4437                 goto commit;
4438
4439         /* See if there is enough pinned space to make this reservation */
4440         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4441                                    bytes) >= 0)
4442                 goto commit;
4443
4444         /*
4445          * See if there is some space in the delayed insertion reservation for
4446          * this reservation.
4447          */
4448         if (space_info != delayed_rsv->space_info)
4449                 return -ENOSPC;
4450
4451         spin_lock(&delayed_rsv->lock);
4452         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4453                                    bytes - delayed_rsv->size) >= 0) {
4454                 spin_unlock(&delayed_rsv->lock);
4455                 return -ENOSPC;
4456         }
4457         spin_unlock(&delayed_rsv->lock);
4458
4459 commit:
4460         trans = btrfs_join_transaction(root);
4461         if (IS_ERR(trans))
4462                 return -ENOSPC;
4463
4464         return btrfs_commit_transaction(trans, root);
4465 }
4466
4467 enum flush_state {
4468         FLUSH_DELAYED_ITEMS_NR  =       1,
4469         FLUSH_DELAYED_ITEMS     =       2,
4470         FLUSH_DELALLOC          =       3,
4471         FLUSH_DELALLOC_WAIT     =       4,
4472         ALLOC_CHUNK             =       5,
4473         COMMIT_TRANS            =       6,
4474 };
4475
4476 static int flush_space(struct btrfs_root *root,
4477                        struct btrfs_space_info *space_info, u64 num_bytes,
4478                        u64 orig_bytes, int state)
4479 {
4480         struct btrfs_trans_handle *trans;
4481         int nr;
4482         int ret = 0;
4483
4484         switch (state) {
4485         case FLUSH_DELAYED_ITEMS_NR:
4486         case FLUSH_DELAYED_ITEMS:
4487                 if (state == FLUSH_DELAYED_ITEMS_NR)
4488                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4489                 else
4490                         nr = -1;
4491
4492                 trans = btrfs_join_transaction(root);
4493                 if (IS_ERR(trans)) {
4494                         ret = PTR_ERR(trans);
4495                         break;
4496                 }
4497                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4498                 btrfs_end_transaction(trans, root);
4499                 break;
4500         case FLUSH_DELALLOC:
4501         case FLUSH_DELALLOC_WAIT:
4502                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4503                                 state == FLUSH_DELALLOC_WAIT);
4504                 break;
4505         case ALLOC_CHUNK:
4506                 trans = btrfs_join_transaction(root);
4507                 if (IS_ERR(trans)) {
4508                         ret = PTR_ERR(trans);
4509                         break;
4510                 }
4511                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4512                                      btrfs_get_alloc_profile(root, 0),
4513                                      CHUNK_ALLOC_NO_FORCE);
4514                 btrfs_end_transaction(trans, root);
4515                 if (ret == -ENOSPC)
4516                         ret = 0;
4517                 break;
4518         case COMMIT_TRANS:
4519                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4520                 break;
4521         default:
4522                 ret = -ENOSPC;
4523                 break;
4524         }
4525
4526         return ret;
4527 }
4528
4529 static inline u64
4530 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4531                                  struct btrfs_space_info *space_info)
4532 {
4533         u64 used;
4534         u64 expected;
4535         u64 to_reclaim;
4536
4537         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4538                                 16 * 1024 * 1024);
4539         spin_lock(&space_info->lock);
4540         if (can_overcommit(root, space_info, to_reclaim,
4541                            BTRFS_RESERVE_FLUSH_ALL)) {
4542                 to_reclaim = 0;
4543                 goto out;
4544         }
4545
4546         used = space_info->bytes_used + space_info->bytes_reserved +
4547                space_info->bytes_pinned + space_info->bytes_readonly +
4548                space_info->bytes_may_use;
4549         if (can_overcommit(root, space_info, 1024 * 1024,
4550                            BTRFS_RESERVE_FLUSH_ALL))
4551                 expected = div_factor_fine(space_info->total_bytes, 95);
4552         else
4553                 expected = div_factor_fine(space_info->total_bytes, 90);
4554
4555         if (used > expected)
4556                 to_reclaim = used - expected;
4557         else
4558                 to_reclaim = 0;
4559         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4560                                      space_info->bytes_reserved);
4561 out:
4562         spin_unlock(&space_info->lock);
4563
4564         return to_reclaim;
4565 }
4566
4567 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4568                                         struct btrfs_fs_info *fs_info, u64 used)
4569 {
4570         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4571
4572         /* If we're just plain full then async reclaim just slows us down. */
4573         if (space_info->bytes_used >= thresh)
4574                 return 0;
4575
4576         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4577                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4578 }
4579
4580 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4581                                        struct btrfs_fs_info *fs_info,
4582                                        int flush_state)
4583 {
4584         u64 used;
4585
4586         spin_lock(&space_info->lock);
4587         /*
4588          * We run out of space and have not got any free space via flush_space,
4589          * so don't bother doing async reclaim.
4590          */
4591         if (flush_state > COMMIT_TRANS && space_info->full) {
4592                 spin_unlock(&space_info->lock);
4593                 return 0;
4594         }
4595
4596         used = space_info->bytes_used + space_info->bytes_reserved +
4597                space_info->bytes_pinned + space_info->bytes_readonly +
4598                space_info->bytes_may_use;
4599         if (need_do_async_reclaim(space_info, fs_info, used)) {
4600                 spin_unlock(&space_info->lock);
4601                 return 1;
4602         }
4603         spin_unlock(&space_info->lock);
4604
4605         return 0;
4606 }
4607
4608 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4609 {
4610         struct btrfs_fs_info *fs_info;
4611         struct btrfs_space_info *space_info;
4612         u64 to_reclaim;
4613         int flush_state;
4614
4615         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4616         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4617
4618         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4619                                                       space_info);
4620         if (!to_reclaim)
4621                 return;
4622
4623         flush_state = FLUSH_DELAYED_ITEMS_NR;
4624         do {
4625                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4626                             to_reclaim, flush_state);
4627                 flush_state++;
4628                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4629                                                  flush_state))
4630                         return;
4631         } while (flush_state < COMMIT_TRANS);
4632 }
4633
4634 void btrfs_init_async_reclaim_work(struct work_struct *work)
4635 {
4636         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4637 }
4638
4639 /**
4640  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4641  * @root - the root we're allocating for
4642  * @block_rsv - the block_rsv we're allocating for
4643  * @orig_bytes - the number of bytes we want
4644  * @flush - whether or not we can flush to make our reservation
4645  *
4646  * This will reserve orgi_bytes number of bytes from the space info associated
4647  * with the block_rsv.  If there is not enough space it will make an attempt to
4648  * flush out space to make room.  It will do this by flushing delalloc if
4649  * possible or committing the transaction.  If flush is 0 then no attempts to
4650  * regain reservations will be made and this will fail if there is not enough
4651  * space already.
4652  */
4653 static int reserve_metadata_bytes(struct btrfs_root *root,
4654                                   struct btrfs_block_rsv *block_rsv,
4655                                   u64 orig_bytes,
4656                                   enum btrfs_reserve_flush_enum flush)
4657 {
4658         struct btrfs_space_info *space_info = block_rsv->space_info;
4659         u64 used;
4660         u64 num_bytes = orig_bytes;
4661         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4662         int ret = 0;
4663         bool flushing = false;
4664
4665 again:
4666         ret = 0;
4667         spin_lock(&space_info->lock);
4668         /*
4669          * We only want to wait if somebody other than us is flushing and we
4670          * are actually allowed to flush all things.
4671          */
4672         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4673                space_info->flush) {
4674                 spin_unlock(&space_info->lock);
4675                 /*
4676                  * If we have a trans handle we can't wait because the flusher
4677                  * may have to commit the transaction, which would mean we would
4678                  * deadlock since we are waiting for the flusher to finish, but
4679                  * hold the current transaction open.
4680                  */
4681                 if (current->journal_info)
4682                         return -EAGAIN;
4683                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4684                 /* Must have been killed, return */
4685                 if (ret)
4686                         return -EINTR;
4687
4688                 spin_lock(&space_info->lock);
4689         }
4690
4691         ret = -ENOSPC;
4692         used = space_info->bytes_used + space_info->bytes_reserved +
4693                 space_info->bytes_pinned + space_info->bytes_readonly +
4694                 space_info->bytes_may_use;
4695
4696         /*
4697          * The idea here is that we've not already over-reserved the block group
4698          * then we can go ahead and save our reservation first and then start
4699          * flushing if we need to.  Otherwise if we've already overcommitted
4700          * lets start flushing stuff first and then come back and try to make
4701          * our reservation.
4702          */
4703         if (used <= space_info->total_bytes) {
4704                 if (used + orig_bytes <= space_info->total_bytes) {
4705                         space_info->bytes_may_use += orig_bytes;
4706                         trace_btrfs_space_reservation(root->fs_info,
4707                                 "space_info", space_info->flags, orig_bytes, 1);
4708                         ret = 0;
4709                 } else {
4710                         /*
4711                          * Ok set num_bytes to orig_bytes since we aren't
4712                          * overocmmitted, this way we only try and reclaim what
4713                          * we need.
4714                          */
4715                         num_bytes = orig_bytes;
4716                 }
4717         } else {
4718                 /*
4719                  * Ok we're over committed, set num_bytes to the overcommitted
4720                  * amount plus the amount of bytes that we need for this
4721                  * reservation.
4722                  */
4723                 num_bytes = used - space_info->total_bytes +
4724                         (orig_bytes * 2);
4725         }
4726
4727         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4728                 space_info->bytes_may_use += orig_bytes;
4729                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4730                                               space_info->flags, orig_bytes,
4731                                               1);
4732                 ret = 0;
4733         }
4734
4735         /*
4736          * Couldn't make our reservation, save our place so while we're trying
4737          * to reclaim space we can actually use it instead of somebody else
4738          * stealing it from us.
4739          *
4740          * We make the other tasks wait for the flush only when we can flush
4741          * all things.
4742          */
4743         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4744                 flushing = true;
4745                 space_info->flush = 1;
4746         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4747                 used += orig_bytes;
4748                 /*
4749                  * We will do the space reservation dance during log replay,
4750                  * which means we won't have fs_info->fs_root set, so don't do
4751                  * the async reclaim as we will panic.
4752                  */
4753                 if (!root->fs_info->log_root_recovering &&
4754                     need_do_async_reclaim(space_info, root->fs_info, used) &&
4755                     !work_busy(&root->fs_info->async_reclaim_work))
4756                         queue_work(system_unbound_wq,
4757                                    &root->fs_info->async_reclaim_work);
4758         }
4759         spin_unlock(&space_info->lock);
4760
4761         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4762                 goto out;
4763
4764         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4765                           flush_state);
4766         flush_state++;
4767
4768         /*
4769          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4770          * would happen. So skip delalloc flush.
4771          */
4772         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4773             (flush_state == FLUSH_DELALLOC ||
4774              flush_state == FLUSH_DELALLOC_WAIT))
4775                 flush_state = ALLOC_CHUNK;
4776
4777         if (!ret)
4778                 goto again;
4779         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4780                  flush_state < COMMIT_TRANS)
4781                 goto again;
4782         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4783                  flush_state <= COMMIT_TRANS)
4784                 goto again;
4785
4786 out:
4787         if (ret == -ENOSPC &&
4788             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4789                 struct btrfs_block_rsv *global_rsv =
4790                         &root->fs_info->global_block_rsv;
4791
4792                 if (block_rsv != global_rsv &&
4793                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4794                         ret = 0;
4795         }
4796         if (ret == -ENOSPC)
4797                 trace_btrfs_space_reservation(root->fs_info,
4798                                               "space_info:enospc",
4799                                               space_info->flags, orig_bytes, 1);
4800         if (flushing) {
4801                 spin_lock(&space_info->lock);
4802                 space_info->flush = 0;
4803                 wake_up_all(&space_info->wait);
4804                 spin_unlock(&space_info->lock);
4805         }
4806         return ret;
4807 }
4808
4809 static struct btrfs_block_rsv *get_block_rsv(
4810                                         const struct btrfs_trans_handle *trans,
4811                                         const struct btrfs_root *root)
4812 {
4813         struct btrfs_block_rsv *block_rsv = NULL;
4814
4815         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4816                 block_rsv = trans->block_rsv;
4817
4818         if (root == root->fs_info->csum_root && trans->adding_csums)
4819                 block_rsv = trans->block_rsv;
4820
4821         if (root == root->fs_info->uuid_root)
4822                 block_rsv = trans->block_rsv;
4823
4824         if (!block_rsv)
4825                 block_rsv = root->block_rsv;
4826
4827         if (!block_rsv)
4828                 block_rsv = &root->fs_info->empty_block_rsv;
4829
4830         return block_rsv;
4831 }
4832
4833 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4834                                u64 num_bytes)
4835 {
4836         int ret = -ENOSPC;
4837         spin_lock(&block_rsv->lock);
4838         if (block_rsv->reserved >= num_bytes) {
4839                 block_rsv->reserved -= num_bytes;
4840                 if (block_rsv->reserved < block_rsv->size)
4841                         block_rsv->full = 0;
4842                 ret = 0;
4843         }
4844         spin_unlock(&block_rsv->lock);
4845         return ret;
4846 }
4847
4848 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4849                                 u64 num_bytes, int update_size)
4850 {
4851         spin_lock(&block_rsv->lock);
4852         block_rsv->reserved += num_bytes;
4853         if (update_size)
4854                 block_rsv->size += num_bytes;
4855         else if (block_rsv->reserved >= block_rsv->size)
4856                 block_rsv->full = 1;
4857         spin_unlock(&block_rsv->lock);
4858 }
4859
4860 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4861                              struct btrfs_block_rsv *dest, u64 num_bytes,
4862                              int min_factor)
4863 {
4864         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4865         u64 min_bytes;
4866
4867         if (global_rsv->space_info != dest->space_info)
4868                 return -ENOSPC;
4869
4870         spin_lock(&global_rsv->lock);
4871         min_bytes = div_factor(global_rsv->size, min_factor);
4872         if (global_rsv->reserved < min_bytes + num_bytes) {
4873                 spin_unlock(&global_rsv->lock);
4874                 return -ENOSPC;
4875         }
4876         global_rsv->reserved -= num_bytes;
4877         if (global_rsv->reserved < global_rsv->size)
4878                 global_rsv->full = 0;
4879         spin_unlock(&global_rsv->lock);
4880
4881         block_rsv_add_bytes(dest, num_bytes, 1);
4882         return 0;
4883 }
4884
4885 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4886                                     struct btrfs_block_rsv *block_rsv,
4887                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4888 {
4889         struct btrfs_space_info *space_info = block_rsv->space_info;
4890
4891         spin_lock(&block_rsv->lock);
4892         if (num_bytes == (u64)-1)
4893                 num_bytes = block_rsv->size;
4894         block_rsv->size -= num_bytes;
4895         if (block_rsv->reserved >= block_rsv->size) {
4896                 num_bytes = block_rsv->reserved - block_rsv->size;
4897                 block_rsv->reserved = block_rsv->size;
4898                 block_rsv->full = 1;
4899         } else {
4900                 num_bytes = 0;
4901         }
4902         spin_unlock(&block_rsv->lock);
4903
4904         if (num_bytes > 0) {
4905                 if (dest) {
4906                         spin_lock(&dest->lock);
4907                         if (!dest->full) {
4908                                 u64 bytes_to_add;
4909
4910                                 bytes_to_add = dest->size - dest->reserved;
4911                                 bytes_to_add = min(num_bytes, bytes_to_add);
4912                                 dest->reserved += bytes_to_add;
4913                                 if (dest->reserved >= dest->size)
4914                                         dest->full = 1;
4915                                 num_bytes -= bytes_to_add;
4916                         }
4917                         spin_unlock(&dest->lock);
4918                 }
4919                 if (num_bytes) {
4920                         spin_lock(&space_info->lock);
4921                         space_info->bytes_may_use -= num_bytes;
4922                         trace_btrfs_space_reservation(fs_info, "space_info",
4923                                         space_info->flags, num_bytes, 0);
4924                         spin_unlock(&space_info->lock);
4925                 }
4926         }
4927 }
4928
4929 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4930                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4931 {
4932         int ret;
4933
4934         ret = block_rsv_use_bytes(src, num_bytes);
4935         if (ret)
4936                 return ret;
4937
4938         block_rsv_add_bytes(dst, num_bytes, 1);
4939         return 0;
4940 }
4941
4942 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4943 {
4944         memset(rsv, 0, sizeof(*rsv));
4945         spin_lock_init(&rsv->lock);
4946         rsv->type = type;
4947 }
4948
4949 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4950                                               unsigned short type)
4951 {
4952         struct btrfs_block_rsv *block_rsv;
4953         struct btrfs_fs_info *fs_info = root->fs_info;
4954
4955         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4956         if (!block_rsv)
4957                 return NULL;
4958
4959         btrfs_init_block_rsv(block_rsv, type);
4960         block_rsv->space_info = __find_space_info(fs_info,
4961                                                   BTRFS_BLOCK_GROUP_METADATA);
4962         return block_rsv;
4963 }
4964
4965 void btrfs_free_block_rsv(struct btrfs_root *root,
4966                           struct btrfs_block_rsv *rsv)
4967 {
4968         if (!rsv)
4969                 return;
4970         btrfs_block_rsv_release(root, rsv, (u64)-1);
4971         kfree(rsv);
4972 }
4973
4974 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
4975 {
4976         kfree(rsv);
4977 }
4978
4979 int btrfs_block_rsv_add(struct btrfs_root *root,
4980                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4981                         enum btrfs_reserve_flush_enum flush)
4982 {
4983         int ret;
4984
4985         if (num_bytes == 0)
4986                 return 0;
4987
4988         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4989         if (!ret) {
4990                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4991                 return 0;
4992         }
4993
4994         return ret;
4995 }
4996
4997 int btrfs_block_rsv_check(struct btrfs_root *root,
4998                           struct btrfs_block_rsv *block_rsv, int min_factor)
4999 {
5000         u64 num_bytes = 0;
5001         int ret = -ENOSPC;
5002
5003         if (!block_rsv)
5004                 return 0;
5005
5006         spin_lock(&block_rsv->lock);
5007         num_bytes = div_factor(block_rsv->size, min_factor);
5008         if (block_rsv->reserved >= num_bytes)
5009                 ret = 0;
5010         spin_unlock(&block_rsv->lock);
5011
5012         return ret;
5013 }
5014
5015 int btrfs_block_rsv_refill(struct btrfs_root *root,
5016                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5017                            enum btrfs_reserve_flush_enum flush)
5018 {
5019         u64 num_bytes = 0;
5020         int ret = -ENOSPC;
5021
5022         if (!block_rsv)
5023                 return 0;
5024
5025         spin_lock(&block_rsv->lock);
5026         num_bytes = min_reserved;
5027         if (block_rsv->reserved >= num_bytes)
5028                 ret = 0;
5029         else
5030                 num_bytes -= block_rsv->reserved;
5031         spin_unlock(&block_rsv->lock);
5032
5033         if (!ret)
5034                 return 0;
5035
5036         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5037         if (!ret) {
5038                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5039                 return 0;
5040         }
5041
5042         return ret;
5043 }
5044
5045 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5046                             struct btrfs_block_rsv *dst_rsv,
5047                             u64 num_bytes)
5048 {
5049         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5050 }
5051
5052 void btrfs_block_rsv_release(struct btrfs_root *root,
5053                              struct btrfs_block_rsv *block_rsv,
5054                              u64 num_bytes)
5055 {
5056         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5057         if (global_rsv == block_rsv ||
5058             block_rsv->space_info != global_rsv->space_info)
5059                 global_rsv = NULL;
5060         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5061                                 num_bytes);
5062 }
5063
5064 /*
5065  * helper to calculate size of global block reservation.
5066  * the desired value is sum of space used by extent tree,
5067  * checksum tree and root tree
5068  */
5069 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5070 {
5071         struct btrfs_space_info *sinfo;
5072         u64 num_bytes;
5073         u64 meta_used;
5074         u64 data_used;
5075         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5076
5077         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5078         spin_lock(&sinfo->lock);
5079         data_used = sinfo->bytes_used;
5080         spin_unlock(&sinfo->lock);
5081
5082         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5083         spin_lock(&sinfo->lock);
5084         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5085                 data_used = 0;
5086         meta_used = sinfo->bytes_used;
5087         spin_unlock(&sinfo->lock);
5088
5089         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5090                     csum_size * 2;
5091         num_bytes += div_u64(data_used + meta_used, 50);
5092
5093         if (num_bytes * 3 > meta_used)
5094                 num_bytes = div_u64(meta_used, 3);
5095
5096         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5097 }
5098
5099 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5100 {
5101         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5102         struct btrfs_space_info *sinfo = block_rsv->space_info;
5103         u64 num_bytes;
5104
5105         num_bytes = calc_global_metadata_size(fs_info);
5106
5107         spin_lock(&sinfo->lock);
5108         spin_lock(&block_rsv->lock);
5109
5110         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
5111
5112         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5113                     sinfo->bytes_reserved + sinfo->bytes_readonly +
5114                     sinfo->bytes_may_use;
5115
5116         if (sinfo->total_bytes > num_bytes) {
5117                 num_bytes = sinfo->total_bytes - num_bytes;
5118                 block_rsv->reserved += num_bytes;
5119                 sinfo->bytes_may_use += num_bytes;
5120                 trace_btrfs_space_reservation(fs_info, "space_info",
5121                                       sinfo->flags, num_bytes, 1);
5122         }
5123
5124         if (block_rsv->reserved >= block_rsv->size) {
5125                 num_bytes = block_rsv->reserved - block_rsv->size;
5126                 sinfo->bytes_may_use -= num_bytes;
5127                 trace_btrfs_space_reservation(fs_info, "space_info",
5128                                       sinfo->flags, num_bytes, 0);
5129                 block_rsv->reserved = block_rsv->size;
5130                 block_rsv->full = 1;
5131         }
5132
5133         spin_unlock(&block_rsv->lock);
5134         spin_unlock(&sinfo->lock);
5135 }
5136
5137 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5138 {
5139         struct btrfs_space_info *space_info;
5140
5141         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5142         fs_info->chunk_block_rsv.space_info = space_info;
5143
5144         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5145         fs_info->global_block_rsv.space_info = space_info;
5146         fs_info->delalloc_block_rsv.space_info = space_info;
5147         fs_info->trans_block_rsv.space_info = space_info;
5148         fs_info->empty_block_rsv.space_info = space_info;
5149         fs_info->delayed_block_rsv.space_info = space_info;
5150
5151         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5152         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5153         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5154         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5155         if (fs_info->quota_root)
5156                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5157         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5158
5159         update_global_block_rsv(fs_info);
5160 }
5161
5162 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5163 {
5164         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5165                                 (u64)-1);
5166         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5167         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5168         WARN_ON(fs_info->trans_block_rsv.size > 0);
5169         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5170         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5171         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5172         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5173         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5174 }
5175
5176 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5177                                   struct btrfs_root *root)
5178 {
5179         if (!trans->block_rsv)
5180                 return;
5181
5182         if (!trans->bytes_reserved)
5183                 return;
5184
5185         trace_btrfs_space_reservation(root->fs_info, "transaction",
5186                                       trans->transid, trans->bytes_reserved, 0);
5187         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5188         trans->bytes_reserved = 0;
5189 }
5190
5191 /* Can only return 0 or -ENOSPC */
5192 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5193                                   struct inode *inode)
5194 {
5195         struct btrfs_root *root = BTRFS_I(inode)->root;
5196         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5197         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5198
5199         /*
5200          * We need to hold space in order to delete our orphan item once we've
5201          * added it, so this takes the reservation so we can release it later
5202          * when we are truly done with the orphan item.
5203          */
5204         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5205         trace_btrfs_space_reservation(root->fs_info, "orphan",
5206                                       btrfs_ino(inode), num_bytes, 1);
5207         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5208 }
5209
5210 void btrfs_orphan_release_metadata(struct inode *inode)
5211 {
5212         struct btrfs_root *root = BTRFS_I(inode)->root;
5213         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5214         trace_btrfs_space_reservation(root->fs_info, "orphan",
5215                                       btrfs_ino(inode), num_bytes, 0);
5216         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5217 }
5218
5219 /*
5220  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5221  * root: the root of the parent directory
5222  * rsv: block reservation
5223  * items: the number of items that we need do reservation
5224  * qgroup_reserved: used to return the reserved size in qgroup
5225  *
5226  * This function is used to reserve the space for snapshot/subvolume
5227  * creation and deletion. Those operations are different with the
5228  * common file/directory operations, they change two fs/file trees
5229  * and root tree, the number of items that the qgroup reserves is
5230  * different with the free space reservation. So we can not use
5231  * the space reseravtion mechanism in start_transaction().
5232  */
5233 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5234                                      struct btrfs_block_rsv *rsv,
5235                                      int items,
5236                                      u64 *qgroup_reserved,
5237                                      bool use_global_rsv)
5238 {
5239         u64 num_bytes;
5240         int ret;
5241         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5242
5243         if (root->fs_info->quota_enabled) {
5244                 /* One for parent inode, two for dir entries */
5245                 num_bytes = 3 * root->nodesize;
5246                 ret = btrfs_qgroup_reserve(root, num_bytes);
5247                 if (ret)
5248                         return ret;
5249         } else {
5250                 num_bytes = 0;
5251         }
5252
5253         *qgroup_reserved = num_bytes;
5254
5255         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5256         rsv->space_info = __find_space_info(root->fs_info,
5257                                             BTRFS_BLOCK_GROUP_METADATA);
5258         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5259                                   BTRFS_RESERVE_FLUSH_ALL);
5260
5261         if (ret == -ENOSPC && use_global_rsv)
5262                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5263
5264         if (ret) {
5265                 if (*qgroup_reserved)
5266                         btrfs_qgroup_free(root, *qgroup_reserved);
5267         }
5268
5269         return ret;
5270 }
5271
5272 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5273                                       struct btrfs_block_rsv *rsv,
5274                                       u64 qgroup_reserved)
5275 {
5276         btrfs_block_rsv_release(root, rsv, (u64)-1);
5277 }
5278
5279 /**
5280  * drop_outstanding_extent - drop an outstanding extent
5281  * @inode: the inode we're dropping the extent for
5282  * @num_bytes: the number of bytes we're relaseing.
5283  *
5284  * This is called when we are freeing up an outstanding extent, either called
5285  * after an error or after an extent is written.  This will return the number of
5286  * reserved extents that need to be freed.  This must be called with
5287  * BTRFS_I(inode)->lock held.
5288  */
5289 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5290 {
5291         unsigned drop_inode_space = 0;
5292         unsigned dropped_extents = 0;
5293         unsigned num_extents = 0;
5294
5295         num_extents = (unsigned)div64_u64(num_bytes +
5296                                           BTRFS_MAX_EXTENT_SIZE - 1,
5297                                           BTRFS_MAX_EXTENT_SIZE);
5298         ASSERT(num_extents);
5299         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5300         BTRFS_I(inode)->outstanding_extents -= num_extents;
5301
5302         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5303             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5304                                &BTRFS_I(inode)->runtime_flags))
5305                 drop_inode_space = 1;
5306
5307         /*
5308          * If we have more or the same amount of outsanding extents than we have
5309          * reserved then we need to leave the reserved extents count alone.
5310          */
5311         if (BTRFS_I(inode)->outstanding_extents >=
5312             BTRFS_I(inode)->reserved_extents)
5313                 return drop_inode_space;
5314
5315         dropped_extents = BTRFS_I(inode)->reserved_extents -
5316                 BTRFS_I(inode)->outstanding_extents;
5317         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5318         return dropped_extents + drop_inode_space;
5319 }
5320
5321 /**
5322  * calc_csum_metadata_size - return the amount of metada space that must be
5323  *      reserved/free'd for the given bytes.
5324  * @inode: the inode we're manipulating
5325  * @num_bytes: the number of bytes in question
5326  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5327  *
5328  * This adjusts the number of csum_bytes in the inode and then returns the
5329  * correct amount of metadata that must either be reserved or freed.  We
5330  * calculate how many checksums we can fit into one leaf and then divide the
5331  * number of bytes that will need to be checksumed by this value to figure out
5332  * how many checksums will be required.  If we are adding bytes then the number
5333  * may go up and we will return the number of additional bytes that must be
5334  * reserved.  If it is going down we will return the number of bytes that must
5335  * be freed.
5336  *
5337  * This must be called with BTRFS_I(inode)->lock held.
5338  */
5339 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5340                                    int reserve)
5341 {
5342         struct btrfs_root *root = BTRFS_I(inode)->root;
5343         u64 old_csums, num_csums;
5344
5345         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5346             BTRFS_I(inode)->csum_bytes == 0)
5347                 return 0;
5348
5349         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5350         if (reserve)
5351                 BTRFS_I(inode)->csum_bytes += num_bytes;
5352         else
5353                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5354         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5355
5356         /* No change, no need to reserve more */
5357         if (old_csums == num_csums)
5358                 return 0;
5359
5360         if (reserve)
5361                 return btrfs_calc_trans_metadata_size(root,
5362                                                       num_csums - old_csums);
5363
5364         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5365 }
5366
5367 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5368 {
5369         struct btrfs_root *root = BTRFS_I(inode)->root;
5370         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5371         u64 to_reserve = 0;
5372         u64 csum_bytes;
5373         unsigned nr_extents = 0;
5374         int extra_reserve = 0;
5375         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5376         int ret = 0;
5377         bool delalloc_lock = true;
5378         u64 to_free = 0;
5379         unsigned dropped;
5380
5381         /* If we are a free space inode we need to not flush since we will be in
5382          * the middle of a transaction commit.  We also don't need the delalloc
5383          * mutex since we won't race with anybody.  We need this mostly to make
5384          * lockdep shut its filthy mouth.
5385          */
5386         if (btrfs_is_free_space_inode(inode)) {
5387                 flush = BTRFS_RESERVE_NO_FLUSH;
5388                 delalloc_lock = false;
5389         }
5390
5391         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5392             btrfs_transaction_in_commit(root->fs_info))
5393                 schedule_timeout(1);
5394
5395         if (delalloc_lock)
5396                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5397
5398         num_bytes = ALIGN(num_bytes, root->sectorsize);
5399
5400         spin_lock(&BTRFS_I(inode)->lock);
5401         nr_extents = (unsigned)div64_u64(num_bytes +
5402                                          BTRFS_MAX_EXTENT_SIZE - 1,
5403                                          BTRFS_MAX_EXTENT_SIZE);
5404         BTRFS_I(inode)->outstanding_extents += nr_extents;
5405         nr_extents = 0;
5406
5407         if (BTRFS_I(inode)->outstanding_extents >
5408             BTRFS_I(inode)->reserved_extents)
5409                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5410                         BTRFS_I(inode)->reserved_extents;
5411
5412         /*
5413          * Add an item to reserve for updating the inode when we complete the
5414          * delalloc io.
5415          */
5416         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5417                       &BTRFS_I(inode)->runtime_flags)) {
5418                 nr_extents++;
5419                 extra_reserve = 1;
5420         }
5421
5422         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5423         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5424         csum_bytes = BTRFS_I(inode)->csum_bytes;
5425         spin_unlock(&BTRFS_I(inode)->lock);
5426
5427         if (root->fs_info->quota_enabled) {
5428                 ret = btrfs_qgroup_reserve(root, nr_extents * root->nodesize);
5429                 if (ret)
5430                         goto out_fail;
5431         }
5432
5433         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5434         if (unlikely(ret)) {
5435                 if (root->fs_info->quota_enabled)
5436                         btrfs_qgroup_free(root, nr_extents * root->nodesize);
5437                 goto out_fail;
5438         }
5439
5440         spin_lock(&BTRFS_I(inode)->lock);
5441         if (extra_reserve) {
5442                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5443                         &BTRFS_I(inode)->runtime_flags);
5444                 nr_extents--;
5445         }
5446         BTRFS_I(inode)->reserved_extents += nr_extents;
5447         spin_unlock(&BTRFS_I(inode)->lock);
5448
5449         if (delalloc_lock)
5450                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5451
5452         if (to_reserve)
5453                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5454                                               btrfs_ino(inode), to_reserve, 1);
5455         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5456
5457         return 0;
5458
5459 out_fail:
5460         spin_lock(&BTRFS_I(inode)->lock);
5461         dropped = drop_outstanding_extent(inode, num_bytes);
5462         /*
5463          * If the inodes csum_bytes is the same as the original
5464          * csum_bytes then we know we haven't raced with any free()ers
5465          * so we can just reduce our inodes csum bytes and carry on.
5466          */
5467         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5468                 calc_csum_metadata_size(inode, num_bytes, 0);
5469         } else {
5470                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5471                 u64 bytes;
5472
5473                 /*
5474                  * This is tricky, but first we need to figure out how much we
5475                  * free'd from any free-ers that occured during this
5476                  * reservation, so we reset ->csum_bytes to the csum_bytes
5477                  * before we dropped our lock, and then call the free for the
5478                  * number of bytes that were freed while we were trying our
5479                  * reservation.
5480                  */
5481                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5482                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5483                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5484
5485
5486                 /*
5487                  * Now we need to see how much we would have freed had we not
5488                  * been making this reservation and our ->csum_bytes were not
5489                  * artificially inflated.
5490                  */
5491                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5492                 bytes = csum_bytes - orig_csum_bytes;
5493                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5494
5495                 /*
5496                  * Now reset ->csum_bytes to what it should be.  If bytes is
5497                  * more than to_free then we would have free'd more space had we
5498                  * not had an artificially high ->csum_bytes, so we need to free
5499                  * the remainder.  If bytes is the same or less then we don't
5500                  * need to do anything, the other free-ers did the correct
5501                  * thing.
5502                  */
5503                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5504                 if (bytes > to_free)
5505                         to_free = bytes - to_free;
5506                 else
5507                         to_free = 0;
5508         }
5509         spin_unlock(&BTRFS_I(inode)->lock);
5510         if (dropped)
5511                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5512
5513         if (to_free) {
5514                 btrfs_block_rsv_release(root, block_rsv, to_free);
5515                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5516                                               btrfs_ino(inode), to_free, 0);
5517         }
5518         if (delalloc_lock)
5519                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5520         return ret;
5521 }
5522
5523 /**
5524  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5525  * @inode: the inode to release the reservation for
5526  * @num_bytes: the number of bytes we're releasing
5527  *
5528  * This will release the metadata reservation for an inode.  This can be called
5529  * once we complete IO for a given set of bytes to release their metadata
5530  * reservations.
5531  */
5532 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5533 {
5534         struct btrfs_root *root = BTRFS_I(inode)->root;
5535         u64 to_free = 0;
5536         unsigned dropped;
5537
5538         num_bytes = ALIGN(num_bytes, root->sectorsize);
5539         spin_lock(&BTRFS_I(inode)->lock);
5540         dropped = drop_outstanding_extent(inode, num_bytes);
5541
5542         if (num_bytes)
5543                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5544         spin_unlock(&BTRFS_I(inode)->lock);
5545         if (dropped > 0)
5546                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5547
5548         if (btrfs_test_is_dummy_root(root))
5549                 return;
5550
5551         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5552                                       btrfs_ino(inode), to_free, 0);
5553
5554         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5555                                 to_free);
5556 }
5557
5558 /**
5559  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5560  * @inode: inode we're writing to
5561  * @num_bytes: the number of bytes we want to allocate
5562  *
5563  * This will do the following things
5564  *
5565  * o reserve space in the data space info for num_bytes
5566  * o reserve space in the metadata space info based on number of outstanding
5567  *   extents and how much csums will be needed
5568  * o add to the inodes ->delalloc_bytes
5569  * o add it to the fs_info's delalloc inodes list.
5570  *
5571  * This will return 0 for success and -ENOSPC if there is no space left.
5572  */
5573 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5574 {
5575         int ret;
5576
5577         ret = btrfs_check_data_free_space(inode, num_bytes, num_bytes);
5578         if (ret)
5579                 return ret;
5580
5581         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5582         if (ret) {
5583                 btrfs_free_reserved_data_space(inode, num_bytes);
5584                 return ret;
5585         }
5586
5587         return 0;
5588 }
5589
5590 /**
5591  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5592  * @inode: inode we're releasing space for
5593  * @num_bytes: the number of bytes we want to free up
5594  *
5595  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5596  * called in the case that we don't need the metadata AND data reservations
5597  * anymore.  So if there is an error or we insert an inline extent.
5598  *
5599  * This function will release the metadata space that was not used and will
5600  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5601  * list if there are no delalloc bytes left.
5602  */
5603 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5604 {
5605         btrfs_delalloc_release_metadata(inode, num_bytes);
5606         btrfs_free_reserved_data_space(inode, num_bytes);
5607 }
5608
5609 static int update_block_group(struct btrfs_trans_handle *trans,
5610                               struct btrfs_root *root, u64 bytenr,
5611                               u64 num_bytes, int alloc)
5612 {
5613         struct btrfs_block_group_cache *cache = NULL;
5614         struct btrfs_fs_info *info = root->fs_info;
5615         u64 total = num_bytes;
5616         u64 old_val;
5617         u64 byte_in_group;
5618         int factor;
5619
5620         /* block accounting for super block */
5621         spin_lock(&info->delalloc_root_lock);
5622         old_val = btrfs_super_bytes_used(info->super_copy);
5623         if (alloc)
5624                 old_val += num_bytes;
5625         else
5626                 old_val -= num_bytes;
5627         btrfs_set_super_bytes_used(info->super_copy, old_val);
5628         spin_unlock(&info->delalloc_root_lock);
5629
5630         while (total) {
5631                 cache = btrfs_lookup_block_group(info, bytenr);
5632                 if (!cache)
5633                         return -ENOENT;
5634                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5635                                     BTRFS_BLOCK_GROUP_RAID1 |
5636                                     BTRFS_BLOCK_GROUP_RAID10))
5637                         factor = 2;
5638                 else
5639                         factor = 1;
5640                 /*
5641                  * If this block group has free space cache written out, we
5642                  * need to make sure to load it if we are removing space.  This
5643                  * is because we need the unpinning stage to actually add the
5644                  * space back to the block group, otherwise we will leak space.
5645                  */
5646                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5647                         cache_block_group(cache, 1);
5648
5649                 byte_in_group = bytenr - cache->key.objectid;
5650                 WARN_ON(byte_in_group > cache->key.offset);
5651
5652                 spin_lock(&cache->space_info->lock);
5653                 spin_lock(&cache->lock);
5654
5655                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5656                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5657                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5658
5659                 old_val = btrfs_block_group_used(&cache->item);
5660                 num_bytes = min(total, cache->key.offset - byte_in_group);
5661                 if (alloc) {
5662                         old_val += num_bytes;
5663                         btrfs_set_block_group_used(&cache->item, old_val);
5664                         cache->reserved -= num_bytes;
5665                         cache->space_info->bytes_reserved -= num_bytes;
5666                         cache->space_info->bytes_used += num_bytes;
5667                         cache->space_info->disk_used += num_bytes * factor;
5668                         spin_unlock(&cache->lock);
5669                         spin_unlock(&cache->space_info->lock);
5670                 } else {
5671                         old_val -= num_bytes;
5672                         btrfs_set_block_group_used(&cache->item, old_val);
5673                         cache->pinned += num_bytes;
5674                         cache->space_info->bytes_pinned += num_bytes;
5675                         cache->space_info->bytes_used -= num_bytes;
5676                         cache->space_info->disk_used -= num_bytes * factor;
5677                         spin_unlock(&cache->lock);
5678                         spin_unlock(&cache->space_info->lock);
5679
5680                         set_extent_dirty(info->pinned_extents,
5681                                          bytenr, bytenr + num_bytes - 1,
5682                                          GFP_NOFS | __GFP_NOFAIL);
5683                         /*
5684                          * No longer have used bytes in this block group, queue
5685                          * it for deletion.
5686                          */
5687                         if (old_val == 0) {
5688                                 spin_lock(&info->unused_bgs_lock);
5689                                 if (list_empty(&cache->bg_list)) {
5690                                         btrfs_get_block_group(cache);
5691                                         list_add_tail(&cache->bg_list,
5692                                                       &info->unused_bgs);
5693                                 }
5694                                 spin_unlock(&info->unused_bgs_lock);
5695                         }
5696                 }
5697
5698                 spin_lock(&trans->transaction->dirty_bgs_lock);
5699                 if (list_empty(&cache->dirty_list)) {
5700                         list_add_tail(&cache->dirty_list,
5701                                       &trans->transaction->dirty_bgs);
5702                                 trans->transaction->num_dirty_bgs++;
5703                         btrfs_get_block_group(cache);
5704                 }
5705                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5706
5707                 btrfs_put_block_group(cache);
5708                 total -= num_bytes;
5709                 bytenr += num_bytes;
5710         }
5711         return 0;
5712 }
5713
5714 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5715 {
5716         struct btrfs_block_group_cache *cache;
5717         u64 bytenr;
5718
5719         spin_lock(&root->fs_info->block_group_cache_lock);
5720         bytenr = root->fs_info->first_logical_byte;
5721         spin_unlock(&root->fs_info->block_group_cache_lock);
5722
5723         if (bytenr < (u64)-1)
5724                 return bytenr;
5725
5726         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5727         if (!cache)
5728                 return 0;
5729
5730         bytenr = cache->key.objectid;
5731         btrfs_put_block_group(cache);
5732
5733         return bytenr;
5734 }
5735
5736 static int pin_down_extent(struct btrfs_root *root,
5737                            struct btrfs_block_group_cache *cache,
5738                            u64 bytenr, u64 num_bytes, int reserved)
5739 {
5740         spin_lock(&cache->space_info->lock);
5741         spin_lock(&cache->lock);
5742         cache->pinned += num_bytes;
5743         cache->space_info->bytes_pinned += num_bytes;
5744         if (reserved) {
5745                 cache->reserved -= num_bytes;
5746                 cache->space_info->bytes_reserved -= num_bytes;
5747         }
5748         spin_unlock(&cache->lock);
5749         spin_unlock(&cache->space_info->lock);
5750
5751         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5752                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5753         if (reserved)
5754                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5755         return 0;
5756 }
5757
5758 /*
5759  * this function must be called within transaction
5760  */
5761 int btrfs_pin_extent(struct btrfs_root *root,
5762                      u64 bytenr, u64 num_bytes, int reserved)
5763 {
5764         struct btrfs_block_group_cache *cache;
5765
5766         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5767         BUG_ON(!cache); /* Logic error */
5768
5769         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5770
5771         btrfs_put_block_group(cache);
5772         return 0;
5773 }
5774
5775 /*
5776  * this function must be called within transaction
5777  */
5778 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5779                                     u64 bytenr, u64 num_bytes)
5780 {
5781         struct btrfs_block_group_cache *cache;
5782         int ret;
5783
5784         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5785         if (!cache)
5786                 return -EINVAL;
5787
5788         /*
5789          * pull in the free space cache (if any) so that our pin
5790          * removes the free space from the cache.  We have load_only set
5791          * to one because the slow code to read in the free extents does check
5792          * the pinned extents.
5793          */
5794         cache_block_group(cache, 1);
5795
5796         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5797
5798         /* remove us from the free space cache (if we're there at all) */
5799         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5800         btrfs_put_block_group(cache);
5801         return ret;
5802 }
5803
5804 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5805 {
5806         int ret;
5807         struct btrfs_block_group_cache *block_group;
5808         struct btrfs_caching_control *caching_ctl;
5809
5810         block_group = btrfs_lookup_block_group(root->fs_info, start);
5811         if (!block_group)
5812                 return -EINVAL;
5813
5814         cache_block_group(block_group, 0);
5815         caching_ctl = get_caching_control(block_group);
5816
5817         if (!caching_ctl) {
5818                 /* Logic error */
5819                 BUG_ON(!block_group_cache_done(block_group));
5820                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5821         } else {
5822                 mutex_lock(&caching_ctl->mutex);
5823
5824                 if (start >= caching_ctl->progress) {
5825                         ret = add_excluded_extent(root, start, num_bytes);
5826                 } else if (start + num_bytes <= caching_ctl->progress) {
5827                         ret = btrfs_remove_free_space(block_group,
5828                                                       start, num_bytes);
5829                 } else {
5830                         num_bytes = caching_ctl->progress - start;
5831                         ret = btrfs_remove_free_space(block_group,
5832                                                       start, num_bytes);
5833                         if (ret)
5834                                 goto out_lock;
5835
5836                         num_bytes = (start + num_bytes) -
5837                                 caching_ctl->progress;
5838                         start = caching_ctl->progress;
5839                         ret = add_excluded_extent(root, start, num_bytes);
5840                 }
5841 out_lock:
5842                 mutex_unlock(&caching_ctl->mutex);
5843                 put_caching_control(caching_ctl);
5844         }
5845         btrfs_put_block_group(block_group);
5846         return ret;
5847 }
5848
5849 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5850                                  struct extent_buffer *eb)
5851 {
5852         struct btrfs_file_extent_item *item;
5853         struct btrfs_key key;
5854         int found_type;
5855         int i;
5856
5857         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5858                 return 0;
5859
5860         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5861                 btrfs_item_key_to_cpu(eb, &key, i);
5862                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5863                         continue;
5864                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5865                 found_type = btrfs_file_extent_type(eb, item);
5866                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5867                         continue;
5868                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5869                         continue;
5870                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5871                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5872                 __exclude_logged_extent(log, key.objectid, key.offset);
5873         }
5874
5875         return 0;
5876 }
5877
5878 /**
5879  * btrfs_update_reserved_bytes - update the block_group and space info counters
5880  * @cache:      The cache we are manipulating
5881  * @num_bytes:  The number of bytes in question
5882  * @reserve:    One of the reservation enums
5883  * @delalloc:   The blocks are allocated for the delalloc write
5884  *
5885  * This is called by the allocator when it reserves space, or by somebody who is
5886  * freeing space that was never actually used on disk.  For example if you
5887  * reserve some space for a new leaf in transaction A and before transaction A
5888  * commits you free that leaf, you call this with reserve set to 0 in order to
5889  * clear the reservation.
5890  *
5891  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5892  * ENOSPC accounting.  For data we handle the reservation through clearing the
5893  * delalloc bits in the io_tree.  We have to do this since we could end up
5894  * allocating less disk space for the amount of data we have reserved in the
5895  * case of compression.
5896  *
5897  * If this is a reservation and the block group has become read only we cannot
5898  * make the reservation and return -EAGAIN, otherwise this function always
5899  * succeeds.
5900  */
5901 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5902                                        u64 num_bytes, int reserve, int delalloc)
5903 {
5904         struct btrfs_space_info *space_info = cache->space_info;
5905         int ret = 0;
5906
5907         spin_lock(&space_info->lock);
5908         spin_lock(&cache->lock);
5909         if (reserve != RESERVE_FREE) {
5910                 if (cache->ro) {
5911                         ret = -EAGAIN;
5912                 } else {
5913                         cache->reserved += num_bytes;
5914                         space_info->bytes_reserved += num_bytes;
5915                         if (reserve == RESERVE_ALLOC) {
5916                                 trace_btrfs_space_reservation(cache->fs_info,
5917                                                 "space_info", space_info->flags,
5918                                                 num_bytes, 0);
5919                                 space_info->bytes_may_use -= num_bytes;
5920                         }
5921
5922                         if (delalloc)
5923                                 cache->delalloc_bytes += num_bytes;
5924                 }
5925         } else {
5926                 if (cache->ro)
5927                         space_info->bytes_readonly += num_bytes;
5928                 cache->reserved -= num_bytes;
5929                 space_info->bytes_reserved -= num_bytes;
5930
5931                 if (delalloc)
5932                         cache->delalloc_bytes -= num_bytes;
5933         }
5934         spin_unlock(&cache->lock);
5935         spin_unlock(&space_info->lock);
5936         return ret;
5937 }
5938
5939 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5940                                 struct btrfs_root *root)
5941 {
5942         struct btrfs_fs_info *fs_info = root->fs_info;
5943         struct btrfs_caching_control *next;
5944         struct btrfs_caching_control *caching_ctl;
5945         struct btrfs_block_group_cache *cache;
5946
5947         down_write(&fs_info->commit_root_sem);
5948
5949         list_for_each_entry_safe(caching_ctl, next,
5950                                  &fs_info->caching_block_groups, list) {
5951                 cache = caching_ctl->block_group;
5952                 if (block_group_cache_done(cache)) {
5953                         cache->last_byte_to_unpin = (u64)-1;
5954                         list_del_init(&caching_ctl->list);
5955                         put_caching_control(caching_ctl);
5956                 } else {
5957                         cache->last_byte_to_unpin = caching_ctl->progress;
5958                 }
5959         }
5960
5961         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5962                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5963         else
5964                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5965
5966         up_write(&fs_info->commit_root_sem);
5967
5968         update_global_block_rsv(fs_info);
5969 }
5970
5971 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
5972                               const bool return_free_space)
5973 {
5974         struct btrfs_fs_info *fs_info = root->fs_info;
5975         struct btrfs_block_group_cache *cache = NULL;
5976         struct btrfs_space_info *space_info;
5977         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5978         u64 len;
5979         bool readonly;
5980
5981         while (start <= end) {
5982                 readonly = false;
5983                 if (!cache ||
5984                     start >= cache->key.objectid + cache->key.offset) {
5985                         if (cache)
5986                                 btrfs_put_block_group(cache);
5987                         cache = btrfs_lookup_block_group(fs_info, start);
5988                         BUG_ON(!cache); /* Logic error */
5989                 }
5990
5991                 len = cache->key.objectid + cache->key.offset - start;
5992                 len = min(len, end + 1 - start);
5993
5994                 if (start < cache->last_byte_to_unpin) {
5995                         len = min(len, cache->last_byte_to_unpin - start);
5996                         if (return_free_space)
5997                                 btrfs_add_free_space(cache, start, len);
5998                 }
5999
6000                 start += len;
6001                 space_info = cache->space_info;
6002
6003                 spin_lock(&space_info->lock);
6004                 spin_lock(&cache->lock);
6005                 cache->pinned -= len;
6006                 space_info->bytes_pinned -= len;
6007                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6008                 if (cache->ro) {
6009                         space_info->bytes_readonly += len;
6010                         readonly = true;
6011                 }
6012                 spin_unlock(&cache->lock);
6013                 if (!readonly && global_rsv->space_info == space_info) {
6014                         spin_lock(&global_rsv->lock);
6015                         if (!global_rsv->full) {
6016                                 len = min(len, global_rsv->size -
6017                                           global_rsv->reserved);
6018                                 global_rsv->reserved += len;
6019                                 space_info->bytes_may_use += len;
6020                                 if (global_rsv->reserved >= global_rsv->size)
6021                                         global_rsv->full = 1;
6022                         }
6023                         spin_unlock(&global_rsv->lock);
6024                 }
6025                 spin_unlock(&space_info->lock);
6026         }
6027
6028         if (cache)
6029                 btrfs_put_block_group(cache);
6030         return 0;
6031 }
6032
6033 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6034                                struct btrfs_root *root)
6035 {
6036         struct btrfs_fs_info *fs_info = root->fs_info;
6037         struct extent_io_tree *unpin;
6038         u64 start;
6039         u64 end;
6040         int ret;
6041
6042         if (trans->aborted)
6043                 return 0;
6044
6045         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6046                 unpin = &fs_info->freed_extents[1];
6047         else
6048                 unpin = &fs_info->freed_extents[0];
6049
6050         while (1) {
6051                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6052                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6053                                             EXTENT_DIRTY, NULL);
6054                 if (ret) {
6055                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6056                         break;
6057                 }
6058
6059                 if (btrfs_test_opt(root, DISCARD))
6060                         ret = btrfs_discard_extent(root, start,
6061                                                    end + 1 - start, NULL);
6062
6063                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
6064                 unpin_extent_range(root, start, end, true);
6065                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6066                 cond_resched();
6067         }
6068
6069         return 0;
6070 }
6071
6072 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6073                              u64 owner, u64 root_objectid)
6074 {
6075         struct btrfs_space_info *space_info;
6076         u64 flags;
6077
6078         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6079                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6080                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6081                 else
6082                         flags = BTRFS_BLOCK_GROUP_METADATA;
6083         } else {
6084                 flags = BTRFS_BLOCK_GROUP_DATA;
6085         }
6086
6087         space_info = __find_space_info(fs_info, flags);
6088         BUG_ON(!space_info); /* Logic bug */
6089         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6090 }
6091
6092
6093 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6094                                 struct btrfs_root *root,
6095                                 u64 bytenr, u64 num_bytes, u64 parent,
6096                                 u64 root_objectid, u64 owner_objectid,
6097                                 u64 owner_offset, int refs_to_drop,
6098                                 struct btrfs_delayed_extent_op *extent_op,
6099                                 int no_quota)
6100 {
6101         struct btrfs_key key;
6102         struct btrfs_path *path;
6103         struct btrfs_fs_info *info = root->fs_info;
6104         struct btrfs_root *extent_root = info->extent_root;
6105         struct extent_buffer *leaf;
6106         struct btrfs_extent_item *ei;
6107         struct btrfs_extent_inline_ref *iref;
6108         int ret;
6109         int is_data;
6110         int extent_slot = 0;
6111         int found_extent = 0;
6112         int num_to_del = 1;
6113         u32 item_size;
6114         u64 refs;
6115         int last_ref = 0;
6116         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_SUB_EXCL;
6117         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6118                                                  SKINNY_METADATA);
6119
6120         if (!info->quota_enabled || !is_fstree(root_objectid))
6121                 no_quota = 1;
6122
6123         path = btrfs_alloc_path();
6124         if (!path)
6125                 return -ENOMEM;
6126
6127         path->reada = 1;
6128         path->leave_spinning = 1;
6129
6130         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6131         BUG_ON(!is_data && refs_to_drop != 1);
6132
6133         if (is_data)
6134                 skinny_metadata = 0;
6135
6136         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6137                                     bytenr, num_bytes, parent,
6138                                     root_objectid, owner_objectid,
6139                                     owner_offset);
6140         if (ret == 0) {
6141                 extent_slot = path->slots[0];
6142                 while (extent_slot >= 0) {
6143                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6144                                               extent_slot);
6145                         if (key.objectid != bytenr)
6146                                 break;
6147                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6148                             key.offset == num_bytes) {
6149                                 found_extent = 1;
6150                                 break;
6151                         }
6152                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6153                             key.offset == owner_objectid) {
6154                                 found_extent = 1;
6155                                 break;
6156                         }
6157                         if (path->slots[0] - extent_slot > 5)
6158                                 break;
6159                         extent_slot--;
6160                 }
6161 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6162                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6163                 if (found_extent && item_size < sizeof(*ei))
6164                         found_extent = 0;
6165 #endif
6166                 if (!found_extent) {
6167                         BUG_ON(iref);
6168                         ret = remove_extent_backref(trans, extent_root, path,
6169                                                     NULL, refs_to_drop,
6170                                                     is_data, &last_ref);
6171                         if (ret) {
6172                                 btrfs_abort_transaction(trans, extent_root, ret);
6173                                 goto out;
6174                         }
6175                         btrfs_release_path(path);
6176                         path->leave_spinning = 1;
6177
6178                         key.objectid = bytenr;
6179                         key.type = BTRFS_EXTENT_ITEM_KEY;
6180                         key.offset = num_bytes;
6181
6182                         if (!is_data && skinny_metadata) {
6183                                 key.type = BTRFS_METADATA_ITEM_KEY;
6184                                 key.offset = owner_objectid;
6185                         }
6186
6187                         ret = btrfs_search_slot(trans, extent_root,
6188                                                 &key, path, -1, 1);
6189                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6190                                 /*
6191                                  * Couldn't find our skinny metadata item,
6192                                  * see if we have ye olde extent item.
6193                                  */
6194                                 path->slots[0]--;
6195                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6196                                                       path->slots[0]);
6197                                 if (key.objectid == bytenr &&
6198                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6199                                     key.offset == num_bytes)
6200                                         ret = 0;
6201                         }
6202
6203                         if (ret > 0 && skinny_metadata) {
6204                                 skinny_metadata = false;
6205                                 key.objectid = bytenr;
6206                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6207                                 key.offset = num_bytes;
6208                                 btrfs_release_path(path);
6209                                 ret = btrfs_search_slot(trans, extent_root,
6210                                                         &key, path, -1, 1);
6211                         }
6212
6213                         if (ret) {
6214                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6215                                         ret, bytenr);
6216                                 if (ret > 0)
6217                                         btrfs_print_leaf(extent_root,
6218                                                          path->nodes[0]);
6219                         }
6220                         if (ret < 0) {
6221                                 btrfs_abort_transaction(trans, extent_root, ret);
6222                                 goto out;
6223                         }
6224                         extent_slot = path->slots[0];
6225                 }
6226         } else if (WARN_ON(ret == -ENOENT)) {
6227                 btrfs_print_leaf(extent_root, path->nodes[0]);
6228                 btrfs_err(info,
6229                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6230                         bytenr, parent, root_objectid, owner_objectid,
6231                         owner_offset);
6232                 btrfs_abort_transaction(trans, extent_root, ret);
6233                 goto out;
6234         } else {
6235                 btrfs_abort_transaction(trans, extent_root, ret);
6236                 goto out;
6237         }
6238
6239         leaf = path->nodes[0];
6240         item_size = btrfs_item_size_nr(leaf, extent_slot);
6241 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6242         if (item_size < sizeof(*ei)) {
6243                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6244                 ret = convert_extent_item_v0(trans, extent_root, path,
6245                                              owner_objectid, 0);
6246                 if (ret < 0) {
6247                         btrfs_abort_transaction(trans, extent_root, ret);
6248                         goto out;
6249                 }
6250
6251                 btrfs_release_path(path);
6252                 path->leave_spinning = 1;
6253
6254                 key.objectid = bytenr;
6255                 key.type = BTRFS_EXTENT_ITEM_KEY;
6256                 key.offset = num_bytes;
6257
6258                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6259                                         -1, 1);
6260                 if (ret) {
6261                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6262                                 ret, bytenr);
6263                         btrfs_print_leaf(extent_root, path->nodes[0]);
6264                 }
6265                 if (ret < 0) {
6266                         btrfs_abort_transaction(trans, extent_root, ret);
6267                         goto out;
6268                 }
6269
6270                 extent_slot = path->slots[0];
6271                 leaf = path->nodes[0];
6272                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6273         }
6274 #endif
6275         BUG_ON(item_size < sizeof(*ei));
6276         ei = btrfs_item_ptr(leaf, extent_slot,
6277                             struct btrfs_extent_item);
6278         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6279             key.type == BTRFS_EXTENT_ITEM_KEY) {
6280                 struct btrfs_tree_block_info *bi;
6281                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6282                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6283                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6284         }
6285
6286         refs = btrfs_extent_refs(leaf, ei);
6287         if (refs < refs_to_drop) {
6288                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6289                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6290                 ret = -EINVAL;
6291                 btrfs_abort_transaction(trans, extent_root, ret);
6292                 goto out;
6293         }
6294         refs -= refs_to_drop;
6295
6296         if (refs > 0) {
6297                 type = BTRFS_QGROUP_OPER_SUB_SHARED;
6298                 if (extent_op)
6299                         __run_delayed_extent_op(extent_op, leaf, ei);
6300                 /*
6301                  * In the case of inline back ref, reference count will
6302                  * be updated by remove_extent_backref
6303                  */
6304                 if (iref) {
6305                         BUG_ON(!found_extent);
6306                 } else {
6307                         btrfs_set_extent_refs(leaf, ei, refs);
6308                         btrfs_mark_buffer_dirty(leaf);
6309                 }
6310                 if (found_extent) {
6311                         ret = remove_extent_backref(trans, extent_root, path,
6312                                                     iref, refs_to_drop,
6313                                                     is_data, &last_ref);
6314                         if (ret) {
6315                                 btrfs_abort_transaction(trans, extent_root, ret);
6316                                 goto out;
6317                         }
6318                 }
6319                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6320                                  root_objectid);
6321         } else {
6322                 if (found_extent) {
6323                         BUG_ON(is_data && refs_to_drop !=
6324                                extent_data_ref_count(root, path, iref));
6325                         if (iref) {
6326                                 BUG_ON(path->slots[0] != extent_slot);
6327                         } else {
6328                                 BUG_ON(path->slots[0] != extent_slot + 1);
6329                                 path->slots[0] = extent_slot;
6330                                 num_to_del = 2;
6331                         }
6332                 }
6333
6334                 last_ref = 1;
6335                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6336                                       num_to_del);
6337                 if (ret) {
6338                         btrfs_abort_transaction(trans, extent_root, ret);
6339                         goto out;
6340                 }
6341                 btrfs_release_path(path);
6342
6343                 if (is_data) {
6344                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6345                         if (ret) {
6346                                 btrfs_abort_transaction(trans, extent_root, ret);
6347                                 goto out;
6348                         }
6349                 }
6350
6351                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6352                 if (ret) {
6353                         btrfs_abort_transaction(trans, extent_root, ret);
6354                         goto out;
6355                 }
6356         }
6357         btrfs_release_path(path);
6358
6359         /* Deal with the quota accounting */
6360         if (!ret && last_ref && !no_quota) {
6361                 int mod_seq = 0;
6362
6363                 if (owner_objectid >= BTRFS_FIRST_FREE_OBJECTID &&
6364                     type == BTRFS_QGROUP_OPER_SUB_SHARED)
6365                         mod_seq = 1;
6366
6367                 ret = btrfs_qgroup_record_ref(trans, info, root_objectid,
6368                                               bytenr, num_bytes, type,
6369                                               mod_seq);
6370         }
6371 out:
6372         btrfs_free_path(path);
6373         return ret;
6374 }
6375
6376 /*
6377  * when we free an block, it is possible (and likely) that we free the last
6378  * delayed ref for that extent as well.  This searches the delayed ref tree for
6379  * a given extent, and if there are no other delayed refs to be processed, it
6380  * removes it from the tree.
6381  */
6382 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6383                                       struct btrfs_root *root, u64 bytenr)
6384 {
6385         struct btrfs_delayed_ref_head *head;
6386         struct btrfs_delayed_ref_root *delayed_refs;
6387         int ret = 0;
6388
6389         delayed_refs = &trans->transaction->delayed_refs;
6390         spin_lock(&delayed_refs->lock);
6391         head = btrfs_find_delayed_ref_head(trans, bytenr);
6392         if (!head)
6393                 goto out_delayed_unlock;
6394
6395         spin_lock(&head->lock);
6396         if (rb_first(&head->ref_root))
6397                 goto out;
6398
6399         if (head->extent_op) {
6400                 if (!head->must_insert_reserved)
6401                         goto out;
6402                 btrfs_free_delayed_extent_op(head->extent_op);
6403                 head->extent_op = NULL;
6404         }
6405
6406         /*
6407          * waiting for the lock here would deadlock.  If someone else has it
6408          * locked they are already in the process of dropping it anyway
6409          */
6410         if (!mutex_trylock(&head->mutex))
6411                 goto out;
6412
6413         /*
6414          * at this point we have a head with no other entries.  Go
6415          * ahead and process it.
6416          */
6417         head->node.in_tree = 0;
6418         rb_erase(&head->href_node, &delayed_refs->href_root);
6419
6420         atomic_dec(&delayed_refs->num_entries);
6421
6422         /*
6423          * we don't take a ref on the node because we're removing it from the
6424          * tree, so we just steal the ref the tree was holding.
6425          */
6426         delayed_refs->num_heads--;
6427         if (head->processing == 0)
6428                 delayed_refs->num_heads_ready--;
6429         head->processing = 0;
6430         spin_unlock(&head->lock);
6431         spin_unlock(&delayed_refs->lock);
6432
6433         BUG_ON(head->extent_op);
6434         if (head->must_insert_reserved)
6435                 ret = 1;
6436
6437         mutex_unlock(&head->mutex);
6438         btrfs_put_delayed_ref(&head->node);
6439         return ret;
6440 out:
6441         spin_unlock(&head->lock);
6442
6443 out_delayed_unlock:
6444         spin_unlock(&delayed_refs->lock);
6445         return 0;
6446 }
6447
6448 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6449                            struct btrfs_root *root,
6450                            struct extent_buffer *buf,
6451                            u64 parent, int last_ref)
6452 {
6453         int pin = 1;
6454         int ret;
6455
6456         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6457                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6458                                         buf->start, buf->len,
6459                                         parent, root->root_key.objectid,
6460                                         btrfs_header_level(buf),
6461                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6462                 BUG_ON(ret); /* -ENOMEM */
6463         }
6464
6465         if (!last_ref)
6466                 return;
6467
6468         if (btrfs_header_generation(buf) == trans->transid) {
6469                 struct btrfs_block_group_cache *cache;
6470
6471                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6472                         ret = check_ref_cleanup(trans, root, buf->start);
6473                         if (!ret)
6474                                 goto out;
6475                 }
6476
6477                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6478
6479                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6480                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6481                         btrfs_put_block_group(cache);
6482                         goto out;
6483                 }
6484
6485                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6486
6487                 btrfs_add_free_space(cache, buf->start, buf->len);
6488                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6489                 btrfs_put_block_group(cache);
6490                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6491                 pin = 0;
6492         }
6493 out:
6494         if (pin)
6495                 add_pinned_bytes(root->fs_info, buf->len,
6496                                  btrfs_header_level(buf),
6497                                  root->root_key.objectid);
6498
6499         /*
6500          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6501          * anymore.
6502          */
6503         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6504 }
6505
6506 /* Can return -ENOMEM */
6507 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6508                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6509                       u64 owner, u64 offset, int no_quota)
6510 {
6511         int ret;
6512         struct btrfs_fs_info *fs_info = root->fs_info;
6513
6514         if (btrfs_test_is_dummy_root(root))
6515                 return 0;
6516
6517         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6518
6519         /*
6520          * tree log blocks never actually go into the extent allocation
6521          * tree, just update pinning info and exit early.
6522          */
6523         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6524                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6525                 /* unlocks the pinned mutex */
6526                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6527                 ret = 0;
6528         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6529                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6530                                         num_bytes,
6531                                         parent, root_objectid, (int)owner,
6532                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6533         } else {
6534                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6535                                                 num_bytes,
6536                                                 parent, root_objectid, owner,
6537                                                 offset, BTRFS_DROP_DELAYED_REF,
6538                                                 NULL, no_quota);
6539         }
6540         return ret;
6541 }
6542
6543 /*
6544  * when we wait for progress in the block group caching, its because
6545  * our allocation attempt failed at least once.  So, we must sleep
6546  * and let some progress happen before we try again.
6547  *
6548  * This function will sleep at least once waiting for new free space to
6549  * show up, and then it will check the block group free space numbers
6550  * for our min num_bytes.  Another option is to have it go ahead
6551  * and look in the rbtree for a free extent of a given size, but this
6552  * is a good start.
6553  *
6554  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6555  * any of the information in this block group.
6556  */
6557 static noinline void
6558 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6559                                 u64 num_bytes)
6560 {
6561         struct btrfs_caching_control *caching_ctl;
6562
6563         caching_ctl = get_caching_control(cache);
6564         if (!caching_ctl)
6565                 return;
6566
6567         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6568                    (cache->free_space_ctl->free_space >= num_bytes));
6569
6570         put_caching_control(caching_ctl);
6571 }
6572
6573 static noinline int
6574 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6575 {
6576         struct btrfs_caching_control *caching_ctl;
6577         int ret = 0;
6578
6579         caching_ctl = get_caching_control(cache);
6580         if (!caching_ctl)
6581                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6582
6583         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6584         if (cache->cached == BTRFS_CACHE_ERROR)
6585                 ret = -EIO;
6586         put_caching_control(caching_ctl);
6587         return ret;
6588 }
6589
6590 int __get_raid_index(u64 flags)
6591 {
6592         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6593                 return BTRFS_RAID_RAID10;
6594         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6595                 return BTRFS_RAID_RAID1;
6596         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6597                 return BTRFS_RAID_DUP;
6598         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6599                 return BTRFS_RAID_RAID0;
6600         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6601                 return BTRFS_RAID_RAID5;
6602         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6603                 return BTRFS_RAID_RAID6;
6604
6605         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6606 }
6607
6608 int get_block_group_index(struct btrfs_block_group_cache *cache)
6609 {
6610         return __get_raid_index(cache->flags);
6611 }
6612
6613 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6614         [BTRFS_RAID_RAID10]     = "raid10",
6615         [BTRFS_RAID_RAID1]      = "raid1",
6616         [BTRFS_RAID_DUP]        = "dup",
6617         [BTRFS_RAID_RAID0]      = "raid0",
6618         [BTRFS_RAID_SINGLE]     = "single",
6619         [BTRFS_RAID_RAID5]      = "raid5",
6620         [BTRFS_RAID_RAID6]      = "raid6",
6621 };
6622
6623 static const char *get_raid_name(enum btrfs_raid_types type)
6624 {
6625         if (type >= BTRFS_NR_RAID_TYPES)
6626                 return NULL;
6627
6628         return btrfs_raid_type_names[type];
6629 }
6630
6631 enum btrfs_loop_type {
6632         LOOP_CACHING_NOWAIT = 0,
6633         LOOP_CACHING_WAIT = 1,
6634         LOOP_ALLOC_CHUNK = 2,
6635         LOOP_NO_EMPTY_SIZE = 3,
6636 };
6637
6638 static inline void
6639 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6640                        int delalloc)
6641 {
6642         if (delalloc)
6643                 down_read(&cache->data_rwsem);
6644 }
6645
6646 static inline void
6647 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6648                        int delalloc)
6649 {
6650         btrfs_get_block_group(cache);
6651         if (delalloc)
6652                 down_read(&cache->data_rwsem);
6653 }
6654
6655 static struct btrfs_block_group_cache *
6656 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6657                    struct btrfs_free_cluster *cluster,
6658                    int delalloc)
6659 {
6660         struct btrfs_block_group_cache *used_bg;
6661         bool locked = false;
6662 again:
6663         spin_lock(&cluster->refill_lock);
6664         if (locked) {
6665                 if (used_bg == cluster->block_group)
6666                         return used_bg;
6667
6668                 up_read(&used_bg->data_rwsem);
6669                 btrfs_put_block_group(used_bg);
6670         }
6671
6672         used_bg = cluster->block_group;
6673         if (!used_bg)
6674                 return NULL;
6675
6676         if (used_bg == block_group)
6677                 return used_bg;
6678
6679         btrfs_get_block_group(used_bg);
6680
6681         if (!delalloc)
6682                 return used_bg;
6683
6684         if (down_read_trylock(&used_bg->data_rwsem))
6685                 return used_bg;
6686
6687         spin_unlock(&cluster->refill_lock);
6688         down_read(&used_bg->data_rwsem);
6689         locked = true;
6690         goto again;
6691 }
6692
6693 static inline void
6694 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6695                          int delalloc)
6696 {
6697         if (delalloc)
6698                 up_read(&cache->data_rwsem);
6699         btrfs_put_block_group(cache);
6700 }
6701
6702 /*
6703  * walks the btree of allocated extents and find a hole of a given size.
6704  * The key ins is changed to record the hole:
6705  * ins->objectid == start position
6706  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6707  * ins->offset == the size of the hole.
6708  * Any available blocks before search_start are skipped.
6709  *
6710  * If there is no suitable free space, we will record the max size of
6711  * the free space extent currently.
6712  */
6713 static noinline int find_free_extent(struct btrfs_root *orig_root,
6714                                      u64 num_bytes, u64 empty_size,
6715                                      u64 hint_byte, struct btrfs_key *ins,
6716                                      u64 flags, int delalloc)
6717 {
6718         int ret = 0;
6719         struct btrfs_root *root = orig_root->fs_info->extent_root;
6720         struct btrfs_free_cluster *last_ptr = NULL;
6721         struct btrfs_block_group_cache *block_group = NULL;
6722         u64 search_start = 0;
6723         u64 max_extent_size = 0;
6724         int empty_cluster = 2 * 1024 * 1024;
6725         struct btrfs_space_info *space_info;
6726         int loop = 0;
6727         int index = __get_raid_index(flags);
6728         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6729                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6730         bool failed_cluster_refill = false;
6731         bool failed_alloc = false;
6732         bool use_cluster = true;
6733         bool have_caching_bg = false;
6734
6735         WARN_ON(num_bytes < root->sectorsize);
6736         ins->type = BTRFS_EXTENT_ITEM_KEY;
6737         ins->objectid = 0;
6738         ins->offset = 0;
6739
6740         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6741
6742         space_info = __find_space_info(root->fs_info, flags);
6743         if (!space_info) {
6744                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6745                 return -ENOSPC;
6746         }
6747
6748         /*
6749          * If the space info is for both data and metadata it means we have a
6750          * small filesystem and we can't use the clustering stuff.
6751          */
6752         if (btrfs_mixed_space_info(space_info))
6753                 use_cluster = false;
6754
6755         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6756                 last_ptr = &root->fs_info->meta_alloc_cluster;
6757                 if (!btrfs_test_opt(root, SSD))
6758                         empty_cluster = 64 * 1024;
6759         }
6760
6761         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6762             btrfs_test_opt(root, SSD)) {
6763                 last_ptr = &root->fs_info->data_alloc_cluster;
6764         }
6765
6766         if (last_ptr) {
6767                 spin_lock(&last_ptr->lock);
6768                 if (last_ptr->block_group)
6769                         hint_byte = last_ptr->window_start;
6770                 spin_unlock(&last_ptr->lock);
6771         }
6772
6773         search_start = max(search_start, first_logical_byte(root, 0));
6774         search_start = max(search_start, hint_byte);
6775
6776         if (!last_ptr)
6777                 empty_cluster = 0;
6778
6779         if (search_start == hint_byte) {
6780                 block_group = btrfs_lookup_block_group(root->fs_info,
6781                                                        search_start);
6782                 /*
6783                  * we don't want to use the block group if it doesn't match our
6784                  * allocation bits, or if its not cached.
6785                  *
6786                  * However if we are re-searching with an ideal block group
6787                  * picked out then we don't care that the block group is cached.
6788                  */
6789                 if (block_group && block_group_bits(block_group, flags) &&
6790                     block_group->cached != BTRFS_CACHE_NO) {
6791                         down_read(&space_info->groups_sem);
6792                         if (list_empty(&block_group->list) ||
6793                             block_group->ro) {
6794                                 /*
6795                                  * someone is removing this block group,
6796                                  * we can't jump into the have_block_group
6797                                  * target because our list pointers are not
6798                                  * valid
6799                                  */
6800                                 btrfs_put_block_group(block_group);
6801                                 up_read(&space_info->groups_sem);
6802                         } else {
6803                                 index = get_block_group_index(block_group);
6804                                 btrfs_lock_block_group(block_group, delalloc);
6805                                 goto have_block_group;
6806                         }
6807                 } else if (block_group) {
6808                         btrfs_put_block_group(block_group);
6809                 }
6810         }
6811 search:
6812         have_caching_bg = false;
6813         down_read(&space_info->groups_sem);
6814         list_for_each_entry(block_group, &space_info->block_groups[index],
6815                             list) {
6816                 u64 offset;
6817                 int cached;
6818
6819                 btrfs_grab_block_group(block_group, delalloc);
6820                 search_start = block_group->key.objectid;
6821
6822                 /*
6823                  * this can happen if we end up cycling through all the
6824                  * raid types, but we want to make sure we only allocate
6825                  * for the proper type.
6826                  */
6827                 if (!block_group_bits(block_group, flags)) {
6828                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6829                                 BTRFS_BLOCK_GROUP_RAID1 |
6830                                 BTRFS_BLOCK_GROUP_RAID5 |
6831                                 BTRFS_BLOCK_GROUP_RAID6 |
6832                                 BTRFS_BLOCK_GROUP_RAID10;
6833
6834                         /*
6835                          * if they asked for extra copies and this block group
6836                          * doesn't provide them, bail.  This does allow us to
6837                          * fill raid0 from raid1.
6838                          */
6839                         if ((flags & extra) && !(block_group->flags & extra))
6840                                 goto loop;
6841                 }
6842
6843 have_block_group:
6844                 cached = block_group_cache_done(block_group);
6845                 if (unlikely(!cached)) {
6846                         ret = cache_block_group(block_group, 0);
6847                         BUG_ON(ret < 0);
6848                         ret = 0;
6849                 }
6850
6851                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6852                         goto loop;
6853                 if (unlikely(block_group->ro))
6854                         goto loop;
6855
6856                 /*
6857                  * Ok we want to try and use the cluster allocator, so
6858                  * lets look there
6859                  */
6860                 if (last_ptr) {
6861                         struct btrfs_block_group_cache *used_block_group;
6862                         unsigned long aligned_cluster;
6863                         /*
6864                          * the refill lock keeps out other
6865                          * people trying to start a new cluster
6866                          */
6867                         used_block_group = btrfs_lock_cluster(block_group,
6868                                                               last_ptr,
6869                                                               delalloc);
6870                         if (!used_block_group)
6871                                 goto refill_cluster;
6872
6873                         if (used_block_group != block_group &&
6874                             (used_block_group->ro ||
6875                              !block_group_bits(used_block_group, flags)))
6876                                 goto release_cluster;
6877
6878                         offset = btrfs_alloc_from_cluster(used_block_group,
6879                                                 last_ptr,
6880                                                 num_bytes,
6881                                                 used_block_group->key.objectid,
6882                                                 &max_extent_size);
6883                         if (offset) {
6884                                 /* we have a block, we're done */
6885                                 spin_unlock(&last_ptr->refill_lock);
6886                                 trace_btrfs_reserve_extent_cluster(root,
6887                                                 used_block_group,
6888                                                 search_start, num_bytes);
6889                                 if (used_block_group != block_group) {
6890                                         btrfs_release_block_group(block_group,
6891                                                                   delalloc);
6892                                         block_group = used_block_group;
6893                                 }
6894                                 goto checks;
6895                         }
6896
6897                         WARN_ON(last_ptr->block_group != used_block_group);
6898 release_cluster:
6899                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6900                          * set up a new clusters, so lets just skip it
6901                          * and let the allocator find whatever block
6902                          * it can find.  If we reach this point, we
6903                          * will have tried the cluster allocator
6904                          * plenty of times and not have found
6905                          * anything, so we are likely way too
6906                          * fragmented for the clustering stuff to find
6907                          * anything.
6908                          *
6909                          * However, if the cluster is taken from the
6910                          * current block group, release the cluster
6911                          * first, so that we stand a better chance of
6912                          * succeeding in the unclustered
6913                          * allocation.  */
6914                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6915                             used_block_group != block_group) {
6916                                 spin_unlock(&last_ptr->refill_lock);
6917                                 btrfs_release_block_group(used_block_group,
6918                                                           delalloc);
6919                                 goto unclustered_alloc;
6920                         }
6921
6922                         /*
6923                          * this cluster didn't work out, free it and
6924                          * start over
6925                          */
6926                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6927
6928                         if (used_block_group != block_group)
6929                                 btrfs_release_block_group(used_block_group,
6930                                                           delalloc);
6931 refill_cluster:
6932                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6933                                 spin_unlock(&last_ptr->refill_lock);
6934                                 goto unclustered_alloc;
6935                         }
6936
6937                         aligned_cluster = max_t(unsigned long,
6938                                                 empty_cluster + empty_size,
6939                                               block_group->full_stripe_len);
6940
6941                         /* allocate a cluster in this block group */
6942                         ret = btrfs_find_space_cluster(root, block_group,
6943                                                        last_ptr, search_start,
6944                                                        num_bytes,
6945                                                        aligned_cluster);
6946                         if (ret == 0) {
6947                                 /*
6948                                  * now pull our allocation out of this
6949                                  * cluster
6950                                  */
6951                                 offset = btrfs_alloc_from_cluster(block_group,
6952                                                         last_ptr,
6953                                                         num_bytes,
6954                                                         search_start,
6955                                                         &max_extent_size);
6956                                 if (offset) {
6957                                         /* we found one, proceed */
6958                                         spin_unlock(&last_ptr->refill_lock);
6959                                         trace_btrfs_reserve_extent_cluster(root,
6960                                                 block_group, search_start,
6961                                                 num_bytes);
6962                                         goto checks;
6963                                 }
6964                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6965                                    && !failed_cluster_refill) {
6966                                 spin_unlock(&last_ptr->refill_lock);
6967
6968                                 failed_cluster_refill = true;
6969                                 wait_block_group_cache_progress(block_group,
6970                                        num_bytes + empty_cluster + empty_size);
6971                                 goto have_block_group;
6972                         }
6973
6974                         /*
6975                          * at this point we either didn't find a cluster
6976                          * or we weren't able to allocate a block from our
6977                          * cluster.  Free the cluster we've been trying
6978                          * to use, and go to the next block group
6979                          */
6980                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6981                         spin_unlock(&last_ptr->refill_lock);
6982                         goto loop;
6983                 }
6984
6985 unclustered_alloc:
6986                 spin_lock(&block_group->free_space_ctl->tree_lock);
6987                 if (cached &&
6988                     block_group->free_space_ctl->free_space <
6989                     num_bytes + empty_cluster + empty_size) {
6990                         if (block_group->free_space_ctl->free_space >
6991                             max_extent_size)
6992                                 max_extent_size =
6993                                         block_group->free_space_ctl->free_space;
6994                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6995                         goto loop;
6996                 }
6997                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6998
6999                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7000                                                     num_bytes, empty_size,
7001                                                     &max_extent_size);
7002                 /*
7003                  * If we didn't find a chunk, and we haven't failed on this
7004                  * block group before, and this block group is in the middle of
7005                  * caching and we are ok with waiting, then go ahead and wait
7006                  * for progress to be made, and set failed_alloc to true.
7007                  *
7008                  * If failed_alloc is true then we've already waited on this
7009                  * block group once and should move on to the next block group.
7010                  */
7011                 if (!offset && !failed_alloc && !cached &&
7012                     loop > LOOP_CACHING_NOWAIT) {
7013                         wait_block_group_cache_progress(block_group,
7014                                                 num_bytes + empty_size);
7015                         failed_alloc = true;
7016                         goto have_block_group;
7017                 } else if (!offset) {
7018                         if (!cached)
7019                                 have_caching_bg = true;
7020                         goto loop;
7021                 }
7022 checks:
7023                 search_start = ALIGN(offset, root->stripesize);
7024
7025                 /* move on to the next group */
7026                 if (search_start + num_bytes >
7027                     block_group->key.objectid + block_group->key.offset) {
7028                         btrfs_add_free_space(block_group, offset, num_bytes);
7029                         goto loop;
7030                 }
7031
7032                 if (offset < search_start)
7033                         btrfs_add_free_space(block_group, offset,
7034                                              search_start - offset);
7035                 BUG_ON(offset > search_start);
7036
7037                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7038                                                   alloc_type, delalloc);
7039                 if (ret == -EAGAIN) {
7040                         btrfs_add_free_space(block_group, offset, num_bytes);
7041                         goto loop;
7042                 }
7043
7044                 /* we are all good, lets return */
7045                 ins->objectid = search_start;
7046                 ins->offset = num_bytes;
7047
7048                 trace_btrfs_reserve_extent(orig_root, block_group,
7049                                            search_start, num_bytes);
7050                 btrfs_release_block_group(block_group, delalloc);
7051                 break;
7052 loop:
7053                 failed_cluster_refill = false;
7054                 failed_alloc = false;
7055                 BUG_ON(index != get_block_group_index(block_group));
7056                 btrfs_release_block_group(block_group, delalloc);
7057         }
7058         up_read(&space_info->groups_sem);
7059
7060         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7061                 goto search;
7062
7063         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7064                 goto search;
7065
7066         /*
7067          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7068          *                      caching kthreads as we move along
7069          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7070          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7071          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7072          *                      again
7073          */
7074         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7075                 index = 0;
7076                 loop++;
7077                 if (loop == LOOP_ALLOC_CHUNK) {
7078                         struct btrfs_trans_handle *trans;
7079                         int exist = 0;
7080
7081                         trans = current->journal_info;
7082                         if (trans)
7083                                 exist = 1;
7084                         else
7085                                 trans = btrfs_join_transaction(root);
7086
7087                         if (IS_ERR(trans)) {
7088                                 ret = PTR_ERR(trans);
7089                                 goto out;
7090                         }
7091
7092                         ret = do_chunk_alloc(trans, root, flags,
7093                                              CHUNK_ALLOC_FORCE);
7094                         /*
7095                          * Do not bail out on ENOSPC since we
7096                          * can do more things.
7097                          */
7098                         if (ret < 0 && ret != -ENOSPC)
7099                                 btrfs_abort_transaction(trans,
7100                                                         root, ret);
7101                         else
7102                                 ret = 0;
7103                         if (!exist)
7104                                 btrfs_end_transaction(trans, root);
7105                         if (ret)
7106                                 goto out;
7107                 }
7108
7109                 if (loop == LOOP_NO_EMPTY_SIZE) {
7110                         empty_size = 0;
7111                         empty_cluster = 0;
7112                 }
7113
7114                 goto search;
7115         } else if (!ins->objectid) {
7116                 ret = -ENOSPC;
7117         } else if (ins->objectid) {
7118                 ret = 0;
7119         }
7120 out:
7121         if (ret == -ENOSPC)
7122                 ins->offset = max_extent_size;
7123         return ret;
7124 }
7125
7126 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7127                             int dump_block_groups)
7128 {
7129         struct btrfs_block_group_cache *cache;
7130         int index = 0;
7131
7132         spin_lock(&info->lock);
7133         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7134                info->flags,
7135                info->total_bytes - info->bytes_used - info->bytes_pinned -
7136                info->bytes_reserved - info->bytes_readonly,
7137                (info->full) ? "" : "not ");
7138         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7139                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7140                info->total_bytes, info->bytes_used, info->bytes_pinned,
7141                info->bytes_reserved, info->bytes_may_use,
7142                info->bytes_readonly);
7143         spin_unlock(&info->lock);
7144
7145         if (!dump_block_groups)
7146                 return;
7147
7148         down_read(&info->groups_sem);
7149 again:
7150         list_for_each_entry(cache, &info->block_groups[index], list) {
7151                 spin_lock(&cache->lock);
7152                 printk(KERN_INFO "BTRFS: "
7153                            "block group %llu has %llu bytes, "
7154                            "%llu used %llu pinned %llu reserved %s\n",
7155                        cache->key.objectid, cache->key.offset,
7156                        btrfs_block_group_used(&cache->item), cache->pinned,
7157                        cache->reserved, cache->ro ? "[readonly]" : "");
7158                 btrfs_dump_free_space(cache, bytes);
7159                 spin_unlock(&cache->lock);
7160         }
7161         if (++index < BTRFS_NR_RAID_TYPES)
7162                 goto again;
7163         up_read(&info->groups_sem);
7164 }
7165
7166 int btrfs_reserve_extent(struct btrfs_root *root,
7167                          u64 num_bytes, u64 min_alloc_size,
7168                          u64 empty_size, u64 hint_byte,
7169                          struct btrfs_key *ins, int is_data, int delalloc)
7170 {
7171         bool final_tried = false;
7172         u64 flags;
7173         int ret;
7174
7175         flags = btrfs_get_alloc_profile(root, is_data);
7176 again:
7177         WARN_ON(num_bytes < root->sectorsize);
7178         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7179                                flags, delalloc);
7180
7181         if (ret == -ENOSPC) {
7182                 if (!final_tried && ins->offset) {
7183                         num_bytes = min(num_bytes >> 1, ins->offset);
7184                         num_bytes = round_down(num_bytes, root->sectorsize);
7185                         num_bytes = max(num_bytes, min_alloc_size);
7186                         if (num_bytes == min_alloc_size)
7187                                 final_tried = true;
7188                         goto again;
7189                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7190                         struct btrfs_space_info *sinfo;
7191
7192                         sinfo = __find_space_info(root->fs_info, flags);
7193                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7194                                 flags, num_bytes);
7195                         if (sinfo)
7196                                 dump_space_info(sinfo, num_bytes, 1);
7197                 }
7198         }
7199
7200         return ret;
7201 }
7202
7203 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7204                                         u64 start, u64 len,
7205                                         int pin, int delalloc)
7206 {
7207         struct btrfs_block_group_cache *cache;
7208         int ret = 0;
7209
7210         cache = btrfs_lookup_block_group(root->fs_info, start);
7211         if (!cache) {
7212                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7213                         start);
7214                 return -ENOSPC;
7215         }
7216
7217         if (pin)
7218                 pin_down_extent(root, cache, start, len, 1);
7219         else {
7220                 if (btrfs_test_opt(root, DISCARD))
7221                         ret = btrfs_discard_extent(root, start, len, NULL);
7222                 btrfs_add_free_space(cache, start, len);
7223                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7224         }
7225
7226         btrfs_put_block_group(cache);
7227
7228         trace_btrfs_reserved_extent_free(root, start, len);
7229
7230         return ret;
7231 }
7232
7233 int btrfs_free_reserved_extent(struct btrfs_root *root,
7234                                u64 start, u64 len, int delalloc)
7235 {
7236         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7237 }
7238
7239 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7240                                        u64 start, u64 len)
7241 {
7242         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7243 }
7244
7245 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7246                                       struct btrfs_root *root,
7247                                       u64 parent, u64 root_objectid,
7248                                       u64 flags, u64 owner, u64 offset,
7249                                       struct btrfs_key *ins, int ref_mod)
7250 {
7251         int ret;
7252         struct btrfs_fs_info *fs_info = root->fs_info;
7253         struct btrfs_extent_item *extent_item;
7254         struct btrfs_extent_inline_ref *iref;
7255         struct btrfs_path *path;
7256         struct extent_buffer *leaf;
7257         int type;
7258         u32 size;
7259
7260         if (parent > 0)
7261                 type = BTRFS_SHARED_DATA_REF_KEY;
7262         else
7263                 type = BTRFS_EXTENT_DATA_REF_KEY;
7264
7265         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7266
7267         path = btrfs_alloc_path();
7268         if (!path)
7269                 return -ENOMEM;
7270
7271         path->leave_spinning = 1;
7272         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7273                                       ins, size);
7274         if (ret) {
7275                 btrfs_free_path(path);
7276                 return ret;
7277         }
7278
7279         leaf = path->nodes[0];
7280         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7281                                      struct btrfs_extent_item);
7282         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7283         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7284         btrfs_set_extent_flags(leaf, extent_item,
7285                                flags | BTRFS_EXTENT_FLAG_DATA);
7286
7287         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7288         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7289         if (parent > 0) {
7290                 struct btrfs_shared_data_ref *ref;
7291                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7292                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7293                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7294         } else {
7295                 struct btrfs_extent_data_ref *ref;
7296                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7297                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7298                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7299                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7300                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7301         }
7302
7303         btrfs_mark_buffer_dirty(path->nodes[0]);
7304         btrfs_free_path(path);
7305
7306         /* Always set parent to 0 here since its exclusive anyway. */
7307         ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7308                                       ins->objectid, ins->offset,
7309                                       BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7310         if (ret)
7311                 return ret;
7312
7313         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7314         if (ret) { /* -ENOENT, logic error */
7315                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7316                         ins->objectid, ins->offset);
7317                 BUG();
7318         }
7319         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7320         return ret;
7321 }
7322
7323 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7324                                      struct btrfs_root *root,
7325                                      u64 parent, u64 root_objectid,
7326                                      u64 flags, struct btrfs_disk_key *key,
7327                                      int level, struct btrfs_key *ins,
7328                                      int no_quota)
7329 {
7330         int ret;
7331         struct btrfs_fs_info *fs_info = root->fs_info;
7332         struct btrfs_extent_item *extent_item;
7333         struct btrfs_tree_block_info *block_info;
7334         struct btrfs_extent_inline_ref *iref;
7335         struct btrfs_path *path;
7336         struct extent_buffer *leaf;
7337         u32 size = sizeof(*extent_item) + sizeof(*iref);
7338         u64 num_bytes = ins->offset;
7339         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7340                                                  SKINNY_METADATA);
7341
7342         if (!skinny_metadata)
7343                 size += sizeof(*block_info);
7344
7345         path = btrfs_alloc_path();
7346         if (!path) {
7347                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7348                                                    root->nodesize);
7349                 return -ENOMEM;
7350         }
7351
7352         path->leave_spinning = 1;
7353         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7354                                       ins, size);
7355         if (ret) {
7356                 btrfs_free_path(path);
7357                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7358                                                    root->nodesize);
7359                 return ret;
7360         }
7361
7362         leaf = path->nodes[0];
7363         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7364                                      struct btrfs_extent_item);
7365         btrfs_set_extent_refs(leaf, extent_item, 1);
7366         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7367         btrfs_set_extent_flags(leaf, extent_item,
7368                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7369
7370         if (skinny_metadata) {
7371                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7372                 num_bytes = root->nodesize;
7373         } else {
7374                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7375                 btrfs_set_tree_block_key(leaf, block_info, key);
7376                 btrfs_set_tree_block_level(leaf, block_info, level);
7377                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7378         }
7379
7380         if (parent > 0) {
7381                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7382                 btrfs_set_extent_inline_ref_type(leaf, iref,
7383                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7384                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7385         } else {
7386                 btrfs_set_extent_inline_ref_type(leaf, iref,
7387                                                  BTRFS_TREE_BLOCK_REF_KEY);
7388                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7389         }
7390
7391         btrfs_mark_buffer_dirty(leaf);
7392         btrfs_free_path(path);
7393
7394         if (!no_quota) {
7395                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7396                                               ins->objectid, num_bytes,
7397                                               BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7398                 if (ret)
7399                         return ret;
7400         }
7401
7402         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7403                                  1);
7404         if (ret) { /* -ENOENT, logic error */
7405                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7406                         ins->objectid, ins->offset);
7407                 BUG();
7408         }
7409
7410         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7411         return ret;
7412 }
7413
7414 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7415                                      struct btrfs_root *root,
7416                                      u64 root_objectid, u64 owner,
7417                                      u64 offset, struct btrfs_key *ins)
7418 {
7419         int ret;
7420
7421         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7422
7423         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7424                                          ins->offset, 0,
7425                                          root_objectid, owner, offset,
7426                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7427         return ret;
7428 }
7429
7430 /*
7431  * this is used by the tree logging recovery code.  It records that
7432  * an extent has been allocated and makes sure to clear the free
7433  * space cache bits as well
7434  */
7435 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7436                                    struct btrfs_root *root,
7437                                    u64 root_objectid, u64 owner, u64 offset,
7438                                    struct btrfs_key *ins)
7439 {
7440         int ret;
7441         struct btrfs_block_group_cache *block_group;
7442
7443         /*
7444          * Mixed block groups will exclude before processing the log so we only
7445          * need to do the exlude dance if this fs isn't mixed.
7446          */
7447         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7448                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7449                 if (ret)
7450                         return ret;
7451         }
7452
7453         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7454         if (!block_group)
7455                 return -EINVAL;
7456
7457         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7458                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7459         BUG_ON(ret); /* logic error */
7460         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7461                                          0, owner, offset, ins, 1);
7462         btrfs_put_block_group(block_group);
7463         return ret;
7464 }
7465
7466 static struct extent_buffer *
7467 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7468                       u64 bytenr, int level)
7469 {
7470         struct extent_buffer *buf;
7471
7472         buf = btrfs_find_create_tree_block(root, bytenr);
7473         if (!buf)
7474                 return ERR_PTR(-ENOMEM);
7475         btrfs_set_header_generation(buf, trans->transid);
7476         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7477         btrfs_tree_lock(buf);
7478         clean_tree_block(trans, root->fs_info, buf);
7479         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7480
7481         btrfs_set_lock_blocking(buf);
7482         btrfs_set_buffer_uptodate(buf);
7483
7484         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7485                 buf->log_index = root->log_transid % 2;
7486                 /*
7487                  * we allow two log transactions at a time, use different
7488                  * EXENT bit to differentiate dirty pages.
7489                  */
7490                 if (buf->log_index == 0)
7491                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7492                                         buf->start + buf->len - 1, GFP_NOFS);
7493                 else
7494                         set_extent_new(&root->dirty_log_pages, buf->start,
7495                                         buf->start + buf->len - 1, GFP_NOFS);
7496         } else {
7497                 buf->log_index = -1;
7498                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7499                          buf->start + buf->len - 1, GFP_NOFS);
7500         }
7501         trans->blocks_used++;
7502         /* this returns a buffer locked for blocking */
7503         return buf;
7504 }
7505
7506 static struct btrfs_block_rsv *
7507 use_block_rsv(struct btrfs_trans_handle *trans,
7508               struct btrfs_root *root, u32 blocksize)
7509 {
7510         struct btrfs_block_rsv *block_rsv;
7511         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7512         int ret;
7513         bool global_updated = false;
7514
7515         block_rsv = get_block_rsv(trans, root);
7516
7517         if (unlikely(block_rsv->size == 0))
7518                 goto try_reserve;
7519 again:
7520         ret = block_rsv_use_bytes(block_rsv, blocksize);
7521         if (!ret)
7522                 return block_rsv;
7523
7524         if (block_rsv->failfast)
7525                 return ERR_PTR(ret);
7526
7527         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7528                 global_updated = true;
7529                 update_global_block_rsv(root->fs_info);
7530                 goto again;
7531         }
7532
7533         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7534                 static DEFINE_RATELIMIT_STATE(_rs,
7535                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7536                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7537                 if (__ratelimit(&_rs))
7538                         WARN(1, KERN_DEBUG
7539                                 "BTRFS: block rsv returned %d\n", ret);
7540         }
7541 try_reserve:
7542         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7543                                      BTRFS_RESERVE_NO_FLUSH);
7544         if (!ret)
7545                 return block_rsv;
7546         /*
7547          * If we couldn't reserve metadata bytes try and use some from
7548          * the global reserve if its space type is the same as the global
7549          * reservation.
7550          */
7551         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7552             block_rsv->space_info == global_rsv->space_info) {
7553                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7554                 if (!ret)
7555                         return global_rsv;
7556         }
7557         return ERR_PTR(ret);
7558 }
7559
7560 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7561                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7562 {
7563         block_rsv_add_bytes(block_rsv, blocksize, 0);
7564         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7565 }
7566
7567 /*
7568  * finds a free extent and does all the dirty work required for allocation
7569  * returns the key for the extent through ins, and a tree buffer for
7570  * the first block of the extent through buf.
7571  *
7572  * returns the tree buffer or an ERR_PTR on error.
7573  */
7574 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7575                                         struct btrfs_root *root,
7576                                         u64 parent, u64 root_objectid,
7577                                         struct btrfs_disk_key *key, int level,
7578                                         u64 hint, u64 empty_size)
7579 {
7580         struct btrfs_key ins;
7581         struct btrfs_block_rsv *block_rsv;
7582         struct extent_buffer *buf;
7583         struct btrfs_delayed_extent_op *extent_op;
7584         u64 flags = 0;
7585         int ret;
7586         u32 blocksize = root->nodesize;
7587         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7588                                                  SKINNY_METADATA);
7589
7590         if (btrfs_test_is_dummy_root(root)) {
7591                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7592                                             level);
7593                 if (!IS_ERR(buf))
7594                         root->alloc_bytenr += blocksize;
7595                 return buf;
7596         }
7597
7598         block_rsv = use_block_rsv(trans, root, blocksize);
7599         if (IS_ERR(block_rsv))
7600                 return ERR_CAST(block_rsv);
7601
7602         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7603                                    empty_size, hint, &ins, 0, 0);
7604         if (ret)
7605                 goto out_unuse;
7606
7607         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
7608         if (IS_ERR(buf)) {
7609                 ret = PTR_ERR(buf);
7610                 goto out_free_reserved;
7611         }
7612
7613         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7614                 if (parent == 0)
7615                         parent = ins.objectid;
7616                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7617         } else
7618                 BUG_ON(parent > 0);
7619
7620         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7621                 extent_op = btrfs_alloc_delayed_extent_op();
7622                 if (!extent_op) {
7623                         ret = -ENOMEM;
7624                         goto out_free_buf;
7625                 }
7626                 if (key)
7627                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7628                 else
7629                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7630                 extent_op->flags_to_set = flags;
7631                 if (skinny_metadata)
7632                         extent_op->update_key = 0;
7633                 else
7634                         extent_op->update_key = 1;
7635                 extent_op->update_flags = 1;
7636                 extent_op->is_data = 0;
7637                 extent_op->level = level;
7638
7639                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7640                                                  ins.objectid, ins.offset,
7641                                                  parent, root_objectid, level,
7642                                                  BTRFS_ADD_DELAYED_EXTENT,
7643                                                  extent_op, 0);
7644                 if (ret)
7645                         goto out_free_delayed;
7646         }
7647         return buf;
7648
7649 out_free_delayed:
7650         btrfs_free_delayed_extent_op(extent_op);
7651 out_free_buf:
7652         free_extent_buffer(buf);
7653 out_free_reserved:
7654         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
7655 out_unuse:
7656         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7657         return ERR_PTR(ret);
7658 }
7659
7660 struct walk_control {
7661         u64 refs[BTRFS_MAX_LEVEL];
7662         u64 flags[BTRFS_MAX_LEVEL];
7663         struct btrfs_key update_progress;
7664         int stage;
7665         int level;
7666         int shared_level;
7667         int update_ref;
7668         int keep_locks;
7669         int reada_slot;
7670         int reada_count;
7671         int for_reloc;
7672 };
7673
7674 #define DROP_REFERENCE  1
7675 #define UPDATE_BACKREF  2
7676
7677 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7678                                      struct btrfs_root *root,
7679                                      struct walk_control *wc,
7680                                      struct btrfs_path *path)
7681 {
7682         u64 bytenr;
7683         u64 generation;
7684         u64 refs;
7685         u64 flags;
7686         u32 nritems;
7687         u32 blocksize;
7688         struct btrfs_key key;
7689         struct extent_buffer *eb;
7690         int ret;
7691         int slot;
7692         int nread = 0;
7693
7694         if (path->slots[wc->level] < wc->reada_slot) {
7695                 wc->reada_count = wc->reada_count * 2 / 3;
7696                 wc->reada_count = max(wc->reada_count, 2);
7697         } else {
7698                 wc->reada_count = wc->reada_count * 3 / 2;
7699                 wc->reada_count = min_t(int, wc->reada_count,
7700                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7701         }
7702
7703         eb = path->nodes[wc->level];
7704         nritems = btrfs_header_nritems(eb);
7705         blocksize = root->nodesize;
7706
7707         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7708                 if (nread >= wc->reada_count)
7709                         break;
7710
7711                 cond_resched();
7712                 bytenr = btrfs_node_blockptr(eb, slot);
7713                 generation = btrfs_node_ptr_generation(eb, slot);
7714
7715                 if (slot == path->slots[wc->level])
7716                         goto reada;
7717
7718                 if (wc->stage == UPDATE_BACKREF &&
7719                     generation <= root->root_key.offset)
7720                         continue;
7721
7722                 /* We don't lock the tree block, it's OK to be racy here */
7723                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7724                                                wc->level - 1, 1, &refs,
7725                                                &flags);
7726                 /* We don't care about errors in readahead. */
7727                 if (ret < 0)
7728                         continue;
7729                 BUG_ON(refs == 0);
7730
7731                 if (wc->stage == DROP_REFERENCE) {
7732                         if (refs == 1)
7733                                 goto reada;
7734
7735                         if (wc->level == 1 &&
7736                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7737                                 continue;
7738                         if (!wc->update_ref ||
7739                             generation <= root->root_key.offset)
7740                                 continue;
7741                         btrfs_node_key_to_cpu(eb, &key, slot);
7742                         ret = btrfs_comp_cpu_keys(&key,
7743                                                   &wc->update_progress);
7744                         if (ret < 0)
7745                                 continue;
7746                 } else {
7747                         if (wc->level == 1 &&
7748                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7749                                 continue;
7750                 }
7751 reada:
7752                 readahead_tree_block(root, bytenr);
7753                 nread++;
7754         }
7755         wc->reada_slot = slot;
7756 }
7757
7758 static int account_leaf_items(struct btrfs_trans_handle *trans,
7759                               struct btrfs_root *root,
7760                               struct extent_buffer *eb)
7761 {
7762         int nr = btrfs_header_nritems(eb);
7763         int i, extent_type, ret;
7764         struct btrfs_key key;
7765         struct btrfs_file_extent_item *fi;
7766         u64 bytenr, num_bytes;
7767
7768         for (i = 0; i < nr; i++) {
7769                 btrfs_item_key_to_cpu(eb, &key, i);
7770
7771                 if (key.type != BTRFS_EXTENT_DATA_KEY)
7772                         continue;
7773
7774                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
7775                 /* filter out non qgroup-accountable extents  */
7776                 extent_type = btrfs_file_extent_type(eb, fi);
7777
7778                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
7779                         continue;
7780
7781                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
7782                 if (!bytenr)
7783                         continue;
7784
7785                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
7786
7787                 ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7788                                               root->objectid,
7789                                               bytenr, num_bytes,
7790                                               BTRFS_QGROUP_OPER_SUB_SUBTREE, 0);
7791                 if (ret)
7792                         return ret;
7793         }
7794         return 0;
7795 }
7796
7797 /*
7798  * Walk up the tree from the bottom, freeing leaves and any interior
7799  * nodes which have had all slots visited. If a node (leaf or
7800  * interior) is freed, the node above it will have it's slot
7801  * incremented. The root node will never be freed.
7802  *
7803  * At the end of this function, we should have a path which has all
7804  * slots incremented to the next position for a search. If we need to
7805  * read a new node it will be NULL and the node above it will have the
7806  * correct slot selected for a later read.
7807  *
7808  * If we increment the root nodes slot counter past the number of
7809  * elements, 1 is returned to signal completion of the search.
7810  */
7811 static int adjust_slots_upwards(struct btrfs_root *root,
7812                                 struct btrfs_path *path, int root_level)
7813 {
7814         int level = 0;
7815         int nr, slot;
7816         struct extent_buffer *eb;
7817
7818         if (root_level == 0)
7819                 return 1;
7820
7821         while (level <= root_level) {
7822                 eb = path->nodes[level];
7823                 nr = btrfs_header_nritems(eb);
7824                 path->slots[level]++;
7825                 slot = path->slots[level];
7826                 if (slot >= nr || level == 0) {
7827                         /*
7828                          * Don't free the root -  we will detect this
7829                          * condition after our loop and return a
7830                          * positive value for caller to stop walking the tree.
7831                          */
7832                         if (level != root_level) {
7833                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7834                                 path->locks[level] = 0;
7835
7836                                 free_extent_buffer(eb);
7837                                 path->nodes[level] = NULL;
7838                                 path->slots[level] = 0;
7839                         }
7840                 } else {
7841                         /*
7842                          * We have a valid slot to walk back down
7843                          * from. Stop here so caller can process these
7844                          * new nodes.
7845                          */
7846                         break;
7847                 }
7848
7849                 level++;
7850         }
7851
7852         eb = path->nodes[root_level];
7853         if (path->slots[root_level] >= btrfs_header_nritems(eb))
7854                 return 1;
7855
7856         return 0;
7857 }
7858
7859 /*
7860  * root_eb is the subtree root and is locked before this function is called.
7861  */
7862 static int account_shared_subtree(struct btrfs_trans_handle *trans,
7863                                   struct btrfs_root *root,
7864                                   struct extent_buffer *root_eb,
7865                                   u64 root_gen,
7866                                   int root_level)
7867 {
7868         int ret = 0;
7869         int level;
7870         struct extent_buffer *eb = root_eb;
7871         struct btrfs_path *path = NULL;
7872
7873         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
7874         BUG_ON(root_eb == NULL);
7875
7876         if (!root->fs_info->quota_enabled)
7877                 return 0;
7878
7879         if (!extent_buffer_uptodate(root_eb)) {
7880                 ret = btrfs_read_buffer(root_eb, root_gen);
7881                 if (ret)
7882                         goto out;
7883         }
7884
7885         if (root_level == 0) {
7886                 ret = account_leaf_items(trans, root, root_eb);
7887                 goto out;
7888         }
7889
7890         path = btrfs_alloc_path();
7891         if (!path)
7892                 return -ENOMEM;
7893
7894         /*
7895          * Walk down the tree.  Missing extent blocks are filled in as
7896          * we go. Metadata is accounted every time we read a new
7897          * extent block.
7898          *
7899          * When we reach a leaf, we account for file extent items in it,
7900          * walk back up the tree (adjusting slot pointers as we go)
7901          * and restart the search process.
7902          */
7903         extent_buffer_get(root_eb); /* For path */
7904         path->nodes[root_level] = root_eb;
7905         path->slots[root_level] = 0;
7906         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
7907 walk_down:
7908         level = root_level;
7909         while (level >= 0) {
7910                 if (path->nodes[level] == NULL) {
7911                         int parent_slot;
7912                         u64 child_gen;
7913                         u64 child_bytenr;
7914
7915                         /* We need to get child blockptr/gen from
7916                          * parent before we can read it. */
7917                         eb = path->nodes[level + 1];
7918                         parent_slot = path->slots[level + 1];
7919                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
7920                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
7921
7922                         eb = read_tree_block(root, child_bytenr, child_gen);
7923                         if (!eb || !extent_buffer_uptodate(eb)) {
7924                                 ret = -EIO;
7925                                 goto out;
7926                         }
7927
7928                         path->nodes[level] = eb;
7929                         path->slots[level] = 0;
7930
7931                         btrfs_tree_read_lock(eb);
7932                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
7933                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
7934
7935                         ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7936                                                 root->objectid,
7937                                                 child_bytenr,
7938                                                 root->nodesize,
7939                                                 BTRFS_QGROUP_OPER_SUB_SUBTREE,
7940                                                 0);
7941                         if (ret)
7942                                 goto out;
7943
7944                 }
7945
7946                 if (level == 0) {
7947                         ret = account_leaf_items(trans, root, path->nodes[level]);
7948                         if (ret)
7949                                 goto out;
7950
7951                         /* Nonzero return here means we completed our search */
7952                         ret = adjust_slots_upwards(root, path, root_level);
7953                         if (ret)
7954                                 break;
7955
7956                         /* Restart search with new slots */
7957                         goto walk_down;
7958                 }
7959
7960                 level--;
7961         }
7962
7963         ret = 0;
7964 out:
7965         btrfs_free_path(path);
7966
7967         return ret;
7968 }
7969
7970 /*
7971  * helper to process tree block while walking down the tree.
7972  *
7973  * when wc->stage == UPDATE_BACKREF, this function updates
7974  * back refs for pointers in the block.
7975  *
7976  * NOTE: return value 1 means we should stop walking down.
7977  */
7978 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7979                                    struct btrfs_root *root,
7980                                    struct btrfs_path *path,
7981                                    struct walk_control *wc, int lookup_info)
7982 {
7983         int level = wc->level;
7984         struct extent_buffer *eb = path->nodes[level];
7985         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7986         int ret;
7987
7988         if (wc->stage == UPDATE_BACKREF &&
7989             btrfs_header_owner(eb) != root->root_key.objectid)
7990                 return 1;
7991
7992         /*
7993          * when reference count of tree block is 1, it won't increase
7994          * again. once full backref flag is set, we never clear it.
7995          */
7996         if (lookup_info &&
7997             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7998              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7999                 BUG_ON(!path->locks[level]);
8000                 ret = btrfs_lookup_extent_info(trans, root,
8001                                                eb->start, level, 1,
8002                                                &wc->refs[level],
8003                                                &wc->flags[level]);
8004                 BUG_ON(ret == -ENOMEM);
8005                 if (ret)
8006                         return ret;
8007                 BUG_ON(wc->refs[level] == 0);
8008         }
8009
8010         if (wc->stage == DROP_REFERENCE) {
8011                 if (wc->refs[level] > 1)
8012                         return 1;
8013
8014                 if (path->locks[level] && !wc->keep_locks) {
8015                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8016                         path->locks[level] = 0;
8017                 }
8018                 return 0;
8019         }
8020
8021         /* wc->stage == UPDATE_BACKREF */
8022         if (!(wc->flags[level] & flag)) {
8023                 BUG_ON(!path->locks[level]);
8024                 ret = btrfs_inc_ref(trans, root, eb, 1);
8025                 BUG_ON(ret); /* -ENOMEM */
8026                 ret = btrfs_dec_ref(trans, root, eb, 0);
8027                 BUG_ON(ret); /* -ENOMEM */
8028                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8029                                                   eb->len, flag,
8030                                                   btrfs_header_level(eb), 0);
8031                 BUG_ON(ret); /* -ENOMEM */
8032                 wc->flags[level] |= flag;
8033         }
8034
8035         /*
8036          * the block is shared by multiple trees, so it's not good to
8037          * keep the tree lock
8038          */
8039         if (path->locks[level] && level > 0) {
8040                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8041                 path->locks[level] = 0;
8042         }
8043         return 0;
8044 }
8045
8046 /*
8047  * helper to process tree block pointer.
8048  *
8049  * when wc->stage == DROP_REFERENCE, this function checks
8050  * reference count of the block pointed to. if the block
8051  * is shared and we need update back refs for the subtree
8052  * rooted at the block, this function changes wc->stage to
8053  * UPDATE_BACKREF. if the block is shared and there is no
8054  * need to update back, this function drops the reference
8055  * to the block.
8056  *
8057  * NOTE: return value 1 means we should stop walking down.
8058  */
8059 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8060                                  struct btrfs_root *root,
8061                                  struct btrfs_path *path,
8062                                  struct walk_control *wc, int *lookup_info)
8063 {
8064         u64 bytenr;
8065         u64 generation;
8066         u64 parent;
8067         u32 blocksize;
8068         struct btrfs_key key;
8069         struct extent_buffer *next;
8070         int level = wc->level;
8071         int reada = 0;
8072         int ret = 0;
8073         bool need_account = false;
8074
8075         generation = btrfs_node_ptr_generation(path->nodes[level],
8076                                                path->slots[level]);
8077         /*
8078          * if the lower level block was created before the snapshot
8079          * was created, we know there is no need to update back refs
8080          * for the subtree
8081          */
8082         if (wc->stage == UPDATE_BACKREF &&
8083             generation <= root->root_key.offset) {
8084                 *lookup_info = 1;
8085                 return 1;
8086         }
8087
8088         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8089         blocksize = root->nodesize;
8090
8091         next = btrfs_find_tree_block(root->fs_info, bytenr);
8092         if (!next) {
8093                 next = btrfs_find_create_tree_block(root, bytenr);
8094                 if (!next)
8095                         return -ENOMEM;
8096                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8097                                                level - 1);
8098                 reada = 1;
8099         }
8100         btrfs_tree_lock(next);
8101         btrfs_set_lock_blocking(next);
8102
8103         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8104                                        &wc->refs[level - 1],
8105                                        &wc->flags[level - 1]);
8106         if (ret < 0) {
8107                 btrfs_tree_unlock(next);
8108                 return ret;
8109         }
8110
8111         if (unlikely(wc->refs[level - 1] == 0)) {
8112                 btrfs_err(root->fs_info, "Missing references.");
8113                 BUG();
8114         }
8115         *lookup_info = 0;
8116
8117         if (wc->stage == DROP_REFERENCE) {
8118                 if (wc->refs[level - 1] > 1) {
8119                         need_account = true;
8120                         if (level == 1 &&
8121                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8122                                 goto skip;
8123
8124                         if (!wc->update_ref ||
8125                             generation <= root->root_key.offset)
8126                                 goto skip;
8127
8128                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8129                                               path->slots[level]);
8130                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8131                         if (ret < 0)
8132                                 goto skip;
8133
8134                         wc->stage = UPDATE_BACKREF;
8135                         wc->shared_level = level - 1;
8136                 }
8137         } else {
8138                 if (level == 1 &&
8139                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8140                         goto skip;
8141         }
8142
8143         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8144                 btrfs_tree_unlock(next);
8145                 free_extent_buffer(next);
8146                 next = NULL;
8147                 *lookup_info = 1;
8148         }
8149
8150         if (!next) {
8151                 if (reada && level == 1)
8152                         reada_walk_down(trans, root, wc, path);
8153                 next = read_tree_block(root, bytenr, generation);
8154                 if (!next || !extent_buffer_uptodate(next)) {
8155                         free_extent_buffer(next);
8156                         return -EIO;
8157                 }
8158                 btrfs_tree_lock(next);
8159                 btrfs_set_lock_blocking(next);
8160         }
8161
8162         level--;
8163         BUG_ON(level != btrfs_header_level(next));
8164         path->nodes[level] = next;
8165         path->slots[level] = 0;
8166         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8167         wc->level = level;
8168         if (wc->level == 1)
8169                 wc->reada_slot = 0;
8170         return 0;
8171 skip:
8172         wc->refs[level - 1] = 0;
8173         wc->flags[level - 1] = 0;
8174         if (wc->stage == DROP_REFERENCE) {
8175                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8176                         parent = path->nodes[level]->start;
8177                 } else {
8178                         BUG_ON(root->root_key.objectid !=
8179                                btrfs_header_owner(path->nodes[level]));
8180                         parent = 0;
8181                 }
8182
8183                 if (need_account) {
8184                         ret = account_shared_subtree(trans, root, next,
8185                                                      generation, level - 1);
8186                         if (ret) {
8187                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8188                                         "%d accounting shared subtree. Quota "
8189                                         "is out of sync, rescan required.\n",
8190                                         root->fs_info->sb->s_id, ret);
8191                         }
8192                 }
8193                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8194                                 root->root_key.objectid, level - 1, 0, 0);
8195                 BUG_ON(ret); /* -ENOMEM */
8196         }
8197         btrfs_tree_unlock(next);
8198         free_extent_buffer(next);
8199         *lookup_info = 1;
8200         return 1;
8201 }
8202
8203 /*
8204  * helper to process tree block while walking up the tree.
8205  *
8206  * when wc->stage == DROP_REFERENCE, this function drops
8207  * reference count on the block.
8208  *
8209  * when wc->stage == UPDATE_BACKREF, this function changes
8210  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8211  * to UPDATE_BACKREF previously while processing the block.
8212  *
8213  * NOTE: return value 1 means we should stop walking up.
8214  */
8215 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8216                                  struct btrfs_root *root,
8217                                  struct btrfs_path *path,
8218                                  struct walk_control *wc)
8219 {
8220         int ret;
8221         int level = wc->level;
8222         struct extent_buffer *eb = path->nodes[level];
8223         u64 parent = 0;
8224
8225         if (wc->stage == UPDATE_BACKREF) {
8226                 BUG_ON(wc->shared_level < level);
8227                 if (level < wc->shared_level)
8228                         goto out;
8229
8230                 ret = find_next_key(path, level + 1, &wc->update_progress);
8231                 if (ret > 0)
8232                         wc->update_ref = 0;
8233
8234                 wc->stage = DROP_REFERENCE;
8235                 wc->shared_level = -1;
8236                 path->slots[level] = 0;
8237
8238                 /*
8239                  * check reference count again if the block isn't locked.
8240                  * we should start walking down the tree again if reference
8241                  * count is one.
8242                  */
8243                 if (!path->locks[level]) {
8244                         BUG_ON(level == 0);
8245                         btrfs_tree_lock(eb);
8246                         btrfs_set_lock_blocking(eb);
8247                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8248
8249                         ret = btrfs_lookup_extent_info(trans, root,
8250                                                        eb->start, level, 1,
8251                                                        &wc->refs[level],
8252                                                        &wc->flags[level]);
8253                         if (ret < 0) {
8254                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8255                                 path->locks[level] = 0;
8256                                 return ret;
8257                         }
8258                         BUG_ON(wc->refs[level] == 0);
8259                         if (wc->refs[level] == 1) {
8260                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8261                                 path->locks[level] = 0;
8262                                 return 1;
8263                         }
8264                 }
8265         }
8266
8267         /* wc->stage == DROP_REFERENCE */
8268         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8269
8270         if (wc->refs[level] == 1) {
8271                 if (level == 0) {
8272                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8273                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8274                         else
8275                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8276                         BUG_ON(ret); /* -ENOMEM */
8277                         ret = account_leaf_items(trans, root, eb);
8278                         if (ret) {
8279                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8280                                         "%d accounting leaf items. Quota "
8281                                         "is out of sync, rescan required.\n",
8282                                         root->fs_info->sb->s_id, ret);
8283                         }
8284                 }
8285                 /* make block locked assertion in clean_tree_block happy */
8286                 if (!path->locks[level] &&
8287                     btrfs_header_generation(eb) == trans->transid) {
8288                         btrfs_tree_lock(eb);
8289                         btrfs_set_lock_blocking(eb);
8290                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8291                 }
8292                 clean_tree_block(trans, root->fs_info, eb);
8293         }
8294
8295         if (eb == root->node) {
8296                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8297                         parent = eb->start;
8298                 else
8299                         BUG_ON(root->root_key.objectid !=
8300                                btrfs_header_owner(eb));
8301         } else {
8302                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8303                         parent = path->nodes[level + 1]->start;
8304                 else
8305                         BUG_ON(root->root_key.objectid !=
8306                                btrfs_header_owner(path->nodes[level + 1]));
8307         }
8308
8309         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8310 out:
8311         wc->refs[level] = 0;
8312         wc->flags[level] = 0;
8313         return 0;
8314 }
8315
8316 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8317                                    struct btrfs_root *root,
8318                                    struct btrfs_path *path,
8319                                    struct walk_control *wc)
8320 {
8321         int level = wc->level;
8322         int lookup_info = 1;
8323         int ret;
8324
8325         while (level >= 0) {
8326                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8327                 if (ret > 0)
8328                         break;
8329
8330                 if (level == 0)
8331                         break;
8332
8333                 if (path->slots[level] >=
8334                     btrfs_header_nritems(path->nodes[level]))
8335                         break;
8336
8337                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8338                 if (ret > 0) {
8339                         path->slots[level]++;
8340                         continue;
8341                 } else if (ret < 0)
8342                         return ret;
8343                 level = wc->level;
8344         }
8345         return 0;
8346 }
8347
8348 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8349                                  struct btrfs_root *root,
8350                                  struct btrfs_path *path,
8351                                  struct walk_control *wc, int max_level)
8352 {
8353         int level = wc->level;
8354         int ret;
8355
8356         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8357         while (level < max_level && path->nodes[level]) {
8358                 wc->level = level;
8359                 if (path->slots[level] + 1 <
8360                     btrfs_header_nritems(path->nodes[level])) {
8361                         path->slots[level]++;
8362                         return 0;
8363                 } else {
8364                         ret = walk_up_proc(trans, root, path, wc);
8365                         if (ret > 0)
8366                                 return 0;
8367
8368                         if (path->locks[level]) {
8369                                 btrfs_tree_unlock_rw(path->nodes[level],
8370                                                      path->locks[level]);
8371                                 path->locks[level] = 0;
8372                         }
8373                         free_extent_buffer(path->nodes[level]);
8374                         path->nodes[level] = NULL;
8375                         level++;
8376                 }
8377         }
8378         return 1;
8379 }
8380
8381 /*
8382  * drop a subvolume tree.
8383  *
8384  * this function traverses the tree freeing any blocks that only
8385  * referenced by the tree.
8386  *
8387  * when a shared tree block is found. this function decreases its
8388  * reference count by one. if update_ref is true, this function
8389  * also make sure backrefs for the shared block and all lower level
8390  * blocks are properly updated.
8391  *
8392  * If called with for_reloc == 0, may exit early with -EAGAIN
8393  */
8394 int btrfs_drop_snapshot(struct btrfs_root *root,
8395                          struct btrfs_block_rsv *block_rsv, int update_ref,
8396                          int for_reloc)
8397 {
8398         struct btrfs_path *path;
8399         struct btrfs_trans_handle *trans;
8400         struct btrfs_root *tree_root = root->fs_info->tree_root;
8401         struct btrfs_root_item *root_item = &root->root_item;
8402         struct walk_control *wc;
8403         struct btrfs_key key;
8404         int err = 0;
8405         int ret;
8406         int level;
8407         bool root_dropped = false;
8408
8409         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8410
8411         path = btrfs_alloc_path();
8412         if (!path) {
8413                 err = -ENOMEM;
8414                 goto out;
8415         }
8416
8417         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8418         if (!wc) {
8419                 btrfs_free_path(path);
8420                 err = -ENOMEM;
8421                 goto out;
8422         }
8423
8424         trans = btrfs_start_transaction(tree_root, 0);
8425         if (IS_ERR(trans)) {
8426                 err = PTR_ERR(trans);
8427                 goto out_free;
8428         }
8429
8430         if (block_rsv)
8431                 trans->block_rsv = block_rsv;
8432
8433         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8434                 level = btrfs_header_level(root->node);
8435                 path->nodes[level] = btrfs_lock_root_node(root);
8436                 btrfs_set_lock_blocking(path->nodes[level]);
8437                 path->slots[level] = 0;
8438                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8439                 memset(&wc->update_progress, 0,
8440                        sizeof(wc->update_progress));
8441         } else {
8442                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8443                 memcpy(&wc->update_progress, &key,
8444                        sizeof(wc->update_progress));
8445
8446                 level = root_item->drop_level;
8447                 BUG_ON(level == 0);
8448                 path->lowest_level = level;
8449                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8450                 path->lowest_level = 0;
8451                 if (ret < 0) {
8452                         err = ret;
8453                         goto out_end_trans;
8454                 }
8455                 WARN_ON(ret > 0);
8456
8457                 /*
8458                  * unlock our path, this is safe because only this
8459                  * function is allowed to delete this snapshot
8460                  */
8461                 btrfs_unlock_up_safe(path, 0);
8462
8463                 level = btrfs_header_level(root->node);
8464                 while (1) {
8465                         btrfs_tree_lock(path->nodes[level]);
8466                         btrfs_set_lock_blocking(path->nodes[level]);
8467                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8468
8469                         ret = btrfs_lookup_extent_info(trans, root,
8470                                                 path->nodes[level]->start,
8471                                                 level, 1, &wc->refs[level],
8472                                                 &wc->flags[level]);
8473                         if (ret < 0) {
8474                                 err = ret;
8475                                 goto out_end_trans;
8476                         }
8477                         BUG_ON(wc->refs[level] == 0);
8478
8479                         if (level == root_item->drop_level)
8480                                 break;
8481
8482                         btrfs_tree_unlock(path->nodes[level]);
8483                         path->locks[level] = 0;
8484                         WARN_ON(wc->refs[level] != 1);
8485                         level--;
8486                 }
8487         }
8488
8489         wc->level = level;
8490         wc->shared_level = -1;
8491         wc->stage = DROP_REFERENCE;
8492         wc->update_ref = update_ref;
8493         wc->keep_locks = 0;
8494         wc->for_reloc = for_reloc;
8495         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8496
8497         while (1) {
8498
8499                 ret = walk_down_tree(trans, root, path, wc);
8500                 if (ret < 0) {
8501                         err = ret;
8502                         break;
8503                 }
8504
8505                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8506                 if (ret < 0) {
8507                         err = ret;
8508                         break;
8509                 }
8510
8511                 if (ret > 0) {
8512                         BUG_ON(wc->stage != DROP_REFERENCE);
8513                         break;
8514                 }
8515
8516                 if (wc->stage == DROP_REFERENCE) {
8517                         level = wc->level;
8518                         btrfs_node_key(path->nodes[level],
8519                                        &root_item->drop_progress,
8520                                        path->slots[level]);
8521                         root_item->drop_level = level;
8522                 }
8523
8524                 BUG_ON(wc->level == 0);
8525                 if (btrfs_should_end_transaction(trans, tree_root) ||
8526                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8527                         ret = btrfs_update_root(trans, tree_root,
8528                                                 &root->root_key,
8529                                                 root_item);
8530                         if (ret) {
8531                                 btrfs_abort_transaction(trans, tree_root, ret);
8532                                 err = ret;
8533                                 goto out_end_trans;
8534                         }
8535
8536                         /*
8537                          * Qgroup update accounting is run from
8538                          * delayed ref handling. This usually works
8539                          * out because delayed refs are normally the
8540                          * only way qgroup updates are added. However,
8541                          * we may have added updates during our tree
8542                          * walk so run qgroups here to make sure we
8543                          * don't lose any updates.
8544                          */
8545                         ret = btrfs_delayed_qgroup_accounting(trans,
8546                                                               root->fs_info);
8547                         if (ret)
8548                                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8549                                                    "running qgroup updates "
8550                                                    "during snapshot delete. "
8551                                                    "Quota is out of sync, "
8552                                                    "rescan required.\n", ret);
8553
8554                         btrfs_end_transaction_throttle(trans, tree_root);
8555                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8556                                 pr_debug("BTRFS: drop snapshot early exit\n");
8557                                 err = -EAGAIN;
8558                                 goto out_free;
8559                         }
8560
8561                         trans = btrfs_start_transaction(tree_root, 0);
8562                         if (IS_ERR(trans)) {
8563                                 err = PTR_ERR(trans);
8564                                 goto out_free;
8565                         }
8566                         if (block_rsv)
8567                                 trans->block_rsv = block_rsv;
8568                 }
8569         }
8570         btrfs_release_path(path);
8571         if (err)
8572                 goto out_end_trans;
8573
8574         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8575         if (ret) {
8576                 btrfs_abort_transaction(trans, tree_root, ret);
8577                 goto out_end_trans;
8578         }
8579
8580         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8581                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8582                                       NULL, NULL);
8583                 if (ret < 0) {
8584                         btrfs_abort_transaction(trans, tree_root, ret);
8585                         err = ret;
8586                         goto out_end_trans;
8587                 } else if (ret > 0) {
8588                         /* if we fail to delete the orphan item this time
8589                          * around, it'll get picked up the next time.
8590                          *
8591                          * The most common failure here is just -ENOENT.
8592                          */
8593                         btrfs_del_orphan_item(trans, tree_root,
8594                                               root->root_key.objectid);
8595                 }
8596         }
8597
8598         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8599                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
8600         } else {
8601                 free_extent_buffer(root->node);
8602                 free_extent_buffer(root->commit_root);
8603                 btrfs_put_fs_root(root);
8604         }
8605         root_dropped = true;
8606 out_end_trans:
8607         ret = btrfs_delayed_qgroup_accounting(trans, tree_root->fs_info);
8608         if (ret)
8609                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8610                                    "running qgroup updates "
8611                                    "during snapshot delete. "
8612                                    "Quota is out of sync, "
8613                                    "rescan required.\n", ret);
8614
8615         btrfs_end_transaction_throttle(trans, tree_root);
8616 out_free:
8617         kfree(wc);
8618         btrfs_free_path(path);
8619 out:
8620         /*
8621          * So if we need to stop dropping the snapshot for whatever reason we
8622          * need to make sure to add it back to the dead root list so that we
8623          * keep trying to do the work later.  This also cleans up roots if we
8624          * don't have it in the radix (like when we recover after a power fail
8625          * or unmount) so we don't leak memory.
8626          */
8627         if (!for_reloc && root_dropped == false)
8628                 btrfs_add_dead_root(root);
8629         if (err && err != -EAGAIN)
8630                 btrfs_std_error(root->fs_info, err);
8631         return err;
8632 }
8633
8634 /*
8635  * drop subtree rooted at tree block 'node'.
8636  *
8637  * NOTE: this function will unlock and release tree block 'node'
8638  * only used by relocation code
8639  */
8640 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8641                         struct btrfs_root *root,
8642                         struct extent_buffer *node,
8643                         struct extent_buffer *parent)
8644 {
8645         struct btrfs_path *path;
8646         struct walk_control *wc;
8647         int level;
8648         int parent_level;
8649         int ret = 0;
8650         int wret;
8651
8652         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8653
8654         path = btrfs_alloc_path();
8655         if (!path)
8656                 return -ENOMEM;
8657
8658         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8659         if (!wc) {
8660                 btrfs_free_path(path);
8661                 return -ENOMEM;
8662         }
8663
8664         btrfs_assert_tree_locked(parent);
8665         parent_level = btrfs_header_level(parent);
8666         extent_buffer_get(parent);
8667         path->nodes[parent_level] = parent;
8668         path->slots[parent_level] = btrfs_header_nritems(parent);
8669
8670         btrfs_assert_tree_locked(node);
8671         level = btrfs_header_level(node);
8672         path->nodes[level] = node;
8673         path->slots[level] = 0;
8674         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8675
8676         wc->refs[parent_level] = 1;
8677         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8678         wc->level = level;
8679         wc->shared_level = -1;
8680         wc->stage = DROP_REFERENCE;
8681         wc->update_ref = 0;
8682         wc->keep_locks = 1;
8683         wc->for_reloc = 1;
8684         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8685
8686         while (1) {
8687                 wret = walk_down_tree(trans, root, path, wc);
8688                 if (wret < 0) {
8689                         ret = wret;
8690                         break;
8691                 }
8692
8693                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8694                 if (wret < 0)
8695                         ret = wret;
8696                 if (wret != 0)
8697                         break;
8698         }
8699
8700         kfree(wc);
8701         btrfs_free_path(path);
8702         return ret;
8703 }
8704
8705 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8706 {
8707         u64 num_devices;
8708         u64 stripped;
8709
8710         /*
8711          * if restripe for this chunk_type is on pick target profile and
8712          * return, otherwise do the usual balance
8713          */
8714         stripped = get_restripe_target(root->fs_info, flags);
8715         if (stripped)
8716                 return extended_to_chunk(stripped);
8717
8718         num_devices = root->fs_info->fs_devices->rw_devices;
8719
8720         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8721                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8722                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8723
8724         if (num_devices == 1) {
8725                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8726                 stripped = flags & ~stripped;
8727
8728                 /* turn raid0 into single device chunks */
8729                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8730                         return stripped;
8731
8732                 /* turn mirroring into duplication */
8733                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8734                              BTRFS_BLOCK_GROUP_RAID10))
8735                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8736         } else {
8737                 /* they already had raid on here, just return */
8738                 if (flags & stripped)
8739                         return flags;
8740
8741                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8742                 stripped = flags & ~stripped;
8743
8744                 /* switch duplicated blocks with raid1 */
8745                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8746                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8747
8748                 /* this is drive concat, leave it alone */
8749         }
8750
8751         return flags;
8752 }
8753
8754 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8755 {
8756         struct btrfs_space_info *sinfo = cache->space_info;
8757         u64 num_bytes;
8758         u64 min_allocable_bytes;
8759         int ret = -ENOSPC;
8760
8761
8762         /*
8763          * We need some metadata space and system metadata space for
8764          * allocating chunks in some corner cases until we force to set
8765          * it to be readonly.
8766          */
8767         if ((sinfo->flags &
8768              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8769             !force)
8770                 min_allocable_bytes = 1 * 1024 * 1024;
8771         else
8772                 min_allocable_bytes = 0;
8773
8774         spin_lock(&sinfo->lock);
8775         spin_lock(&cache->lock);
8776
8777         if (cache->ro) {
8778                 ret = 0;
8779                 goto out;
8780         }
8781
8782         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8783                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8784
8785         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8786             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8787             min_allocable_bytes <= sinfo->total_bytes) {
8788                 sinfo->bytes_readonly += num_bytes;
8789                 cache->ro = 1;
8790                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
8791                 ret = 0;
8792         }
8793 out:
8794         spin_unlock(&cache->lock);
8795         spin_unlock(&sinfo->lock);
8796         return ret;
8797 }
8798
8799 int btrfs_set_block_group_ro(struct btrfs_root *root,
8800                              struct btrfs_block_group_cache *cache)
8801
8802 {
8803         struct btrfs_trans_handle *trans;
8804         u64 alloc_flags;
8805         int ret;
8806
8807         BUG_ON(cache->ro);
8808
8809 again:
8810         trans = btrfs_join_transaction(root);
8811         if (IS_ERR(trans))
8812                 return PTR_ERR(trans);
8813
8814         /*
8815          * we're not allowed to set block groups readonly after the dirty
8816          * block groups cache has started writing.  If it already started,
8817          * back off and let this transaction commit
8818          */
8819         mutex_lock(&root->fs_info->ro_block_group_mutex);
8820         if (trans->transaction->dirty_bg_run) {
8821                 u64 transid = trans->transid;
8822
8823                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
8824                 btrfs_end_transaction(trans, root);
8825
8826                 ret = btrfs_wait_for_commit(root, transid);
8827                 if (ret)
8828                         return ret;
8829                 goto again;
8830         }
8831
8832
8833         ret = set_block_group_ro(cache, 0);
8834         if (!ret)
8835                 goto out;
8836         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8837         ret = do_chunk_alloc(trans, root, alloc_flags,
8838                              CHUNK_ALLOC_FORCE);
8839         if (ret < 0)
8840                 goto out;
8841         ret = set_block_group_ro(cache, 0);
8842 out:
8843         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
8844                 alloc_flags = update_block_group_flags(root, cache->flags);
8845                 check_system_chunk(trans, root, alloc_flags);
8846         }
8847         mutex_unlock(&root->fs_info->ro_block_group_mutex);
8848
8849         btrfs_end_transaction(trans, root);
8850         return ret;
8851 }
8852
8853 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8854                             struct btrfs_root *root, u64 type)
8855 {
8856         u64 alloc_flags = get_alloc_profile(root, type);
8857         return do_chunk_alloc(trans, root, alloc_flags,
8858                               CHUNK_ALLOC_FORCE);
8859 }
8860
8861 /*
8862  * helper to account the unused space of all the readonly block group in the
8863  * space_info. takes mirrors into account.
8864  */
8865 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8866 {
8867         struct btrfs_block_group_cache *block_group;
8868         u64 free_bytes = 0;
8869         int factor;
8870
8871         /* It's df, we don't care if it's racey */
8872         if (list_empty(&sinfo->ro_bgs))
8873                 return 0;
8874
8875         spin_lock(&sinfo->lock);
8876         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
8877                 spin_lock(&block_group->lock);
8878
8879                 if (!block_group->ro) {
8880                         spin_unlock(&block_group->lock);
8881                         continue;
8882                 }
8883
8884                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8885                                           BTRFS_BLOCK_GROUP_RAID10 |
8886                                           BTRFS_BLOCK_GROUP_DUP))
8887                         factor = 2;
8888                 else
8889                         factor = 1;
8890
8891                 free_bytes += (block_group->key.offset -
8892                                btrfs_block_group_used(&block_group->item)) *
8893                                factor;
8894
8895                 spin_unlock(&block_group->lock);
8896         }
8897         spin_unlock(&sinfo->lock);
8898
8899         return free_bytes;
8900 }
8901
8902 void btrfs_set_block_group_rw(struct btrfs_root *root,
8903                               struct btrfs_block_group_cache *cache)
8904 {
8905         struct btrfs_space_info *sinfo = cache->space_info;
8906         u64 num_bytes;
8907
8908         BUG_ON(!cache->ro);
8909
8910         spin_lock(&sinfo->lock);
8911         spin_lock(&cache->lock);
8912         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8913                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8914         sinfo->bytes_readonly -= num_bytes;
8915         cache->ro = 0;
8916         list_del_init(&cache->ro_list);
8917         spin_unlock(&cache->lock);
8918         spin_unlock(&sinfo->lock);
8919 }
8920
8921 /*
8922  * checks to see if its even possible to relocate this block group.
8923  *
8924  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8925  * ok to go ahead and try.
8926  */
8927 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8928 {
8929         struct btrfs_block_group_cache *block_group;
8930         struct btrfs_space_info *space_info;
8931         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8932         struct btrfs_device *device;
8933         struct btrfs_trans_handle *trans;
8934         u64 min_free;
8935         u64 dev_min = 1;
8936         u64 dev_nr = 0;
8937         u64 target;
8938         int index;
8939         int full = 0;
8940         int ret = 0;
8941
8942         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8943
8944         /* odd, couldn't find the block group, leave it alone */
8945         if (!block_group)
8946                 return -1;
8947
8948         min_free = btrfs_block_group_used(&block_group->item);
8949
8950         /* no bytes used, we're good */
8951         if (!min_free)
8952                 goto out;
8953
8954         space_info = block_group->space_info;
8955         spin_lock(&space_info->lock);
8956
8957         full = space_info->full;
8958
8959         /*
8960          * if this is the last block group we have in this space, we can't
8961          * relocate it unless we're able to allocate a new chunk below.
8962          *
8963          * Otherwise, we need to make sure we have room in the space to handle
8964          * all of the extents from this block group.  If we can, we're good
8965          */
8966         if ((space_info->total_bytes != block_group->key.offset) &&
8967             (space_info->bytes_used + space_info->bytes_reserved +
8968              space_info->bytes_pinned + space_info->bytes_readonly +
8969              min_free < space_info->total_bytes)) {
8970                 spin_unlock(&space_info->lock);
8971                 goto out;
8972         }
8973         spin_unlock(&space_info->lock);
8974
8975         /*
8976          * ok we don't have enough space, but maybe we have free space on our
8977          * devices to allocate new chunks for relocation, so loop through our
8978          * alloc devices and guess if we have enough space.  if this block
8979          * group is going to be restriped, run checks against the target
8980          * profile instead of the current one.
8981          */
8982         ret = -1;
8983
8984         /*
8985          * index:
8986          *      0: raid10
8987          *      1: raid1
8988          *      2: dup
8989          *      3: raid0
8990          *      4: single
8991          */
8992         target = get_restripe_target(root->fs_info, block_group->flags);
8993         if (target) {
8994                 index = __get_raid_index(extended_to_chunk(target));
8995         } else {
8996                 /*
8997                  * this is just a balance, so if we were marked as full
8998                  * we know there is no space for a new chunk
8999                  */
9000                 if (full)
9001                         goto out;
9002
9003                 index = get_block_group_index(block_group);
9004         }
9005
9006         if (index == BTRFS_RAID_RAID10) {
9007                 dev_min = 4;
9008                 /* Divide by 2 */
9009                 min_free >>= 1;
9010         } else if (index == BTRFS_RAID_RAID1) {
9011                 dev_min = 2;
9012         } else if (index == BTRFS_RAID_DUP) {
9013                 /* Multiply by 2 */
9014                 min_free <<= 1;
9015         } else if (index == BTRFS_RAID_RAID0) {
9016                 dev_min = fs_devices->rw_devices;
9017                 min_free = div64_u64(min_free, dev_min);
9018         }
9019
9020         /* We need to do this so that we can look at pending chunks */
9021         trans = btrfs_join_transaction(root);
9022         if (IS_ERR(trans)) {
9023                 ret = PTR_ERR(trans);
9024                 goto out;
9025         }
9026
9027         mutex_lock(&root->fs_info->chunk_mutex);
9028         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9029                 u64 dev_offset;
9030
9031                 /*
9032                  * check to make sure we can actually find a chunk with enough
9033                  * space to fit our block group in.
9034                  */
9035                 if (device->total_bytes > device->bytes_used + min_free &&
9036                     !device->is_tgtdev_for_dev_replace) {
9037                         ret = find_free_dev_extent(trans, device, min_free,
9038                                                    &dev_offset, NULL);
9039                         if (!ret)
9040                                 dev_nr++;
9041
9042                         if (dev_nr >= dev_min)
9043                                 break;
9044
9045                         ret = -1;
9046                 }
9047         }
9048         mutex_unlock(&root->fs_info->chunk_mutex);
9049         btrfs_end_transaction(trans, root);
9050 out:
9051         btrfs_put_block_group(block_group);
9052         return ret;
9053 }
9054
9055 static int find_first_block_group(struct btrfs_root *root,
9056                 struct btrfs_path *path, struct btrfs_key *key)
9057 {
9058         int ret = 0;
9059         struct btrfs_key found_key;
9060         struct extent_buffer *leaf;
9061         int slot;
9062
9063         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9064         if (ret < 0)
9065                 goto out;
9066
9067         while (1) {
9068                 slot = path->slots[0];
9069                 leaf = path->nodes[0];
9070                 if (slot >= btrfs_header_nritems(leaf)) {
9071                         ret = btrfs_next_leaf(root, path);
9072                         if (ret == 0)
9073                                 continue;
9074                         if (ret < 0)
9075                                 goto out;
9076                         break;
9077                 }
9078                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9079
9080                 if (found_key.objectid >= key->objectid &&
9081                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9082                         ret = 0;
9083                         goto out;
9084                 }
9085                 path->slots[0]++;
9086         }
9087 out:
9088         return ret;
9089 }
9090
9091 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9092 {
9093         struct btrfs_block_group_cache *block_group;
9094         u64 last = 0;
9095
9096         while (1) {
9097                 struct inode *inode;
9098
9099                 block_group = btrfs_lookup_first_block_group(info, last);
9100                 while (block_group) {
9101                         spin_lock(&block_group->lock);
9102                         if (block_group->iref)
9103                                 break;
9104                         spin_unlock(&block_group->lock);
9105                         block_group = next_block_group(info->tree_root,
9106                                                        block_group);
9107                 }
9108                 if (!block_group) {
9109                         if (last == 0)
9110                                 break;
9111                         last = 0;
9112                         continue;
9113                 }
9114
9115                 inode = block_group->inode;
9116                 block_group->iref = 0;
9117                 block_group->inode = NULL;
9118                 spin_unlock(&block_group->lock);
9119                 iput(inode);
9120                 last = block_group->key.objectid + block_group->key.offset;
9121                 btrfs_put_block_group(block_group);
9122         }
9123 }
9124
9125 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9126 {
9127         struct btrfs_block_group_cache *block_group;
9128         struct btrfs_space_info *space_info;
9129         struct btrfs_caching_control *caching_ctl;
9130         struct rb_node *n;
9131
9132         down_write(&info->commit_root_sem);
9133         while (!list_empty(&info->caching_block_groups)) {
9134                 caching_ctl = list_entry(info->caching_block_groups.next,
9135                                          struct btrfs_caching_control, list);
9136                 list_del(&caching_ctl->list);
9137                 put_caching_control(caching_ctl);
9138         }
9139         up_write(&info->commit_root_sem);
9140
9141         spin_lock(&info->unused_bgs_lock);
9142         while (!list_empty(&info->unused_bgs)) {
9143                 block_group = list_first_entry(&info->unused_bgs,
9144                                                struct btrfs_block_group_cache,
9145                                                bg_list);
9146                 list_del_init(&block_group->bg_list);
9147                 btrfs_put_block_group(block_group);
9148         }
9149         spin_unlock(&info->unused_bgs_lock);
9150
9151         spin_lock(&info->block_group_cache_lock);
9152         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9153                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9154                                        cache_node);
9155                 rb_erase(&block_group->cache_node,
9156                          &info->block_group_cache_tree);
9157                 RB_CLEAR_NODE(&block_group->cache_node);
9158                 spin_unlock(&info->block_group_cache_lock);
9159
9160                 down_write(&block_group->space_info->groups_sem);
9161                 list_del(&block_group->list);
9162                 up_write(&block_group->space_info->groups_sem);
9163
9164                 if (block_group->cached == BTRFS_CACHE_STARTED)
9165                         wait_block_group_cache_done(block_group);
9166
9167                 /*
9168                  * We haven't cached this block group, which means we could
9169                  * possibly have excluded extents on this block group.
9170                  */
9171                 if (block_group->cached == BTRFS_CACHE_NO ||
9172                     block_group->cached == BTRFS_CACHE_ERROR)
9173                         free_excluded_extents(info->extent_root, block_group);
9174
9175                 btrfs_remove_free_space_cache(block_group);
9176                 btrfs_put_block_group(block_group);
9177
9178                 spin_lock(&info->block_group_cache_lock);
9179         }
9180         spin_unlock(&info->block_group_cache_lock);
9181
9182         /* now that all the block groups are freed, go through and
9183          * free all the space_info structs.  This is only called during
9184          * the final stages of unmount, and so we know nobody is
9185          * using them.  We call synchronize_rcu() once before we start,
9186          * just to be on the safe side.
9187          */
9188         synchronize_rcu();
9189
9190         release_global_block_rsv(info);
9191
9192         while (!list_empty(&info->space_info)) {
9193                 int i;
9194
9195                 space_info = list_entry(info->space_info.next,
9196                                         struct btrfs_space_info,
9197                                         list);
9198                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9199                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9200                             space_info->bytes_reserved > 0 ||
9201                             space_info->bytes_may_use > 0)) {
9202                                 dump_space_info(space_info, 0, 0);
9203                         }
9204                 }
9205                 list_del(&space_info->list);
9206                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9207                         struct kobject *kobj;
9208                         kobj = space_info->block_group_kobjs[i];
9209                         space_info->block_group_kobjs[i] = NULL;
9210                         if (kobj) {
9211                                 kobject_del(kobj);
9212                                 kobject_put(kobj);
9213                         }
9214                 }
9215                 kobject_del(&space_info->kobj);
9216                 kobject_put(&space_info->kobj);
9217         }
9218         return 0;
9219 }
9220
9221 static void __link_block_group(struct btrfs_space_info *space_info,
9222                                struct btrfs_block_group_cache *cache)
9223 {
9224         int index = get_block_group_index(cache);
9225         bool first = false;
9226
9227         down_write(&space_info->groups_sem);
9228         if (list_empty(&space_info->block_groups[index]))
9229                 first = true;
9230         list_add_tail(&cache->list, &space_info->block_groups[index]);
9231         up_write(&space_info->groups_sem);
9232
9233         if (first) {
9234                 struct raid_kobject *rkobj;
9235                 int ret;
9236
9237                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9238                 if (!rkobj)
9239                         goto out_err;
9240                 rkobj->raid_type = index;
9241                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9242                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9243                                   "%s", get_raid_name(index));
9244                 if (ret) {
9245                         kobject_put(&rkobj->kobj);
9246                         goto out_err;
9247                 }
9248                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9249         }
9250
9251         return;
9252 out_err:
9253         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9254 }
9255
9256 static struct btrfs_block_group_cache *
9257 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9258 {
9259         struct btrfs_block_group_cache *cache;
9260
9261         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9262         if (!cache)
9263                 return NULL;
9264
9265         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9266                                         GFP_NOFS);
9267         if (!cache->free_space_ctl) {
9268                 kfree(cache);
9269                 return NULL;
9270         }
9271
9272         cache->key.objectid = start;
9273         cache->key.offset = size;
9274         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9275
9276         cache->sectorsize = root->sectorsize;
9277         cache->fs_info = root->fs_info;
9278         cache->full_stripe_len = btrfs_full_stripe_len(root,
9279                                                &root->fs_info->mapping_tree,
9280                                                start);
9281         atomic_set(&cache->count, 1);
9282         spin_lock_init(&cache->lock);
9283         init_rwsem(&cache->data_rwsem);
9284         INIT_LIST_HEAD(&cache->list);
9285         INIT_LIST_HEAD(&cache->cluster_list);
9286         INIT_LIST_HEAD(&cache->bg_list);
9287         INIT_LIST_HEAD(&cache->ro_list);
9288         INIT_LIST_HEAD(&cache->dirty_list);
9289         INIT_LIST_HEAD(&cache->io_list);
9290         btrfs_init_free_space_ctl(cache);
9291         atomic_set(&cache->trimming, 0);
9292
9293         return cache;
9294 }
9295
9296 int btrfs_read_block_groups(struct btrfs_root *root)
9297 {
9298         struct btrfs_path *path;
9299         int ret;
9300         struct btrfs_block_group_cache *cache;
9301         struct btrfs_fs_info *info = root->fs_info;
9302         struct btrfs_space_info *space_info;
9303         struct btrfs_key key;
9304         struct btrfs_key found_key;
9305         struct extent_buffer *leaf;
9306         int need_clear = 0;
9307         u64 cache_gen;
9308
9309         root = info->extent_root;
9310         key.objectid = 0;
9311         key.offset = 0;
9312         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9313         path = btrfs_alloc_path();
9314         if (!path)
9315                 return -ENOMEM;
9316         path->reada = 1;
9317
9318         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9319         if (btrfs_test_opt(root, SPACE_CACHE) &&
9320             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9321                 need_clear = 1;
9322         if (btrfs_test_opt(root, CLEAR_CACHE))
9323                 need_clear = 1;
9324
9325         while (1) {
9326                 ret = find_first_block_group(root, path, &key);
9327                 if (ret > 0)
9328                         break;
9329                 if (ret != 0)
9330                         goto error;
9331
9332                 leaf = path->nodes[0];
9333                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9334
9335                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9336                                                        found_key.offset);
9337                 if (!cache) {
9338                         ret = -ENOMEM;
9339                         goto error;
9340                 }
9341
9342                 if (need_clear) {
9343                         /*
9344                          * When we mount with old space cache, we need to
9345                          * set BTRFS_DC_CLEAR and set dirty flag.
9346                          *
9347                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9348                          *    truncate the old free space cache inode and
9349                          *    setup a new one.
9350                          * b) Setting 'dirty flag' makes sure that we flush
9351                          *    the new space cache info onto disk.
9352                          */
9353                         if (btrfs_test_opt(root, SPACE_CACHE))
9354                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9355                 }
9356
9357                 read_extent_buffer(leaf, &cache->item,
9358                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9359                                    sizeof(cache->item));
9360                 cache->flags = btrfs_block_group_flags(&cache->item);
9361
9362                 key.objectid = found_key.objectid + found_key.offset;
9363                 btrfs_release_path(path);
9364
9365                 /*
9366                  * We need to exclude the super stripes now so that the space
9367                  * info has super bytes accounted for, otherwise we'll think
9368                  * we have more space than we actually do.
9369                  */
9370                 ret = exclude_super_stripes(root, cache);
9371                 if (ret) {
9372                         /*
9373                          * We may have excluded something, so call this just in
9374                          * case.
9375                          */
9376                         free_excluded_extents(root, cache);
9377                         btrfs_put_block_group(cache);
9378                         goto error;
9379                 }
9380
9381                 /*
9382                  * check for two cases, either we are full, and therefore
9383                  * don't need to bother with the caching work since we won't
9384                  * find any space, or we are empty, and we can just add all
9385                  * the space in and be done with it.  This saves us _alot_ of
9386                  * time, particularly in the full case.
9387                  */
9388                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9389                         cache->last_byte_to_unpin = (u64)-1;
9390                         cache->cached = BTRFS_CACHE_FINISHED;
9391                         free_excluded_extents(root, cache);
9392                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9393                         cache->last_byte_to_unpin = (u64)-1;
9394                         cache->cached = BTRFS_CACHE_FINISHED;
9395                         add_new_free_space(cache, root->fs_info,
9396                                            found_key.objectid,
9397                                            found_key.objectid +
9398                                            found_key.offset);
9399                         free_excluded_extents(root, cache);
9400                 }
9401
9402                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9403                 if (ret) {
9404                         btrfs_remove_free_space_cache(cache);
9405                         btrfs_put_block_group(cache);
9406                         goto error;
9407                 }
9408
9409                 ret = update_space_info(info, cache->flags, found_key.offset,
9410                                         btrfs_block_group_used(&cache->item),
9411                                         &space_info);
9412                 if (ret) {
9413                         btrfs_remove_free_space_cache(cache);
9414                         spin_lock(&info->block_group_cache_lock);
9415                         rb_erase(&cache->cache_node,
9416                                  &info->block_group_cache_tree);
9417                         RB_CLEAR_NODE(&cache->cache_node);
9418                         spin_unlock(&info->block_group_cache_lock);
9419                         btrfs_put_block_group(cache);
9420                         goto error;
9421                 }
9422
9423                 cache->space_info = space_info;
9424                 spin_lock(&cache->space_info->lock);
9425                 cache->space_info->bytes_readonly += cache->bytes_super;
9426                 spin_unlock(&cache->space_info->lock);
9427
9428                 __link_block_group(space_info, cache);
9429
9430                 set_avail_alloc_bits(root->fs_info, cache->flags);
9431                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9432                         set_block_group_ro(cache, 1);
9433                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9434                         spin_lock(&info->unused_bgs_lock);
9435                         /* Should always be true but just in case. */
9436                         if (list_empty(&cache->bg_list)) {
9437                                 btrfs_get_block_group(cache);
9438                                 list_add_tail(&cache->bg_list,
9439                                               &info->unused_bgs);
9440                         }
9441                         spin_unlock(&info->unused_bgs_lock);
9442                 }
9443         }
9444
9445         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9446                 if (!(get_alloc_profile(root, space_info->flags) &
9447                       (BTRFS_BLOCK_GROUP_RAID10 |
9448                        BTRFS_BLOCK_GROUP_RAID1 |
9449                        BTRFS_BLOCK_GROUP_RAID5 |
9450                        BTRFS_BLOCK_GROUP_RAID6 |
9451                        BTRFS_BLOCK_GROUP_DUP)))
9452                         continue;
9453                 /*
9454                  * avoid allocating from un-mirrored block group if there are
9455                  * mirrored block groups.
9456                  */
9457                 list_for_each_entry(cache,
9458                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9459                                 list)
9460                         set_block_group_ro(cache, 1);
9461                 list_for_each_entry(cache,
9462                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9463                                 list)
9464                         set_block_group_ro(cache, 1);
9465         }
9466
9467         init_global_block_rsv(info);
9468         ret = 0;
9469 error:
9470         btrfs_free_path(path);
9471         return ret;
9472 }
9473
9474 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9475                                        struct btrfs_root *root)
9476 {
9477         struct btrfs_block_group_cache *block_group, *tmp;
9478         struct btrfs_root *extent_root = root->fs_info->extent_root;
9479         struct btrfs_block_group_item item;
9480         struct btrfs_key key;
9481         int ret = 0;
9482
9483         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9484                 if (ret)
9485                         goto next;
9486
9487                 spin_lock(&block_group->lock);
9488                 memcpy(&item, &block_group->item, sizeof(item));
9489                 memcpy(&key, &block_group->key, sizeof(key));
9490                 spin_unlock(&block_group->lock);
9491
9492                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9493                                         sizeof(item));
9494                 if (ret)
9495                         btrfs_abort_transaction(trans, extent_root, ret);
9496                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9497                                                key.objectid, key.offset);
9498                 if (ret)
9499                         btrfs_abort_transaction(trans, extent_root, ret);
9500 next:
9501                 list_del_init(&block_group->bg_list);
9502         }
9503 }
9504
9505 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9506                            struct btrfs_root *root, u64 bytes_used,
9507                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9508                            u64 size)
9509 {
9510         int ret;
9511         struct btrfs_root *extent_root;
9512         struct btrfs_block_group_cache *cache;
9513
9514         extent_root = root->fs_info->extent_root;
9515
9516         btrfs_set_log_full_commit(root->fs_info, trans);
9517
9518         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9519         if (!cache)
9520                 return -ENOMEM;
9521
9522         btrfs_set_block_group_used(&cache->item, bytes_used);
9523         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9524         btrfs_set_block_group_flags(&cache->item, type);
9525
9526         cache->flags = type;
9527         cache->last_byte_to_unpin = (u64)-1;
9528         cache->cached = BTRFS_CACHE_FINISHED;
9529         ret = exclude_super_stripes(root, cache);
9530         if (ret) {
9531                 /*
9532                  * We may have excluded something, so call this just in
9533                  * case.
9534                  */
9535                 free_excluded_extents(root, cache);
9536                 btrfs_put_block_group(cache);
9537                 return ret;
9538         }
9539
9540         add_new_free_space(cache, root->fs_info, chunk_offset,
9541                            chunk_offset + size);
9542
9543         free_excluded_extents(root, cache);
9544
9545         ret = btrfs_add_block_group_cache(root->fs_info, cache);
9546         if (ret) {
9547                 btrfs_remove_free_space_cache(cache);
9548                 btrfs_put_block_group(cache);
9549                 return ret;
9550         }
9551
9552         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
9553                                 &cache->space_info);
9554         if (ret) {
9555                 btrfs_remove_free_space_cache(cache);
9556                 spin_lock(&root->fs_info->block_group_cache_lock);
9557                 rb_erase(&cache->cache_node,
9558                          &root->fs_info->block_group_cache_tree);
9559                 RB_CLEAR_NODE(&cache->cache_node);
9560                 spin_unlock(&root->fs_info->block_group_cache_lock);
9561                 btrfs_put_block_group(cache);
9562                 return ret;
9563         }
9564         update_global_block_rsv(root->fs_info);
9565
9566         spin_lock(&cache->space_info->lock);
9567         cache->space_info->bytes_readonly += cache->bytes_super;
9568         spin_unlock(&cache->space_info->lock);
9569
9570         __link_block_group(cache->space_info, cache);
9571
9572         list_add_tail(&cache->bg_list, &trans->new_bgs);
9573
9574         set_avail_alloc_bits(extent_root->fs_info, type);
9575
9576         return 0;
9577 }
9578
9579 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9580 {
9581         u64 extra_flags = chunk_to_extended(flags) &
9582                                 BTRFS_EXTENDED_PROFILE_MASK;
9583
9584         write_seqlock(&fs_info->profiles_lock);
9585         if (flags & BTRFS_BLOCK_GROUP_DATA)
9586                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9587         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9588                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9589         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9590                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9591         write_sequnlock(&fs_info->profiles_lock);
9592 }
9593
9594 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9595                              struct btrfs_root *root, u64 group_start,
9596                              struct extent_map *em)
9597 {
9598         struct btrfs_path *path;
9599         struct btrfs_block_group_cache *block_group;
9600         struct btrfs_free_cluster *cluster;
9601         struct btrfs_root *tree_root = root->fs_info->tree_root;
9602         struct btrfs_key key;
9603         struct inode *inode;
9604         struct kobject *kobj = NULL;
9605         int ret;
9606         int index;
9607         int factor;
9608         struct btrfs_caching_control *caching_ctl = NULL;
9609         bool remove_em;
9610
9611         root = root->fs_info->extent_root;
9612
9613         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9614         BUG_ON(!block_group);
9615         BUG_ON(!block_group->ro);
9616
9617         /*
9618          * Free the reserved super bytes from this block group before
9619          * remove it.
9620          */
9621         free_excluded_extents(root, block_group);
9622
9623         memcpy(&key, &block_group->key, sizeof(key));
9624         index = get_block_group_index(block_group);
9625         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9626                                   BTRFS_BLOCK_GROUP_RAID1 |
9627                                   BTRFS_BLOCK_GROUP_RAID10))
9628                 factor = 2;
9629         else
9630                 factor = 1;
9631
9632         /* make sure this block group isn't part of an allocation cluster */
9633         cluster = &root->fs_info->data_alloc_cluster;
9634         spin_lock(&cluster->refill_lock);
9635         btrfs_return_cluster_to_free_space(block_group, cluster);
9636         spin_unlock(&cluster->refill_lock);
9637
9638         /*
9639          * make sure this block group isn't part of a metadata
9640          * allocation cluster
9641          */
9642         cluster = &root->fs_info->meta_alloc_cluster;
9643         spin_lock(&cluster->refill_lock);
9644         btrfs_return_cluster_to_free_space(block_group, cluster);
9645         spin_unlock(&cluster->refill_lock);
9646
9647         path = btrfs_alloc_path();
9648         if (!path) {
9649                 ret = -ENOMEM;
9650                 goto out;
9651         }
9652
9653         /*
9654          * get the inode first so any iput calls done for the io_list
9655          * aren't the final iput (no unlinks allowed now)
9656          */
9657         inode = lookup_free_space_inode(tree_root, block_group, path);
9658
9659         mutex_lock(&trans->transaction->cache_write_mutex);
9660         /*
9661          * make sure our free spache cache IO is done before remove the
9662          * free space inode
9663          */
9664         spin_lock(&trans->transaction->dirty_bgs_lock);
9665         if (!list_empty(&block_group->io_list)) {
9666                 list_del_init(&block_group->io_list);
9667
9668                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
9669
9670                 spin_unlock(&trans->transaction->dirty_bgs_lock);
9671                 btrfs_wait_cache_io(root, trans, block_group,
9672                                     &block_group->io_ctl, path,
9673                                     block_group->key.objectid);
9674                 btrfs_put_block_group(block_group);
9675                 spin_lock(&trans->transaction->dirty_bgs_lock);
9676         }
9677
9678         if (!list_empty(&block_group->dirty_list)) {
9679                 list_del_init(&block_group->dirty_list);
9680                 btrfs_put_block_group(block_group);
9681         }
9682         spin_unlock(&trans->transaction->dirty_bgs_lock);
9683         mutex_unlock(&trans->transaction->cache_write_mutex);
9684
9685         if (!IS_ERR(inode)) {
9686                 ret = btrfs_orphan_add(trans, inode);
9687                 if (ret) {
9688                         btrfs_add_delayed_iput(inode);
9689                         goto out;
9690                 }
9691                 clear_nlink(inode);
9692                 /* One for the block groups ref */
9693                 spin_lock(&block_group->lock);
9694                 if (block_group->iref) {
9695                         block_group->iref = 0;
9696                         block_group->inode = NULL;
9697                         spin_unlock(&block_group->lock);
9698                         iput(inode);
9699                 } else {
9700                         spin_unlock(&block_group->lock);
9701                 }
9702                 /* One for our lookup ref */
9703                 btrfs_add_delayed_iput(inode);
9704         }
9705
9706         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9707         key.offset = block_group->key.objectid;
9708         key.type = 0;
9709
9710         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9711         if (ret < 0)
9712                 goto out;
9713         if (ret > 0)
9714                 btrfs_release_path(path);
9715         if (ret == 0) {
9716                 ret = btrfs_del_item(trans, tree_root, path);
9717                 if (ret)
9718                         goto out;
9719                 btrfs_release_path(path);
9720         }
9721
9722         spin_lock(&root->fs_info->block_group_cache_lock);
9723         rb_erase(&block_group->cache_node,
9724                  &root->fs_info->block_group_cache_tree);
9725         RB_CLEAR_NODE(&block_group->cache_node);
9726
9727         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9728                 root->fs_info->first_logical_byte = (u64)-1;
9729         spin_unlock(&root->fs_info->block_group_cache_lock);
9730
9731         down_write(&block_group->space_info->groups_sem);
9732         /*
9733          * we must use list_del_init so people can check to see if they
9734          * are still on the list after taking the semaphore
9735          */
9736         list_del_init(&block_group->list);
9737         if (list_empty(&block_group->space_info->block_groups[index])) {
9738                 kobj = block_group->space_info->block_group_kobjs[index];
9739                 block_group->space_info->block_group_kobjs[index] = NULL;
9740                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9741         }
9742         up_write(&block_group->space_info->groups_sem);
9743         if (kobj) {
9744                 kobject_del(kobj);
9745                 kobject_put(kobj);
9746         }
9747
9748         if (block_group->has_caching_ctl)
9749                 caching_ctl = get_caching_control(block_group);
9750         if (block_group->cached == BTRFS_CACHE_STARTED)
9751                 wait_block_group_cache_done(block_group);
9752         if (block_group->has_caching_ctl) {
9753                 down_write(&root->fs_info->commit_root_sem);
9754                 if (!caching_ctl) {
9755                         struct btrfs_caching_control *ctl;
9756
9757                         list_for_each_entry(ctl,
9758                                     &root->fs_info->caching_block_groups, list)
9759                                 if (ctl->block_group == block_group) {
9760                                         caching_ctl = ctl;
9761                                         atomic_inc(&caching_ctl->count);
9762                                         break;
9763                                 }
9764                 }
9765                 if (caching_ctl)
9766                         list_del_init(&caching_ctl->list);
9767                 up_write(&root->fs_info->commit_root_sem);
9768                 if (caching_ctl) {
9769                         /* Once for the caching bgs list and once for us. */
9770                         put_caching_control(caching_ctl);
9771                         put_caching_control(caching_ctl);
9772                 }
9773         }
9774
9775         spin_lock(&trans->transaction->dirty_bgs_lock);
9776         if (!list_empty(&block_group->dirty_list)) {
9777                 WARN_ON(1);
9778         }
9779         if (!list_empty(&block_group->io_list)) {
9780                 WARN_ON(1);
9781         }
9782         spin_unlock(&trans->transaction->dirty_bgs_lock);
9783         btrfs_remove_free_space_cache(block_group);
9784
9785         spin_lock(&block_group->space_info->lock);
9786         list_del_init(&block_group->ro_list);
9787
9788         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
9789                 WARN_ON(block_group->space_info->total_bytes
9790                         < block_group->key.offset);
9791                 WARN_ON(block_group->space_info->bytes_readonly
9792                         < block_group->key.offset);
9793                 WARN_ON(block_group->space_info->disk_total
9794                         < block_group->key.offset * factor);
9795         }
9796         block_group->space_info->total_bytes -= block_group->key.offset;
9797         block_group->space_info->bytes_readonly -= block_group->key.offset;
9798         block_group->space_info->disk_total -= block_group->key.offset * factor;
9799
9800         spin_unlock(&block_group->space_info->lock);
9801
9802         memcpy(&key, &block_group->key, sizeof(key));
9803
9804         lock_chunks(root);
9805         if (!list_empty(&em->list)) {
9806                 /* We're in the transaction->pending_chunks list. */
9807                 free_extent_map(em);
9808         }
9809         spin_lock(&block_group->lock);
9810         block_group->removed = 1;
9811         /*
9812          * At this point trimming can't start on this block group, because we
9813          * removed the block group from the tree fs_info->block_group_cache_tree
9814          * so no one can't find it anymore and even if someone already got this
9815          * block group before we removed it from the rbtree, they have already
9816          * incremented block_group->trimming - if they didn't, they won't find
9817          * any free space entries because we already removed them all when we
9818          * called btrfs_remove_free_space_cache().
9819          *
9820          * And we must not remove the extent map from the fs_info->mapping_tree
9821          * to prevent the same logical address range and physical device space
9822          * ranges from being reused for a new block group. This is because our
9823          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
9824          * completely transactionless, so while it is trimming a range the
9825          * currently running transaction might finish and a new one start,
9826          * allowing for new block groups to be created that can reuse the same
9827          * physical device locations unless we take this special care.
9828          */
9829         remove_em = (atomic_read(&block_group->trimming) == 0);
9830         /*
9831          * Make sure a trimmer task always sees the em in the pinned_chunks list
9832          * if it sees block_group->removed == 1 (needs to lock block_group->lock
9833          * before checking block_group->removed).
9834          */
9835         if (!remove_em) {
9836                 /*
9837                  * Our em might be in trans->transaction->pending_chunks which
9838                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
9839                  * and so is the fs_info->pinned_chunks list.
9840                  *
9841                  * So at this point we must be holding the chunk_mutex to avoid
9842                  * any races with chunk allocation (more specifically at
9843                  * volumes.c:contains_pending_extent()), to ensure it always
9844                  * sees the em, either in the pending_chunks list or in the
9845                  * pinned_chunks list.
9846                  */
9847                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
9848         }
9849         spin_unlock(&block_group->lock);
9850
9851         if (remove_em) {
9852                 struct extent_map_tree *em_tree;
9853
9854                 em_tree = &root->fs_info->mapping_tree.map_tree;
9855                 write_lock(&em_tree->lock);
9856                 /*
9857                  * The em might be in the pending_chunks list, so make sure the
9858                  * chunk mutex is locked, since remove_extent_mapping() will
9859                  * delete us from that list.
9860                  */
9861                 remove_extent_mapping(em_tree, em);
9862                 write_unlock(&em_tree->lock);
9863                 /* once for the tree */
9864                 free_extent_map(em);
9865         }
9866
9867         unlock_chunks(root);
9868
9869         btrfs_put_block_group(block_group);
9870         btrfs_put_block_group(block_group);
9871
9872         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9873         if (ret > 0)
9874                 ret = -EIO;
9875         if (ret < 0)
9876                 goto out;
9877
9878         ret = btrfs_del_item(trans, root, path);
9879 out:
9880         btrfs_free_path(path);
9881         return ret;
9882 }
9883
9884 /*
9885  * Process the unused_bgs list and remove any that don't have any allocated
9886  * space inside of them.
9887  */
9888 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
9889 {
9890         struct btrfs_block_group_cache *block_group;
9891         struct btrfs_space_info *space_info;
9892         struct btrfs_root *root = fs_info->extent_root;
9893         struct btrfs_trans_handle *trans;
9894         int ret = 0;
9895
9896         if (!fs_info->open)
9897                 return;
9898
9899         spin_lock(&fs_info->unused_bgs_lock);
9900         while (!list_empty(&fs_info->unused_bgs)) {
9901                 u64 start, end;
9902
9903                 block_group = list_first_entry(&fs_info->unused_bgs,
9904                                                struct btrfs_block_group_cache,
9905                                                bg_list);
9906                 space_info = block_group->space_info;
9907                 list_del_init(&block_group->bg_list);
9908                 if (ret || btrfs_mixed_space_info(space_info)) {
9909                         btrfs_put_block_group(block_group);
9910                         continue;
9911                 }
9912                 spin_unlock(&fs_info->unused_bgs_lock);
9913
9914                 /* Don't want to race with allocators so take the groups_sem */
9915                 down_write(&space_info->groups_sem);
9916                 spin_lock(&block_group->lock);
9917                 if (block_group->reserved ||
9918                     btrfs_block_group_used(&block_group->item) ||
9919                     block_group->ro) {
9920                         /*
9921                          * We want to bail if we made new allocations or have
9922                          * outstanding allocations in this block group.  We do
9923                          * the ro check in case balance is currently acting on
9924                          * this block group.
9925                          */
9926                         spin_unlock(&block_group->lock);
9927                         up_write(&space_info->groups_sem);
9928                         goto next;
9929                 }
9930                 spin_unlock(&block_group->lock);
9931
9932                 /* We don't want to force the issue, only flip if it's ok. */
9933                 ret = set_block_group_ro(block_group, 0);
9934                 up_write(&space_info->groups_sem);
9935                 if (ret < 0) {
9936                         ret = 0;
9937                         goto next;
9938                 }
9939
9940                 /*
9941                  * Want to do this before we do anything else so we can recover
9942                  * properly if we fail to join the transaction.
9943                  */
9944                 /* 1 for btrfs_orphan_reserve_metadata() */
9945                 trans = btrfs_start_transaction(root, 1);
9946                 if (IS_ERR(trans)) {
9947                         btrfs_set_block_group_rw(root, block_group);
9948                         ret = PTR_ERR(trans);
9949                         goto next;
9950                 }
9951
9952                 /*
9953                  * We could have pending pinned extents for this block group,
9954                  * just delete them, we don't care about them anymore.
9955                  */
9956                 start = block_group->key.objectid;
9957                 end = start + block_group->key.offset - 1;
9958                 /*
9959                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
9960                  * btrfs_finish_extent_commit(). If we are at transaction N,
9961                  * another task might be running finish_extent_commit() for the
9962                  * previous transaction N - 1, and have seen a range belonging
9963                  * to the block group in freed_extents[] before we were able to
9964                  * clear the whole block group range from freed_extents[]. This
9965                  * means that task can lookup for the block group after we
9966                  * unpinned it from freed_extents[] and removed it, leading to
9967                  * a BUG_ON() at btrfs_unpin_extent_range().
9968                  */
9969                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
9970                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
9971                                   EXTENT_DIRTY, GFP_NOFS);
9972                 if (ret) {
9973                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9974                         btrfs_set_block_group_rw(root, block_group);
9975                         goto end_trans;
9976                 }
9977                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
9978                                   EXTENT_DIRTY, GFP_NOFS);
9979                 if (ret) {
9980                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9981                         btrfs_set_block_group_rw(root, block_group);
9982                         goto end_trans;
9983                 }
9984                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9985
9986                 /* Reset pinned so btrfs_put_block_group doesn't complain */
9987                 spin_lock(&space_info->lock);
9988                 spin_lock(&block_group->lock);
9989
9990                 space_info->bytes_pinned -= block_group->pinned;
9991                 space_info->bytes_readonly += block_group->pinned;
9992                 percpu_counter_add(&space_info->total_bytes_pinned,
9993                                    -block_group->pinned);
9994                 block_group->pinned = 0;
9995
9996                 spin_unlock(&block_group->lock);
9997                 spin_unlock(&space_info->lock);
9998
9999                 /*
10000                  * Btrfs_remove_chunk will abort the transaction if things go
10001                  * horribly wrong.
10002                  */
10003                 ret = btrfs_remove_chunk(trans, root,
10004                                          block_group->key.objectid);
10005 end_trans:
10006                 btrfs_end_transaction(trans, root);
10007 next:
10008                 btrfs_put_block_group(block_group);
10009                 spin_lock(&fs_info->unused_bgs_lock);
10010         }
10011         spin_unlock(&fs_info->unused_bgs_lock);
10012 }
10013
10014 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10015 {
10016         struct btrfs_space_info *space_info;
10017         struct btrfs_super_block *disk_super;
10018         u64 features;
10019         u64 flags;
10020         int mixed = 0;
10021         int ret;
10022
10023         disk_super = fs_info->super_copy;
10024         if (!btrfs_super_root(disk_super))
10025                 return 1;
10026
10027         features = btrfs_super_incompat_flags(disk_super);
10028         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10029                 mixed = 1;
10030
10031         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10032         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10033         if (ret)
10034                 goto out;
10035
10036         if (mixed) {
10037                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10038                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10039         } else {
10040                 flags = BTRFS_BLOCK_GROUP_METADATA;
10041                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10042                 if (ret)
10043                         goto out;
10044
10045                 flags = BTRFS_BLOCK_GROUP_DATA;
10046                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10047         }
10048 out:
10049         return ret;
10050 }
10051
10052 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10053 {
10054         return unpin_extent_range(root, start, end, false);
10055 }
10056
10057 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10058 {
10059         struct btrfs_fs_info *fs_info = root->fs_info;
10060         struct btrfs_block_group_cache *cache = NULL;
10061         u64 group_trimmed;
10062         u64 start;
10063         u64 end;
10064         u64 trimmed = 0;
10065         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10066         int ret = 0;
10067
10068         /*
10069          * try to trim all FS space, our block group may start from non-zero.
10070          */
10071         if (range->len == total_bytes)
10072                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10073         else
10074                 cache = btrfs_lookup_block_group(fs_info, range->start);
10075
10076         while (cache) {
10077                 if (cache->key.objectid >= (range->start + range->len)) {
10078                         btrfs_put_block_group(cache);
10079                         break;
10080                 }
10081
10082                 start = max(range->start, cache->key.objectid);
10083                 end = min(range->start + range->len,
10084                                 cache->key.objectid + cache->key.offset);
10085
10086                 if (end - start >= range->minlen) {
10087                         if (!block_group_cache_done(cache)) {
10088                                 ret = cache_block_group(cache, 0);
10089                                 if (ret) {
10090                                         btrfs_put_block_group(cache);
10091                                         break;
10092                                 }
10093                                 ret = wait_block_group_cache_done(cache);
10094                                 if (ret) {
10095                                         btrfs_put_block_group(cache);
10096                                         break;
10097                                 }
10098                         }
10099                         ret = btrfs_trim_block_group(cache,
10100                                                      &group_trimmed,
10101                                                      start,
10102                                                      end,
10103                                                      range->minlen);
10104
10105                         trimmed += group_trimmed;
10106                         if (ret) {
10107                                 btrfs_put_block_group(cache);
10108                                 break;
10109                         }
10110                 }
10111
10112                 cache = next_block_group(fs_info->tree_root, cache);
10113         }
10114
10115         range->len = trimmed;
10116         return ret;
10117 }
10118
10119 /*
10120  * btrfs_{start,end}_write_no_snapshoting() are similar to
10121  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10122  * data into the page cache through nocow before the subvolume is snapshoted,
10123  * but flush the data into disk after the snapshot creation, or to prevent
10124  * operations while snapshoting is ongoing and that cause the snapshot to be
10125  * inconsistent (writes followed by expanding truncates for example).
10126  */
10127 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10128 {
10129         percpu_counter_dec(&root->subv_writers->counter);
10130         /*
10131          * Make sure counter is updated before we wake up
10132          * waiters.
10133          */
10134         smp_mb();
10135         if (waitqueue_active(&root->subv_writers->wait))
10136                 wake_up(&root->subv_writers->wait);
10137 }
10138
10139 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10140 {
10141         if (atomic_read(&root->will_be_snapshoted))
10142                 return 0;
10143
10144         percpu_counter_inc(&root->subv_writers->counter);
10145         /*
10146          * Make sure counter is updated before we check for snapshot creation.
10147          */
10148         smp_mb();
10149         if (atomic_read(&root->will_be_snapshoted)) {
10150                 btrfs_end_write_no_snapshoting(root);
10151                 return 0;
10152         }
10153         return 1;
10154 }