fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include <linux/percpu_counter.h>
  28 #include "hash.h"
  29 #include "tree-log.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "volumes.h"
  33 #include "raid56.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36 #include "math.h"
  37 #include "sysfs.h"
  38 #include "qgroup.h"
  39
  40 #undef SCRAMBLE_DELAYED_REFS
  41
  42 /*
  43  * control flags for do_chunk_alloc's force field
  44  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  45  * if we really need one.
  46  *
  47  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  48  * if we have very few chunks already allocated.  This is
  49  * used as part of the clustering code to help make sure
  50  * we have a good pool of storage to cluster in, without
  51  * filling the FS with empty chunks
  52  *
  53  * CHUNK_ALLOC_FORCE means it must try to allocate one
  54  *
  55  */
  56 enum {
  57         CHUNK_ALLOC_NO_FORCE = 0,
  58         CHUNK_ALLOC_LIMITED = 1,
  59         CHUNK_ALLOC_FORCE = 2,
  60 };
  61
  62 /*
  63  * Control how reservations are dealt with.
  64  *
  65  * RESERVE_FREE - freeing a reservation.
  66  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  67  *   ENOSPC accounting
  68  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  69  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  70  */
  71 enum {
  72         RESERVE_FREE = 0,
  73         RESERVE_ALLOC = 1,
  74         RESERVE_ALLOC_NO_ACCOUNT = 2,
  75 };
  76
  77 static int update_block_group(struct btrfs_trans_handle *trans,
  78                               struct btrfs_root *root, u64 bytenr,
  79                               u64 num_bytes, int alloc);
  80 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  81                                 struct btrfs_root *root,
  82                                 struct btrfs_delayed_ref_node *node, u64 parent,
  83                                 u64 root_objectid, u64 owner_objectid,
  84                                 u64 owner_offset, int refs_to_drop,
  85                                 struct btrfs_delayed_extent_op *extra_op);
  86 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  87                                     struct extent_buffer *leaf,
  88                                     struct btrfs_extent_item *ei);
  89 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  90                                       struct btrfs_root *root,
  91                                       u64 parent, u64 root_objectid,
  92                                       u64 flags, u64 owner, u64 offset,
  93                                       struct btrfs_key *ins, int ref_mod);
  94 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  95                                      struct btrfs_root *root,
  96                                      u64 parent, u64 root_objectid,
  97                                      u64 flags, struct btrfs_disk_key *key,
  98                                      int level, struct btrfs_key *ins,
  99                                      int no_quota);
 100 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
 101                           struct btrfs_root *extent_root, u64 flags,
 102                           int force);
 103 static int find_next_key(struct btrfs_path *path, int level,
 104                          struct btrfs_key *key);
 105 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 106                             int dump_block_groups);
 107 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 108                                        u64 num_bytes, int reserve,
 109                                        int delalloc);
 110 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 111                                u64 num_bytes);
 112 int btrfs_pin_extent(struct btrfs_root *root,
 113                      u64 bytenr, u64 num_bytes, int reserved);
 114
 115 static noinline int
 116 block_group_cache_done(struct btrfs_block_group_cache *cache)
 117 {
 118         smp_mb();
 119         return cache->cached == BTRFS_CACHE_FINISHED ||
 120                 cache->cached == BTRFS_CACHE_ERROR;
 121 }
 122
 123 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 124 {
 125         return (cache->flags & bits) == bits;
 126 }
 127
 128 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 129 {
 130         atomic_inc(&cache->count);
 131 }
 132
 133 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 134 {
 135         if (atomic_dec_and_test(&cache->count)) {
 136                 WARN_ON(cache->pinned > 0);
 137                 WARN_ON(cache->reserved > 0);
 138                 kfree(cache->free_space_ctl);
 139                 kfree(cache);
 140         }
 141 }
 142
 143 /*
 144  * this adds the block group to the fs_info rb tree for the block group
 145  * cache
 146  */
 147 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 148                                 struct btrfs_block_group_cache *block_group)
 149 {
 150         struct rb_node **p;
 151         struct rb_node *parent = NULL;
 152         struct btrfs_block_group_cache *cache;
 153
 154         spin_lock(&info->block_group_cache_lock);
 155         p = &info->block_group_cache_tree.rb_node;
 156
 157         while (*p) {
 158                 parent = *p;
 159                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 160                                  cache_node);
 161                 if (block_group->key.objectid < cache->key.objectid) {
 162                         p = &(*p)->rb_left;
 163                 } else if (block_group->key.objectid > cache->key.objectid) {
 164                         p = &(*p)->rb_right;
 165                 } else {
 166                         spin_unlock(&info->block_group_cache_lock);
 167                         return -EEXIST;
 168                 }
 169         }
 170
 171         rb_link_node(&block_group->cache_node, parent, p);
 172         rb_insert_color(&block_group->cache_node,
 173                         &info->block_group_cache_tree);
 174
 175         if (info->first_logical_byte > block_group->key.objectid)
 176                 info->first_logical_byte = block_group->key.objectid;
 177
 178         spin_unlock(&info->block_group_cache_lock);
 179
 180         return 0;
 181 }
 182
 183 /*
 184  * This will return the block group at or after bytenr if contains is 0, else
 185  * it will return the block group that contains the bytenr
 186  */
 187 static struct btrfs_block_group_cache *
 188 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 189                               int contains)
 190 {
 191         struct btrfs_block_group_cache *cache, *ret = NULL;
 192         struct rb_node *n;
 193         u64 end, start;
 194
 195         spin_lock(&info->block_group_cache_lock);
 196         n = info->block_group_cache_tree.rb_node;
 197
 198         while (n) {
 199                 cache = rb_entry(n, struct btrfs_block_group_cache,
 200                                  cache_node);
 201                 end = cache->key.objectid + cache->key.offset - 1;
 202                 start = cache->key.objectid;
 203
 204                 if (bytenr < start) {
 205                         if (!contains && (!ret || start < ret->key.objectid))
 206                                 ret = cache;
 207                         n = n->rb_left;
 208                 } else if (bytenr > start) {
 209                         if (contains && bytenr <= end) {
 210                                 ret = cache;
 211                                 break;
 212                         }
 213                         n = n->rb_right;
 214                 } else {
 215                         ret = cache;
 216                         break;
 217                 }
 218         }
 219         if (ret) {
 220                 btrfs_get_block_group(ret);
 221                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 222                         info->first_logical_byte = ret->key.objectid;
 223         }
 224         spin_unlock(&info->block_group_cache_lock);
 225
 226         return ret;
 227 }
 228
 229 static int add_excluded_extent(struct btrfs_root *root,
 230                                u64 start, u64 num_bytes)
 231 {
 232         u64 end = start + num_bytes - 1;
 233         set_extent_bits(&root->fs_info->freed_extents[0],
 234                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 235         set_extent_bits(&root->fs_info->freed_extents[1],
 236                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 237         return 0;
 238 }
 239
 240 static void free_excluded_extents(struct btrfs_root *root,
 241                                   struct btrfs_block_group_cache *cache)
 242 {
 243         u64 start, end;
 244
 245         start = cache->key.objectid;
 246         end = start + cache->key.offset - 1;
 247
 248         clear_extent_bits(&root->fs_info->freed_extents[0],
 249                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 250         clear_extent_bits(&root->fs_info->freed_extents[1],
 251                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 252 }
 253
 254 static int exclude_super_stripes(struct btrfs_root *root,
 255                                  struct btrfs_block_group_cache *cache)
 256 {
 257         u64 bytenr;
 258         u64 *logical;
 259         int stripe_len;
 260         int i, nr, ret;
 261
 262         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 263                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 264                 cache->bytes_super += stripe_len;
 265                 ret = add_excluded_extent(root, cache->key.objectid,
 266                                           stripe_len);
 267                 if (ret)
 268                         return ret;
 269         }
 270
 271         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 272                 bytenr = btrfs_sb_offset(i);
 273                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 274                                        cache->key.objectid, bytenr,
 275                                        0, &logical, &nr, &stripe_len);
 276                 if (ret)
 277                         return ret;
 278
 279                 while (nr--) {
 280                         u64 start, len;
 281
 282                         if (logical[nr] > cache->key.objectid +
 283                             cache->key.offset)
 284                                 continue;
 285
 286                         if (logical[nr] + stripe_len <= cache->key.objectid)
 287                                 continue;
 288
 289                         start = logical[nr];
 290                         if (start < cache->key.objectid) {
 291                                 start = cache->key.objectid;
 292                                 len = (logical[nr] + stripe_len) - start;
 293                         } else {
 294                                 len = min_t(u64, stripe_len,
 295                                             cache->key.objectid +
 296                                             cache->key.offset - start);
 297                         }
 298
 299                         cache->bytes_super += len;
 300                         ret = add_excluded_extent(root, start, len);
 301                         if (ret) {
 302                                 kfree(logical);
 303                                 return ret;
 304                         }
 305                 }
 306
 307                 kfree(logical);
 308         }
 309         return 0;
 310 }
 311
 312 static struct btrfs_caching_control *
 313 get_caching_control(struct btrfs_block_group_cache *cache)
 314 {
 315         struct btrfs_caching_control *ctl;
 316
 317         spin_lock(&cache->lock);
 318         if (!cache->caching_ctl) {
 319                 spin_unlock(&cache->lock);
 320                 return NULL;
 321         }
 322
 323         ctl = cache->caching_ctl;
 324         atomic_inc(&ctl->count);
 325         spin_unlock(&cache->lock);
 326         return ctl;
 327 }
 328
 329 static void put_caching_control(struct btrfs_caching_control *ctl)
 330 {
 331         if (atomic_dec_and_test(&ctl->count))
 332                 kfree(ctl);
 333 }
 334
 335 /*
 336  * this is only called by cache_block_group, since we could have freed extents
 337  * we need to check the pinned_extents for any extents that can't be used yet
 338  * since their free space will be released as soon as the transaction commits.
 339  */
 340 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 341                               struct btrfs_fs_info *info, u64 start, u64 end)
 342 {
 343         u64 extent_start, extent_end, size, total_added = 0;
 344         int ret;
 345
 346         while (start < end) {
 347                 ret = find_first_extent_bit(info->pinned_extents, start,
 348                                             &extent_start, &extent_end,
 349                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 350                                             NULL);
 351                 if (ret)
 352                         break;
 353
 354                 if (extent_start <= start) {
 355                         start = extent_end + 1;
 356                 } else if (extent_start > start && extent_start < end) {
 357                         size = extent_start - start;
 358                         total_added += size;
 359                         ret = btrfs_add_free_space(block_group, start,
 360                                                    size);
 361                         BUG_ON(ret); /* -ENOMEM or logic error */
 362                         start = extent_end + 1;
 363                 } else {
 364                         break;
 365                 }
 366         }
 367
 368         if (start < end) {
 369                 size = end - start;
 370                 total_added += size;
 371                 ret = btrfs_add_free_space(block_group, start, size);
 372                 BUG_ON(ret); /* -ENOMEM or logic error */
 373         }
 374
 375         return total_added;
 376 }
 377
 378 static noinline void caching_thread(struct btrfs_work *work)
 379 {
 380         struct btrfs_block_group_cache *block_group;
 381         struct btrfs_fs_info *fs_info;
 382         struct btrfs_caching_control *caching_ctl;
 383         struct btrfs_root *extent_root;
 384         struct btrfs_path *path;
 385         struct extent_buffer *leaf;
 386         struct btrfs_key key;
 387         u64 total_found = 0;
 388         u64 last = 0;
 389         u32 nritems;
 390         int ret = -ENOMEM;
 391
 392         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 393         block_group = caching_ctl->block_group;
 394         fs_info = block_group->fs_info;
 395         extent_root = fs_info->extent_root;
 396
 397         path = btrfs_alloc_path();
 398         if (!path)
 399                 goto out;
 400
 401         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 402
 403         /*
 404          * We don't want to deadlock with somebody trying to allocate a new
 405          * extent for the extent root while also trying to search the extent
 406          * root to add free space.  So we skip locking and search the commit
 407          * root, since its read-only
 408          */
 409         path->skip_locking = 1;
 410         path->search_commit_root = 1;
 411         path->reada = 1;
 412
 413         key.objectid = last;
 414         key.offset = 0;
 415         key.type = BTRFS_EXTENT_ITEM_KEY;
 416 again:
 417         mutex_lock(&caching_ctl->mutex);
 418         /* need to make sure the commit_root doesn't disappear */
 419         down_read(&fs_info->commit_root_sem);
 420
 421 next:
 422         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 423         if (ret < 0)
 424                 goto err;
 425
 426         leaf = path->nodes[0];
 427         nritems = btrfs_header_nritems(leaf);
 428
 429         while (1) {
 430                 if (btrfs_fs_closing(fs_info) > 1) {
 431                         last = (u64)-1;
 432                         break;
 433                 }
 434
 435                 if (path->slots[0] < nritems) {
 436                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 437                 } else {
 438                         ret = find_next_key(path, 0, &key);
 439                         if (ret)
 440                                 break;
 441
 442                         if (need_resched() ||
 443                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 444                                 caching_ctl->progress = last;
 445                                 btrfs_release_path(path);
 446                                 up_read(&fs_info->commit_root_sem);
 447                                 mutex_unlock(&caching_ctl->mutex);
 448                                 cond_resched();
 449                                 goto again;
 450                         }
 451
 452                         ret = btrfs_next_leaf(extent_root, path);
 453                         if (ret < 0)
 454                                 goto err;
 455                         if (ret)
 456                                 break;
 457                         leaf = path->nodes[0];
 458                         nritems = btrfs_header_nritems(leaf);
 459                         continue;
 460                 }
 461
 462                 if (key.objectid < last) {
 463                         key.objectid = last;
 464                         key.offset = 0;
 465                         key.type = BTRFS_EXTENT_ITEM_KEY;
 466
 467                         caching_ctl->progress = last;
 468                         btrfs_release_path(path);
 469                         goto next;
 470                 }
 471
 472                 if (key.objectid < block_group->key.objectid) {
 473                         path->slots[0]++;
 474                         continue;
 475                 }
 476
 477                 if (key.objectid >= block_group->key.objectid +
 478                     block_group->key.offset)
 479                         break;
 480
 481                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 482                     key.type == BTRFS_METADATA_ITEM_KEY) {
 483                         total_found += add_new_free_space(block_group,
 484                                                           fs_info, last,
 485                                                           key.objectid);
 486                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 487                                 last = key.objectid +
 488                                         fs_info->tree_root->nodesize;
 489                         else
 490                                 last = key.objectid + key.offset;
 491
 492                         if (total_found > (1024 * 1024 * 2)) {
 493                                 total_found = 0;
 494                                 wake_up(&caching_ctl->wait);
 495                         }
 496                 }
 497                 path->slots[0]++;
 498         }
 499         ret = 0;
 500
 501         total_found += add_new_free_space(block_group, fs_info, last,
 502                                           block_group->key.objectid +
 503                                           block_group->key.offset);
 504         caching_ctl->progress = (u64)-1;
 505
 506         spin_lock(&block_group->lock);
 507         block_group->caching_ctl = NULL;
 508         block_group->cached = BTRFS_CACHE_FINISHED;
 509         spin_unlock(&block_group->lock);
 510
 511 err:
 512         btrfs_free_path(path);
 513         up_read(&fs_info->commit_root_sem);
 514
 515         free_excluded_extents(extent_root, block_group);
 516
 517         mutex_unlock(&caching_ctl->mutex);
 518 out:
 519         if (ret) {
 520                 spin_lock(&block_group->lock);
 521                 block_group->caching_ctl = NULL;
 522                 block_group->cached = BTRFS_CACHE_ERROR;
 523                 spin_unlock(&block_group->lock);
 524         }
 525         wake_up(&caching_ctl->wait);
 526
 527         put_caching_control(caching_ctl);
 528         btrfs_put_block_group(block_group);
 529 }
 530
 531 static int cache_block_group(struct btrfs_block_group_cache *cache,
 532                              int load_cache_only)
 533 {
 534         DEFINE_WAIT(wait);
 535         struct btrfs_fs_info *fs_info = cache->fs_info;
 536         struct btrfs_caching_control *caching_ctl;
 537         int ret = 0;
 538
 539         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 540         if (!caching_ctl)
 541                 return -ENOMEM;
 542
 543         INIT_LIST_HEAD(&caching_ctl->list);
 544         mutex_init(&caching_ctl->mutex);
 545         init_waitqueue_head(&caching_ctl->wait);
 546         caching_ctl->block_group = cache;
 547         caching_ctl->progress = cache->key.objectid;
 548         atomic_set(&caching_ctl->count, 1);
 549         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 550                         caching_thread, NULL, NULL);
 551
 552         spin_lock(&cache->lock);
 553         /*
 554          * This should be a rare occasion, but this could happen I think in the
 555          * case where one thread starts to load the space cache info, and then
 556          * some other thread starts a transaction commit which tries to do an
 557          * allocation while the other thread is still loading the space cache
 558          * info.  The previous loop should have kept us from choosing this block
 559          * group, but if we've moved to the state where we will wait on caching
 560          * block groups we need to first check if we're doing a fast load here,
 561          * so we can wait for it to finish, otherwise we could end up allocating
 562          * from a block group who's cache gets evicted for one reason or
 563          * another.
 564          */
 565         while (cache->cached == BTRFS_CACHE_FAST) {
 566                 struct btrfs_caching_control *ctl;
 567
 568                 ctl = cache->caching_ctl;
 569                 atomic_inc(&ctl->count);
 570                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 571                 spin_unlock(&cache->lock);
 572
 573                 schedule();
 574
 575                 finish_wait(&ctl->wait, &wait);
 576                 put_caching_control(ctl);
 577                 spin_lock(&cache->lock);
 578         }
 579
 580         if (cache->cached != BTRFS_CACHE_NO) {
 581                 spin_unlock(&cache->lock);
 582                 kfree(caching_ctl);
 583                 return 0;
 584         }
 585         WARN_ON(cache->caching_ctl);
 586         cache->caching_ctl = caching_ctl;
 587         cache->cached = BTRFS_CACHE_FAST;
 588         spin_unlock(&cache->lock);
 589
 590         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 591                 mutex_lock(&caching_ctl->mutex);
 592                 ret = load_free_space_cache(fs_info, cache);
 593
 594                 spin_lock(&cache->lock);
 595                 if (ret == 1) {
 596                         cache->caching_ctl = NULL;
 597                         cache->cached = BTRFS_CACHE_FINISHED;
 598                         cache->last_byte_to_unpin = (u64)-1;
 599                         caching_ctl->progress = (u64)-1;
 600                 } else {
 601                         if (load_cache_only) {
 602                                 cache->caching_ctl = NULL;
 603                                 cache->cached = BTRFS_CACHE_NO;
 604                         } else {
 605                                 cache->cached = BTRFS_CACHE_STARTED;
 606                                 cache->has_caching_ctl = 1;
 607                         }
 608                 }
 609                 spin_unlock(&cache->lock);
 610                 mutex_unlock(&caching_ctl->mutex);
 611
 612                 wake_up(&caching_ctl->wait);
 613                 if (ret == 1) {
 614                         put_caching_control(caching_ctl);
 615                         free_excluded_extents(fs_info->extent_root, cache);
 616                         return 0;
 617                 }
 618         } else {
 619                 /*
 620                  * We are not going to do the fast caching, set cached to the
 621                  * appropriate value and wakeup any waiters.
 622                  */
 623                 spin_lock(&cache->lock);
 624                 if (load_cache_only) {
 625                         cache->caching_ctl = NULL;
 626                         cache->cached = BTRFS_CACHE_NO;
 627                 } else {
 628                         cache->cached = BTRFS_CACHE_STARTED;
 629                         cache->has_caching_ctl = 1;
 630                 }
 631                 spin_unlock(&cache->lock);
 632                 wake_up(&caching_ctl->wait);
 633         }
 634
 635         if (load_cache_only) {
 636                 put_caching_control(caching_ctl);
 637                 return 0;
 638         }
 639
 640         down_write(&fs_info->commit_root_sem);
 641         atomic_inc(&caching_ctl->count);
 642         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 643         up_write(&fs_info->commit_root_sem);
 644
 645         btrfs_get_block_group(cache);
 646
 647         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 648
 649         return ret;
 650 }
 651
 652 /*
 653  * return the block group that starts at or after bytenr
 654  */
 655 static struct btrfs_block_group_cache *
 656 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 657 {
 658         struct btrfs_block_group_cache *cache;
 659
 660         cache = block_group_cache_tree_search(info, bytenr, 0);
 661
 662         return cache;
 663 }
 664
 665 /*
 666  * return the block group that contains the given bytenr
 667  */
 668 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 669                                                  struct btrfs_fs_info *info,
 670                                                  u64 bytenr)
 671 {
 672         struct btrfs_block_group_cache *cache;
 673
 674         cache = block_group_cache_tree_search(info, bytenr, 1);
 675
 676         return cache;
 677 }
 678
 679 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 680                                                   u64 flags)
 681 {
 682         struct list_head *head = &info->space_info;
 683         struct btrfs_space_info *found;
 684
 685         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 686
 687         rcu_read_lock();
 688         list_for_each_entry_rcu(found, head, list) {
 689                 if (found->flags & flags) {
 690                         rcu_read_unlock();
 691                         return found;
 692                 }
 693         }
 694         rcu_read_unlock();
 695         return NULL;
 696 }
 697
 698 /*
 699  * after adding space to the filesystem, we need to clear the full flags
 700  * on all the space infos.
 701  */
 702 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 703 {
 704         struct list_head *head = &info->space_info;
 705         struct btrfs_space_info *found;
 706
 707         rcu_read_lock();
 708         list_for_each_entry_rcu(found, head, list)
 709                 found->full = 0;
 710         rcu_read_unlock();
 711 }
 712
 713 /* simple helper to search for an existing data extent at a given offset */
 714 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 715 {
 716         int ret;
 717         struct btrfs_key key;
 718         struct btrfs_path *path;
 719
 720         path = btrfs_alloc_path();
 721         if (!path)
 722                 return -ENOMEM;
 723
 724         key.objectid = start;
 725         key.offset = len;
 726         key.type = BTRFS_EXTENT_ITEM_KEY;
 727         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 728                                 0, 0);
 729         btrfs_free_path(path);
 730         return ret;
 731 }
 732
 733 /*
 734  * helper function to lookup reference count and flags of a tree block.
 735  *
 736  * the head node for delayed ref is used to store the sum of all the
 737  * reference count modifications queued up in the rbtree. the head
 738  * node may also store the extent flags to set. This way you can check
 739  * to see what the reference count and extent flags would be if all of
 740  * the delayed refs are not processed.
 741  */
 742 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 743                              struct btrfs_root *root, u64 bytenr,
 744                              u64 offset, int metadata, u64 *refs, u64 *flags)
 745 {
 746         struct btrfs_delayed_ref_head *head;
 747         struct btrfs_delayed_ref_root *delayed_refs;
 748         struct btrfs_path *path;
 749         struct btrfs_extent_item *ei;
 750         struct extent_buffer *leaf;
 751         struct btrfs_key key;
 752         u32 item_size;
 753         u64 num_refs;
 754         u64 extent_flags;
 755         int ret;
 756
 757         /*
 758          * If we don't have skinny metadata, don't bother doing anything
 759          * different
 760          */
 761         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 762                 offset = root->nodesize;
 763                 metadata = 0;
 764         }
 765
 766         path = btrfs_alloc_path();
 767         if (!path)
 768                 return -ENOMEM;
 769
 770         if (!trans) {
 771                 path->skip_locking = 1;
 772                 path->search_commit_root = 1;
 773         }
 774
 775 search_again:
 776         key.objectid = bytenr;
 777         key.offset = offset;
 778         if (metadata)
 779                 key.type = BTRFS_METADATA_ITEM_KEY;
 780         else
 781                 key.type = BTRFS_EXTENT_ITEM_KEY;
 782
 783         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 784                                 &key, path, 0, 0);
 785         if (ret < 0)
 786                 goto out_free;
 787
 788         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 789                 if (path->slots[0]) {
 790                         path->slots[0]--;
 791                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 792                                               path->slots[0]);
 793                         if (key.objectid == bytenr &&
 794                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 795                             key.offset == root->nodesize)
 796                                 ret = 0;
 797                 }
 798         }
 799
 800         if (ret == 0) {
 801                 leaf = path->nodes[0];
 802                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 803                 if (item_size >= sizeof(*ei)) {
 804                         ei = btrfs_item_ptr(leaf, path->slots[0],
 805                                             struct btrfs_extent_item);
 806                         num_refs = btrfs_extent_refs(leaf, ei);
 807                         extent_flags = btrfs_extent_flags(leaf, ei);
 808                 } else {
 809 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 810                         struct btrfs_extent_item_v0 *ei0;
 811                         BUG_ON(item_size != sizeof(*ei0));
 812                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 813                                              struct btrfs_extent_item_v0);
 814                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 815                         /* FIXME: this isn't correct for data */
 816                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 817 #else
 818                         BUG();
 819 #endif
 820                 }
 821                 BUG_ON(num_refs == 0);
 822         } else {
 823                 num_refs = 0;
 824                 extent_flags = 0;
 825                 ret = 0;
 826         }
 827
 828         if (!trans)
 829                 goto out;
 830
 831         delayed_refs = &trans->transaction->delayed_refs;
 832         spin_lock(&delayed_refs->lock);
 833         head = btrfs_find_delayed_ref_head(trans, bytenr);
 834         if (head) {
 835                 if (!mutex_trylock(&head->mutex)) {
 836                         atomic_inc(&head->node.refs);
 837                         spin_unlock(&delayed_refs->lock);
 838
 839                         btrfs_release_path(path);
 840
 841                         /*
 842                          * Mutex was contended, block until it's released and try
 843                          * again
 844                          */
 845                         mutex_lock(&head->mutex);
 846                         mutex_unlock(&head->mutex);
 847                         btrfs_put_delayed_ref(&head->node);
 848                         goto search_again;
 849                 }
 850                 spin_lock(&head->lock);
 851                 if (head->extent_op && head->extent_op->update_flags)
 852                         extent_flags |= head->extent_op->flags_to_set;
 853                 else
 854                         BUG_ON(num_refs == 0);
 855
 856                 num_refs += head->node.ref_mod;
 857                 spin_unlock(&head->lock);
 858                 mutex_unlock(&head->mutex);
 859         }
 860         spin_unlock(&delayed_refs->lock);
 861 out:
 862         WARN_ON(num_refs == 0);
 863         if (refs)
 864                 *refs = num_refs;
 865         if (flags)
 866                 *flags = extent_flags;
 867 out_free:
 868         btrfs_free_path(path);
 869         return ret;
 870 }
 871
 872 /*
 873  * Back reference rules.  Back refs have three main goals:
 874  *
 875  * 1) differentiate between all holders of references to an extent so that
 876  *    when a reference is dropped we can make sure it was a valid reference
 877  *    before freeing the extent.
 878  *
 879  * 2) Provide enough information to quickly find the holders of an extent
 880  *    if we notice a given block is corrupted or bad.
 881  *
 882  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 883  *    maintenance.  This is actually the same as #2, but with a slightly
 884  *    different use case.
 885  *
 886  * There are two kinds of back refs. The implicit back refs is optimized
 887  * for pointers in non-shared tree blocks. For a given pointer in a block,
 888  * back refs of this kind provide information about the block's owner tree
 889  * and the pointer's key. These information allow us to find the block by
 890  * b-tree searching. The full back refs is for pointers in tree blocks not
 891  * referenced by their owner trees. The location of tree block is recorded
 892  * in the back refs. Actually the full back refs is generic, and can be
 893  * used in all cases the implicit back refs is used. The major shortcoming
 894  * of the full back refs is its overhead. Every time a tree block gets
 895  * COWed, we have to update back refs entry for all pointers in it.
 896  *
 897  * For a newly allocated tree block, we use implicit back refs for
 898  * pointers in it. This means most tree related operations only involve
 899  * implicit back refs. For a tree block created in old transaction, the
 900  * only way to drop a reference to it is COW it. So we can detect the
 901  * event that tree block loses its owner tree's reference and do the
 902  * back refs conversion.
 903  *
 904  * When a tree block is COW'd through a tree, there are four cases:
 905  *
 906  * The reference count of the block is one and the tree is the block's
 907  * owner tree. Nothing to do in this case.
 908  *
 909  * The reference count of the block is one and the tree is not the
 910  * block's owner tree. In this case, full back refs is used for pointers
 911  * in the block. Remove these full back refs, add implicit back refs for
 912  * every pointers in the new block.
 913  *
 914  * The reference count of the block is greater than one and the tree is
 915  * the block's owner tree. In this case, implicit back refs is used for
 916  * pointers in the block. Add full back refs for every pointers in the
 917  * block, increase lower level extents' reference counts. The original
 918  * implicit back refs are entailed to the new block.
 919  *
 920  * The reference count of the block is greater than one and the tree is
 921  * not the block's owner tree. Add implicit back refs for every pointer in
 922  * the new block, increase lower level extents' reference count.
 923  *
 924  * Back Reference Key composing:
 925  *
 926  * The key objectid corresponds to the first byte in the extent,
 927  * The key type is used to differentiate between types of back refs.
 928  * There are different meanings of the key offset for different types
 929  * of back refs.
 930  *
 931  * File extents can be referenced by:
 932  *
 933  * - multiple snapshots, subvolumes, or different generations in one subvol
 934  * - different files inside a single subvolume
 935  * - different offsets inside a file (bookend extents in file.c)
 936  *
 937  * The extent ref structure for the implicit back refs has fields for:
 938  *
 939  * - Objectid of the subvolume root
 940  * - objectid of the file holding the reference
 941  * - original offset in the file
 942  * - how many bookend extents
 943  *
 944  * The key offset for the implicit back refs is hash of the first
 945  * three fields.
 946  *
 947  * The extent ref structure for the full back refs has field for:
 948  *
 949  * - number of pointers in the tree leaf
 950  *
 951  * The key offset for the implicit back refs is the first byte of
 952  * the tree leaf
 953  *
 954  * When a file extent is allocated, The implicit back refs is used.
 955  * the fields are filled in:
 956  *
 957  *     (root_key.objectid, inode objectid, offset in file, 1)
 958  *
 959  * When a file extent is removed file truncation, we find the
 960  * corresponding implicit back refs and check the following fields:
 961  *
 962  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 963  *
 964  * Btree extents can be referenced by:
 965  *
 966  * - Different subvolumes
 967  *
 968  * Both the implicit back refs and the full back refs for tree blocks
 969  * only consist of key. The key offset for the implicit back refs is
 970  * objectid of block's owner tree. The key offset for the full back refs
 971  * is the first byte of parent block.
 972  *
 973  * When implicit back refs is used, information about the lowest key and
 974  * level of the tree block are required. These information are stored in
 975  * tree block info structure.
 976  */
 977
 978 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 979 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 980                                   struct btrfs_root *root,
 981                                   struct btrfs_path *path,
 982                                   u64 owner, u32 extra_size)
 983 {
 984         struct btrfs_extent_item *item;
 985         struct btrfs_extent_item_v0 *ei0;
 986         struct btrfs_extent_ref_v0 *ref0;
 987         struct btrfs_tree_block_info *bi;
 988         struct extent_buffer *leaf;
 989         struct btrfs_key key;
 990         struct btrfs_key found_key;
 991         u32 new_size = sizeof(*item);
 992         u64 refs;
 993         int ret;
 994
 995         leaf = path->nodes[0];
 996         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 997
 998         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 999         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1000                              struct btrfs_extent_item_v0);
1001         refs = btrfs_extent_refs_v0(leaf, ei0);
1002
1003         if (owner == (u64)-1) {
1004                 while (1) {
1005                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1006                                 ret = btrfs_next_leaf(root, path);
1007                                 if (ret < 0)
1008                                         return ret;
1009                                 BUG_ON(ret > 0); /* Corruption */
1010                                 leaf = path->nodes[0];
1011                         }
1012                         btrfs_item_key_to_cpu(leaf, &found_key,
1013                                               path->slots[0]);
1014                         BUG_ON(key.objectid != found_key.objectid);
1015                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1016                                 path->slots[0]++;
1017                                 continue;
1018                         }
1019                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1020                                               struct btrfs_extent_ref_v0);
1021                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1022                         break;
1023                 }
1024         }
1025         btrfs_release_path(path);
1026
1027         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1028                 new_size += sizeof(*bi);
1029
1030         new_size -= sizeof(*ei0);
1031         ret = btrfs_search_slot(trans, root, &key, path,
1032                                 new_size + extra_size, 1);
1033         if (ret < 0)
1034                 return ret;
1035         BUG_ON(ret); /* Corruption */
1036
1037         btrfs_extend_item(root, path, new_size);
1038
1039         leaf = path->nodes[0];
1040         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1041         btrfs_set_extent_refs(leaf, item, refs);
1042         /* FIXME: get real generation */
1043         btrfs_set_extent_generation(leaf, item, 0);
1044         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1045                 btrfs_set_extent_flags(leaf, item,
1046                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1047                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1048                 bi = (struct btrfs_tree_block_info *)(item + 1);
1049                 /* FIXME: get first key of the block */
1050                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1051                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1052         } else {
1053                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1054         }
1055         btrfs_mark_buffer_dirty(leaf);
1056         return 0;
1057 }
1058 #endif
1059
1060 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1061 {
1062         u32 high_crc = ~(u32)0;
1063         u32 low_crc = ~(u32)0;
1064         __le64 lenum;
1065
1066         lenum = cpu_to_le64(root_objectid);
1067         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1068         lenum = cpu_to_le64(owner);
1069         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1070         lenum = cpu_to_le64(offset);
1071         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1072
1073         return ((u64)high_crc << 31) ^ (u64)low_crc;
1074 }
1075
1076 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1077                                      struct btrfs_extent_data_ref *ref)
1078 {
1079         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1080                                     btrfs_extent_data_ref_objectid(leaf, ref),
1081                                     btrfs_extent_data_ref_offset(leaf, ref));
1082 }
1083
1084 static int match_extent_data_ref(struct extent_buffer *leaf,
1085                                  struct btrfs_extent_data_ref *ref,
1086                                  u64 root_objectid, u64 owner, u64 offset)
1087 {
1088         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1089             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1090             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1091                 return 0;
1092         return 1;
1093 }
1094
1095 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1096                                            struct btrfs_root *root,
1097                                            struct btrfs_path *path,
1098                                            u64 bytenr, u64 parent,
1099                                            u64 root_objectid,
1100                                            u64 owner, u64 offset)
1101 {
1102         struct btrfs_key key;
1103         struct btrfs_extent_data_ref *ref;
1104         struct extent_buffer *leaf;
1105         u32 nritems;
1106         int ret;
1107         int recow;
1108         int err = -ENOENT;
1109
1110         key.objectid = bytenr;
1111         if (parent) {
1112                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1113                 key.offset = parent;
1114         } else {
1115                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1116                 key.offset = hash_extent_data_ref(root_objectid,
1117                                                   owner, offset);
1118         }
1119 again:
1120         recow = 0;
1121         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1122         if (ret < 0) {
1123                 err = ret;
1124                 goto fail;
1125         }
1126
1127         if (parent) {
1128                 if (!ret)
1129                         return 0;
1130 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1131                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1132                 btrfs_release_path(path);
1133                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1134                 if (ret < 0) {
1135                         err = ret;
1136                         goto fail;
1137                 }
1138                 if (!ret)
1139                         return 0;
1140 #endif
1141                 goto fail;
1142         }
1143
1144         leaf = path->nodes[0];
1145         nritems = btrfs_header_nritems(leaf);
1146         while (1) {
1147                 if (path->slots[0] >= nritems) {
1148                         ret = btrfs_next_leaf(root, path);
1149                         if (ret < 0)
1150                                 err = ret;
1151                         if (ret)
1152                                 goto fail;
1153
1154                         leaf = path->nodes[0];
1155                         nritems = btrfs_header_nritems(leaf);
1156                         recow = 1;
1157                 }
1158
1159                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1160                 if (key.objectid != bytenr ||
1161                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1162                         goto fail;
1163
1164                 ref = btrfs_item_ptr(leaf, path->slots[0],
1165                                      struct btrfs_extent_data_ref);
1166
1167                 if (match_extent_data_ref(leaf, ref, root_objectid,
1168                                           owner, offset)) {
1169                         if (recow) {
1170                                 btrfs_release_path(path);
1171                                 goto again;
1172                         }
1173                         err = 0;
1174                         break;
1175                 }
1176                 path->slots[0]++;
1177         }
1178 fail:
1179         return err;
1180 }
1181
1182 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1183                                            struct btrfs_root *root,
1184                                            struct btrfs_path *path,
1185                                            u64 bytenr, u64 parent,
1186                                            u64 root_objectid, u64 owner,
1187                                            u64 offset, int refs_to_add)
1188 {
1189         struct btrfs_key key;
1190         struct extent_buffer *leaf;
1191         u32 size;
1192         u32 num_refs;
1193         int ret;
1194
1195         key.objectid = bytenr;
1196         if (parent) {
1197                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1198                 key.offset = parent;
1199                 size = sizeof(struct btrfs_shared_data_ref);
1200         } else {
1201                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1202                 key.offset = hash_extent_data_ref(root_objectid,
1203                                                   owner, offset);
1204                 size = sizeof(struct btrfs_extent_data_ref);
1205         }
1206
1207         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1208         if (ret && ret != -EEXIST)
1209                 goto fail;
1210
1211         leaf = path->nodes[0];
1212         if (parent) {
1213                 struct btrfs_shared_data_ref *ref;
1214                 ref = btrfs_item_ptr(leaf, path->slots[0],
1215                                      struct btrfs_shared_data_ref);
1216                 if (ret == 0) {
1217                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1218                 } else {
1219                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1220                         num_refs += refs_to_add;
1221                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1222                 }
1223         } else {
1224                 struct btrfs_extent_data_ref *ref;
1225                 while (ret == -EEXIST) {
1226                         ref = btrfs_item_ptr(leaf, path->slots[0],
1227                                              struct btrfs_extent_data_ref);
1228                         if (match_extent_data_ref(leaf, ref, root_objectid,
1229                                                   owner, offset))
1230                                 break;
1231                         btrfs_release_path(path);
1232                         key.offset++;
1233                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1234                                                       size);
1235                         if (ret && ret != -EEXIST)
1236                                 goto fail;
1237
1238                         leaf = path->nodes[0];
1239                 }
1240                 ref = btrfs_item_ptr(leaf, path->slots[0],
1241                                      struct btrfs_extent_data_ref);
1242                 if (ret == 0) {
1243                         btrfs_set_extent_data_ref_root(leaf, ref,
1244                                                        root_objectid);
1245                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1246                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1247                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1248                 } else {
1249                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1250                         num_refs += refs_to_add;
1251                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1252                 }
1253         }
1254         btrfs_mark_buffer_dirty(leaf);
1255         ret = 0;
1256 fail:
1257         btrfs_release_path(path);
1258         return ret;
1259 }
1260
1261 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1262                                            struct btrfs_root *root,
1263                                            struct btrfs_path *path,
1264                                            int refs_to_drop, int *last_ref)
1265 {
1266         struct btrfs_key key;
1267         struct btrfs_extent_data_ref *ref1 = NULL;
1268         struct btrfs_shared_data_ref *ref2 = NULL;
1269         struct extent_buffer *leaf;
1270         u32 num_refs = 0;
1271         int ret = 0;
1272
1273         leaf = path->nodes[0];
1274         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1275
1276         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1277                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1278                                       struct btrfs_extent_data_ref);
1279                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1280         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1281                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1282                                       struct btrfs_shared_data_ref);
1283                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1284 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1285         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1286                 struct btrfs_extent_ref_v0 *ref0;
1287                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1288                                       struct btrfs_extent_ref_v0);
1289                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1290 #endif
1291         } else {
1292                 BUG();
1293         }
1294
1295         BUG_ON(num_refs < refs_to_drop);
1296         num_refs -= refs_to_drop;
1297
1298         if (num_refs == 0) {
1299                 ret = btrfs_del_item(trans, root, path);
1300                 *last_ref = 1;
1301         } else {
1302                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1303                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1304                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1305                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1306 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1307                 else {
1308                         struct btrfs_extent_ref_v0 *ref0;
1309                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1310                                         struct btrfs_extent_ref_v0);
1311                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1312                 }
1313 #endif
1314                 btrfs_mark_buffer_dirty(leaf);
1315         }
1316         return ret;
1317 }
1318
1319 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1320                                           struct btrfs_path *path,
1321                                           struct btrfs_extent_inline_ref *iref)
1322 {
1323         struct btrfs_key key;
1324         struct extent_buffer *leaf;
1325         struct btrfs_extent_data_ref *ref1;
1326         struct btrfs_shared_data_ref *ref2;
1327         u32 num_refs = 0;
1328
1329         leaf = path->nodes[0];
1330         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1331         if (iref) {
1332                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1333                     BTRFS_EXTENT_DATA_REF_KEY) {
1334                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1335                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1336                 } else {
1337                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1338                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1339                 }
1340         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1341                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1342                                       struct btrfs_extent_data_ref);
1343                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1344         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1345                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1346                                       struct btrfs_shared_data_ref);
1347                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1348 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1349         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1350                 struct btrfs_extent_ref_v0 *ref0;
1351                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1352                                       struct btrfs_extent_ref_v0);
1353                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1354 #endif
1355         } else {
1356                 WARN_ON(1);
1357         }
1358         return num_refs;
1359 }
1360
1361 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1362                                           struct btrfs_root *root,
1363                                           struct btrfs_path *path,
1364                                           u64 bytenr, u64 parent,
1365                                           u64 root_objectid)
1366 {
1367         struct btrfs_key key;
1368         int ret;
1369
1370         key.objectid = bytenr;
1371         if (parent) {
1372                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1373                 key.offset = parent;
1374         } else {
1375                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1376                 key.offset = root_objectid;
1377         }
1378
1379         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1380         if (ret > 0)
1381                 ret = -ENOENT;
1382 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1383         if (ret == -ENOENT && parent) {
1384                 btrfs_release_path(path);
1385                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1386                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1387                 if (ret > 0)
1388                         ret = -ENOENT;
1389         }
1390 #endif
1391         return ret;
1392 }
1393
1394 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1395                                           struct btrfs_root *root,
1396                                           struct btrfs_path *path,
1397                                           u64 bytenr, u64 parent,
1398                                           u64 root_objectid)
1399 {
1400         struct btrfs_key key;
1401         int ret;
1402
1403         key.objectid = bytenr;
1404         if (parent) {
1405                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1406                 key.offset = parent;
1407         } else {
1408                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1409                 key.offset = root_objectid;
1410         }
1411
1412         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1413         btrfs_release_path(path);
1414         return ret;
1415 }
1416
1417 static inline int extent_ref_type(u64 parent, u64 owner)
1418 {
1419         int type;
1420         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1421                 if (parent > 0)
1422                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1423                 else
1424                         type = BTRFS_TREE_BLOCK_REF_KEY;
1425         } else {
1426                 if (parent > 0)
1427                         type = BTRFS_SHARED_DATA_REF_KEY;
1428                 else
1429                         type = BTRFS_EXTENT_DATA_REF_KEY;
1430         }
1431         return type;
1432 }
1433
1434 static int find_next_key(struct btrfs_path *path, int level,
1435                          struct btrfs_key *key)
1436
1437 {
1438         for (; level < BTRFS_MAX_LEVEL; level++) {
1439                 if (!path->nodes[level])
1440                         break;
1441                 if (path->slots[level] + 1 >=
1442                     btrfs_header_nritems(path->nodes[level]))
1443                         continue;
1444                 if (level == 0)
1445                         btrfs_item_key_to_cpu(path->nodes[level], key,
1446                                               path->slots[level] + 1);
1447                 else
1448                         btrfs_node_key_to_cpu(path->nodes[level], key,
1449                                               path->slots[level] + 1);
1450                 return 0;
1451         }
1452         return 1;
1453 }
1454
1455 /*
1456  * look for inline back ref. if back ref is found, *ref_ret is set
1457  * to the address of inline back ref, and 0 is returned.
1458  *
1459  * if back ref isn't found, *ref_ret is set to the address where it
1460  * should be inserted, and -ENOENT is returned.
1461  *
1462  * if insert is true and there are too many inline back refs, the path
1463  * points to the extent item, and -EAGAIN is returned.
1464  *
1465  * NOTE: inline back refs are ordered in the same way that back ref
1466  *       items in the tree are ordered.
1467  */
1468 static noinline_for_stack
1469 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1470                                  struct btrfs_root *root,
1471                                  struct btrfs_path *path,
1472                                  struct btrfs_extent_inline_ref **ref_ret,
1473                                  u64 bytenr, u64 num_bytes,
1474                                  u64 parent, u64 root_objectid,
1475                                  u64 owner, u64 offset, int insert)
1476 {
1477         struct btrfs_key key;
1478         struct extent_buffer *leaf;
1479         struct btrfs_extent_item *ei;
1480         struct btrfs_extent_inline_ref *iref;
1481         u64 flags;
1482         u64 item_size;
1483         unsigned long ptr;
1484         unsigned long end;
1485         int extra_size;
1486         int type;
1487         int want;
1488         int ret;
1489         int err = 0;
1490         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1491                                                  SKINNY_METADATA);
1492
1493         key.objectid = bytenr;
1494         key.type = BTRFS_EXTENT_ITEM_KEY;
1495         key.offset = num_bytes;
1496
1497         want = extent_ref_type(parent, owner);
1498         if (insert) {
1499                 extra_size = btrfs_extent_inline_ref_size(want);
1500                 path->keep_locks = 1;
1501         } else
1502                 extra_size = -1;
1503
1504         /*
1505          * Owner is our parent level, so we can just add one to get the level
1506          * for the block we are interested in.
1507          */
1508         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1509                 key.type = BTRFS_METADATA_ITEM_KEY;
1510                 key.offset = owner;
1511         }
1512
1513 again:
1514         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1515         if (ret < 0) {
1516                 err = ret;
1517                 goto out;
1518         }
1519
1520         /*
1521          * We may be a newly converted file system which still has the old fat
1522          * extent entries for metadata, so try and see if we have one of those.
1523          */
1524         if (ret > 0 && skinny_metadata) {
1525                 skinny_metadata = false;
1526                 if (path->slots[0]) {
1527                         path->slots[0]--;
1528                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1529                                               path->slots[0]);
1530                         if (key.objectid == bytenr &&
1531                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1532                             key.offset == num_bytes)
1533                                 ret = 0;
1534                 }
1535                 if (ret) {
1536                         key.objectid = bytenr;
1537                         key.type = BTRFS_EXTENT_ITEM_KEY;
1538                         key.offset = num_bytes;
1539                         btrfs_release_path(path);
1540                         goto again;
1541                 }
1542         }
1543
1544         if (ret && !insert) {
1545                 err = -ENOENT;
1546                 goto out;
1547         } else if (WARN_ON(ret)) {
1548                 err = -EIO;
1549                 goto out;
1550         }
1551
1552         leaf = path->nodes[0];
1553         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1554 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1555         if (item_size < sizeof(*ei)) {
1556                 if (!insert) {
1557                         err = -ENOENT;
1558                         goto out;
1559                 }
1560                 ret = convert_extent_item_v0(trans, root, path, owner,
1561                                              extra_size);
1562                 if (ret < 0) {
1563                         err = ret;
1564                         goto out;
1565                 }
1566                 leaf = path->nodes[0];
1567                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1568         }
1569 #endif
1570         BUG_ON(item_size < sizeof(*ei));
1571
1572         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1573         flags = btrfs_extent_flags(leaf, ei);
1574
1575         ptr = (unsigned long)(ei + 1);
1576         end = (unsigned long)ei + item_size;
1577
1578         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1579                 ptr += sizeof(struct btrfs_tree_block_info);
1580                 BUG_ON(ptr > end);
1581         }
1582
1583         err = -ENOENT;
1584         while (1) {
1585                 if (ptr >= end) {
1586                         WARN_ON(ptr > end);
1587                         break;
1588                 }
1589                 iref = (struct btrfs_extent_inline_ref *)ptr;
1590                 type = btrfs_extent_inline_ref_type(leaf, iref);
1591                 if (want < type)
1592                         break;
1593                 if (want > type) {
1594                         ptr += btrfs_extent_inline_ref_size(type);
1595                         continue;
1596                 }
1597
1598                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1599                         struct btrfs_extent_data_ref *dref;
1600                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1601                         if (match_extent_data_ref(leaf, dref, root_objectid,
1602                                                   owner, offset)) {
1603                                 err = 0;
1604                                 break;
1605                         }
1606                         if (hash_extent_data_ref_item(leaf, dref) <
1607                             hash_extent_data_ref(root_objectid, owner, offset))
1608                                 break;
1609                 } else {
1610                         u64 ref_offset;
1611                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1612                         if (parent > 0) {
1613                                 if (parent == ref_offset) {
1614                                         err = 0;
1615                                         break;
1616                                 }
1617                                 if (ref_offset < parent)
1618                                         break;
1619                         } else {
1620                                 if (root_objectid == ref_offset) {
1621                                         err = 0;
1622                                         break;
1623                                 }
1624                                 if (ref_offset < root_objectid)
1625                                         break;
1626                         }
1627                 }
1628                 ptr += btrfs_extent_inline_ref_size(type);
1629         }
1630         if (err == -ENOENT && insert) {
1631                 if (item_size + extra_size >=
1632                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1633                         err = -EAGAIN;
1634                         goto out;
1635                 }
1636                 /*
1637                  * To add new inline back ref, we have to make sure
1638                  * there is no corresponding back ref item.
1639                  * For simplicity, we just do not add new inline back
1640                  * ref if there is any kind of item for this block
1641                  */
1642                 if (find_next_key(path, 0, &key) == 0 &&
1643                     key.objectid == bytenr &&
1644                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1645                         err = -EAGAIN;
1646                         goto out;
1647                 }
1648         }
1649         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1650 out:
1651         if (insert) {
1652                 path->keep_locks = 0;
1653                 btrfs_unlock_up_safe(path, 1);
1654         }
1655         return err;
1656 }
1657
1658 /*
1659  * helper to add new inline back ref
1660  */
1661 static noinline_for_stack
1662 void setup_inline_extent_backref(struct btrfs_root *root,
1663                                  struct btrfs_path *path,
1664                                  struct btrfs_extent_inline_ref *iref,
1665                                  u64 parent, u64 root_objectid,
1666                                  u64 owner, u64 offset, int refs_to_add,
1667                                  struct btrfs_delayed_extent_op *extent_op)
1668 {
1669         struct extent_buffer *leaf;
1670         struct btrfs_extent_item *ei;
1671         unsigned long ptr;
1672         unsigned long end;
1673         unsigned long item_offset;
1674         u64 refs;
1675         int size;
1676         int type;
1677
1678         leaf = path->nodes[0];
1679         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1680         item_offset = (unsigned long)iref - (unsigned long)ei;
1681
1682         type = extent_ref_type(parent, owner);
1683         size = btrfs_extent_inline_ref_size(type);
1684
1685         btrfs_extend_item(root, path, size);
1686
1687         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1688         refs = btrfs_extent_refs(leaf, ei);
1689         refs += refs_to_add;
1690         btrfs_set_extent_refs(leaf, ei, refs);
1691         if (extent_op)
1692                 __run_delayed_extent_op(extent_op, leaf, ei);
1693
1694         ptr = (unsigned long)ei + item_offset;
1695         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1696         if (ptr < end - size)
1697                 memmove_extent_buffer(leaf, ptr + size, ptr,
1698                                       end - size - ptr);
1699
1700         iref = (struct btrfs_extent_inline_ref *)ptr;
1701         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1702         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1703                 struct btrfs_extent_data_ref *dref;
1704                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1705                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1706                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1707                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1708                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1709         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1710                 struct btrfs_shared_data_ref *sref;
1711                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1712                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1713                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1714         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1715                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1716         } else {
1717                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1718         }
1719         btrfs_mark_buffer_dirty(leaf);
1720 }
1721
1722 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1723                                  struct btrfs_root *root,
1724                                  struct btrfs_path *path,
1725                                  struct btrfs_extent_inline_ref **ref_ret,
1726                                  u64 bytenr, u64 num_bytes, u64 parent,
1727                                  u64 root_objectid, u64 owner, u64 offset)
1728 {
1729         int ret;
1730
1731         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1732                                            bytenr, num_bytes, parent,
1733                                            root_objectid, owner, offset, 0);
1734         if (ret != -ENOENT)
1735                 return ret;
1736
1737         btrfs_release_path(path);
1738         *ref_ret = NULL;
1739
1740         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1741                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1742                                             root_objectid);
1743         } else {
1744                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1745                                              root_objectid, owner, offset);
1746         }
1747         return ret;
1748 }
1749
1750 /*
1751  * helper to update/remove inline back ref
1752  */
1753 static noinline_for_stack
1754 void update_inline_extent_backref(struct btrfs_root *root,
1755                                   struct btrfs_path *path,
1756                                   struct btrfs_extent_inline_ref *iref,
1757                                   int refs_to_mod,
1758                                   struct btrfs_delayed_extent_op *extent_op,
1759                                   int *last_ref)
1760 {
1761         struct extent_buffer *leaf;
1762         struct btrfs_extent_item *ei;
1763         struct btrfs_extent_data_ref *dref = NULL;
1764         struct btrfs_shared_data_ref *sref = NULL;
1765         unsigned long ptr;
1766         unsigned long end;
1767         u32 item_size;
1768         int size;
1769         int type;
1770         u64 refs;
1771
1772         leaf = path->nodes[0];
1773         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1774         refs = btrfs_extent_refs(leaf, ei);
1775         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1776         refs += refs_to_mod;
1777         btrfs_set_extent_refs(leaf, ei, refs);
1778         if (extent_op)
1779                 __run_delayed_extent_op(extent_op, leaf, ei);
1780
1781         type = btrfs_extent_inline_ref_type(leaf, iref);
1782
1783         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1784                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1785                 refs = btrfs_extent_data_ref_count(leaf, dref);
1786         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1787                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1788                 refs = btrfs_shared_data_ref_count(leaf, sref);
1789         } else {
1790                 refs = 1;
1791                 BUG_ON(refs_to_mod != -1);
1792         }
1793
1794         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1795         refs += refs_to_mod;
1796
1797         if (refs > 0) {
1798                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1799                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1800                 else
1801                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1802         } else {
1803                 *last_ref = 1;
1804                 size =  btrfs_extent_inline_ref_size(type);
1805                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1806                 ptr = (unsigned long)iref;
1807                 end = (unsigned long)ei + item_size;
1808                 if (ptr + size < end)
1809                         memmove_extent_buffer(leaf, ptr, ptr + size,
1810                                               end - ptr - size);
1811                 item_size -= size;
1812                 btrfs_truncate_item(root, path, item_size, 1);
1813         }
1814         btrfs_mark_buffer_dirty(leaf);
1815 }
1816
1817 static noinline_for_stack
1818 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1819                                  struct btrfs_root *root,
1820                                  struct btrfs_path *path,
1821                                  u64 bytenr, u64 num_bytes, u64 parent,
1822                                  u64 root_objectid, u64 owner,
1823                                  u64 offset, int refs_to_add,
1824                                  struct btrfs_delayed_extent_op *extent_op)
1825 {
1826         struct btrfs_extent_inline_ref *iref;
1827         int ret;
1828
1829         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1830                                            bytenr, num_bytes, parent,
1831                                            root_objectid, owner, offset, 1);
1832         if (ret == 0) {
1833                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1834                 update_inline_extent_backref(root, path, iref,
1835                                              refs_to_add, extent_op, NULL);
1836         } else if (ret == -ENOENT) {
1837                 setup_inline_extent_backref(root, path, iref, parent,
1838                                             root_objectid, owner, offset,
1839                                             refs_to_add, extent_op);
1840                 ret = 0;
1841         }
1842         return ret;
1843 }
1844
1845 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1846                                  struct btrfs_root *root,
1847                                  struct btrfs_path *path,
1848                                  u64 bytenr, u64 parent, u64 root_objectid,
1849                                  u64 owner, u64 offset, int refs_to_add)
1850 {
1851         int ret;
1852         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1853                 BUG_ON(refs_to_add != 1);
1854                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1855                                             parent, root_objectid);
1856         } else {
1857                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1858                                              parent, root_objectid,
1859                                              owner, offset, refs_to_add);
1860         }
1861         return ret;
1862 }
1863
1864 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1865                                  struct btrfs_root *root,
1866                                  struct btrfs_path *path,
1867                                  struct btrfs_extent_inline_ref *iref,
1868                                  int refs_to_drop, int is_data, int *last_ref)
1869 {
1870         int ret = 0;
1871
1872         BUG_ON(!is_data && refs_to_drop != 1);
1873         if (iref) {
1874                 update_inline_extent_backref(root, path, iref,
1875                                              -refs_to_drop, NULL, last_ref);
1876         } else if (is_data) {
1877                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1878                                              last_ref);
1879         } else {
1880                 *last_ref = 1;
1881                 ret = btrfs_del_item(trans, root, path);
1882         }
1883         return ret;
1884 }
1885
1886 static int btrfs_issue_discard(struct block_device *bdev,
1887                                 u64 start, u64 len)
1888 {
1889         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1890 }
1891
1892 int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1893                          u64 num_bytes, u64 *actual_bytes)
1894 {
1895         int ret;
1896         u64 discarded_bytes = 0;
1897         struct btrfs_bio *bbio = NULL;
1898
1899
1900         /* Tell the block device(s) that the sectors can be discarded */
1901         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1902                               bytenr, &num_bytes, &bbio, 0);
1903         /* Error condition is -ENOMEM */
1904         if (!ret) {
1905                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1906                 int i;
1907
1908
1909                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1910                         if (!stripe->dev->can_discard)
1911                                 continue;
1912
1913                         ret = btrfs_issue_discard(stripe->dev->bdev,
1914                                                   stripe->physical,
1915                                                   stripe->length);
1916                         if (!ret)
1917                                 discarded_bytes += stripe->length;
1918                         else if (ret != -EOPNOTSUPP)
1919                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1920
1921                         /*
1922                          * Just in case we get back EOPNOTSUPP for some reason,
1923                          * just ignore the return value so we don't screw up
1924                          * people calling discard_extent.
1925                          */
1926                         ret = 0;
1927                 }
1928                 btrfs_put_bbio(bbio);
1929         }
1930
1931         if (actual_bytes)
1932                 *actual_bytes = discarded_bytes;
1933
1934
1935         if (ret == -EOPNOTSUPP)
1936                 ret = 0;
1937         return ret;
1938 }
1939
1940 /* Can return -ENOMEM */
1941 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1942                          struct btrfs_root *root,
1943                          u64 bytenr, u64 num_bytes, u64 parent,
1944                          u64 root_objectid, u64 owner, u64 offset,
1945                          int no_quota)
1946 {
1947         int ret;
1948         struct btrfs_fs_info *fs_info = root->fs_info;
1949
1950         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1951                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1952
1953         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1954                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1955                                         num_bytes,
1956                                         parent, root_objectid, (int)owner,
1957                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1958         } else {
1959                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1960                                         num_bytes,
1961                                         parent, root_objectid, owner, offset,
1962                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1963         }
1964         return ret;
1965 }
1966
1967 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1968                                   struct btrfs_root *root,
1969                                   struct btrfs_delayed_ref_node *node,
1970                                   u64 parent, u64 root_objectid,
1971                                   u64 owner, u64 offset, int refs_to_add,
1972                                   struct btrfs_delayed_extent_op *extent_op)
1973 {
1974         struct btrfs_fs_info *fs_info = root->fs_info;
1975         struct btrfs_path *path;
1976         struct extent_buffer *leaf;
1977         struct btrfs_extent_item *item;
1978         struct btrfs_key key;
1979         u64 bytenr = node->bytenr;
1980         u64 num_bytes = node->num_bytes;
1981         u64 refs;
1982         int ret;
1983         int no_quota = node->no_quota;
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)
2000                 goto out;
2001
2002         /*
2003          * Ok we had -EAGAIN which means we didn't have space to insert and
2004          * inline extent ref, so just update the reference count and add a
2005          * normal backref.
2006          */
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], struct btrfs_extent_item);
2010         refs = btrfs_extent_refs(leaf, item);
2011         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2012         if (extent_op)
2013                 __run_delayed_extent_op(extent_op, leaf, item);
2014
2015         btrfs_mark_buffer_dirty(leaf);
2016         btrfs_release_path(path);
2017
2018         path->reada = 1;
2019         path->leave_spinning = 1;
2020         /* now insert the actual backref */
2021         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2022                                     path, bytenr, parent, root_objectid,
2023                                     owner, offset, refs_to_add);
2024         if (ret)
2025                 btrfs_abort_transaction(trans, root, ret);
2026 out:
2027         btrfs_free_path(path);
2028         return ret;
2029 }
2030
2031 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2032                                 struct btrfs_root *root,
2033                                 struct btrfs_delayed_ref_node *node,
2034                                 struct btrfs_delayed_extent_op *extent_op,
2035                                 int insert_reserved)
2036 {
2037         int ret = 0;
2038         struct btrfs_delayed_data_ref *ref;
2039         struct btrfs_key ins;
2040         u64 parent = 0;
2041         u64 ref_root = 0;
2042         u64 flags = 0;
2043
2044         ins.objectid = node->bytenr;
2045         ins.offset = node->num_bytes;
2046         ins.type = BTRFS_EXTENT_ITEM_KEY;
2047
2048         ref = btrfs_delayed_node_to_data_ref(node);
2049         trace_run_delayed_data_ref(node, ref, node->action);
2050
2051         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2052                 parent = ref->parent;
2053         ref_root = ref->root;
2054
2055         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2056                 if (extent_op)
2057                         flags |= extent_op->flags_to_set;
2058                 ret = alloc_reserved_file_extent(trans, root,
2059                                                  parent, ref_root, flags,
2060                                                  ref->objectid, ref->offset,
2061                                                  &ins, node->ref_mod);
2062         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2063                 ret = __btrfs_inc_extent_ref(trans, root, node, parent,
2064                                              ref_root, ref->objectid,
2065                                              ref->offset, node->ref_mod,
2066                                              extent_op);
2067         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2068                 ret = __btrfs_free_extent(trans, root, node, parent,
2069                                           ref_root, ref->objectid,
2070                                           ref->offset, node->ref_mod,
2071                                           extent_op);
2072         } else {
2073                 BUG();
2074         }
2075         return ret;
2076 }
2077
2078 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2079                                     struct extent_buffer *leaf,
2080                                     struct btrfs_extent_item *ei)
2081 {
2082         u64 flags = btrfs_extent_flags(leaf, ei);
2083         if (extent_op->update_flags) {
2084                 flags |= extent_op->flags_to_set;
2085                 btrfs_set_extent_flags(leaf, ei, flags);
2086         }
2087
2088         if (extent_op->update_key) {
2089                 struct btrfs_tree_block_info *bi;
2090                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2091                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2092                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2093         }
2094 }
2095
2096 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2097                                  struct btrfs_root *root,
2098                                  struct btrfs_delayed_ref_node *node,
2099                                  struct btrfs_delayed_extent_op *extent_op)
2100 {
2101         struct btrfs_key key;
2102         struct btrfs_path *path;
2103         struct btrfs_extent_item *ei;
2104         struct extent_buffer *leaf;
2105         u32 item_size;
2106         int ret;
2107         int err = 0;
2108         int metadata = !extent_op->is_data;
2109
2110         if (trans->aborted)
2111                 return 0;
2112
2113         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2114                 metadata = 0;
2115
2116         path = btrfs_alloc_path();
2117         if (!path)
2118                 return -ENOMEM;
2119
2120         key.objectid = node->bytenr;
2121
2122         if (metadata) {
2123                 key.type = BTRFS_METADATA_ITEM_KEY;
2124                 key.offset = extent_op->level;
2125         } else {
2126                 key.type = BTRFS_EXTENT_ITEM_KEY;
2127                 key.offset = node->num_bytes;
2128         }
2129
2130 again:
2131         path->reada = 1;
2132         path->leave_spinning = 1;
2133         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2134                                 path, 0, 1);
2135         if (ret < 0) {
2136                 err = ret;
2137                 goto out;
2138         }
2139         if (ret > 0) {
2140                 if (metadata) {
2141                         if (path->slots[0] > 0) {
2142                                 path->slots[0]--;
2143                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2144                                                       path->slots[0]);
2145                                 if (key.objectid == node->bytenr &&
2146                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2147                                     key.offset == node->num_bytes)
2148                                         ret = 0;
2149                         }
2150                         if (ret > 0) {
2151                                 btrfs_release_path(path);
2152                                 metadata = 0;
2153
2154                                 key.objectid = node->bytenr;
2155                                 key.offset = node->num_bytes;
2156                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2157                                 goto again;
2158                         }
2159                 } else {
2160                         err = -EIO;
2161                         goto out;
2162                 }
2163         }
2164
2165         leaf = path->nodes[0];
2166         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2167 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2168         if (item_size < sizeof(*ei)) {
2169                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2170                                              path, (u64)-1, 0);
2171                 if (ret < 0) {
2172                         err = ret;
2173                         goto out;
2174                 }
2175                 leaf = path->nodes[0];
2176                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2177         }
2178 #endif
2179         BUG_ON(item_size < sizeof(*ei));
2180         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2181         __run_delayed_extent_op(extent_op, leaf, ei);
2182
2183         btrfs_mark_buffer_dirty(leaf);
2184 out:
2185         btrfs_free_path(path);
2186         return err;
2187 }
2188
2189 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2190                                 struct btrfs_root *root,
2191                                 struct btrfs_delayed_ref_node *node,
2192                                 struct btrfs_delayed_extent_op *extent_op,
2193                                 int insert_reserved)
2194 {
2195         int ret = 0;
2196         struct btrfs_delayed_tree_ref *ref;
2197         struct btrfs_key ins;
2198         u64 parent = 0;
2199         u64 ref_root = 0;
2200         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2201                                                  SKINNY_METADATA);
2202
2203         ref = btrfs_delayed_node_to_tree_ref(node);
2204         trace_run_delayed_tree_ref(node, ref, node->action);
2205
2206         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2207                 parent = ref->parent;
2208         ref_root = ref->root;
2209
2210         ins.objectid = node->bytenr;
2211         if (skinny_metadata) {
2212                 ins.offset = ref->level;
2213                 ins.type = BTRFS_METADATA_ITEM_KEY;
2214         } else {
2215                 ins.offset = node->num_bytes;
2216                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2217         }
2218
2219         BUG_ON(node->ref_mod != 1);
2220         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2221                 BUG_ON(!extent_op || !extent_op->update_flags);
2222                 ret = alloc_reserved_tree_block(trans, root,
2223                                                 parent, ref_root,
2224                                                 extent_op->flags_to_set,
2225                                                 &extent_op->key,
2226                                                 ref->level, &ins,
2227                                                 node->no_quota);
2228         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2229                 ret = __btrfs_inc_extent_ref(trans, root, node,
2230                                              parent, ref_root,
2231                                              ref->level, 0, 1,
2232                                              extent_op);
2233         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2234                 ret = __btrfs_free_extent(trans, root, node,
2235                                           parent, ref_root,
2236                                           ref->level, 0, 1, extent_op);
2237         } else {
2238                 BUG();
2239         }
2240         return ret;
2241 }
2242
2243 /* helper function to actually process a single delayed ref entry */
2244 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2245                                struct btrfs_root *root,
2246                                struct btrfs_delayed_ref_node *node,
2247                                struct btrfs_delayed_extent_op *extent_op,
2248                                int insert_reserved)
2249 {
2250         int ret = 0;
2251
2252         if (trans->aborted) {
2253                 if (insert_reserved)
2254                         btrfs_pin_extent(root, node->bytenr,
2255                                          node->num_bytes, 1);
2256                 return 0;
2257         }
2258
2259         if (btrfs_delayed_ref_is_head(node)) {
2260                 struct btrfs_delayed_ref_head *head;
2261                 /*
2262                  * we've hit the end of the chain and we were supposed
2263                  * to insert this extent into the tree.  But, it got
2264                  * deleted before we ever needed to insert it, so all
2265                  * we have to do is clean up the accounting
2266                  */
2267                 BUG_ON(extent_op);
2268                 head = btrfs_delayed_node_to_head(node);
2269                 trace_run_delayed_ref_head(node, head, node->action);
2270
2271                 if (insert_reserved) {
2272                         btrfs_pin_extent(root, node->bytenr,
2273                                          node->num_bytes, 1);
2274                         if (head->is_data) {
2275                                 ret = btrfs_del_csums(trans, root,
2276                                                       node->bytenr,
2277                                                       node->num_bytes);
2278                         }
2279                 }
2280                 return ret;
2281         }
2282
2283         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2284             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2285                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2286                                            insert_reserved);
2287         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2288                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2289                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2290                                            insert_reserved);
2291         else
2292                 BUG();
2293         return ret;
2294 }
2295
2296 static inline struct btrfs_delayed_ref_node *
2297 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2298 {
2299         struct btrfs_delayed_ref_node *ref;
2300
2301         if (list_empty(&head->ref_list))
2302                 return NULL;
2303
2304         /*
2305          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2306          * This is to prevent a ref count from going down to zero, which deletes
2307          * the extent item from the extent tree, when there still are references
2308          * to add, which would fail because they would not find the extent item.
2309          */
2310         list_for_each_entry(ref, &head->ref_list, list) {
2311                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2312                         return ref;
2313         }
2314
2315         return list_entry(head->ref_list.next, struct btrfs_delayed_ref_node,
2316                           list);
2317 }
2318
2319 /*
2320  * Returns 0 on success or if called with an already aborted transaction.
2321  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2322  */
2323 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2324                                              struct btrfs_root *root,
2325                                              unsigned long nr)
2326 {
2327         struct btrfs_delayed_ref_root *delayed_refs;
2328         struct btrfs_delayed_ref_node *ref;
2329         struct btrfs_delayed_ref_head *locked_ref = NULL;
2330         struct btrfs_delayed_extent_op *extent_op;
2331         struct btrfs_fs_info *fs_info = root->fs_info;
2332         ktime_t start = ktime_get();
2333         int ret;
2334         unsigned long count = 0;
2335         unsigned long actual_count = 0;
2336         int must_insert_reserved = 0;
2337
2338         delayed_refs = &trans->transaction->delayed_refs;
2339         while (1) {
2340                 if (!locked_ref) {
2341                         if (count >= nr)
2342                                 break;
2343
2344                         spin_lock(&delayed_refs->lock);
2345                         locked_ref = btrfs_select_ref_head(trans);
2346                         if (!locked_ref) {
2347                                 spin_unlock(&delayed_refs->lock);
2348                                 break;
2349                         }
2350
2351                         /* grab the lock that says we are going to process
2352                          * all the refs for this head */
2353                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2354                         spin_unlock(&delayed_refs->lock);
2355                         /*
2356                          * we may have dropped the spin lock to get the head
2357                          * mutex lock, and that might have given someone else
2358                          * time to free the head.  If that's true, it has been
2359                          * removed from our list and we can move on.
2360                          */
2361                         if (ret == -EAGAIN) {
2362                                 locked_ref = NULL;
2363                                 count++;
2364                                 continue;
2365                         }
2366                 }
2367
2368                 spin_lock(&locked_ref->lock);
2369
2370                 /*
2371                  * locked_ref is the head node, so we have to go one
2372                  * node back for any delayed ref updates
2373                  */
2374                 ref = select_delayed_ref(locked_ref);
2375
2376                 if (ref && ref->seq &&
2377                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2378                         spin_unlock(&locked_ref->lock);
2379                         btrfs_delayed_ref_unlock(locked_ref);
2380                         spin_lock(&delayed_refs->lock);
2381                         locked_ref->processing = 0;
2382                         delayed_refs->num_heads_ready++;
2383                         spin_unlock(&delayed_refs->lock);
2384                         locked_ref = NULL;
2385                         cond_resched();
2386                         count++;
2387                         continue;
2388                 }
2389
2390                 /*
2391                  * record the must insert reserved flag before we
2392                  * drop the spin lock.
2393                  */
2394                 must_insert_reserved = locked_ref->must_insert_reserved;
2395                 locked_ref->must_insert_reserved = 0;
2396
2397                 extent_op = locked_ref->extent_op;
2398                 locked_ref->extent_op = NULL;
2399
2400                 if (!ref) {
2401
2402
2403                         /* All delayed refs have been processed, Go ahead
2404                          * and send the head node to run_one_delayed_ref,
2405                          * so that any accounting fixes can happen
2406                          */
2407                         ref = &locked_ref->node;
2408
2409                         if (extent_op && must_insert_reserved) {
2410                                 btrfs_free_delayed_extent_op(extent_op);
2411                                 extent_op = NULL;
2412                         }
2413
2414                         if (extent_op) {
2415                                 spin_unlock(&locked_ref->lock);
2416                                 ret = run_delayed_extent_op(trans, root,
2417                                                             ref, extent_op);
2418                                 btrfs_free_delayed_extent_op(extent_op);
2419
2420                                 if (ret) {
2421                                         /*
2422                                          * Need to reset must_insert_reserved if
2423                                          * there was an error so the abort stuff
2424                                          * can cleanup the reserved space
2425                                          * properly.
2426                                          */
2427                                         if (must_insert_reserved)
2428                                                 locked_ref->must_insert_reserved = 1;
2429                                         locked_ref->processing = 0;
2430                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2431                                         btrfs_delayed_ref_unlock(locked_ref);
2432                                         return ret;
2433                                 }
2434                                 continue;
2435                         }
2436
2437                         /*
2438                          * Need to drop our head ref lock and re-aqcuire the
2439                          * delayed ref lock and then re-check to make sure
2440                          * nobody got added.
2441                          */
2442                         spin_unlock(&locked_ref->lock);
2443                         spin_lock(&delayed_refs->lock);
2444                         spin_lock(&locked_ref->lock);
2445                         if (!list_empty(&locked_ref->ref_list) ||
2446                             locked_ref->extent_op) {
2447                                 spin_unlock(&locked_ref->lock);
2448                                 spin_unlock(&delayed_refs->lock);
2449                                 continue;
2450                         }
2451                         ref->in_tree = 0;
2452                         delayed_refs->num_heads--;
2453                         rb_erase(&locked_ref->href_node,
2454                                  &delayed_refs->href_root);
2455                         spin_unlock(&delayed_refs->lock);
2456                 } else {
2457                         actual_count++;
2458                         ref->in_tree = 0;
2459                         list_del(&ref->list);
2460                 }
2461                 atomic_dec(&delayed_refs->num_entries);
2462
2463                 if (!btrfs_delayed_ref_is_head(ref)) {
2464                         /*
2465                          * when we play the delayed ref, also correct the
2466                          * ref_mod on head
2467                          */
2468                         switch (ref->action) {
2469                         case BTRFS_ADD_DELAYED_REF:
2470                         case BTRFS_ADD_DELAYED_EXTENT:
2471                                 locked_ref->node.ref_mod -= ref->ref_mod;
2472                                 break;
2473                         case BTRFS_DROP_DELAYED_REF:
2474                                 locked_ref->node.ref_mod += ref->ref_mod;
2475                                 break;
2476                         default:
2477                                 WARN_ON(1);
2478                         }
2479                 }
2480                 spin_unlock(&locked_ref->lock);
2481
2482                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2483                                           must_insert_reserved);
2484
2485                 btrfs_free_delayed_extent_op(extent_op);
2486                 if (ret) {
2487                         locked_ref->processing = 0;
2488                         btrfs_delayed_ref_unlock(locked_ref);
2489                         btrfs_put_delayed_ref(ref);
2490                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2491                         return ret;
2492                 }
2493
2494                 /*
2495                  * If this node is a head, that means all the refs in this head
2496                  * have been dealt with, and we will pick the next head to deal
2497                  * with, so we must unlock the head and drop it from the cluster
2498                  * list before we release it.
2499                  */
2500                 if (btrfs_delayed_ref_is_head(ref)) {
2501                         if (locked_ref->is_data &&
2502                             locked_ref->total_ref_mod < 0) {
2503                                 spin_lock(&delayed_refs->lock);
2504                                 delayed_refs->pending_csums -= ref->num_bytes;
2505                                 spin_unlock(&delayed_refs->lock);
2506                         }
2507                         btrfs_delayed_ref_unlock(locked_ref);
2508                         locked_ref = NULL;
2509                 }
2510                 btrfs_put_delayed_ref(ref);
2511                 count++;
2512                 cond_resched();
2513         }
2514
2515         /*
2516          * We don't want to include ref heads since we can have empty ref heads
2517          * and those will drastically skew our runtime down since we just do
2518          * accounting, no actual extent tree updates.
2519          */
2520         if (actual_count > 0) {
2521                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2522                 u64 avg;
2523
2524                 /*
2525                  * We weigh the current average higher than our current runtime
2526                  * to avoid large swings in the average.
2527                  */
2528                 spin_lock(&delayed_refs->lock);
2529                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2530                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2531                 spin_unlock(&delayed_refs->lock);
2532         }
2533         return 0;
2534 }
2535
2536 #ifdef SCRAMBLE_DELAYED_REFS
2537 /*
2538  * Normally delayed refs get processed in ascending bytenr order. This
2539  * correlates in most cases to the order added. To expose dependencies on this
2540  * order, we start to process the tree in the middle instead of the beginning
2541  */
2542 static u64 find_middle(struct rb_root *root)
2543 {
2544         struct rb_node *n = root->rb_node;
2545         struct btrfs_delayed_ref_node *entry;
2546         int alt = 1;
2547         u64 middle;
2548         u64 first = 0, last = 0;
2549
2550         n = rb_first(root);
2551         if (n) {
2552                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2553                 first = entry->bytenr;
2554         }
2555         n = rb_last(root);
2556         if (n) {
2557                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2558                 last = entry->bytenr;
2559         }
2560         n = root->rb_node;
2561
2562         while (n) {
2563                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2564                 WARN_ON(!entry->in_tree);
2565
2566                 middle = entry->bytenr;
2567
2568                 if (alt)
2569                         n = n->rb_left;
2570                 else
2571                         n = n->rb_right;
2572
2573                 alt = 1 - alt;
2574         }
2575         return middle;
2576 }
2577 #endif
2578
2579 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2580 {
2581         u64 num_bytes;
2582
2583         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2584                              sizeof(struct btrfs_extent_inline_ref));
2585         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2586                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2587
2588         /*
2589          * We don't ever fill up leaves all the way so multiply by 2 just to be
2590          * closer to what we're really going to want to ouse.
2591          */
2592         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2593 }
2594
2595 /*
2596  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2597  * would require to store the csums for that many bytes.
2598  */
2599 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2600 {
2601         u64 csum_size;
2602         u64 num_csums_per_leaf;
2603         u64 num_csums;
2604
2605         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
2606         num_csums_per_leaf = div64_u64(csum_size,
2607                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2608         num_csums = div64_u64(csum_bytes, root->sectorsize);
2609         num_csums += num_csums_per_leaf - 1;
2610         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2611         return num_csums;
2612 }
2613
2614 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2615                                        struct btrfs_root *root)
2616 {
2617         struct btrfs_block_rsv *global_rsv;
2618         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2619         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2620         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2621         u64 num_bytes, num_dirty_bgs_bytes;
2622         int ret = 0;
2623
2624         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2625         num_heads = heads_to_leaves(root, num_heads);
2626         if (num_heads > 1)
2627                 num_bytes += (num_heads - 1) * root->nodesize;
2628         num_bytes <<= 1;
2629         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2630         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2631                                                              num_dirty_bgs);
2632         global_rsv = &root->fs_info->global_block_rsv;
2633
2634         /*
2635          * If we can't allocate any more chunks lets make sure we have _lots_ of
2636          * wiggle room since running delayed refs can create more delayed refs.
2637          */
2638         if (global_rsv->space_info->full) {
2639                 num_dirty_bgs_bytes <<= 1;
2640                 num_bytes <<= 1;
2641         }
2642
2643         spin_lock(&global_rsv->lock);
2644         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2645                 ret = 1;
2646         spin_unlock(&global_rsv->lock);
2647         return ret;
2648 }
2649
2650 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2651                                        struct btrfs_root *root)
2652 {
2653         struct btrfs_fs_info *fs_info = root->fs_info;
2654         u64 num_entries =
2655                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2656         u64 avg_runtime;
2657         u64 val;
2658
2659         smp_mb();
2660         avg_runtime = fs_info->avg_delayed_ref_runtime;
2661         val = num_entries * avg_runtime;
2662         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2663                 return 1;
2664         if (val >= NSEC_PER_SEC / 2)
2665                 return 2;
2666
2667         return btrfs_check_space_for_delayed_refs(trans, root);
2668 }
2669
2670 struct async_delayed_refs {
2671         struct btrfs_root *root;
2672         int count;
2673         int error;
2674         int sync;
2675         struct completion wait;
2676         struct btrfs_work work;
2677 };
2678
2679 static void delayed_ref_async_start(struct btrfs_work *work)
2680 {
2681         struct async_delayed_refs *async;
2682         struct btrfs_trans_handle *trans;
2683         int ret;
2684
2685         async = container_of(work, struct async_delayed_refs, work);
2686
2687         trans = btrfs_join_transaction(async->root);
2688         if (IS_ERR(trans)) {
2689                 async->error = PTR_ERR(trans);
2690                 goto done;
2691         }
2692
2693         /*
2694          * trans->sync means that when we call end_transaciton, we won't
2695          * wait on delayed refs
2696          */
2697         trans->sync = true;
2698         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2699         if (ret)
2700                 async->error = ret;
2701
2702         ret = btrfs_end_transaction(trans, async->root);
2703         if (ret && !async->error)
2704                 async->error = ret;
2705 done:
2706         if (async->sync)
2707                 complete(&async->wait);
2708         else
2709                 kfree(async);
2710 }
2711
2712 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2713                                  unsigned long count, int wait)
2714 {
2715         struct async_delayed_refs *async;
2716         int ret;
2717
2718         async = kmalloc(sizeof(*async), GFP_NOFS);
2719         if (!async)
2720                 return -ENOMEM;
2721
2722         async->root = root->fs_info->tree_root;
2723         async->count = count;
2724         async->error = 0;
2725         if (wait)
2726                 async->sync = 1;
2727         else
2728                 async->sync = 0;
2729         init_completion(&async->wait);
2730
2731         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2732                         delayed_ref_async_start, NULL, NULL);
2733
2734         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2735
2736         if (wait) {
2737                 wait_for_completion(&async->wait);
2738                 ret = async->error;
2739                 kfree(async);
2740                 return ret;
2741         }
2742         return 0;
2743 }
2744
2745 /*
2746  * this starts processing the delayed reference count updates and
2747  * extent insertions we have queued up so far.  count can be
2748  * 0, which means to process everything in the tree at the start
2749  * of the run (but not newly added entries), or it can be some target
2750  * number you'd like to process.
2751  *
2752  * Returns 0 on success or if called with an aborted transaction
2753  * Returns <0 on error and aborts the transaction
2754  */
2755 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2756                            struct btrfs_root *root, unsigned long count)
2757 {
2758         struct rb_node *node;
2759         struct btrfs_delayed_ref_root *delayed_refs;
2760         struct btrfs_delayed_ref_head *head;
2761         int ret;
2762         int run_all = count == (unsigned long)-1;
2763
2764         /* We'll clean this up in btrfs_cleanup_transaction */
2765         if (trans->aborted)
2766                 return 0;
2767
2768         if (root == root->fs_info->extent_root)
2769                 root = root->fs_info->tree_root;
2770
2771         delayed_refs = &trans->transaction->delayed_refs;
2772         if (count == 0)
2773                 count = atomic_read(&delayed_refs->num_entries) * 2;
2774
2775 again:
2776 #ifdef SCRAMBLE_DELAYED_REFS
2777         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2778 #endif
2779         ret = __btrfs_run_delayed_refs(trans, root, count);
2780         if (ret < 0) {
2781                 btrfs_abort_transaction(trans, root, ret);
2782                 return ret;
2783         }
2784
2785         if (run_all) {
2786                 if (!list_empty(&trans->new_bgs))
2787                         btrfs_create_pending_block_groups(trans, root);
2788
2789                 spin_lock(&delayed_refs->lock);
2790                 node = rb_first(&delayed_refs->href_root);
2791                 if (!node) {
2792                         spin_unlock(&delayed_refs->lock);
2793                         goto out;
2794                 }
2795                 count = (unsigned long)-1;
2796
2797                 while (node) {
2798                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2799                                         href_node);
2800                         if (btrfs_delayed_ref_is_head(&head->node)) {
2801                                 struct btrfs_delayed_ref_node *ref;
2802
2803                                 ref = &head->node;
2804                                 atomic_inc(&ref->refs);
2805
2806                                 spin_unlock(&delayed_refs->lock);
2807                                 /*
2808                                  * Mutex was contended, block until it's
2809                                  * released and try again
2810                                  */
2811                                 mutex_lock(&head->mutex);
2812                                 mutex_unlock(&head->mutex);
2813
2814                                 btrfs_put_delayed_ref(ref);
2815                                 cond_resched();
2816                                 goto again;
2817                         } else {
2818                                 WARN_ON(1);
2819                         }
2820                         node = rb_next(node);
2821                 }
2822                 spin_unlock(&delayed_refs->lock);
2823                 cond_resched();
2824                 goto again;
2825         }
2826 out:
2827         assert_qgroups_uptodate(trans);
2828         return 0;
2829 }
2830
2831 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2832                                 struct btrfs_root *root,
2833                                 u64 bytenr, u64 num_bytes, u64 flags,
2834                                 int level, int is_data)
2835 {
2836         struct btrfs_delayed_extent_op *extent_op;
2837         int ret;
2838
2839         extent_op = btrfs_alloc_delayed_extent_op();
2840         if (!extent_op)
2841                 return -ENOMEM;
2842
2843         extent_op->flags_to_set = flags;
2844         extent_op->update_flags = 1;
2845         extent_op->update_key = 0;
2846         extent_op->is_data = is_data ? 1 : 0;
2847         extent_op->level = level;
2848
2849         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2850                                           num_bytes, extent_op);
2851         if (ret)
2852                 btrfs_free_delayed_extent_op(extent_op);
2853         return ret;
2854 }
2855
2856 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2857                                       struct btrfs_root *root,
2858                                       struct btrfs_path *path,
2859                                       u64 objectid, u64 offset, u64 bytenr)
2860 {
2861         struct btrfs_delayed_ref_head *head;
2862         struct btrfs_delayed_ref_node *ref;
2863         struct btrfs_delayed_data_ref *data_ref;
2864         struct btrfs_delayed_ref_root *delayed_refs;
2865         int ret = 0;
2866
2867         delayed_refs = &trans->transaction->delayed_refs;
2868         spin_lock(&delayed_refs->lock);
2869         head = btrfs_find_delayed_ref_head(trans, bytenr);
2870         if (!head) {
2871                 spin_unlock(&delayed_refs->lock);
2872                 return 0;
2873         }
2874
2875         if (!mutex_trylock(&head->mutex)) {
2876                 atomic_inc(&head->node.refs);
2877                 spin_unlock(&delayed_refs->lock);
2878
2879                 btrfs_release_path(path);
2880
2881                 /*
2882                  * Mutex was contended, block until it's released and let
2883                  * caller try again
2884                  */
2885                 mutex_lock(&head->mutex);
2886                 mutex_unlock(&head->mutex);
2887                 btrfs_put_delayed_ref(&head->node);
2888                 return -EAGAIN;
2889         }
2890         spin_unlock(&delayed_refs->lock);
2891
2892         spin_lock(&head->lock);
2893         list_for_each_entry(ref, &head->ref_list, list) {
2894                 /* If it's a shared ref we know a cross reference exists */
2895                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2896                         ret = 1;
2897                         break;
2898                 }
2899
2900                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2901
2902                 /*
2903                  * If our ref doesn't match the one we're currently looking at
2904                  * then we have a cross reference.
2905                  */
2906                 if (data_ref->root != root->root_key.objectid ||
2907                     data_ref->objectid != objectid ||
2908                     data_ref->offset != offset) {
2909                         ret = 1;
2910                         break;
2911                 }
2912         }
2913         spin_unlock(&head->lock);
2914         mutex_unlock(&head->mutex);
2915         return ret;
2916 }
2917
2918 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2919                                         struct btrfs_root *root,
2920                                         struct btrfs_path *path,
2921                                         u64 objectid, u64 offset, u64 bytenr)
2922 {
2923         struct btrfs_root *extent_root = root->fs_info->extent_root;
2924         struct extent_buffer *leaf;
2925         struct btrfs_extent_data_ref *ref;
2926         struct btrfs_extent_inline_ref *iref;
2927         struct btrfs_extent_item *ei;
2928         struct btrfs_key key;
2929         u32 item_size;
2930         int ret;
2931
2932         key.objectid = bytenr;
2933         key.offset = (u64)-1;
2934         key.type = BTRFS_EXTENT_ITEM_KEY;
2935
2936         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2937         if (ret < 0)
2938                 goto out;
2939         BUG_ON(ret == 0); /* Corruption */
2940
2941         ret = -ENOENT;
2942         if (path->slots[0] == 0)
2943                 goto out;
2944
2945         path->slots[0]--;
2946         leaf = path->nodes[0];
2947         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2948
2949         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2950                 goto out;
2951
2952         ret = 1;
2953         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2954 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2955         if (item_size < sizeof(*ei)) {
2956                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2957                 goto out;
2958         }
2959 #endif
2960         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2961
2962         if (item_size != sizeof(*ei) +
2963             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2964                 goto out;
2965
2966         if (btrfs_extent_generation(leaf, ei) <=
2967             btrfs_root_last_snapshot(&root->root_item))
2968                 goto out;
2969
2970         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2971         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2972             BTRFS_EXTENT_DATA_REF_KEY)
2973                 goto out;
2974
2975         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2976         if (btrfs_extent_refs(leaf, ei) !=
2977             btrfs_extent_data_ref_count(leaf, ref) ||
2978             btrfs_extent_data_ref_root(leaf, ref) !=
2979             root->root_key.objectid ||
2980             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2981             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2982                 goto out;
2983
2984         ret = 0;
2985 out:
2986         return ret;
2987 }
2988
2989 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2990                           struct btrfs_root *root,
2991                           u64 objectid, u64 offset, u64 bytenr)
2992 {
2993         struct btrfs_path *path;
2994         int ret;
2995         int ret2;
2996
2997         path = btrfs_alloc_path();
2998         if (!path)
2999                 return -ENOENT;
3000
3001         do {
3002                 ret = check_committed_ref(trans, root, path, objectid,
3003                                           offset, bytenr);
3004                 if (ret && ret != -ENOENT)
3005                         goto out;
3006
3007                 ret2 = check_delayed_ref(trans, root, path, objectid,
3008                                          offset, bytenr);
3009         } while (ret2 == -EAGAIN);
3010
3011         if (ret2 && ret2 != -ENOENT) {
3012                 ret = ret2;
3013                 goto out;
3014         }
3015
3016         if (ret != -ENOENT || ret2 != -ENOENT)
3017                 ret = 0;
3018 out:
3019         btrfs_free_path(path);
3020         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3021                 WARN_ON(ret > 0);
3022         return ret;
3023 }
3024
3025 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3026                            struct btrfs_root *root,
3027                            struct extent_buffer *buf,
3028                            int full_backref, int inc)
3029 {
3030         u64 bytenr;
3031         u64 num_bytes;
3032         u64 parent;
3033         u64 ref_root;
3034         u32 nritems;
3035         struct btrfs_key key;
3036         struct btrfs_file_extent_item *fi;
3037         int i;
3038         int level;
3039         int ret = 0;
3040         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3041                             u64, u64, u64, u64, u64, u64, int);
3042
3043
3044         if (btrfs_test_is_dummy_root(root))
3045                 return 0;
3046
3047         ref_root = btrfs_header_owner(buf);
3048         nritems = btrfs_header_nritems(buf);
3049         level = btrfs_header_level(buf);
3050
3051         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3052                 return 0;
3053
3054         if (inc)
3055                 process_func = btrfs_inc_extent_ref;
3056         else
3057                 process_func = btrfs_free_extent;
3058
3059         if (full_backref)
3060                 parent = buf->start;
3061         else
3062                 parent = 0;
3063
3064         for (i = 0; i < nritems; i++) {
3065                 if (level == 0) {
3066                         btrfs_item_key_to_cpu(buf, &key, i);
3067                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3068                                 continue;
3069                         fi = btrfs_item_ptr(buf, i,
3070                                             struct btrfs_file_extent_item);
3071                         if (btrfs_file_extent_type(buf, fi) ==
3072                             BTRFS_FILE_EXTENT_INLINE)
3073                                 continue;
3074                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3075                         if (bytenr == 0)
3076                                 continue;
3077
3078                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3079                         key.offset -= btrfs_file_extent_offset(buf, fi);
3080                         ret = process_func(trans, root, bytenr, num_bytes,
3081                                            parent, ref_root, key.objectid,
3082                                            key.offset, 1);
3083                         if (ret)
3084                                 goto fail;
3085                 } else {
3086                         bytenr = btrfs_node_blockptr(buf, i);
3087                         num_bytes = root->nodesize;
3088                         ret = process_func(trans, root, bytenr, num_bytes,
3089                                            parent, ref_root, level - 1, 0,
3090                                            1);
3091                         if (ret)
3092                                 goto fail;
3093                 }
3094         }
3095         return 0;
3096 fail:
3097         return ret;
3098 }
3099
3100 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3101                   struct extent_buffer *buf, int full_backref)
3102 {
3103         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3104 }
3105
3106 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3107                   struct extent_buffer *buf, int full_backref)
3108 {
3109         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3110 }
3111
3112 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3113                                  struct btrfs_root *root,
3114                                  struct btrfs_path *path,
3115                                  struct btrfs_block_group_cache *cache)
3116 {
3117         int ret;
3118         struct btrfs_root *extent_root = root->fs_info->extent_root;
3119         unsigned long bi;
3120         struct extent_buffer *leaf;
3121
3122         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3123         if (ret) {
3124                 if (ret > 0)
3125                         ret = -ENOENT;
3126                 goto fail;
3127         }
3128
3129         leaf = path->nodes[0];
3130         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3131         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3132         btrfs_mark_buffer_dirty(leaf);
3133 fail:
3134         btrfs_release_path(path);
3135         return ret;
3136
3137 }
3138
3139 static struct btrfs_block_group_cache *
3140 next_block_group(struct btrfs_root *root,
3141                  struct btrfs_block_group_cache *cache)
3142 {
3143         struct rb_node *node;
3144
3145         spin_lock(&root->fs_info->block_group_cache_lock);
3146
3147         /* If our block group was removed, we need a full search. */
3148         if (RB_EMPTY_NODE(&cache->cache_node)) {
3149                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3150
3151                 spin_unlock(&root->fs_info->block_group_cache_lock);
3152                 btrfs_put_block_group(cache);
3153                 cache = btrfs_lookup_first_block_group(root->fs_info,
3154                                                        next_bytenr);
3155                 return cache;
3156         }
3157         node = rb_next(&cache->cache_node);
3158         btrfs_put_block_group(cache);
3159         if (node) {
3160                 cache = rb_entry(node, struct btrfs_block_group_cache,
3161                                  cache_node);
3162                 btrfs_get_block_group(cache);
3163         } else
3164                 cache = NULL;
3165         spin_unlock(&root->fs_info->block_group_cache_lock);
3166         return cache;
3167 }
3168
3169 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3170                             struct btrfs_trans_handle *trans,
3171                             struct btrfs_path *path)
3172 {
3173         struct btrfs_root *root = block_group->fs_info->tree_root;
3174         struct inode *inode = NULL;
3175         u64 alloc_hint = 0;
3176         int dcs = BTRFS_DC_ERROR;
3177         u64 num_pages = 0;
3178         int retries = 0;
3179         int ret = 0;
3180
3181         /*
3182          * If this block group is smaller than 100 megs don't bother caching the
3183          * block group.
3184          */
3185         if (block_group->key.offset < (100 * 1024 * 1024)) {
3186                 spin_lock(&block_group->lock);
3187                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3188                 spin_unlock(&block_group->lock);
3189                 return 0;
3190         }
3191
3192         if (trans->aborted)
3193                 return 0;
3194 again:
3195         inode = lookup_free_space_inode(root, block_group, path);
3196         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3197                 ret = PTR_ERR(inode);
3198                 btrfs_release_path(path);
3199                 goto out;
3200         }
3201
3202         if (IS_ERR(inode)) {
3203                 BUG_ON(retries);
3204                 retries++;
3205
3206                 if (block_group->ro)
3207                         goto out_free;
3208
3209                 ret = create_free_space_inode(root, trans, block_group, path);
3210                 if (ret)
3211                         goto out_free;
3212                 goto again;
3213         }
3214
3215         /* We've already setup this transaction, go ahead and exit */
3216         if (block_group->cache_generation == trans->transid &&
3217             i_size_read(inode)) {
3218                 dcs = BTRFS_DC_SETUP;
3219                 goto out_put;
3220         }
3221
3222         /*
3223          * We want to set the generation to 0, that way if anything goes wrong
3224          * from here on out we know not to trust this cache when we load up next
3225          * time.
3226          */
3227         BTRFS_I(inode)->generation = 0;
3228         ret = btrfs_update_inode(trans, root, inode);
3229         if (ret) {
3230                 /*
3231                  * So theoretically we could recover from this, simply set the
3232                  * super cache generation to 0 so we know to invalidate the
3233                  * cache, but then we'd have to keep track of the block groups
3234                  * that fail this way so we know we _have_ to reset this cache
3235                  * before the next commit or risk reading stale cache.  So to
3236                  * limit our exposure to horrible edge cases lets just abort the
3237                  * transaction, this only happens in really bad situations
3238                  * anyway.
3239                  */
3240                 btrfs_abort_transaction(trans, root, ret);
3241                 goto out_put;
3242         }
3243         WARN_ON(ret);
3244
3245         if (i_size_read(inode) > 0) {
3246                 ret = btrfs_check_trunc_cache_free_space(root,
3247                                         &root->fs_info->global_block_rsv);
3248                 if (ret)
3249                         goto out_put;
3250
3251                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3252                 if (ret)
3253                         goto out_put;
3254         }
3255
3256         spin_lock(&block_group->lock);
3257         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3258             !btrfs_test_opt(root, SPACE_CACHE)) {
3259                 /*
3260                  * don't bother trying to write stuff out _if_
3261                  * a) we're not cached,
3262                  * b) we're with nospace_cache mount option.
3263                  */
3264                 dcs = BTRFS_DC_WRITTEN;
3265                 spin_unlock(&block_group->lock);
3266                 goto out_put;
3267         }
3268         spin_unlock(&block_group->lock);
3269
3270         /*
3271          * Try to preallocate enough space based on how big the block group is.
3272          * Keep in mind this has to include any pinned space which could end up
3273          * taking up quite a bit since it's not folded into the other space
3274          * cache.
3275          */
3276         num_pages = div_u64(block_group->key.offset, 256 * 1024 * 1024);
3277         if (!num_pages)
3278                 num_pages = 1;
3279
3280         num_pages *= 16;
3281         num_pages *= PAGE_CACHE_SIZE;
3282
3283         ret = btrfs_check_data_free_space(inode, num_pages, num_pages);
3284         if (ret)
3285                 goto out_put;
3286
3287         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3288                                               num_pages, num_pages,
3289                                               &alloc_hint);
3290         if (!ret)
3291                 dcs = BTRFS_DC_SETUP;
3292         btrfs_free_reserved_data_space(inode, num_pages);
3293
3294 out_put:
3295         iput(inode);
3296 out_free:
3297         btrfs_release_path(path);
3298 out:
3299         spin_lock(&block_group->lock);
3300         if (!ret && dcs == BTRFS_DC_SETUP)
3301                 block_group->cache_generation = trans->transid;
3302         block_group->disk_cache_state = dcs;
3303         spin_unlock(&block_group->lock);
3304
3305         return ret;
3306 }
3307
3308 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3309                             struct btrfs_root *root)
3310 {
3311         struct btrfs_block_group_cache *cache, *tmp;
3312         struct btrfs_transaction *cur_trans = trans->transaction;
3313         struct btrfs_path *path;
3314
3315         if (list_empty(&cur_trans->dirty_bgs) ||
3316             !btrfs_test_opt(root, SPACE_CACHE))
3317                 return 0;
3318
3319         path = btrfs_alloc_path();
3320         if (!path)
3321                 return -ENOMEM;
3322
3323         /* Could add new block groups, use _safe just in case */
3324         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3325                                  dirty_list) {
3326                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3327                         cache_save_setup(cache, trans, path);
3328         }
3329
3330         btrfs_free_path(path);
3331         return 0;
3332 }
3333
3334 /*
3335  * transaction commit does final block group cache writeback during a
3336  * critical section where nothing is allowed to change the FS.  This is
3337  * required in order for the cache to actually match the block group,
3338  * but can introduce a lot of latency into the commit.
3339  *
3340  * So, btrfs_start_dirty_block_groups is here to kick off block group
3341  * cache IO.  There's a chance we'll have to redo some of it if the
3342  * block group changes again during the commit, but it greatly reduces
3343  * the commit latency by getting rid of the easy block groups while
3344  * we're still allowing others to join the commit.
3345  */
3346 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3347                                    struct btrfs_root *root)
3348 {
3349         struct btrfs_block_group_cache *cache;
3350         struct btrfs_transaction *cur_trans = trans->transaction;
3351         int ret = 0;
3352         int should_put;
3353         struct btrfs_path *path = NULL;
3354         LIST_HEAD(dirty);
3355         struct list_head *io = &cur_trans->io_bgs;
3356         int num_started = 0;
3357         int loops = 0;
3358
3359         spin_lock(&cur_trans->dirty_bgs_lock);
3360         if (list_empty(&cur_trans->dirty_bgs)) {
3361                 spin_unlock(&cur_trans->dirty_bgs_lock);
3362                 return 0;
3363         }
3364         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3365         spin_unlock(&cur_trans->dirty_bgs_lock);
3366
3367 again:
3368         /*
3369          * make sure all the block groups on our dirty list actually
3370          * exist
3371          */
3372         btrfs_create_pending_block_groups(trans, root);
3373
3374         if (!path) {
3375                 path = btrfs_alloc_path();
3376                 if (!path)
3377                         return -ENOMEM;
3378         }
3379
3380         /*
3381          * cache_write_mutex is here only to save us from balance or automatic
3382          * removal of empty block groups deleting this block group while we are
3383          * writing out the cache
3384          */
3385         mutex_lock(&trans->transaction->cache_write_mutex);
3386         while (!list_empty(&dirty)) {
3387                 cache = list_first_entry(&dirty,
3388                                          struct btrfs_block_group_cache,
3389                                          dirty_list);
3390                 /*
3391                  * this can happen if something re-dirties a block
3392                  * group that is already under IO.  Just wait for it to
3393                  * finish and then do it all again
3394                  */
3395                 if (!list_empty(&cache->io_list)) {
3396                         list_del_init(&cache->io_list);
3397                         btrfs_wait_cache_io(root, trans, cache,
3398                                             &cache->io_ctl, path,
3399                                             cache->key.objectid);
3400                         btrfs_put_block_group(cache);
3401                 }
3402
3403
3404                 /*
3405                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3406                  * if it should update the cache_state.  Don't delete
3407                  * until after we wait.
3408                  *
3409                  * Since we're not running in the commit critical section
3410                  * we need the dirty_bgs_lock to protect from update_block_group
3411                  */
3412                 spin_lock(&cur_trans->dirty_bgs_lock);
3413                 list_del_init(&cache->dirty_list);
3414                 spin_unlock(&cur_trans->dirty_bgs_lock);
3415
3416                 should_put = 1;
3417
3418                 cache_save_setup(cache, trans, path);
3419
3420                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3421                         cache->io_ctl.inode = NULL;
3422                         ret = btrfs_write_out_cache(root, trans, cache, path);
3423                         if (ret == 0 && cache->io_ctl.inode) {
3424                                 num_started++;
3425                                 should_put = 0;
3426
3427                                 /*
3428                                  * the cache_write_mutex is protecting
3429                                  * the io_list
3430                                  */
3431                                 list_add_tail(&cache->io_list, io);
3432                         } else {
3433                                 /*
3434                                  * if we failed to write the cache, the
3435                                  * generation will be bad and life goes on
3436                                  */
3437                                 ret = 0;
3438                         }
3439                 }
3440                 if (!ret) {
3441                         ret = write_one_cache_group(trans, root, path, cache);
3442                         /*
3443                          * Our block group might still be attached to the list
3444                          * of new block groups in the transaction handle of some
3445                          * other task (struct btrfs_trans_handle->new_bgs). This
3446                          * means its block group item isn't yet in the extent
3447                          * tree. If this happens ignore the error, as we will
3448                          * try again later in the critical section of the
3449                          * transaction commit.
3450                          */
3451                         if (ret == -ENOENT) {
3452                                 ret = 0;
3453                                 spin_lock(&cur_trans->dirty_bgs_lock);
3454                                 if (list_empty(&cache->dirty_list)) {
3455                                         list_add_tail(&cache->dirty_list,
3456                                                       &cur_trans->dirty_bgs);
3457                                         btrfs_get_block_group(cache);
3458                                 }
3459                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3460                         } else if (ret) {
3461                                 btrfs_abort_transaction(trans, root, ret);
3462                         }
3463                 }
3464
3465                 /* if its not on the io list, we need to put the block group */
3466                 if (should_put)
3467                         btrfs_put_block_group(cache);
3468
3469                 if (ret)
3470                         break;
3471
3472                 /*
3473                  * Avoid blocking other tasks for too long. It might even save
3474                  * us from writing caches for block groups that are going to be
3475                  * removed.
3476                  */
3477                 mutex_unlock(&trans->transaction->cache_write_mutex);
3478                 mutex_lock(&trans->transaction->cache_write_mutex);
3479         }
3480         mutex_unlock(&trans->transaction->cache_write_mutex);
3481
3482         /*
3483          * go through delayed refs for all the stuff we've just kicked off
3484          * and then loop back (just once)
3485          */
3486         ret = btrfs_run_delayed_refs(trans, root, 0);
3487         if (!ret && loops == 0) {
3488                 loops++;
3489                 spin_lock(&cur_trans->dirty_bgs_lock);
3490                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3491                 /*
3492                  * dirty_bgs_lock protects us from concurrent block group
3493                  * deletes too (not just cache_write_mutex).
3494                  */
3495                 if (!list_empty(&dirty)) {
3496                         spin_unlock(&cur_trans->dirty_bgs_lock);
3497                         goto again;
3498                 }
3499                 spin_unlock(&cur_trans->dirty_bgs_lock);
3500         }
3501
3502         btrfs_free_path(path);
3503         return ret;
3504 }
3505
3506 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3507                                    struct btrfs_root *root)
3508 {
3509         struct btrfs_block_group_cache *cache;
3510         struct btrfs_transaction *cur_trans = trans->transaction;
3511         int ret = 0;
3512         int should_put;
3513         struct btrfs_path *path;
3514         struct list_head *io = &cur_trans->io_bgs;
3515         int num_started = 0;
3516
3517         path = btrfs_alloc_path();
3518         if (!path)
3519                 return -ENOMEM;
3520
3521         /*
3522          * We don't need the lock here since we are protected by the transaction
3523          * commit.  We want to do the cache_save_setup first and then run the
3524          * delayed refs to make sure we have the best chance at doing this all
3525          * in one shot.
3526          */
3527         while (!list_empty(&cur_trans->dirty_bgs)) {
3528                 cache = list_first_entry(&cur_trans->dirty_bgs,
3529                                          struct btrfs_block_group_cache,
3530                                          dirty_list);
3531
3532                 /*
3533                  * this can happen if cache_save_setup re-dirties a block
3534                  * group that is already under IO.  Just wait for it to
3535                  * finish and then do it all again
3536                  */
3537                 if (!list_empty(&cache->io_list)) {
3538                         list_del_init(&cache->io_list);
3539                         btrfs_wait_cache_io(root, trans, cache,
3540                                             &cache->io_ctl, path,
3541                                             cache->key.objectid);
3542                         btrfs_put_block_group(cache);
3543                 }
3544
3545                 /*
3546                  * don't remove from the dirty list until after we've waited
3547                  * on any pending IO
3548                  */
3549                 list_del_init(&cache->dirty_list);
3550                 should_put = 1;
3551
3552                 cache_save_setup(cache, trans, path);
3553
3554                 if (!ret)
3555                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3556
3557                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3558                         cache->io_ctl.inode = NULL;
3559                         ret = btrfs_write_out_cache(root, trans, cache, path);
3560                         if (ret == 0 && cache->io_ctl.inode) {
3561                                 num_started++;
3562                                 should_put = 0;
3563                                 list_add_tail(&cache->io_list, io);
3564                         } else {
3565                                 /*
3566                                  * if we failed to write the cache, the
3567                                  * generation will be bad and life goes on
3568                                  */
3569                                 ret = 0;
3570                         }
3571                 }
3572                 if (!ret) {
3573                         ret = write_one_cache_group(trans, root, path, cache);
3574                         if (ret)
3575                                 btrfs_abort_transaction(trans, root, ret);
3576                 }
3577
3578                 /* if its not on the io list, we need to put the block group */
3579                 if (should_put)
3580                         btrfs_put_block_group(cache);
3581         }
3582
3583         while (!list_empty(io)) {
3584                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3585                                          io_list);
3586                 list_del_init(&cache->io_list);
3587                 btrfs_wait_cache_io(root, trans, cache,
3588                                     &cache->io_ctl, path, cache->key.objectid);
3589                 btrfs_put_block_group(cache);
3590         }
3591
3592         btrfs_free_path(path);
3593         return ret;
3594 }
3595
3596 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3597 {
3598         struct btrfs_block_group_cache *block_group;
3599         int readonly = 0;
3600
3601         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3602         if (!block_group || block_group->ro)
3603                 readonly = 1;
3604         if (block_group)
3605                 btrfs_put_block_group(block_group);
3606         return readonly;
3607 }
3608
3609 static const char *alloc_name(u64 flags)
3610 {
3611         switch (flags) {
3612         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3613                 return "mixed";
3614         case BTRFS_BLOCK_GROUP_METADATA:
3615                 return "metadata";
3616         case BTRFS_BLOCK_GROUP_DATA:
3617                 return "data";
3618         case BTRFS_BLOCK_GROUP_SYSTEM:
3619                 return "system";
3620         default:
3621                 WARN_ON(1);
3622                 return "invalid-combination";
3623         };
3624 }
3625
3626 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3627                              u64 total_bytes, u64 bytes_used,
3628                              struct btrfs_space_info **space_info)
3629 {
3630         struct btrfs_space_info *found;
3631         int i;
3632         int factor;
3633         int ret;
3634
3635         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3636                      BTRFS_BLOCK_GROUP_RAID10))
3637                 factor = 2;
3638         else
3639                 factor = 1;
3640
3641         found = __find_space_info(info, flags);
3642         if (found) {
3643                 spin_lock(&found->lock);
3644                 found->total_bytes += total_bytes;
3645                 found->disk_total += total_bytes * factor;
3646                 found->bytes_used += bytes_used;
3647                 found->disk_used += bytes_used * factor;
3648                 if (total_bytes > 0)
3649                         found->full = 0;
3650                 spin_unlock(&found->lock);
3651                 *space_info = found;
3652                 return 0;
3653         }
3654         found = kzalloc(sizeof(*found), GFP_NOFS);
3655         if (!found)
3656                 return -ENOMEM;
3657
3658         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3659         if (ret) {
3660                 kfree(found);
3661                 return ret;
3662         }
3663
3664         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3665                 INIT_LIST_HEAD(&found->block_groups[i]);
3666         init_rwsem(&found->groups_sem);
3667         spin_lock_init(&found->lock);
3668         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3669         found->total_bytes = total_bytes;
3670         found->disk_total = total_bytes * factor;
3671         found->bytes_used = bytes_used;
3672         found->disk_used = bytes_used * factor;
3673         found->bytes_pinned = 0;
3674         found->bytes_reserved = 0;
3675         found->bytes_readonly = 0;
3676         found->bytes_may_use = 0;
3677         if (total_bytes > 0)
3678                 found->full = 0;
3679         else
3680                 found->full = 1;
3681         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3682         found->chunk_alloc = 0;
3683         found->flush = 0;
3684         init_waitqueue_head(&found->wait);
3685         INIT_LIST_HEAD(&found->ro_bgs);
3686
3687         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3688                                     info->space_info_kobj, "%s",
3689                                     alloc_name(found->flags));
3690         if (ret) {
3691                 kfree(found);
3692                 return ret;
3693         }
3694
3695         *space_info = found;
3696         list_add_rcu(&found->list, &info->space_info);
3697         if (flags & BTRFS_BLOCK_GROUP_DATA)
3698                 info->data_sinfo = found;
3699
3700         return ret;
3701 }
3702
3703 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3704 {
3705         u64 extra_flags = chunk_to_extended(flags) &
3706                                 BTRFS_EXTENDED_PROFILE_MASK;
3707
3708         write_seqlock(&fs_info->profiles_lock);
3709         if (flags & BTRFS_BLOCK_GROUP_DATA)
3710                 fs_info->avail_data_alloc_bits |= extra_flags;
3711         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3712                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3713         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3714                 fs_info->avail_system_alloc_bits |= extra_flags;
3715         write_sequnlock(&fs_info->profiles_lock);
3716 }
3717
3718 /*
3719  * returns target flags in extended format or 0 if restripe for this
3720  * chunk_type is not in progress
3721  *
3722  * should be called with either volume_mutex or balance_lock held
3723  */
3724 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3725 {
3726         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3727         u64 target = 0;
3728
3729         if (!bctl)
3730                 return 0;
3731
3732         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3733             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3734                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3735         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3736                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3737                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3738         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3739                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3740                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3741         }
3742
3743         return target;
3744 }
3745
3746 /*
3747  * @flags: available profiles in extended format (see ctree.h)
3748  *
3749  * Returns reduced profile in chunk format.  If profile changing is in
3750  * progress (either running or paused) picks the target profile (if it's
3751  * already available), otherwise falls back to plain reducing.
3752  */
3753 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3754 {
3755         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3756         u64 target;
3757         u64 tmp;
3758
3759         /*
3760          * see if restripe for this chunk_type is in progress, if so
3761          * try to reduce to the target profile
3762          */
3763         spin_lock(&root->fs_info->balance_lock);
3764         target = get_restripe_target(root->fs_info, flags);
3765         if (target) {
3766                 /* pick target profile only if it's already available */
3767                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3768                         spin_unlock(&root->fs_info->balance_lock);
3769                         return extended_to_chunk(target);
3770                 }
3771         }
3772         spin_unlock(&root->fs_info->balance_lock);
3773
3774         /* First, mask out the RAID levels which aren't possible */
3775         if (num_devices == 1)
3776                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3777                            BTRFS_BLOCK_GROUP_RAID5);
3778         if (num_devices < 3)
3779                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3780         if (num_devices < 4)
3781                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3782
3783         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3784                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3785                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3786         flags &= ~tmp;
3787
3788         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3789                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3790         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3791                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3792         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3793                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3794         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3795                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3796         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3797                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3798
3799         return extended_to_chunk(flags | tmp);
3800 }
3801
3802 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3803 {
3804         unsigned seq;
3805         u64 flags;
3806
3807         do {
3808                 flags = orig_flags;
3809                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3810
3811                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3812                         flags |= root->fs_info->avail_data_alloc_bits;
3813                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3814                         flags |= root->fs_info->avail_system_alloc_bits;
3815                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3816                         flags |= root->fs_info->avail_metadata_alloc_bits;
3817         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3818
3819         return btrfs_reduce_alloc_profile(root, flags);
3820 }
3821
3822 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3823 {
3824         u64 flags;
3825         u64 ret;
3826
3827         if (data)
3828                 flags = BTRFS_BLOCK_GROUP_DATA;
3829         else if (root == root->fs_info->chunk_root)
3830                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3831         else
3832                 flags = BTRFS_BLOCK_GROUP_METADATA;
3833
3834         ret = get_alloc_profile(root, flags);
3835         return ret;
3836 }
3837
3838 /*
3839  * This will check the space that the inode allocates from to make sure we have
3840  * enough space for bytes.
3841  */
3842 int btrfs_check_data_free_space(struct inode *inode, u64 bytes, u64 write_bytes)
3843 {
3844         struct btrfs_space_info *data_sinfo;
3845         struct btrfs_root *root = BTRFS_I(inode)->root;
3846         struct btrfs_fs_info *fs_info = root->fs_info;
3847         u64 used;
3848         int ret = 0;
3849         int need_commit = 2;
3850         int have_pinned_space;
3851
3852         /* make sure bytes are sectorsize aligned */
3853         bytes = ALIGN(bytes, root->sectorsize);
3854
3855         if (btrfs_is_free_space_inode(inode)) {
3856                 need_commit = 0;
3857                 ASSERT(current->journal_info);
3858         }
3859
3860         data_sinfo = fs_info->data_sinfo;
3861         if (!data_sinfo)
3862                 goto alloc;
3863
3864 again:
3865         /* make sure we have enough space to handle the data first */
3866         spin_lock(&data_sinfo->lock);
3867         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3868                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3869                 data_sinfo->bytes_may_use;
3870
3871         if (used + bytes > data_sinfo->total_bytes) {
3872                 struct btrfs_trans_handle *trans;
3873
3874                 /*
3875                  * if we don't have enough free bytes in this space then we need
3876                  * to alloc a new chunk.
3877                  */
3878                 if (!data_sinfo->full) {
3879                         u64 alloc_target;
3880
3881                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3882                         spin_unlock(&data_sinfo->lock);
3883 alloc:
3884                         alloc_target = btrfs_get_alloc_profile(root, 1);
3885                         /*
3886                          * It is ugly that we don't call nolock join
3887                          * transaction for the free space inode case here.
3888                          * But it is safe because we only do the data space
3889                          * reservation for the free space cache in the
3890                          * transaction context, the common join transaction
3891                          * just increase the counter of the current transaction
3892                          * handler, doesn't try to acquire the trans_lock of
3893                          * the fs.
3894                          */
3895                         trans = btrfs_join_transaction(root);
3896                         if (IS_ERR(trans))
3897                                 return PTR_ERR(trans);
3898
3899                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3900                                              alloc_target,
3901                                              CHUNK_ALLOC_NO_FORCE);
3902                         btrfs_end_transaction(trans, root);
3903                         if (ret < 0) {
3904                                 if (ret != -ENOSPC)
3905                                         return ret;
3906                                 else {
3907                                         have_pinned_space = 1;
3908                                         goto commit_trans;
3909                                 }
3910                         }
3911
3912                         if (!data_sinfo)
3913                                 data_sinfo = fs_info->data_sinfo;
3914
3915                         goto again;
3916                 }
3917
3918                 /*
3919                  * If we don't have enough pinned space to deal with this
3920                  * allocation, and no removed chunk in current transaction,
3921                  * don't bother committing the transaction.
3922                  */
3923                 have_pinned_space = percpu_counter_compare(
3924                         &data_sinfo->total_bytes_pinned,
3925                         used + bytes - data_sinfo->total_bytes);
3926                 spin_unlock(&data_sinfo->lock);
3927
3928                 /* commit the current transaction and try again */
3929 commit_trans:
3930                 if (need_commit &&
3931                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3932                         need_commit--;
3933
3934                         if (need_commit > 0)
3935                                 btrfs_wait_ordered_roots(fs_info, -1);
3936
3937                         trans = btrfs_join_transaction(root);
3938                         if (IS_ERR(trans))
3939                                 return PTR_ERR(trans);
3940                         if (have_pinned_space >= 0 ||
3941                             trans->transaction->have_free_bgs ||
3942                             need_commit > 0) {
3943                                 ret = btrfs_commit_transaction(trans, root);
3944                                 if (ret)
3945                                         return ret;
3946                                 /*
3947                                  * make sure that all running delayed iput are
3948                                  * done
3949                                  */
3950                                 down_write(&root->fs_info->delayed_iput_sem);
3951                                 up_write(&root->fs_info->delayed_iput_sem);
3952                                 goto again;
3953                         } else {
3954                                 btrfs_end_transaction(trans, root);
3955                         }
3956                 }
3957
3958                 trace_btrfs_space_reservation(root->fs_info,
3959                                               "space_info:enospc",
3960                                               data_sinfo->flags, bytes, 1);
3961                 return -ENOSPC;
3962         }
3963         ret = btrfs_qgroup_reserve(root, write_bytes);
3964         if (ret)
3965                 goto out;
3966         data_sinfo->bytes_may_use += bytes;
3967         trace_btrfs_space_reservation(root->fs_info, "space_info",
3968                                       data_sinfo->flags, bytes, 1);
3969 out:
3970         spin_unlock(&data_sinfo->lock);
3971
3972         return ret;
3973 }
3974
3975 /*
3976  * Called if we need to clear a data reservation for this inode.
3977  */
3978 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3979 {
3980         struct btrfs_root *root = BTRFS_I(inode)->root;
3981         struct btrfs_space_info *data_sinfo;
3982
3983         /* make sure bytes are sectorsize aligned */
3984         bytes = ALIGN(bytes, root->sectorsize);
3985
3986         data_sinfo = root->fs_info->data_sinfo;
3987         spin_lock(&data_sinfo->lock);
3988         WARN_ON(data_sinfo->bytes_may_use < bytes);
3989         data_sinfo->bytes_may_use -= bytes;
3990         trace_btrfs_space_reservation(root->fs_info, "space_info",
3991                                       data_sinfo->flags, bytes, 0);
3992         spin_unlock(&data_sinfo->lock);
3993 }
3994
3995 static void force_metadata_allocation(struct btrfs_fs_info *info)
3996 {
3997         struct list_head *head = &info->space_info;
3998         struct btrfs_space_info *found;
3999
4000         rcu_read_lock();
4001         list_for_each_entry_rcu(found, head, list) {
4002                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4003                         found->force_alloc = CHUNK_ALLOC_FORCE;
4004         }
4005         rcu_read_unlock();
4006 }
4007
4008 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4009 {
4010         return (global->size << 1);
4011 }
4012
4013 static int should_alloc_chunk(struct btrfs_root *root,
4014                               struct btrfs_space_info *sinfo, int force)
4015 {
4016         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4017         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4018         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4019         u64 thresh;
4020
4021         if (force == CHUNK_ALLOC_FORCE)
4022                 return 1;
4023
4024         /*
4025          * We need to take into account the global rsv because for all intents
4026          * and purposes it's used space.  Don't worry about locking the
4027          * global_rsv, it doesn't change except when the transaction commits.
4028          */
4029         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4030                 num_allocated += calc_global_rsv_need_space(global_rsv);
4031
4032         /*
4033          * in limited mode, we want to have some free space up to
4034          * about 1% of the FS size.
4035          */
4036         if (force == CHUNK_ALLOC_LIMITED) {
4037                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4038                 thresh = max_t(u64, 64 * 1024 * 1024,
4039                                div_factor_fine(thresh, 1));
4040
4041                 if (num_bytes - num_allocated < thresh)
4042                         return 1;
4043         }
4044
4045         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
4046                 return 0;
4047         return 1;
4048 }
4049
4050 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4051 {
4052         u64 num_dev;
4053
4054         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4055                     BTRFS_BLOCK_GROUP_RAID0 |
4056                     BTRFS_BLOCK_GROUP_RAID5 |
4057                     BTRFS_BLOCK_GROUP_RAID6))
4058                 num_dev = root->fs_info->fs_devices->rw_devices;
4059         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4060                 num_dev = 2;
4061         else
4062                 num_dev = 1;    /* DUP or single */
4063
4064         return num_dev;
4065 }
4066
4067 /*
4068  * If @is_allocation is true, reserve space in the system space info necessary
4069  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4070  * removing a chunk.
4071  */
4072 void check_system_chunk(struct btrfs_trans_handle *trans,
4073                         struct btrfs_root *root,
4074                         u64 type)
4075 {
4076         struct btrfs_space_info *info;
4077         u64 left;
4078         u64 thresh;
4079         int ret = 0;
4080         u64 num_devs;
4081
4082         /*
4083          * Needed because we can end up allocating a system chunk and for an
4084          * atomic and race free space reservation in the chunk block reserve.
4085          */
4086         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4087
4088         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4089         spin_lock(&info->lock);
4090         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4091                 info->bytes_reserved - info->bytes_readonly -
4092                 info->bytes_may_use;
4093         spin_unlock(&info->lock);
4094
4095         num_devs = get_profile_num_devs(root, type);
4096
4097         /* num_devs device items to update and 1 chunk item to add or remove */
4098         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4099                 btrfs_calc_trans_metadata_size(root, 1);
4100
4101         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4102                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4103                         left, thresh, type);
4104                 dump_space_info(info, 0, 0);
4105         }
4106
4107         if (left < thresh) {
4108                 u64 flags;
4109
4110                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4111                 /*
4112                  * Ignore failure to create system chunk. We might end up not
4113                  * needing it, as we might not need to COW all nodes/leafs from
4114                  * the paths we visit in the chunk tree (they were already COWed
4115                  * or created in the current transaction for example).
4116                  */
4117                 ret = btrfs_alloc_chunk(trans, root, flags);
4118         }
4119
4120         if (!ret) {
4121                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4122                                           &root->fs_info->chunk_block_rsv,
4123                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4124                 if (!ret)
4125                         trans->chunk_bytes_reserved += thresh;
4126         }
4127 }
4128
4129 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4130                           struct btrfs_root *extent_root, u64 flags, int force)
4131 {
4132         struct btrfs_space_info *space_info;
4133         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4134         int wait_for_alloc = 0;
4135         int ret = 0;
4136
4137         /* Don't re-enter if we're already allocating a chunk */
4138         if (trans->allocating_chunk)
4139                 return -ENOSPC;
4140
4141         space_info = __find_space_info(extent_root->fs_info, flags);
4142         if (!space_info) {
4143                 ret = update_space_info(extent_root->fs_info, flags,
4144                                         0, 0, &space_info);
4145                 BUG_ON(ret); /* -ENOMEM */
4146         }
4147         BUG_ON(!space_info); /* Logic error */
4148
4149 again:
4150         spin_lock(&space_info->lock);
4151         if (force < space_info->force_alloc)
4152                 force = space_info->force_alloc;
4153         if (space_info->full) {
4154                 if (should_alloc_chunk(extent_root, space_info, force))
4155                         ret = -ENOSPC;
4156                 else
4157                         ret = 0;
4158                 spin_unlock(&space_info->lock);
4159                 return ret;
4160         }
4161
4162         if (!should_alloc_chunk(extent_root, space_info, force)) {
4163                 spin_unlock(&space_info->lock);
4164                 return 0;
4165         } else if (space_info->chunk_alloc) {
4166                 wait_for_alloc = 1;
4167         } else {
4168                 space_info->chunk_alloc = 1;
4169         }
4170
4171         spin_unlock(&space_info->lock);
4172
4173         mutex_lock(&fs_info->chunk_mutex);
4174
4175         /*
4176          * The chunk_mutex is held throughout the entirety of a chunk
4177          * allocation, so once we've acquired the chunk_mutex we know that the
4178          * other guy is done and we need to recheck and see if we should
4179          * allocate.
4180          */
4181         if (wait_for_alloc) {
4182                 mutex_unlock(&fs_info->chunk_mutex);
4183                 wait_for_alloc = 0;
4184                 goto again;
4185         }
4186
4187         trans->allocating_chunk = true;
4188
4189         /*
4190          * If we have mixed data/metadata chunks we want to make sure we keep
4191          * allocating mixed chunks instead of individual chunks.
4192          */
4193         if (btrfs_mixed_space_info(space_info))
4194                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4195
4196         /*
4197          * if we're doing a data chunk, go ahead and make sure that
4198          * we keep a reasonable number of metadata chunks allocated in the
4199          * FS as well.
4200          */
4201         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4202                 fs_info->data_chunk_allocations++;
4203                 if (!(fs_info->data_chunk_allocations %
4204                       fs_info->metadata_ratio))
4205                         force_metadata_allocation(fs_info);
4206         }
4207
4208         /*
4209          * Check if we have enough space in SYSTEM chunk because we may need
4210          * to update devices.
4211          */
4212         check_system_chunk(trans, extent_root, flags);
4213
4214         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4215         trans->allocating_chunk = false;
4216
4217         spin_lock(&space_info->lock);
4218         if (ret < 0 && ret != -ENOSPC)
4219                 goto out;
4220         if (ret)
4221                 space_info->full = 1;
4222         else
4223                 ret = 1;
4224
4225         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4226 out:
4227         space_info->chunk_alloc = 0;
4228         spin_unlock(&space_info->lock);
4229         mutex_unlock(&fs_info->chunk_mutex);
4230         return ret;
4231 }
4232
4233 static int can_overcommit(struct btrfs_root *root,
4234                           struct btrfs_space_info *space_info, u64 bytes,
4235                           enum btrfs_reserve_flush_enum flush)
4236 {
4237         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4238         u64 profile = btrfs_get_alloc_profile(root, 0);
4239         u64 space_size;
4240         u64 avail;
4241         u64 used;
4242
4243         used = space_info->bytes_used + space_info->bytes_reserved +
4244                 space_info->bytes_pinned + space_info->bytes_readonly;
4245
4246         /*
4247          * We only want to allow over committing if we have lots of actual space
4248          * free, but if we don't have enough space to handle the global reserve
4249          * space then we could end up having a real enospc problem when trying
4250          * to allocate a chunk or some other such important allocation.
4251          */
4252         spin_lock(&global_rsv->lock);
4253         space_size = calc_global_rsv_need_space(global_rsv);
4254         spin_unlock(&global_rsv->lock);
4255         if (used + space_size >= space_info->total_bytes)
4256                 return 0;
4257
4258         used += space_info->bytes_may_use;
4259
4260         spin_lock(&root->fs_info->free_chunk_lock);
4261         avail = root->fs_info->free_chunk_space;
4262         spin_unlock(&root->fs_info->free_chunk_lock);
4263
4264         /*
4265          * If we have dup, raid1 or raid10 then only half of the free
4266          * space is actually useable.  For raid56, the space info used
4267          * doesn't include the parity drive, so we don't have to
4268          * change the math
4269          */
4270         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4271                        BTRFS_BLOCK_GROUP_RAID1 |
4272                        BTRFS_BLOCK_GROUP_RAID10))
4273                 avail >>= 1;
4274
4275         /*
4276          * If we aren't flushing all things, let us overcommit up to
4277          * 1/2th of the space. If we can flush, don't let us overcommit
4278          * too much, let it overcommit up to 1/8 of the space.
4279          */
4280         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4281                 avail >>= 3;
4282         else
4283                 avail >>= 1;
4284
4285         if (used + bytes < space_info->total_bytes + avail)
4286                 return 1;
4287         return 0;
4288 }
4289
4290 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4291                                          unsigned long nr_pages, int nr_items)
4292 {
4293         struct super_block *sb = root->fs_info->sb;
4294
4295         if (down_read_trylock(&sb->s_umount)) {
4296                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4297                 up_read(&sb->s_umount);
4298         } else {
4299                 /*
4300                  * We needn't worry the filesystem going from r/w to r/o though
4301                  * we don't acquire ->s_umount mutex, because the filesystem
4302                  * should guarantee the delalloc inodes list be empty after
4303                  * the filesystem is readonly(all dirty pages are written to
4304                  * the disk).
4305                  */
4306                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4307                 if (!current->journal_info)
4308                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4309         }
4310 }
4311
4312 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4313 {
4314         u64 bytes;
4315         int nr;
4316
4317         bytes = btrfs_calc_trans_metadata_size(root, 1);
4318         nr = (int)div64_u64(to_reclaim, bytes);
4319         if (!nr)
4320                 nr = 1;
4321         return nr;
4322 }
4323
4324 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4325
4326 /*
4327  * shrink metadata reservation for delalloc
4328  */
4329 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4330                             bool wait_ordered)
4331 {
4332         struct btrfs_block_rsv *block_rsv;
4333         struct btrfs_space_info *space_info;
4334         struct btrfs_trans_handle *trans;
4335         u64 delalloc_bytes;
4336         u64 max_reclaim;
4337         long time_left;
4338         unsigned long nr_pages;
4339         int loops;
4340         int items;
4341         enum btrfs_reserve_flush_enum flush;
4342
4343         /* Calc the number of the pages we need flush for space reservation */
4344         items = calc_reclaim_items_nr(root, to_reclaim);
4345         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4346
4347         trans = (struct btrfs_trans_handle *)current->journal_info;
4348         block_rsv = &root->fs_info->delalloc_block_rsv;
4349         space_info = block_rsv->space_info;
4350
4351         delalloc_bytes = percpu_counter_sum_positive(
4352                                                 &root->fs_info->delalloc_bytes);
4353         if (delalloc_bytes == 0) {
4354                 if (trans)
4355                         return;
4356                 if (wait_ordered)
4357                         btrfs_wait_ordered_roots(root->fs_info, items);
4358                 return;
4359         }
4360
4361         loops = 0;
4362         while (delalloc_bytes && loops < 3) {
4363                 max_reclaim = min(delalloc_bytes, to_reclaim);
4364                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4365                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4366                 /*
4367                  * We need to wait for the async pages to actually start before
4368                  * we do anything.
4369                  */
4370                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4371                 if (!max_reclaim)
4372                         goto skip_async;
4373
4374                 if (max_reclaim <= nr_pages)
4375                         max_reclaim = 0;
4376                 else
4377                         max_reclaim -= nr_pages;
4378
4379                 wait_event(root->fs_info->async_submit_wait,
4380                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4381                            (int)max_reclaim);
4382 skip_async:
4383                 if (!trans)
4384                         flush = BTRFS_RESERVE_FLUSH_ALL;
4385                 else
4386                         flush = BTRFS_RESERVE_NO_FLUSH;
4387                 spin_lock(&space_info->lock);
4388                 if (can_overcommit(root, space_info, orig, flush)) {
4389                         spin_unlock(&space_info->lock);
4390                         break;
4391                 }
4392                 spin_unlock(&space_info->lock);
4393
4394                 loops++;
4395                 if (wait_ordered && !trans) {
4396                         btrfs_wait_ordered_roots(root->fs_info, items);
4397                 } else {
4398                         time_left = schedule_timeout_killable(1);
4399                         if (time_left)
4400                                 break;
4401                 }
4402                 delalloc_bytes = percpu_counter_sum_positive(
4403                                                 &root->fs_info->delalloc_bytes);
4404         }
4405 }
4406
4407 /**
4408  * maybe_commit_transaction - possibly commit the transaction if its ok to
4409  * @root - the root we're allocating for
4410  * @bytes - the number of bytes we want to reserve
4411  * @force - force the commit
4412  *
4413  * This will check to make sure that committing the transaction will actually
4414  * get us somewhere and then commit the transaction if it does.  Otherwise it
4415  * will return -ENOSPC.
4416  */
4417 static int may_commit_transaction(struct btrfs_root *root,
4418                                   struct btrfs_space_info *space_info,
4419                                   u64 bytes, int force)
4420 {
4421         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4422         struct btrfs_trans_handle *trans;
4423
4424         trans = (struct btrfs_trans_handle *)current->journal_info;
4425         if (trans)
4426                 return -EAGAIN;
4427
4428         if (force)
4429                 goto commit;
4430
4431         /* See if there is enough pinned space to make this reservation */
4432         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4433                                    bytes) >= 0)
4434                 goto commit;
4435
4436         /*
4437          * See if there is some space in the delayed insertion reservation for
4438          * this reservation.
4439          */
4440         if (space_info != delayed_rsv->space_info)
4441                 return -ENOSPC;
4442
4443         spin_lock(&delayed_rsv->lock);
4444         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4445                                    bytes - delayed_rsv->size) >= 0) {
4446                 spin_unlock(&delayed_rsv->lock);
4447                 return -ENOSPC;
4448         }
4449         spin_unlock(&delayed_rsv->lock);
4450
4451 commit:
4452         trans = btrfs_join_transaction(root);
4453         if (IS_ERR(trans))
4454                 return -ENOSPC;
4455
4456         return btrfs_commit_transaction(trans, root);
4457 }
4458
4459 enum flush_state {
4460         FLUSH_DELAYED_ITEMS_NR  =       1,
4461         FLUSH_DELAYED_ITEMS     =       2,
4462         FLUSH_DELALLOC          =       3,
4463         FLUSH_DELALLOC_WAIT     =       4,
4464         ALLOC_CHUNK             =       5,
4465         COMMIT_TRANS            =       6,
4466 };
4467
4468 static int flush_space(struct btrfs_root *root,
4469                        struct btrfs_space_info *space_info, u64 num_bytes,
4470                        u64 orig_bytes, int state)
4471 {
4472         struct btrfs_trans_handle *trans;
4473         int nr;
4474         int ret = 0;
4475
4476         switch (state) {
4477         case FLUSH_DELAYED_ITEMS_NR:
4478         case FLUSH_DELAYED_ITEMS:
4479                 if (state == FLUSH_DELAYED_ITEMS_NR)
4480                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4481                 else
4482                         nr = -1;
4483
4484                 trans = btrfs_join_transaction(root);
4485                 if (IS_ERR(trans)) {
4486                         ret = PTR_ERR(trans);
4487                         break;
4488                 }
4489                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4490                 btrfs_end_transaction(trans, root);
4491                 break;
4492         case FLUSH_DELALLOC:
4493         case FLUSH_DELALLOC_WAIT:
4494                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4495                                 state == FLUSH_DELALLOC_WAIT);
4496                 break;
4497         case ALLOC_CHUNK:
4498                 trans = btrfs_join_transaction(root);
4499                 if (IS_ERR(trans)) {
4500                         ret = PTR_ERR(trans);
4501                         break;
4502                 }
4503                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4504                                      btrfs_get_alloc_profile(root, 0),
4505                                      CHUNK_ALLOC_NO_FORCE);
4506                 btrfs_end_transaction(trans, root);
4507                 if (ret == -ENOSPC)
4508                         ret = 0;
4509                 break;
4510         case COMMIT_TRANS:
4511                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4512                 break;
4513         default:
4514                 ret = -ENOSPC;
4515                 break;
4516         }
4517
4518         return ret;
4519 }
4520
4521 static inline u64
4522 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4523                                  struct btrfs_space_info *space_info)
4524 {
4525         u64 used;
4526         u64 expected;
4527         u64 to_reclaim;
4528
4529         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4530                                 16 * 1024 * 1024);
4531         spin_lock(&space_info->lock);
4532         if (can_overcommit(root, space_info, to_reclaim,
4533                            BTRFS_RESERVE_FLUSH_ALL)) {
4534                 to_reclaim = 0;
4535                 goto out;
4536         }
4537
4538         used = space_info->bytes_used + space_info->bytes_reserved +
4539                space_info->bytes_pinned + space_info->bytes_readonly +
4540                space_info->bytes_may_use;
4541         if (can_overcommit(root, space_info, 1024 * 1024,
4542                            BTRFS_RESERVE_FLUSH_ALL))
4543                 expected = div_factor_fine(space_info->total_bytes, 95);
4544         else
4545                 expected = div_factor_fine(space_info->total_bytes, 90);
4546
4547         if (used > expected)
4548                 to_reclaim = used - expected;
4549         else
4550                 to_reclaim = 0;
4551         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4552                                      space_info->bytes_reserved);
4553 out:
4554         spin_unlock(&space_info->lock);
4555
4556         return to_reclaim;
4557 }
4558
4559 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4560                                         struct btrfs_fs_info *fs_info, u64 used)
4561 {
4562         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4563
4564         /* If we're just plain full then async reclaim just slows us down. */
4565         if (space_info->bytes_used >= thresh)
4566                 return 0;
4567
4568         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4569                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4570 }
4571
4572 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4573                                        struct btrfs_fs_info *fs_info,
4574                                        int flush_state)
4575 {
4576         u64 used;
4577
4578         spin_lock(&space_info->lock);
4579         /*
4580          * We run out of space and have not got any free space via flush_space,
4581          * so don't bother doing async reclaim.
4582          */
4583         if (flush_state > COMMIT_TRANS && space_info->full) {
4584                 spin_unlock(&space_info->lock);
4585                 return 0;
4586         }
4587
4588         used = space_info->bytes_used + space_info->bytes_reserved +
4589                space_info->bytes_pinned + space_info->bytes_readonly +
4590                space_info->bytes_may_use;
4591         if (need_do_async_reclaim(space_info, fs_info, used)) {
4592                 spin_unlock(&space_info->lock);
4593                 return 1;
4594         }
4595         spin_unlock(&space_info->lock);
4596
4597         return 0;
4598 }
4599
4600 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4601 {
4602         struct btrfs_fs_info *fs_info;
4603         struct btrfs_space_info *space_info;
4604         u64 to_reclaim;
4605         int flush_state;
4606
4607         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4608         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4609
4610         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4611                                                       space_info);
4612         if (!to_reclaim)
4613                 return;
4614
4615         flush_state = FLUSH_DELAYED_ITEMS_NR;
4616         do {
4617                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4618                             to_reclaim, flush_state);
4619                 flush_state++;
4620                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4621                                                  flush_state))
4622                         return;
4623         } while (flush_state < COMMIT_TRANS);
4624 }
4625
4626 void btrfs_init_async_reclaim_work(struct work_struct *work)
4627 {
4628         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4629 }
4630
4631 /**
4632  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4633  * @root - the root we're allocating for
4634  * @block_rsv - the block_rsv we're allocating for
4635  * @orig_bytes - the number of bytes we want
4636  * @flush - whether or not we can flush to make our reservation
4637  *
4638  * This will reserve orgi_bytes number of bytes from the space info associated
4639  * with the block_rsv.  If there is not enough space it will make an attempt to
4640  * flush out space to make room.  It will do this by flushing delalloc if
4641  * possible or committing the transaction.  If flush is 0 then no attempts to
4642  * regain reservations will be made and this will fail if there is not enough
4643  * space already.
4644  */
4645 static int reserve_metadata_bytes(struct btrfs_root *root,
4646                                   struct btrfs_block_rsv *block_rsv,
4647                                   u64 orig_bytes,
4648                                   enum btrfs_reserve_flush_enum flush)
4649 {
4650         struct btrfs_space_info *space_info = block_rsv->space_info;
4651         u64 used;
4652         u64 num_bytes = orig_bytes;
4653         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4654         int ret = 0;
4655         bool flushing = false;
4656
4657 again:
4658         ret = 0;
4659         spin_lock(&space_info->lock);
4660         /*
4661          * We only want to wait if somebody other than us is flushing and we
4662          * are actually allowed to flush all things.
4663          */
4664         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4665                space_info->flush) {
4666                 spin_unlock(&space_info->lock);
4667                 /*
4668                  * If we have a trans handle we can't wait because the flusher
4669                  * may have to commit the transaction, which would mean we would
4670                  * deadlock since we are waiting for the flusher to finish, but
4671                  * hold the current transaction open.
4672                  */
4673                 if (current->journal_info)
4674                         return -EAGAIN;
4675                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4676                 /* Must have been killed, return */
4677                 if (ret)
4678                         return -EINTR;
4679
4680                 spin_lock(&space_info->lock);
4681         }
4682
4683         ret = -ENOSPC;
4684         used = space_info->bytes_used + space_info->bytes_reserved +
4685                 space_info->bytes_pinned + space_info->bytes_readonly +
4686                 space_info->bytes_may_use;
4687
4688         /*
4689          * The idea here is that we've not already over-reserved the block group
4690          * then we can go ahead and save our reservation first and then start
4691          * flushing if we need to.  Otherwise if we've already overcommitted
4692          * lets start flushing stuff first and then come back and try to make
4693          * our reservation.
4694          */
4695         if (used <= space_info->total_bytes) {
4696                 if (used + orig_bytes <= space_info->total_bytes) {
4697                         space_info->bytes_may_use += orig_bytes;
4698                         trace_btrfs_space_reservation(root->fs_info,
4699                                 "space_info", space_info->flags, orig_bytes, 1);
4700                         ret = 0;
4701                 } else {
4702                         /*
4703                          * Ok set num_bytes to orig_bytes since we aren't
4704                          * overocmmitted, this way we only try and reclaim what
4705                          * we need.
4706                          */
4707                         num_bytes = orig_bytes;
4708                 }
4709         } else {
4710                 /*
4711                  * Ok we're over committed, set num_bytes to the overcommitted
4712                  * amount plus the amount of bytes that we need for this
4713                  * reservation.
4714                  */
4715                 num_bytes = used - space_info->total_bytes +
4716                         (orig_bytes * 2);
4717         }
4718
4719         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4720                 space_info->bytes_may_use += orig_bytes;
4721                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4722                                               space_info->flags, orig_bytes,
4723                                               1);
4724                 ret = 0;
4725         }
4726
4727         /*
4728          * Couldn't make our reservation, save our place so while we're trying
4729          * to reclaim space we can actually use it instead of somebody else
4730          * stealing it from us.
4731          *
4732          * We make the other tasks wait for the flush only when we can flush
4733          * all things.
4734          */
4735         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4736                 flushing = true;
4737                 space_info->flush = 1;
4738         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4739                 used += orig_bytes;
4740                 /*
4741                  * We will do the space reservation dance during log replay,
4742                  * which means we won't have fs_info->fs_root set, so don't do
4743                  * the async reclaim as we will panic.
4744                  */
4745                 if (!root->fs_info->log_root_recovering &&
4746                     need_do_async_reclaim(space_info, root->fs_info, used) &&
4747                     !work_busy(&root->fs_info->async_reclaim_work))
4748                         queue_work(system_unbound_wq,
4749                                    &root->fs_info->async_reclaim_work);
4750         }
4751         spin_unlock(&space_info->lock);
4752
4753         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4754                 goto out;
4755
4756         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4757                           flush_state);
4758         flush_state++;
4759
4760         /*
4761          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4762          * would happen. So skip delalloc flush.
4763          */
4764         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4765             (flush_state == FLUSH_DELALLOC ||
4766              flush_state == FLUSH_DELALLOC_WAIT))
4767                 flush_state = ALLOC_CHUNK;
4768
4769         if (!ret)
4770                 goto again;
4771         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4772                  flush_state < COMMIT_TRANS)
4773                 goto again;
4774         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4775                  flush_state <= COMMIT_TRANS)
4776                 goto again;
4777
4778 out:
4779         if (ret == -ENOSPC &&
4780             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4781                 struct btrfs_block_rsv *global_rsv =
4782                         &root->fs_info->global_block_rsv;
4783
4784                 if (block_rsv != global_rsv &&
4785                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4786                         ret = 0;
4787         }
4788         if (ret == -ENOSPC)
4789                 trace_btrfs_space_reservation(root->fs_info,
4790                                               "space_info:enospc",
4791                                               space_info->flags, orig_bytes, 1);
4792         if (flushing) {
4793                 spin_lock(&space_info->lock);
4794                 space_info->flush = 0;
4795                 wake_up_all(&space_info->wait);
4796                 spin_unlock(&space_info->lock);
4797         }
4798         return ret;
4799 }
4800
4801 static struct btrfs_block_rsv *get_block_rsv(
4802                                         const struct btrfs_trans_handle *trans,
4803                                         const struct btrfs_root *root)
4804 {
4805         struct btrfs_block_rsv *block_rsv = NULL;
4806
4807         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4808                 block_rsv = trans->block_rsv;
4809
4810         if (root == root->fs_info->csum_root && trans->adding_csums)
4811                 block_rsv = trans->block_rsv;
4812
4813         if (root == root->fs_info->uuid_root)
4814                 block_rsv = trans->block_rsv;
4815
4816         if (!block_rsv)
4817                 block_rsv = root->block_rsv;
4818
4819         if (!block_rsv)
4820                 block_rsv = &root->fs_info->empty_block_rsv;
4821
4822         return block_rsv;
4823 }
4824
4825 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4826                                u64 num_bytes)
4827 {
4828         int ret = -ENOSPC;
4829         spin_lock(&block_rsv->lock);
4830         if (block_rsv->reserved >= num_bytes) {
4831                 block_rsv->reserved -= num_bytes;
4832                 if (block_rsv->reserved < block_rsv->size)
4833                         block_rsv->full = 0;
4834                 ret = 0;
4835         }
4836         spin_unlock(&block_rsv->lock);
4837         return ret;
4838 }
4839
4840 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4841                                 u64 num_bytes, int update_size)
4842 {
4843         spin_lock(&block_rsv->lock);
4844         block_rsv->reserved += num_bytes;
4845         if (update_size)
4846                 block_rsv->size += num_bytes;
4847         else if (block_rsv->reserved >= block_rsv->size)
4848                 block_rsv->full = 1;
4849         spin_unlock(&block_rsv->lock);
4850 }
4851
4852 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4853                              struct btrfs_block_rsv *dest, u64 num_bytes,
4854                              int min_factor)
4855 {
4856         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4857         u64 min_bytes;
4858
4859         if (global_rsv->space_info != dest->space_info)
4860                 return -ENOSPC;
4861
4862         spin_lock(&global_rsv->lock);
4863         min_bytes = div_factor(global_rsv->size, min_factor);
4864         if (global_rsv->reserved < min_bytes + num_bytes) {
4865                 spin_unlock(&global_rsv->lock);
4866                 return -ENOSPC;
4867         }
4868         global_rsv->reserved -= num_bytes;
4869         if (global_rsv->reserved < global_rsv->size)
4870                 global_rsv->full = 0;
4871         spin_unlock(&global_rsv->lock);
4872
4873         block_rsv_add_bytes(dest, num_bytes, 1);
4874         return 0;
4875 }
4876
4877 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4878                                     struct btrfs_block_rsv *block_rsv,
4879                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4880 {
4881         struct btrfs_space_info *space_info = block_rsv->space_info;
4882
4883         spin_lock(&block_rsv->lock);
4884         if (num_bytes == (u64)-1)
4885                 num_bytes = block_rsv->size;
4886         block_rsv->size -= num_bytes;
4887         if (block_rsv->reserved >= block_rsv->size) {
4888                 num_bytes = block_rsv->reserved - block_rsv->size;
4889                 block_rsv->reserved = block_rsv->size;
4890                 block_rsv->full = 1;
4891         } else {
4892                 num_bytes = 0;
4893         }
4894         spin_unlock(&block_rsv->lock);
4895
4896         if (num_bytes > 0) {
4897                 if (dest) {
4898                         spin_lock(&dest->lock);
4899                         if (!dest->full) {
4900                                 u64 bytes_to_add;
4901
4902                                 bytes_to_add = dest->size - dest->reserved;
4903                                 bytes_to_add = min(num_bytes, bytes_to_add);
4904                                 dest->reserved += bytes_to_add;
4905                                 if (dest->reserved >= dest->size)
4906                                         dest->full = 1;
4907                                 num_bytes -= bytes_to_add;
4908                         }
4909                         spin_unlock(&dest->lock);
4910                 }
4911                 if (num_bytes) {
4912                         spin_lock(&space_info->lock);
4913                         space_info->bytes_may_use -= num_bytes;
4914                         trace_btrfs_space_reservation(fs_info, "space_info",
4915                                         space_info->flags, num_bytes, 0);
4916                         spin_unlock(&space_info->lock);
4917                 }
4918         }
4919 }
4920
4921 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4922                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4923 {
4924         int ret;
4925
4926         ret = block_rsv_use_bytes(src, num_bytes);
4927         if (ret)
4928                 return ret;
4929
4930         block_rsv_add_bytes(dst, num_bytes, 1);
4931         return 0;
4932 }
4933
4934 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4935 {
4936         memset(rsv, 0, sizeof(*rsv));
4937         spin_lock_init(&rsv->lock);
4938         rsv->type = type;
4939 }
4940
4941 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4942                                               unsigned short type)
4943 {
4944         struct btrfs_block_rsv *block_rsv;
4945         struct btrfs_fs_info *fs_info = root->fs_info;
4946
4947         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4948         if (!block_rsv)
4949                 return NULL;
4950
4951         btrfs_init_block_rsv(block_rsv, type);
4952         block_rsv->space_info = __find_space_info(fs_info,
4953                                                   BTRFS_BLOCK_GROUP_METADATA);
4954         return block_rsv;
4955 }
4956
4957 void btrfs_free_block_rsv(struct btrfs_root *root,
4958                           struct btrfs_block_rsv *rsv)
4959 {
4960         if (!rsv)
4961                 return;
4962         btrfs_block_rsv_release(root, rsv, (u64)-1);
4963         kfree(rsv);
4964 }
4965
4966 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
4967 {
4968         kfree(rsv);
4969 }
4970
4971 int btrfs_block_rsv_add(struct btrfs_root *root,
4972                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4973                         enum btrfs_reserve_flush_enum flush)
4974 {
4975         int ret;
4976
4977         if (num_bytes == 0)
4978                 return 0;
4979
4980         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4981         if (!ret) {
4982                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4983                 return 0;
4984         }
4985
4986         return ret;
4987 }
4988
4989 int btrfs_block_rsv_check(struct btrfs_root *root,
4990                           struct btrfs_block_rsv *block_rsv, int min_factor)
4991 {
4992         u64 num_bytes = 0;
4993         int ret = -ENOSPC;
4994
4995         if (!block_rsv)
4996                 return 0;
4997
4998         spin_lock(&block_rsv->lock);
4999         num_bytes = div_factor(block_rsv->size, min_factor);
5000         if (block_rsv->reserved >= num_bytes)
5001                 ret = 0;
5002         spin_unlock(&block_rsv->lock);
5003
5004         return ret;
5005 }
5006
5007 int btrfs_block_rsv_refill(struct btrfs_root *root,
5008                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5009                            enum btrfs_reserve_flush_enum flush)
5010 {
5011         u64 num_bytes = 0;
5012         int ret = -ENOSPC;
5013
5014         if (!block_rsv)
5015                 return 0;
5016
5017         spin_lock(&block_rsv->lock);
5018         num_bytes = min_reserved;
5019         if (block_rsv->reserved >= num_bytes)
5020                 ret = 0;
5021         else
5022                 num_bytes -= block_rsv->reserved;
5023         spin_unlock(&block_rsv->lock);
5024
5025         if (!ret)
5026                 return 0;
5027
5028         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5029         if (!ret) {
5030                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5031                 return 0;
5032         }
5033
5034         return ret;
5035 }
5036
5037 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5038                             struct btrfs_block_rsv *dst_rsv,
5039                             u64 num_bytes)
5040 {
5041         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5042 }
5043
5044 void btrfs_block_rsv_release(struct btrfs_root *root,
5045                              struct btrfs_block_rsv *block_rsv,
5046                              u64 num_bytes)
5047 {
5048         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5049         if (global_rsv == block_rsv ||
5050             block_rsv->space_info != global_rsv->space_info)
5051                 global_rsv = NULL;
5052         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5053                                 num_bytes);
5054 }
5055
5056 /*
5057  * helper to calculate size of global block reservation.
5058  * the desired value is sum of space used by extent tree,
5059  * checksum tree and root tree
5060  */
5061 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5062 {
5063         struct btrfs_space_info *sinfo;
5064         u64 num_bytes;
5065         u64 meta_used;
5066         u64 data_used;
5067         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5068
5069         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5070         spin_lock(&sinfo->lock);
5071         data_used = sinfo->bytes_used;
5072         spin_unlock(&sinfo->lock);
5073
5074         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5075         spin_lock(&sinfo->lock);
5076         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5077                 data_used = 0;
5078         meta_used = sinfo->bytes_used;
5079         spin_unlock(&sinfo->lock);
5080
5081         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5082                     csum_size * 2;
5083         num_bytes += div_u64(data_used + meta_used, 50);
5084
5085         if (num_bytes * 3 > meta_used)
5086                 num_bytes = div_u64(meta_used, 3);
5087
5088         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5089 }
5090
5091 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5092 {
5093         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5094         struct btrfs_space_info *sinfo = block_rsv->space_info;
5095         u64 num_bytes;
5096
5097         num_bytes = calc_global_metadata_size(fs_info);
5098
5099         spin_lock(&sinfo->lock);
5100         spin_lock(&block_rsv->lock);
5101
5102         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
5103
5104         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5105                     sinfo->bytes_reserved + sinfo->bytes_readonly +
5106                     sinfo->bytes_may_use;
5107
5108         if (sinfo->total_bytes > num_bytes) {
5109                 num_bytes = sinfo->total_bytes - num_bytes;
5110                 block_rsv->reserved += num_bytes;
5111                 sinfo->bytes_may_use += num_bytes;
5112                 trace_btrfs_space_reservation(fs_info, "space_info",
5113                                       sinfo->flags, num_bytes, 1);
5114         }
5115
5116         if (block_rsv->reserved >= block_rsv->size) {
5117                 num_bytes = block_rsv->reserved - block_rsv->size;
5118                 sinfo->bytes_may_use -= num_bytes;
5119                 trace_btrfs_space_reservation(fs_info, "space_info",
5120                                       sinfo->flags, num_bytes, 0);
5121                 block_rsv->reserved = block_rsv->size;
5122                 block_rsv->full = 1;
5123         }
5124
5125         spin_unlock(&block_rsv->lock);
5126         spin_unlock(&sinfo->lock);
5127 }
5128
5129 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5130 {
5131         struct btrfs_space_info *space_info;
5132
5133         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5134         fs_info->chunk_block_rsv.space_info = space_info;
5135
5136         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5137         fs_info->global_block_rsv.space_info = space_info;
5138         fs_info->delalloc_block_rsv.space_info = space_info;
5139         fs_info->trans_block_rsv.space_info = space_info;
5140         fs_info->empty_block_rsv.space_info = space_info;
5141         fs_info->delayed_block_rsv.space_info = space_info;
5142
5143         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5144         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5145         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5146         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5147         if (fs_info->quota_root)
5148                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5149         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5150
5151         update_global_block_rsv(fs_info);
5152 }
5153
5154 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5155 {
5156         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5157                                 (u64)-1);
5158         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5159         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5160         WARN_ON(fs_info->trans_block_rsv.size > 0);
5161         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5162         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5163         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5164         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5165         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5166 }
5167
5168 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5169                                   struct btrfs_root *root)
5170 {
5171         if (!trans->block_rsv)
5172                 return;
5173
5174         if (!trans->bytes_reserved)
5175                 return;
5176
5177         trace_btrfs_space_reservation(root->fs_info, "transaction",
5178                                       trans->transid, trans->bytes_reserved, 0);
5179         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5180         trans->bytes_reserved = 0;
5181 }
5182
5183 /*
5184  * To be called after all the new block groups attached to the transaction
5185  * handle have been created (btrfs_create_pending_block_groups()).
5186  */
5187 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5188 {
5189         struct btrfs_fs_info *fs_info = trans->root->fs_info;
5190
5191         if (!trans->chunk_bytes_reserved)
5192                 return;
5193
5194         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5195
5196         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5197                                 trans->chunk_bytes_reserved);
5198         trans->chunk_bytes_reserved = 0;
5199 }
5200
5201 /* Can only return 0 or -ENOSPC */
5202 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5203                                   struct inode *inode)
5204 {
5205         struct btrfs_root *root = BTRFS_I(inode)->root;
5206         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5207         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5208
5209         /*
5210          * We need to hold space in order to delete our orphan item once we've
5211          * added it, so this takes the reservation so we can release it later
5212          * when we are truly done with the orphan item.
5213          */
5214         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5215         trace_btrfs_space_reservation(root->fs_info, "orphan",
5216                                       btrfs_ino(inode), num_bytes, 1);
5217         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5218 }
5219
5220 void btrfs_orphan_release_metadata(struct inode *inode)
5221 {
5222         struct btrfs_root *root = BTRFS_I(inode)->root;
5223         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5224         trace_btrfs_space_reservation(root->fs_info, "orphan",
5225                                       btrfs_ino(inode), num_bytes, 0);
5226         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5227 }
5228
5229 /*
5230  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5231  * root: the root of the parent directory
5232  * rsv: block reservation
5233  * items: the number of items that we need do reservation
5234  * qgroup_reserved: used to return the reserved size in qgroup
5235  *
5236  * This function is used to reserve the space for snapshot/subvolume
5237  * creation and deletion. Those operations are different with the
5238  * common file/directory operations, they change two fs/file trees
5239  * and root tree, the number of items that the qgroup reserves is
5240  * different with the free space reservation. So we can not use
5241  * the space reseravtion mechanism in start_transaction().
5242  */
5243 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5244                                      struct btrfs_block_rsv *rsv,
5245                                      int items,
5246                                      u64 *qgroup_reserved,
5247                                      bool use_global_rsv)
5248 {
5249         u64 num_bytes;
5250         int ret;
5251         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5252
5253         if (root->fs_info->quota_enabled) {
5254                 /* One for parent inode, two for dir entries */
5255                 num_bytes = 3 * root->nodesize;
5256                 ret = btrfs_qgroup_reserve(root, num_bytes);
5257                 if (ret)
5258                         return ret;
5259         } else {
5260                 num_bytes = 0;
5261         }
5262
5263         *qgroup_reserved = num_bytes;
5264
5265         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5266         rsv->space_info = __find_space_info(root->fs_info,
5267                                             BTRFS_BLOCK_GROUP_METADATA);
5268         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5269                                   BTRFS_RESERVE_FLUSH_ALL);
5270
5271         if (ret == -ENOSPC && use_global_rsv)
5272                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5273
5274         if (ret) {
5275                 if (*qgroup_reserved)
5276                         btrfs_qgroup_free(root, *qgroup_reserved);
5277         }
5278
5279         return ret;
5280 }
5281
5282 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5283                                       struct btrfs_block_rsv *rsv,
5284                                       u64 qgroup_reserved)
5285 {
5286         btrfs_block_rsv_release(root, rsv, (u64)-1);
5287 }
5288
5289 /**
5290  * drop_outstanding_extent - drop an outstanding extent
5291  * @inode: the inode we're dropping the extent for
5292  * @num_bytes: the number of bytes we're relaseing.
5293  *
5294  * This is called when we are freeing up an outstanding extent, either called
5295  * after an error or after an extent is written.  This will return the number of
5296  * reserved extents that need to be freed.  This must be called with
5297  * BTRFS_I(inode)->lock held.
5298  */
5299 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5300 {
5301         unsigned drop_inode_space = 0;
5302         unsigned dropped_extents = 0;
5303         unsigned num_extents = 0;
5304
5305         num_extents = (unsigned)div64_u64(num_bytes +
5306                                           BTRFS_MAX_EXTENT_SIZE - 1,
5307                                           BTRFS_MAX_EXTENT_SIZE);
5308         ASSERT(num_extents);
5309         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5310         BTRFS_I(inode)->outstanding_extents -= num_extents;
5311
5312         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5313             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5314                                &BTRFS_I(inode)->runtime_flags))
5315                 drop_inode_space = 1;
5316
5317         /*
5318          * If we have more or the same amount of outsanding extents than we have
5319          * reserved then we need to leave the reserved extents count alone.
5320          */
5321         if (BTRFS_I(inode)->outstanding_extents >=
5322             BTRFS_I(inode)->reserved_extents)
5323                 return drop_inode_space;
5324
5325         dropped_extents = BTRFS_I(inode)->reserved_extents -
5326                 BTRFS_I(inode)->outstanding_extents;
5327         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5328         return dropped_extents + drop_inode_space;
5329 }
5330
5331 /**
5332  * calc_csum_metadata_size - return the amount of metada space that must be
5333  *      reserved/free'd for the given bytes.
5334  * @inode: the inode we're manipulating
5335  * @num_bytes: the number of bytes in question
5336  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5337  *
5338  * This adjusts the number of csum_bytes in the inode and then returns the
5339  * correct amount of metadata that must either be reserved or freed.  We
5340  * calculate how many checksums we can fit into one leaf and then divide the
5341  * number of bytes that will need to be checksumed by this value to figure out
5342  * how many checksums will be required.  If we are adding bytes then the number
5343  * may go up and we will return the number of additional bytes that must be
5344  * reserved.  If it is going down we will return the number of bytes that must
5345  * be freed.
5346  *
5347  * This must be called with BTRFS_I(inode)->lock held.
5348  */
5349 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5350                                    int reserve)
5351 {
5352         struct btrfs_root *root = BTRFS_I(inode)->root;
5353         u64 old_csums, num_csums;
5354
5355         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5356             BTRFS_I(inode)->csum_bytes == 0)
5357                 return 0;
5358
5359         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5360         if (reserve)
5361                 BTRFS_I(inode)->csum_bytes += num_bytes;
5362         else
5363                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5364         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5365
5366         /* No change, no need to reserve more */
5367         if (old_csums == num_csums)
5368                 return 0;
5369
5370         if (reserve)
5371                 return btrfs_calc_trans_metadata_size(root,
5372                                                       num_csums - old_csums);
5373
5374         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5375 }
5376
5377 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5378 {
5379         struct btrfs_root *root = BTRFS_I(inode)->root;
5380         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5381         u64 to_reserve = 0;
5382         u64 csum_bytes;
5383         unsigned nr_extents = 0;
5384         int extra_reserve = 0;
5385         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5386         int ret = 0;
5387         bool delalloc_lock = true;
5388         u64 to_free = 0;
5389         unsigned dropped;
5390
5391         /* If we are a free space inode we need to not flush since we will be in
5392          * the middle of a transaction commit.  We also don't need the delalloc
5393          * mutex since we won't race with anybody.  We need this mostly to make
5394          * lockdep shut its filthy mouth.
5395          */
5396         if (btrfs_is_free_space_inode(inode)) {
5397                 flush = BTRFS_RESERVE_NO_FLUSH;
5398                 delalloc_lock = false;
5399         }
5400
5401         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5402             btrfs_transaction_in_commit(root->fs_info))
5403                 schedule_timeout(1);
5404
5405         if (delalloc_lock)
5406                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5407
5408         num_bytes = ALIGN(num_bytes, root->sectorsize);
5409
5410         spin_lock(&BTRFS_I(inode)->lock);
5411         nr_extents = (unsigned)div64_u64(num_bytes +
5412                                          BTRFS_MAX_EXTENT_SIZE - 1,
5413                                          BTRFS_MAX_EXTENT_SIZE);
5414         BTRFS_I(inode)->outstanding_extents += nr_extents;
5415         nr_extents = 0;
5416
5417         if (BTRFS_I(inode)->outstanding_extents >
5418             BTRFS_I(inode)->reserved_extents)
5419                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5420                         BTRFS_I(inode)->reserved_extents;
5421
5422         /*
5423          * Add an item to reserve for updating the inode when we complete the
5424          * delalloc io.
5425          */
5426         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5427                       &BTRFS_I(inode)->runtime_flags)) {
5428                 nr_extents++;
5429                 extra_reserve = 1;
5430         }
5431
5432         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5433         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5434         csum_bytes = BTRFS_I(inode)->csum_bytes;
5435         spin_unlock(&BTRFS_I(inode)->lock);
5436
5437         if (root->fs_info->quota_enabled) {
5438                 ret = btrfs_qgroup_reserve(root, nr_extents * root->nodesize);
5439                 if (ret)
5440                         goto out_fail;
5441         }
5442
5443         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5444         if (unlikely(ret)) {
5445                 if (root->fs_info->quota_enabled)
5446                         btrfs_qgroup_free(root, nr_extents * root->nodesize);
5447                 goto out_fail;
5448         }
5449
5450         spin_lock(&BTRFS_I(inode)->lock);
5451         if (extra_reserve) {
5452                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5453                         &BTRFS_I(inode)->runtime_flags);
5454                 nr_extents--;
5455         }
5456         BTRFS_I(inode)->reserved_extents += nr_extents;
5457         spin_unlock(&BTRFS_I(inode)->lock);
5458
5459         if (delalloc_lock)
5460                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5461
5462         if (to_reserve)
5463                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5464                                               btrfs_ino(inode), to_reserve, 1);
5465         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5466
5467         return 0;
5468
5469 out_fail:
5470         spin_lock(&BTRFS_I(inode)->lock);
5471         dropped = drop_outstanding_extent(inode, num_bytes);
5472         /*
5473          * If the inodes csum_bytes is the same as the original
5474          * csum_bytes then we know we haven't raced with any free()ers
5475          * so we can just reduce our inodes csum bytes and carry on.
5476          */
5477         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5478                 calc_csum_metadata_size(inode, num_bytes, 0);
5479         } else {
5480                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5481                 u64 bytes;
5482
5483                 /*
5484                  * This is tricky, but first we need to figure out how much we
5485                  * free'd from any free-ers that occured during this
5486                  * reservation, so we reset ->csum_bytes to the csum_bytes
5487                  * before we dropped our lock, and then call the free for the
5488                  * number of bytes that were freed while we were trying our
5489                  * reservation.
5490                  */
5491                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5492                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5493                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5494
5495
5496                 /*
5497                  * Now we need to see how much we would have freed had we not
5498                  * been making this reservation and our ->csum_bytes were not
5499                  * artificially inflated.
5500                  */
5501                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5502                 bytes = csum_bytes - orig_csum_bytes;
5503                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5504
5505                 /*
5506                  * Now reset ->csum_bytes to what it should be.  If bytes is
5507                  * more than to_free then we would have free'd more space had we
5508                  * not had an artificially high ->csum_bytes, so we need to free
5509                  * the remainder.  If bytes is the same or less then we don't
5510                  * need to do anything, the other free-ers did the correct
5511                  * thing.
5512                  */
5513                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5514                 if (bytes > to_free)
5515                         to_free = bytes - to_free;
5516                 else
5517                         to_free = 0;
5518         }
5519         spin_unlock(&BTRFS_I(inode)->lock);
5520         if (dropped)
5521                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5522
5523         if (to_free) {
5524                 btrfs_block_rsv_release(root, block_rsv, to_free);
5525                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5526                                               btrfs_ino(inode), to_free, 0);
5527         }
5528         if (delalloc_lock)
5529                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5530         return ret;
5531 }
5532
5533 /**
5534  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5535  * @inode: the inode to release the reservation for
5536  * @num_bytes: the number of bytes we're releasing
5537  *
5538  * This will release the metadata reservation for an inode.  This can be called
5539  * once we complete IO for a given set of bytes to release their metadata
5540  * reservations.
5541  */
5542 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5543 {
5544         struct btrfs_root *root = BTRFS_I(inode)->root;
5545         u64 to_free = 0;
5546         unsigned dropped;
5547
5548         num_bytes = ALIGN(num_bytes, root->sectorsize);
5549         spin_lock(&BTRFS_I(inode)->lock);
5550         dropped = drop_outstanding_extent(inode, num_bytes);
5551
5552         if (num_bytes)
5553                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5554         spin_unlock(&BTRFS_I(inode)->lock);
5555         if (dropped > 0)
5556                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5557
5558         if (btrfs_test_is_dummy_root(root))
5559                 return;
5560
5561         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5562                                       btrfs_ino(inode), to_free, 0);
5563
5564         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5565                                 to_free);
5566 }
5567
5568 /**
5569  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5570  * @inode: inode we're writing to
5571  * @num_bytes: the number of bytes we want to allocate
5572  *
5573  * This will do the following things
5574  *
5575  * o reserve space in the data space info for num_bytes
5576  * o reserve space in the metadata space info based on number of outstanding
5577  *   extents and how much csums will be needed
5578  * o add to the inodes ->delalloc_bytes
5579  * o add it to the fs_info's delalloc inodes list.
5580  *
5581  * This will return 0 for success and -ENOSPC if there is no space left.
5582  */
5583 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5584 {
5585         int ret;
5586
5587         ret = btrfs_check_data_free_space(inode, num_bytes, num_bytes);
5588         if (ret)
5589                 return ret;
5590
5591         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5592         if (ret) {
5593                 btrfs_free_reserved_data_space(inode, num_bytes);
5594                 return ret;
5595         }
5596
5597         return 0;
5598 }
5599
5600 /**
5601  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5602  * @inode: inode we're releasing space for
5603  * @num_bytes: the number of bytes we want to free up
5604  *
5605  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5606  * called in the case that we don't need the metadata AND data reservations
5607  * anymore.  So if there is an error or we insert an inline extent.
5608  *
5609  * This function will release the metadata space that was not used and will
5610  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5611  * list if there are no delalloc bytes left.
5612  */
5613 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5614 {
5615         btrfs_delalloc_release_metadata(inode, num_bytes);
5616         btrfs_free_reserved_data_space(inode, num_bytes);
5617 }
5618
5619 static int update_block_group(struct btrfs_trans_handle *trans,
5620                               struct btrfs_root *root, u64 bytenr,
5621                               u64 num_bytes, int alloc)
5622 {
5623         struct btrfs_block_group_cache *cache = NULL;
5624         struct btrfs_fs_info *info = root->fs_info;
5625         u64 total = num_bytes;
5626         u64 old_val;
5627         u64 byte_in_group;
5628         int factor;
5629
5630         /* block accounting for super block */
5631         spin_lock(&info->delalloc_root_lock);
5632         old_val = btrfs_super_bytes_used(info->super_copy);
5633         if (alloc)
5634                 old_val += num_bytes;
5635         else
5636                 old_val -= num_bytes;
5637         btrfs_set_super_bytes_used(info->super_copy, old_val);
5638         spin_unlock(&info->delalloc_root_lock);
5639
5640         while (total) {
5641                 cache = btrfs_lookup_block_group(info, bytenr);
5642                 if (!cache)
5643                         return -ENOENT;
5644                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5645                                     BTRFS_BLOCK_GROUP_RAID1 |
5646                                     BTRFS_BLOCK_GROUP_RAID10))
5647                         factor = 2;
5648                 else
5649                         factor = 1;
5650                 /*
5651                  * If this block group has free space cache written out, we
5652                  * need to make sure to load it if we are removing space.  This
5653                  * is because we need the unpinning stage to actually add the
5654                  * space back to the block group, otherwise we will leak space.
5655                  */
5656                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5657                         cache_block_group(cache, 1);
5658
5659                 byte_in_group = bytenr - cache->key.objectid;
5660                 WARN_ON(byte_in_group > cache->key.offset);
5661
5662                 spin_lock(&cache->space_info->lock);
5663                 spin_lock(&cache->lock);
5664
5665                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5666                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5667                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5668
5669                 old_val = btrfs_block_group_used(&cache->item);
5670                 num_bytes = min(total, cache->key.offset - byte_in_group);
5671                 if (alloc) {
5672                         old_val += num_bytes;
5673                         btrfs_set_block_group_used(&cache->item, old_val);
5674                         cache->reserved -= num_bytes;
5675                         cache->space_info->bytes_reserved -= num_bytes;
5676                         cache->space_info->bytes_used += num_bytes;
5677                         cache->space_info->disk_used += num_bytes * factor;
5678                         spin_unlock(&cache->lock);
5679                         spin_unlock(&cache->space_info->lock);
5680                 } else {
5681                         old_val -= num_bytes;
5682                         btrfs_set_block_group_used(&cache->item, old_val);
5683                         cache->pinned += num_bytes;
5684                         cache->space_info->bytes_pinned += num_bytes;
5685                         cache->space_info->bytes_used -= num_bytes;
5686                         cache->space_info->disk_used -= num_bytes * factor;
5687                         spin_unlock(&cache->lock);
5688                         spin_unlock(&cache->space_info->lock);
5689
5690                         set_extent_dirty(info->pinned_extents,
5691                                          bytenr, bytenr + num_bytes - 1,
5692                                          GFP_NOFS | __GFP_NOFAIL);
5693                         /*
5694                          * No longer have used bytes in this block group, queue
5695                          * it for deletion.
5696                          */
5697                         if (old_val == 0) {
5698                                 spin_lock(&info->unused_bgs_lock);
5699                                 if (list_empty(&cache->bg_list)) {
5700                                         btrfs_get_block_group(cache);
5701                                         list_add_tail(&cache->bg_list,
5702                                                       &info->unused_bgs);
5703                                 }
5704                                 spin_unlock(&info->unused_bgs_lock);
5705                         }
5706                 }
5707
5708                 spin_lock(&trans->transaction->dirty_bgs_lock);
5709                 if (list_empty(&cache->dirty_list)) {
5710                         list_add_tail(&cache->dirty_list,
5711                                       &trans->transaction->dirty_bgs);
5712                                 trans->transaction->num_dirty_bgs++;
5713                         btrfs_get_block_group(cache);
5714                 }
5715                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5716
5717                 btrfs_put_block_group(cache);
5718                 total -= num_bytes;
5719                 bytenr += num_bytes;
5720         }
5721         return 0;
5722 }
5723
5724 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5725 {
5726         struct btrfs_block_group_cache *cache;
5727         u64 bytenr;
5728
5729         spin_lock(&root->fs_info->block_group_cache_lock);
5730         bytenr = root->fs_info->first_logical_byte;
5731         spin_unlock(&root->fs_info->block_group_cache_lock);
5732
5733         if (bytenr < (u64)-1)
5734                 return bytenr;
5735
5736         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5737         if (!cache)
5738                 return 0;
5739
5740         bytenr = cache->key.objectid;
5741         btrfs_put_block_group(cache);
5742
5743         return bytenr;
5744 }
5745
5746 static int pin_down_extent(struct btrfs_root *root,
5747                            struct btrfs_block_group_cache *cache,
5748                            u64 bytenr, u64 num_bytes, int reserved)
5749 {
5750         spin_lock(&cache->space_info->lock);
5751         spin_lock(&cache->lock);
5752         cache->pinned += num_bytes;
5753         cache->space_info->bytes_pinned += num_bytes;
5754         if (reserved) {
5755                 cache->reserved -= num_bytes;
5756                 cache->space_info->bytes_reserved -= num_bytes;
5757         }
5758         spin_unlock(&cache->lock);
5759         spin_unlock(&cache->space_info->lock);
5760
5761         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5762                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5763         if (reserved)
5764                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5765         return 0;
5766 }
5767
5768 /*
5769  * this function must be called within transaction
5770  */
5771 int btrfs_pin_extent(struct btrfs_root *root,
5772                      u64 bytenr, u64 num_bytes, int reserved)
5773 {
5774         struct btrfs_block_group_cache *cache;
5775
5776         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5777         BUG_ON(!cache); /* Logic error */
5778
5779         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5780
5781         btrfs_put_block_group(cache);
5782         return 0;
5783 }
5784
5785 /*
5786  * this function must be called within transaction
5787  */
5788 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5789                                     u64 bytenr, u64 num_bytes)
5790 {
5791         struct btrfs_block_group_cache *cache;
5792         int ret;
5793
5794         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5795         if (!cache)
5796                 return -EINVAL;
5797
5798         /*
5799          * pull in the free space cache (if any) so that our pin
5800          * removes the free space from the cache.  We have load_only set
5801          * to one because the slow code to read in the free extents does check
5802          * the pinned extents.
5803          */
5804         cache_block_group(cache, 1);
5805
5806         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5807
5808         /* remove us from the free space cache (if we're there at all) */
5809         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5810         btrfs_put_block_group(cache);
5811         return ret;
5812 }
5813
5814 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5815 {
5816         int ret;
5817         struct btrfs_block_group_cache *block_group;
5818         struct btrfs_caching_control *caching_ctl;
5819
5820         block_group = btrfs_lookup_block_group(root->fs_info, start);
5821         if (!block_group)
5822                 return -EINVAL;
5823
5824         cache_block_group(block_group, 0);
5825         caching_ctl = get_caching_control(block_group);
5826
5827         if (!caching_ctl) {
5828                 /* Logic error */
5829                 BUG_ON(!block_group_cache_done(block_group));
5830                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5831         } else {
5832                 mutex_lock(&caching_ctl->mutex);
5833
5834                 if (start >= caching_ctl->progress) {
5835                         ret = add_excluded_extent(root, start, num_bytes);
5836                 } else if (start + num_bytes <= caching_ctl->progress) {
5837                         ret = btrfs_remove_free_space(block_group,
5838                                                       start, num_bytes);
5839                 } else {
5840                         num_bytes = caching_ctl->progress - start;
5841                         ret = btrfs_remove_free_space(block_group,
5842                                                       start, num_bytes);
5843                         if (ret)
5844                                 goto out_lock;
5845
5846                         num_bytes = (start + num_bytes) -
5847                                 caching_ctl->progress;
5848                         start = caching_ctl->progress;
5849                         ret = add_excluded_extent(root, start, num_bytes);
5850                 }
5851 out_lock:
5852                 mutex_unlock(&caching_ctl->mutex);
5853                 put_caching_control(caching_ctl);
5854         }
5855         btrfs_put_block_group(block_group);
5856         return ret;
5857 }
5858
5859 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5860                                  struct extent_buffer *eb)
5861 {
5862         struct btrfs_file_extent_item *item;
5863         struct btrfs_key key;
5864         int found_type;
5865         int i;
5866
5867         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5868                 return 0;
5869
5870         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5871                 btrfs_item_key_to_cpu(eb, &key, i);
5872                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5873                         continue;
5874                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5875                 found_type = btrfs_file_extent_type(eb, item);
5876                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5877                         continue;
5878                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5879                         continue;
5880                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5881                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5882                 __exclude_logged_extent(log, key.objectid, key.offset);
5883         }
5884
5885         return 0;
5886 }
5887
5888 /**
5889  * btrfs_update_reserved_bytes - update the block_group and space info counters
5890  * @cache:      The cache we are manipulating
5891  * @num_bytes:  The number of bytes in question
5892  * @reserve:    One of the reservation enums
5893  * @delalloc:   The blocks are allocated for the delalloc write
5894  *
5895  * This is called by the allocator when it reserves space, or by somebody who is
5896  * freeing space that was never actually used on disk.  For example if you
5897  * reserve some space for a new leaf in transaction A and before transaction A
5898  * commits you free that leaf, you call this with reserve set to 0 in order to
5899  * clear the reservation.
5900  *
5901  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5902  * ENOSPC accounting.  For data we handle the reservation through clearing the
5903  * delalloc bits in the io_tree.  We have to do this since we could end up
5904  * allocating less disk space for the amount of data we have reserved in the
5905  * case of compression.
5906  *
5907  * If this is a reservation and the block group has become read only we cannot
5908  * make the reservation and return -EAGAIN, otherwise this function always
5909  * succeeds.
5910  */
5911 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5912                                        u64 num_bytes, int reserve, int delalloc)
5913 {
5914         struct btrfs_space_info *space_info = cache->space_info;
5915         int ret = 0;
5916
5917         spin_lock(&space_info->lock);
5918         spin_lock(&cache->lock);
5919         if (reserve != RESERVE_FREE) {
5920                 if (cache->ro) {
5921                         ret = -EAGAIN;
5922                 } else {
5923                         cache->reserved += num_bytes;
5924                         space_info->bytes_reserved += num_bytes;
5925                         if (reserve == RESERVE_ALLOC) {
5926                                 trace_btrfs_space_reservation(cache->fs_info,
5927                                                 "space_info", space_info->flags,
5928                                                 num_bytes, 0);
5929                                 space_info->bytes_may_use -= num_bytes;
5930                         }
5931
5932                         if (delalloc)
5933                                 cache->delalloc_bytes += num_bytes;
5934                 }
5935         } else {
5936                 if (cache->ro)
5937                         space_info->bytes_readonly += num_bytes;
5938                 cache->reserved -= num_bytes;
5939                 space_info->bytes_reserved -= num_bytes;
5940
5941                 if (delalloc)
5942                         cache->delalloc_bytes -= num_bytes;
5943         }
5944         spin_unlock(&cache->lock);
5945         spin_unlock(&space_info->lock);
5946         return ret;
5947 }
5948
5949 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5950                                 struct btrfs_root *root)
5951 {
5952         struct btrfs_fs_info *fs_info = root->fs_info;
5953         struct btrfs_caching_control *next;
5954         struct btrfs_caching_control *caching_ctl;
5955         struct btrfs_block_group_cache *cache;
5956
5957         down_write(&fs_info->commit_root_sem);
5958
5959         list_for_each_entry_safe(caching_ctl, next,
5960                                  &fs_info->caching_block_groups, list) {
5961                 cache = caching_ctl->block_group;
5962                 if (block_group_cache_done(cache)) {
5963                         cache->last_byte_to_unpin = (u64)-1;
5964                         list_del_init(&caching_ctl->list);
5965                         put_caching_control(caching_ctl);
5966                 } else {
5967                         cache->last_byte_to_unpin = caching_ctl->progress;
5968                 }
5969         }
5970
5971         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5972                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5973         else
5974                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5975
5976         up_write(&fs_info->commit_root_sem);
5977
5978         update_global_block_rsv(fs_info);
5979 }
5980
5981 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
5982                               const bool return_free_space)
5983 {
5984         struct btrfs_fs_info *fs_info = root->fs_info;
5985         struct btrfs_block_group_cache *cache = NULL;
5986         struct btrfs_space_info *space_info;
5987         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5988         u64 len;
5989         bool readonly;
5990
5991         while (start <= end) {
5992                 readonly = false;
5993                 if (!cache ||
5994                     start >= cache->key.objectid + cache->key.offset) {
5995                         if (cache)
5996                                 btrfs_put_block_group(cache);
5997                         cache = btrfs_lookup_block_group(fs_info, start);
5998                         BUG_ON(!cache); /* Logic error */
5999                 }
6000
6001                 len = cache->key.objectid + cache->key.offset - start;
6002                 len = min(len, end + 1 - start);
6003
6004                 if (start < cache->last_byte_to_unpin) {
6005                         len = min(len, cache->last_byte_to_unpin - start);
6006                         if (return_free_space)
6007                                 btrfs_add_free_space(cache, start, len);
6008                 }
6009
6010                 start += len;
6011                 space_info = cache->space_info;
6012
6013                 spin_lock(&space_info->lock);
6014                 spin_lock(&cache->lock);
6015                 cache->pinned -= len;
6016                 space_info->bytes_pinned -= len;
6017                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6018                 if (cache->ro) {
6019                         space_info->bytes_readonly += len;
6020                         readonly = true;
6021                 }
6022                 spin_unlock(&cache->lock);
6023                 if (!readonly && global_rsv->space_info == space_info) {
6024                         spin_lock(&global_rsv->lock);
6025                         if (!global_rsv->full) {
6026                                 len = min(len, global_rsv->size -
6027                                           global_rsv->reserved);
6028                                 global_rsv->reserved += len;
6029                                 space_info->bytes_may_use += len;
6030                                 if (global_rsv->reserved >= global_rsv->size)
6031                                         global_rsv->full = 1;
6032                         }
6033                         spin_unlock(&global_rsv->lock);
6034                 }
6035                 spin_unlock(&space_info->lock);
6036         }
6037
6038         if (cache)
6039                 btrfs_put_block_group(cache);
6040         return 0;
6041 }
6042
6043 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6044                                struct btrfs_root *root)
6045 {
6046         struct btrfs_fs_info *fs_info = root->fs_info;
6047         struct extent_io_tree *unpin;
6048         u64 start;
6049         u64 end;
6050         int ret;
6051
6052         if (trans->aborted)
6053                 return 0;
6054
6055         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6056                 unpin = &fs_info->freed_extents[1];
6057         else
6058                 unpin = &fs_info->freed_extents[0];
6059
6060         while (1) {
6061                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6062                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6063                                             EXTENT_DIRTY, NULL);
6064                 if (ret) {
6065                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6066                         break;
6067                 }
6068
6069                 if (btrfs_test_opt(root, DISCARD))
6070                         ret = btrfs_discard_extent(root, start,
6071                                                    end + 1 - start, NULL);
6072
6073                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
6074                 unpin_extent_range(root, start, end, true);
6075                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6076                 cond_resched();
6077         }
6078
6079         return 0;
6080 }
6081
6082 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6083                              u64 owner, u64 root_objectid)
6084 {
6085         struct btrfs_space_info *space_info;
6086         u64 flags;
6087
6088         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6089                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6090                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6091                 else
6092                         flags = BTRFS_BLOCK_GROUP_METADATA;
6093         } else {
6094                 flags = BTRFS_BLOCK_GROUP_DATA;
6095         }
6096
6097         space_info = __find_space_info(fs_info, flags);
6098         BUG_ON(!space_info); /* Logic bug */
6099         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6100 }
6101
6102
6103 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6104                                 struct btrfs_root *root,
6105                                 struct btrfs_delayed_ref_node *node, u64 parent,
6106                                 u64 root_objectid, u64 owner_objectid,
6107                                 u64 owner_offset, int refs_to_drop,
6108                                 struct btrfs_delayed_extent_op *extent_op)
6109 {
6110         struct btrfs_key key;
6111         struct btrfs_path *path;
6112         struct btrfs_fs_info *info = root->fs_info;
6113         struct btrfs_root *extent_root = info->extent_root;
6114         struct extent_buffer *leaf;
6115         struct btrfs_extent_item *ei;
6116         struct btrfs_extent_inline_ref *iref;
6117         int ret;
6118         int is_data;
6119         int extent_slot = 0;
6120         int found_extent = 0;
6121         int num_to_del = 1;
6122         int no_quota = node->no_quota;
6123         u32 item_size;
6124         u64 refs;
6125         u64 bytenr = node->bytenr;
6126         u64 num_bytes = node->num_bytes;
6127         int last_ref = 0;
6128         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6129                                                  SKINNY_METADATA);
6130
6131         if (!info->quota_enabled || !is_fstree(root_objectid))
6132                 no_quota = 1;
6133
6134         path = btrfs_alloc_path();
6135         if (!path)
6136                 return -ENOMEM;
6137
6138         path->reada = 1;
6139         path->leave_spinning = 1;
6140
6141         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6142         BUG_ON(!is_data && refs_to_drop != 1);
6143
6144         if (is_data)
6145                 skinny_metadata = 0;
6146
6147         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6148                                     bytenr, num_bytes, parent,
6149                                     root_objectid, owner_objectid,
6150                                     owner_offset);
6151         if (ret == 0) {
6152                 extent_slot = path->slots[0];
6153                 while (extent_slot >= 0) {
6154                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6155                                               extent_slot);
6156                         if (key.objectid != bytenr)
6157                                 break;
6158                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6159                             key.offset == num_bytes) {
6160                                 found_extent = 1;
6161                                 break;
6162                         }
6163                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6164                             key.offset == owner_objectid) {
6165                                 found_extent = 1;
6166                                 break;
6167                         }
6168                         if (path->slots[0] - extent_slot > 5)
6169                                 break;
6170                         extent_slot--;
6171                 }
6172 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6173                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6174                 if (found_extent && item_size < sizeof(*ei))
6175                         found_extent = 0;
6176 #endif
6177                 if (!found_extent) {
6178                         BUG_ON(iref);
6179                         ret = remove_extent_backref(trans, extent_root, path,
6180                                                     NULL, refs_to_drop,
6181                                                     is_data, &last_ref);
6182                         if (ret) {
6183                                 btrfs_abort_transaction(trans, extent_root, ret);
6184                                 goto out;
6185                         }
6186                         btrfs_release_path(path);
6187                         path->leave_spinning = 1;
6188
6189                         key.objectid = bytenr;
6190                         key.type = BTRFS_EXTENT_ITEM_KEY;
6191                         key.offset = num_bytes;
6192
6193                         if (!is_data && skinny_metadata) {
6194                                 key.type = BTRFS_METADATA_ITEM_KEY;
6195                                 key.offset = owner_objectid;
6196                         }
6197
6198                         ret = btrfs_search_slot(trans, extent_root,
6199                                                 &key, path, -1, 1);
6200                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6201                                 /*
6202                                  * Couldn't find our skinny metadata item,
6203                                  * see if we have ye olde extent item.
6204                                  */
6205                                 path->slots[0]--;
6206                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6207                                                       path->slots[0]);
6208                                 if (key.objectid == bytenr &&
6209                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6210                                     key.offset == num_bytes)
6211                                         ret = 0;
6212                         }
6213
6214                         if (ret > 0 && skinny_metadata) {
6215                                 skinny_metadata = false;
6216                                 key.objectid = bytenr;
6217                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6218                                 key.offset = num_bytes;
6219                                 btrfs_release_path(path);
6220                                 ret = btrfs_search_slot(trans, extent_root,
6221                                                         &key, path, -1, 1);
6222                         }
6223
6224                         if (ret) {
6225                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6226                                         ret, bytenr);
6227                                 if (ret > 0)
6228                                         btrfs_print_leaf(extent_root,
6229                                                          path->nodes[0]);
6230                         }
6231                         if (ret < 0) {
6232                                 btrfs_abort_transaction(trans, extent_root, ret);
6233                                 goto out;
6234                         }
6235                         extent_slot = path->slots[0];
6236                 }
6237         } else if (WARN_ON(ret == -ENOENT)) {
6238                 btrfs_print_leaf(extent_root, path->nodes[0]);
6239                 btrfs_err(info,
6240                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6241                         bytenr, parent, root_objectid, owner_objectid,
6242                         owner_offset);
6243                 btrfs_abort_transaction(trans, extent_root, ret);
6244                 goto out;
6245         } else {
6246                 btrfs_abort_transaction(trans, extent_root, ret);
6247                 goto out;
6248         }
6249
6250         leaf = path->nodes[0];
6251         item_size = btrfs_item_size_nr(leaf, extent_slot);
6252 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6253         if (item_size < sizeof(*ei)) {
6254                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6255                 ret = convert_extent_item_v0(trans, extent_root, path,
6256                                              owner_objectid, 0);
6257                 if (ret < 0) {
6258                         btrfs_abort_transaction(trans, extent_root, ret);
6259                         goto out;
6260                 }
6261
6262                 btrfs_release_path(path);
6263                 path->leave_spinning = 1;
6264
6265                 key.objectid = bytenr;
6266                 key.type = BTRFS_EXTENT_ITEM_KEY;
6267                 key.offset = num_bytes;
6268
6269                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6270                                         -1, 1);
6271                 if (ret) {
6272                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6273                                 ret, bytenr);
6274                         btrfs_print_leaf(extent_root, path->nodes[0]);
6275                 }
6276                 if (ret < 0) {
6277                         btrfs_abort_transaction(trans, extent_root, ret);
6278                         goto out;
6279                 }
6280
6281                 extent_slot = path->slots[0];
6282                 leaf = path->nodes[0];
6283                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6284         }
6285 #endif
6286         BUG_ON(item_size < sizeof(*ei));
6287         ei = btrfs_item_ptr(leaf, extent_slot,
6288                             struct btrfs_extent_item);
6289         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6290             key.type == BTRFS_EXTENT_ITEM_KEY) {
6291                 struct btrfs_tree_block_info *bi;
6292                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6293                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6294                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6295         }
6296
6297         refs = btrfs_extent_refs(leaf, ei);
6298         if (refs < refs_to_drop) {
6299                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6300                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6301                 ret = -EINVAL;
6302                 btrfs_abort_transaction(trans, extent_root, ret);
6303                 goto out;
6304         }
6305         refs -= refs_to_drop;
6306
6307         if (refs > 0) {
6308                 if (extent_op)
6309                         __run_delayed_extent_op(extent_op, leaf, ei);
6310                 /*
6311                  * In the case of inline back ref, reference count will
6312                  * be updated by remove_extent_backref
6313                  */
6314                 if (iref) {
6315                         BUG_ON(!found_extent);
6316                 } else {
6317                         btrfs_set_extent_refs(leaf, ei, refs);
6318                         btrfs_mark_buffer_dirty(leaf);
6319                 }
6320                 if (found_extent) {
6321                         ret = remove_extent_backref(trans, extent_root, path,
6322                                                     iref, refs_to_drop,
6323                                                     is_data, &last_ref);
6324                         if (ret) {
6325                                 btrfs_abort_transaction(trans, extent_root, ret);
6326                                 goto out;
6327                         }
6328                 }
6329                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6330                                  root_objectid);
6331         } else {
6332                 if (found_extent) {
6333                         BUG_ON(is_data && refs_to_drop !=
6334                                extent_data_ref_count(root, path, iref));
6335                         if (iref) {
6336                                 BUG_ON(path->slots[0] != extent_slot);
6337                         } else {
6338                                 BUG_ON(path->slots[0] != extent_slot + 1);
6339                                 path->slots[0] = extent_slot;
6340                                 num_to_del = 2;
6341                         }
6342                 }
6343
6344                 last_ref = 1;
6345                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6346                                       num_to_del);
6347                 if (ret) {
6348                         btrfs_abort_transaction(trans, extent_root, ret);
6349                         goto out;
6350                 }
6351                 btrfs_release_path(path);
6352
6353                 if (is_data) {
6354                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6355                         if (ret) {
6356                                 btrfs_abort_transaction(trans, extent_root, ret);
6357                                 goto out;
6358                         }
6359                 }
6360
6361                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6362                 if (ret) {
6363                         btrfs_abort_transaction(trans, extent_root, ret);
6364                         goto out;
6365                 }
6366         }
6367         btrfs_release_path(path);
6368
6369 out:
6370         btrfs_free_path(path);
6371         return ret;
6372 }
6373
6374 /*
6375  * when we free an block, it is possible (and likely) that we free the last
6376  * delayed ref for that extent as well.  This searches the delayed ref tree for
6377  * a given extent, and if there are no other delayed refs to be processed, it
6378  * removes it from the tree.
6379  */
6380 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6381                                       struct btrfs_root *root, u64 bytenr)
6382 {
6383         struct btrfs_delayed_ref_head *head;
6384         struct btrfs_delayed_ref_root *delayed_refs;
6385         int ret = 0;
6386
6387         delayed_refs = &trans->transaction->delayed_refs;
6388         spin_lock(&delayed_refs->lock);
6389         head = btrfs_find_delayed_ref_head(trans, bytenr);
6390         if (!head)
6391                 goto out_delayed_unlock;
6392
6393         spin_lock(&head->lock);
6394         if (!list_empty(&head->ref_list))
6395                 goto out;
6396
6397         if (head->extent_op) {
6398                 if (!head->must_insert_reserved)
6399                         goto out;
6400                 btrfs_free_delayed_extent_op(head->extent_op);
6401                 head->extent_op = NULL;
6402         }
6403
6404         /*
6405          * waiting for the lock here would deadlock.  If someone else has it
6406          * locked they are already in the process of dropping it anyway
6407          */
6408         if (!mutex_trylock(&head->mutex))
6409                 goto out;
6410
6411         /*
6412          * at this point we have a head with no other entries.  Go
6413          * ahead and process it.
6414          */
6415         head->node.in_tree = 0;
6416         rb_erase(&head->href_node, &delayed_refs->href_root);
6417
6418         atomic_dec(&delayed_refs->num_entries);
6419
6420         /*
6421          * we don't take a ref on the node because we're removing it from the
6422          * tree, so we just steal the ref the tree was holding.
6423          */
6424         delayed_refs->num_heads--;
6425         if (head->processing == 0)
6426                 delayed_refs->num_heads_ready--;
6427         head->processing = 0;
6428         spin_unlock(&head->lock);
6429         spin_unlock(&delayed_refs->lock);
6430
6431         BUG_ON(head->extent_op);
6432         if (head->must_insert_reserved)
6433                 ret = 1;
6434
6435         mutex_unlock(&head->mutex);
6436         btrfs_put_delayed_ref(&head->node);
6437         return ret;
6438 out:
6439         spin_unlock(&head->lock);
6440
6441 out_delayed_unlock:
6442         spin_unlock(&delayed_refs->lock);
6443         return 0;
6444 }
6445
6446 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6447                            struct btrfs_root *root,
6448                            struct extent_buffer *buf,
6449                            u64 parent, int last_ref)
6450 {
6451         int pin = 1;
6452         int ret;
6453
6454         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6455                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6456                                         buf->start, buf->len,
6457                                         parent, root->root_key.objectid,
6458                                         btrfs_header_level(buf),
6459                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6460                 BUG_ON(ret); /* -ENOMEM */
6461         }
6462
6463         if (!last_ref)
6464                 return;
6465
6466         if (btrfs_header_generation(buf) == trans->transid) {
6467                 struct btrfs_block_group_cache *cache;
6468
6469                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6470                         ret = check_ref_cleanup(trans, root, buf->start);
6471                         if (!ret)
6472                                 goto out;
6473                 }
6474
6475                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6476
6477                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6478                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6479                         btrfs_put_block_group(cache);
6480                         goto out;
6481                 }
6482
6483                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6484
6485                 btrfs_add_free_space(cache, buf->start, buf->len);
6486                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6487                 btrfs_put_block_group(cache);
6488                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6489                 pin = 0;
6490         }
6491 out:
6492         if (pin)
6493                 add_pinned_bytes(root->fs_info, buf->len,
6494                                  btrfs_header_level(buf),
6495                                  root->root_key.objectid);
6496
6497         /*
6498          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6499          * anymore.
6500          */
6501         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6502 }
6503
6504 /* Can return -ENOMEM */
6505 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6506                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6507                       u64 owner, u64 offset, int no_quota)
6508 {
6509         int ret;
6510         struct btrfs_fs_info *fs_info = root->fs_info;
6511
6512         if (btrfs_test_is_dummy_root(root))
6513                 return 0;
6514
6515         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6516
6517         /*
6518          * tree log blocks never actually go into the extent allocation
6519          * tree, just update pinning info and exit early.
6520          */
6521         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6522                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6523                 /* unlocks the pinned mutex */
6524                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6525                 ret = 0;
6526         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6527                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6528                                         num_bytes,
6529                                         parent, root_objectid, (int)owner,
6530                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6531         } else {
6532                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6533                                                 num_bytes,
6534                                                 parent, root_objectid, owner,
6535                                                 offset, BTRFS_DROP_DELAYED_REF,
6536                                                 NULL, no_quota);
6537         }
6538         return ret;
6539 }
6540
6541 /*
6542  * when we wait for progress in the block group caching, its because
6543  * our allocation attempt failed at least once.  So, we must sleep
6544  * and let some progress happen before we try again.
6545  *
6546  * This function will sleep at least once waiting for new free space to
6547  * show up, and then it will check the block group free space numbers
6548  * for our min num_bytes.  Another option is to have it go ahead
6549  * and look in the rbtree for a free extent of a given size, but this
6550  * is a good start.
6551  *
6552  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6553  * any of the information in this block group.
6554  */
6555 static noinline void
6556 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6557                                 u64 num_bytes)
6558 {
6559         struct btrfs_caching_control *caching_ctl;
6560
6561         caching_ctl = get_caching_control(cache);
6562         if (!caching_ctl)
6563                 return;
6564
6565         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6566                    (cache->free_space_ctl->free_space >= num_bytes));
6567
6568         put_caching_control(caching_ctl);
6569 }
6570
6571 static noinline int
6572 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6573 {
6574         struct btrfs_caching_control *caching_ctl;
6575         int ret = 0;
6576
6577         caching_ctl = get_caching_control(cache);
6578         if (!caching_ctl)
6579                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6580
6581         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6582         if (cache->cached == BTRFS_CACHE_ERROR)
6583                 ret = -EIO;
6584         put_caching_control(caching_ctl);
6585         return ret;
6586 }
6587
6588 int __get_raid_index(u64 flags)
6589 {
6590         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6591                 return BTRFS_RAID_RAID10;
6592         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6593                 return BTRFS_RAID_RAID1;
6594         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6595                 return BTRFS_RAID_DUP;
6596         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6597                 return BTRFS_RAID_RAID0;
6598         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6599                 return BTRFS_RAID_RAID5;
6600         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6601                 return BTRFS_RAID_RAID6;
6602
6603         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6604 }
6605
6606 int get_block_group_index(struct btrfs_block_group_cache *cache)
6607 {
6608         return __get_raid_index(cache->flags);
6609 }
6610
6611 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6612         [BTRFS_RAID_RAID10]     = "raid10",
6613         [BTRFS_RAID_RAID1]      = "raid1",
6614         [BTRFS_RAID_DUP]        = "dup",
6615         [BTRFS_RAID_RAID0]      = "raid0",
6616         [BTRFS_RAID_SINGLE]     = "single",
6617         [BTRFS_RAID_RAID5]      = "raid5",
6618         [BTRFS_RAID_RAID6]      = "raid6",
6619 };
6620
6621 static const char *get_raid_name(enum btrfs_raid_types type)
6622 {
6623         if (type >= BTRFS_NR_RAID_TYPES)
6624                 return NULL;
6625
6626         return btrfs_raid_type_names[type];
6627 }
6628
6629 enum btrfs_loop_type {
6630         LOOP_CACHING_NOWAIT = 0,
6631         LOOP_CACHING_WAIT = 1,
6632         LOOP_ALLOC_CHUNK = 2,
6633         LOOP_NO_EMPTY_SIZE = 3,
6634 };
6635
6636 static inline void
6637 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6638                        int delalloc)
6639 {
6640         if (delalloc)
6641                 down_read(&cache->data_rwsem);
6642 }
6643
6644 static inline void
6645 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6646                        int delalloc)
6647 {
6648         btrfs_get_block_group(cache);
6649         if (delalloc)
6650                 down_read(&cache->data_rwsem);
6651 }
6652
6653 static struct btrfs_block_group_cache *
6654 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6655                    struct btrfs_free_cluster *cluster,
6656                    int delalloc)
6657 {
6658         struct btrfs_block_group_cache *used_bg;
6659         bool locked = false;
6660 again:
6661         spin_lock(&cluster->refill_lock);
6662         if (locked) {
6663                 if (used_bg == cluster->block_group)
6664                         return used_bg;
6665
6666                 up_read(&used_bg->data_rwsem);
6667                 btrfs_put_block_group(used_bg);
6668         }
6669
6670         used_bg = cluster->block_group;
6671         if (!used_bg)
6672                 return NULL;
6673
6674         if (used_bg == block_group)
6675                 return used_bg;
6676
6677         btrfs_get_block_group(used_bg);
6678
6679         if (!delalloc)
6680                 return used_bg;
6681
6682         if (down_read_trylock(&used_bg->data_rwsem))
6683                 return used_bg;
6684
6685         spin_unlock(&cluster->refill_lock);
6686         down_read(&used_bg->data_rwsem);
6687         locked = true;
6688         goto again;
6689 }
6690
6691 static inline void
6692 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6693                          int delalloc)
6694 {
6695         if (delalloc)
6696                 up_read(&cache->data_rwsem);
6697         btrfs_put_block_group(cache);
6698 }
6699
6700 /*
6701  * walks the btree of allocated extents and find a hole of a given size.
6702  * The key ins is changed to record the hole:
6703  * ins->objectid == start position
6704  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6705  * ins->offset == the size of the hole.
6706  * Any available blocks before search_start are skipped.
6707  *
6708  * If there is no suitable free space, we will record the max size of
6709  * the free space extent currently.
6710  */
6711 static noinline int find_free_extent(struct btrfs_root *orig_root,
6712                                      u64 num_bytes, u64 empty_size,
6713                                      u64 hint_byte, struct btrfs_key *ins,
6714                                      u64 flags, int delalloc)
6715 {
6716         int ret = 0;
6717         struct btrfs_root *root = orig_root->fs_info->extent_root;
6718         struct btrfs_free_cluster *last_ptr = NULL;
6719         struct btrfs_block_group_cache *block_group = NULL;
6720         u64 search_start = 0;
6721         u64 max_extent_size = 0;
6722         int empty_cluster = 2 * 1024 * 1024;
6723         struct btrfs_space_info *space_info;
6724         int loop = 0;
6725         int index = __get_raid_index(flags);
6726         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6727                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6728         bool failed_cluster_refill = false;
6729         bool failed_alloc = false;
6730         bool use_cluster = true;
6731         bool have_caching_bg = false;
6732
6733         WARN_ON(num_bytes < root->sectorsize);
6734         ins->type = BTRFS_EXTENT_ITEM_KEY;
6735         ins->objectid = 0;
6736         ins->offset = 0;
6737
6738         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6739
6740         space_info = __find_space_info(root->fs_info, flags);
6741         if (!space_info) {
6742                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6743                 return -ENOSPC;
6744         }
6745
6746         /*
6747          * If the space info is for both data and metadata it means we have a
6748          * small filesystem and we can't use the clustering stuff.
6749          */
6750         if (btrfs_mixed_space_info(space_info))
6751                 use_cluster = false;
6752
6753         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6754                 last_ptr = &root->fs_info->meta_alloc_cluster;
6755                 if (!btrfs_test_opt(root, SSD))
6756                         empty_cluster = 64 * 1024;
6757         }
6758
6759         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6760             btrfs_test_opt(root, SSD)) {
6761                 last_ptr = &root->fs_info->data_alloc_cluster;
6762         }
6763
6764         if (last_ptr) {
6765                 spin_lock(&last_ptr->lock);
6766                 if (last_ptr->block_group)
6767                         hint_byte = last_ptr->window_start;
6768                 spin_unlock(&last_ptr->lock);
6769         }
6770
6771         search_start = max(search_start, first_logical_byte(root, 0));
6772         search_start = max(search_start, hint_byte);
6773
6774         if (!last_ptr)
6775                 empty_cluster = 0;
6776
6777         if (search_start == hint_byte) {
6778                 block_group = btrfs_lookup_block_group(root->fs_info,
6779                                                        search_start);
6780                 /*
6781                  * we don't want to use the block group if it doesn't match our
6782                  * allocation bits, or if its not cached.
6783                  *
6784                  * However if we are re-searching with an ideal block group
6785                  * picked out then we don't care that the block group is cached.
6786                  */
6787                 if (block_group && block_group_bits(block_group, flags) &&
6788                     block_group->cached != BTRFS_CACHE_NO) {
6789                         down_read(&space_info->groups_sem);
6790                         if (list_empty(&block_group->list) ||
6791                             block_group->ro) {
6792                                 /*
6793                                  * someone is removing this block group,
6794                                  * we can't jump into the have_block_group
6795                                  * target because our list pointers are not
6796                                  * valid
6797                                  */
6798                                 btrfs_put_block_group(block_group);
6799                                 up_read(&space_info->groups_sem);
6800                         } else {
6801                                 index = get_block_group_index(block_group);
6802                                 btrfs_lock_block_group(block_group, delalloc);
6803                                 goto have_block_group;
6804                         }
6805                 } else if (block_group) {
6806                         btrfs_put_block_group(block_group);
6807                 }
6808         }
6809 search:
6810         have_caching_bg = false;
6811         down_read(&space_info->groups_sem);
6812         list_for_each_entry(block_group, &space_info->block_groups[index],
6813                             list) {
6814                 u64 offset;
6815                 int cached;
6816
6817                 btrfs_grab_block_group(block_group, delalloc);
6818                 search_start = block_group->key.objectid;
6819
6820                 /*
6821                  * this can happen if we end up cycling through all the
6822                  * raid types, but we want to make sure we only allocate
6823                  * for the proper type.
6824                  */
6825                 if (!block_group_bits(block_group, flags)) {
6826                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6827                                 BTRFS_BLOCK_GROUP_RAID1 |
6828                                 BTRFS_BLOCK_GROUP_RAID5 |
6829                                 BTRFS_BLOCK_GROUP_RAID6 |
6830                                 BTRFS_BLOCK_GROUP_RAID10;
6831
6832                         /*
6833                          * if they asked for extra copies and this block group
6834                          * doesn't provide them, bail.  This does allow us to
6835                          * fill raid0 from raid1.
6836                          */
6837                         if ((flags & extra) && !(block_group->flags & extra))
6838                                 goto loop;
6839                 }
6840
6841 have_block_group:
6842                 cached = block_group_cache_done(block_group);
6843                 if (unlikely(!cached)) {
6844                         ret = cache_block_group(block_group, 0);
6845                         BUG_ON(ret < 0);
6846                         ret = 0;
6847                 }
6848
6849                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6850                         goto loop;
6851                 if (unlikely(block_group->ro))
6852                         goto loop;
6853
6854                 /*
6855                  * Ok we want to try and use the cluster allocator, so
6856                  * lets look there
6857                  */
6858                 if (last_ptr) {
6859                         struct btrfs_block_group_cache *used_block_group;
6860                         unsigned long aligned_cluster;
6861                         /*
6862                          * the refill lock keeps out other
6863                          * people trying to start a new cluster
6864                          */
6865                         used_block_group = btrfs_lock_cluster(block_group,
6866                                                               last_ptr,
6867                                                               delalloc);
6868                         if (!used_block_group)
6869                                 goto refill_cluster;
6870
6871                         if (used_block_group != block_group &&
6872                             (used_block_group->ro ||
6873                              !block_group_bits(used_block_group, flags)))
6874                                 goto release_cluster;
6875
6876                         offset = btrfs_alloc_from_cluster(used_block_group,
6877                                                 last_ptr,
6878                                                 num_bytes,
6879                                                 used_block_group->key.objectid,
6880                                                 &max_extent_size);
6881                         if (offset) {
6882                                 /* we have a block, we're done */
6883                                 spin_unlock(&last_ptr->refill_lock);
6884                                 trace_btrfs_reserve_extent_cluster(root,
6885                                                 used_block_group,
6886                                                 search_start, num_bytes);
6887                                 if (used_block_group != block_group) {
6888                                         btrfs_release_block_group(block_group,
6889                                                                   delalloc);
6890                                         block_group = used_block_group;
6891                                 }
6892                                 goto checks;
6893                         }
6894
6895                         WARN_ON(last_ptr->block_group != used_block_group);
6896 release_cluster:
6897                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6898                          * set up a new clusters, so lets just skip it
6899                          * and let the allocator find whatever block
6900                          * it can find.  If we reach this point, we
6901                          * will have tried the cluster allocator
6902                          * plenty of times and not have found
6903                          * anything, so we are likely way too
6904                          * fragmented for the clustering stuff to find
6905                          * anything.
6906                          *
6907                          * However, if the cluster is taken from the
6908                          * current block group, release the cluster
6909                          * first, so that we stand a better chance of
6910                          * succeeding in the unclustered
6911                          * allocation.  */
6912                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6913                             used_block_group != block_group) {
6914                                 spin_unlock(&last_ptr->refill_lock);
6915                                 btrfs_release_block_group(used_block_group,
6916                                                           delalloc);
6917                                 goto unclustered_alloc;
6918                         }
6919
6920                         /*
6921                          * this cluster didn't work out, free it and
6922                          * start over
6923                          */
6924                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6925
6926                         if (used_block_group != block_group)
6927                                 btrfs_release_block_group(used_block_group,
6928                                                           delalloc);
6929 refill_cluster:
6930                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6931                                 spin_unlock(&last_ptr->refill_lock);
6932                                 goto unclustered_alloc;
6933                         }
6934
6935                         aligned_cluster = max_t(unsigned long,
6936                                                 empty_cluster + empty_size,
6937                                               block_group->full_stripe_len);
6938
6939                         /* allocate a cluster in this block group */
6940                         ret = btrfs_find_space_cluster(root, block_group,
6941                                                        last_ptr, search_start,
6942                                                        num_bytes,
6943                                                        aligned_cluster);
6944                         if (ret == 0) {
6945                                 /*
6946                                  * now pull our allocation out of this
6947                                  * cluster
6948                                  */
6949                                 offset = btrfs_alloc_from_cluster(block_group,
6950                                                         last_ptr,
6951                                                         num_bytes,
6952                                                         search_start,
6953                                                         &max_extent_size);
6954                                 if (offset) {
6955                                         /* we found one, proceed */
6956                                         spin_unlock(&last_ptr->refill_lock);
6957                                         trace_btrfs_reserve_extent_cluster(root,
6958                                                 block_group, search_start,
6959                                                 num_bytes);
6960                                         goto checks;
6961                                 }
6962                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6963                                    && !failed_cluster_refill) {
6964                                 spin_unlock(&last_ptr->refill_lock);
6965
6966                                 failed_cluster_refill = true;
6967                                 wait_block_group_cache_progress(block_group,
6968                                        num_bytes + empty_cluster + empty_size);
6969                                 goto have_block_group;
6970                         }
6971
6972                         /*
6973                          * at this point we either didn't find a cluster
6974                          * or we weren't able to allocate a block from our
6975                          * cluster.  Free the cluster we've been trying
6976                          * to use, and go to the next block group
6977                          */
6978                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6979                         spin_unlock(&last_ptr->refill_lock);
6980                         goto loop;
6981                 }
6982
6983 unclustered_alloc:
6984                 spin_lock(&block_group->free_space_ctl->tree_lock);
6985                 if (cached &&
6986                     block_group->free_space_ctl->free_space <
6987                     num_bytes + empty_cluster + empty_size) {
6988                         if (block_group->free_space_ctl->free_space >
6989                             max_extent_size)
6990                                 max_extent_size =
6991                                         block_group->free_space_ctl->free_space;
6992                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6993                         goto loop;
6994                 }
6995                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6996
6997                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6998                                                     num_bytes, empty_size,
6999                                                     &max_extent_size);
7000                 /*
7001                  * If we didn't find a chunk, and we haven't failed on this
7002                  * block group before, and this block group is in the middle of
7003                  * caching and we are ok with waiting, then go ahead and wait
7004                  * for progress to be made, and set failed_alloc to true.
7005                  *
7006                  * If failed_alloc is true then we've already waited on this
7007                  * block group once and should move on to the next block group.
7008                  */
7009                 if (!offset && !failed_alloc && !cached &&
7010                     loop > LOOP_CACHING_NOWAIT) {
7011                         wait_block_group_cache_progress(block_group,
7012                                                 num_bytes + empty_size);
7013                         failed_alloc = true;
7014                         goto have_block_group;
7015                 } else if (!offset) {
7016                         if (!cached)
7017                                 have_caching_bg = true;
7018                         goto loop;
7019                 }
7020 checks:
7021                 search_start = ALIGN(offset, root->stripesize);
7022
7023                 /* move on to the next group */
7024                 if (search_start + num_bytes >
7025                     block_group->key.objectid + block_group->key.offset) {
7026                         btrfs_add_free_space(block_group, offset, num_bytes);
7027                         goto loop;
7028                 }
7029
7030                 if (offset < search_start)
7031                         btrfs_add_free_space(block_group, offset,
7032                                              search_start - offset);
7033                 BUG_ON(offset > search_start);
7034
7035                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7036                                                   alloc_type, delalloc);
7037                 if (ret == -EAGAIN) {
7038                         btrfs_add_free_space(block_group, offset, num_bytes);
7039                         goto loop;
7040                 }
7041
7042                 /* we are all good, lets return */
7043                 ins->objectid = search_start;
7044                 ins->offset = num_bytes;
7045
7046                 trace_btrfs_reserve_extent(orig_root, block_group,
7047                                            search_start, num_bytes);
7048                 btrfs_release_block_group(block_group, delalloc);
7049                 break;
7050 loop:
7051                 failed_cluster_refill = false;
7052                 failed_alloc = false;
7053                 BUG_ON(index != get_block_group_index(block_group));
7054                 btrfs_release_block_group(block_group, delalloc);
7055         }
7056         up_read(&space_info->groups_sem);
7057
7058         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7059                 goto search;
7060
7061         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7062                 goto search;
7063
7064         /*
7065          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7066          *                      caching kthreads as we move along
7067          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7068          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7069          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7070          *                      again
7071          */
7072         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7073                 index = 0;
7074                 loop++;
7075                 if (loop == LOOP_ALLOC_CHUNK) {
7076                         struct btrfs_trans_handle *trans;
7077                         int exist = 0;
7078
7079                         trans = current->journal_info;
7080                         if (trans)
7081                                 exist = 1;
7082                         else
7083                                 trans = btrfs_join_transaction(root);
7084
7085                         if (IS_ERR(trans)) {
7086                                 ret = PTR_ERR(trans);
7087                                 goto out;
7088                         }
7089
7090                         ret = do_chunk_alloc(trans, root, flags,
7091                                              CHUNK_ALLOC_FORCE);
7092                         /*
7093                          * Do not bail out on ENOSPC since we
7094                          * can do more things.
7095                          */
7096                         if (ret < 0 && ret != -ENOSPC)
7097                                 btrfs_abort_transaction(trans,
7098                                                         root, ret);
7099                         else
7100                                 ret = 0;
7101                         if (!exist)
7102                                 btrfs_end_transaction(trans, root);
7103                         if (ret)
7104                                 goto out;
7105                 }
7106
7107                 if (loop == LOOP_NO_EMPTY_SIZE) {
7108                         empty_size = 0;
7109                         empty_cluster = 0;
7110                 }
7111
7112                 goto search;
7113         } else if (!ins->objectid) {
7114                 ret = -ENOSPC;
7115         } else if (ins->objectid) {
7116                 ret = 0;
7117         }
7118 out:
7119         if (ret == -ENOSPC)
7120                 ins->offset = max_extent_size;
7121         return ret;
7122 }
7123
7124 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7125                             int dump_block_groups)
7126 {
7127         struct btrfs_block_group_cache *cache;
7128         int index = 0;
7129
7130         spin_lock(&info->lock);
7131         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7132                info->flags,
7133                info->total_bytes - info->bytes_used - info->bytes_pinned -
7134                info->bytes_reserved - info->bytes_readonly,
7135                (info->full) ? "" : "not ");
7136         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7137                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7138                info->total_bytes, info->bytes_used, info->bytes_pinned,
7139                info->bytes_reserved, info->bytes_may_use,
7140                info->bytes_readonly);
7141         spin_unlock(&info->lock);
7142
7143         if (!dump_block_groups)
7144                 return;
7145
7146         down_read(&info->groups_sem);
7147 again:
7148         list_for_each_entry(cache, &info->block_groups[index], list) {
7149                 spin_lock(&cache->lock);
7150                 printk(KERN_INFO "BTRFS: "
7151                            "block group %llu has %llu bytes, "
7152                            "%llu used %llu pinned %llu reserved %s\n",
7153                        cache->key.objectid, cache->key.offset,
7154                        btrfs_block_group_used(&cache->item), cache->pinned,
7155                        cache->reserved, cache->ro ? "[readonly]" : "");
7156                 btrfs_dump_free_space(cache, bytes);
7157                 spin_unlock(&cache->lock);
7158         }
7159         if (++index < BTRFS_NR_RAID_TYPES)
7160                 goto again;
7161         up_read(&info->groups_sem);
7162 }
7163
7164 int btrfs_reserve_extent(struct btrfs_root *root,
7165                          u64 num_bytes, u64 min_alloc_size,
7166                          u64 empty_size, u64 hint_byte,
7167                          struct btrfs_key *ins, int is_data, int delalloc)
7168 {
7169         bool final_tried = false;
7170         u64 flags;
7171         int ret;
7172
7173         flags = btrfs_get_alloc_profile(root, is_data);
7174 again:
7175         WARN_ON(num_bytes < root->sectorsize);
7176         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7177                                flags, delalloc);
7178
7179         if (ret == -ENOSPC) {
7180                 if (!final_tried && ins->offset) {
7181                         num_bytes = min(num_bytes >> 1, ins->offset);
7182                         num_bytes = round_down(num_bytes, root->sectorsize);
7183                         num_bytes = max(num_bytes, min_alloc_size);
7184                         if (num_bytes == min_alloc_size)
7185                                 final_tried = true;
7186                         goto again;
7187                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7188                         struct btrfs_space_info *sinfo;
7189
7190                         sinfo = __find_space_info(root->fs_info, flags);
7191                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7192                                 flags, num_bytes);
7193                         if (sinfo)
7194                                 dump_space_info(sinfo, num_bytes, 1);
7195                 }
7196         }
7197
7198         return ret;
7199 }
7200
7201 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7202                                         u64 start, u64 len,
7203                                         int pin, int delalloc)
7204 {
7205         struct btrfs_block_group_cache *cache;
7206         int ret = 0;
7207
7208         cache = btrfs_lookup_block_group(root->fs_info, start);
7209         if (!cache) {
7210                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7211                         start);
7212                 return -ENOSPC;
7213         }
7214
7215         if (pin)
7216                 pin_down_extent(root, cache, start, len, 1);
7217         else {
7218                 if (btrfs_test_opt(root, DISCARD))
7219                         ret = btrfs_discard_extent(root, start, len, NULL);
7220                 btrfs_add_free_space(cache, start, len);
7221                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7222         }
7223
7224         btrfs_put_block_group(cache);
7225
7226         trace_btrfs_reserved_extent_free(root, start, len);
7227
7228         return ret;
7229 }
7230
7231 int btrfs_free_reserved_extent(struct btrfs_root *root,
7232                                u64 start, u64 len, int delalloc)
7233 {
7234         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7235 }
7236
7237 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7238                                        u64 start, u64 len)
7239 {
7240         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7241 }
7242
7243 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7244                                       struct btrfs_root *root,
7245                                       u64 parent, u64 root_objectid,
7246                                       u64 flags, u64 owner, u64 offset,
7247                                       struct btrfs_key *ins, int ref_mod)
7248 {
7249         int ret;
7250         struct btrfs_fs_info *fs_info = root->fs_info;
7251         struct btrfs_extent_item *extent_item;
7252         struct btrfs_extent_inline_ref *iref;
7253         struct btrfs_path *path;
7254         struct extent_buffer *leaf;
7255         int type;
7256         u32 size;
7257
7258         if (parent > 0)
7259                 type = BTRFS_SHARED_DATA_REF_KEY;
7260         else
7261                 type = BTRFS_EXTENT_DATA_REF_KEY;
7262
7263         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7264
7265         path = btrfs_alloc_path();
7266         if (!path)
7267                 return -ENOMEM;
7268
7269         path->leave_spinning = 1;
7270         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7271                                       ins, size);
7272         if (ret) {
7273                 btrfs_free_path(path);
7274                 return ret;
7275         }
7276
7277         leaf = path->nodes[0];
7278         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7279                                      struct btrfs_extent_item);
7280         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7281         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7282         btrfs_set_extent_flags(leaf, extent_item,
7283                                flags | BTRFS_EXTENT_FLAG_DATA);
7284
7285         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7286         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7287         if (parent > 0) {
7288                 struct btrfs_shared_data_ref *ref;
7289                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7290                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7291                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7292         } else {
7293                 struct btrfs_extent_data_ref *ref;
7294                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7295                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7296                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7297                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7298                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7299         }
7300
7301         btrfs_mark_buffer_dirty(path->nodes[0]);
7302         btrfs_free_path(path);
7303
7304         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7305         if (ret) { /* -ENOENT, logic error */
7306                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7307                         ins->objectid, ins->offset);
7308                 BUG();
7309         }
7310         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7311         return ret;
7312 }
7313
7314 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7315                                      struct btrfs_root *root,
7316                                      u64 parent, u64 root_objectid,
7317                                      u64 flags, struct btrfs_disk_key *key,
7318                                      int level, struct btrfs_key *ins,
7319                                      int no_quota)
7320 {
7321         int ret;
7322         struct btrfs_fs_info *fs_info = root->fs_info;
7323         struct btrfs_extent_item *extent_item;
7324         struct btrfs_tree_block_info *block_info;
7325         struct btrfs_extent_inline_ref *iref;
7326         struct btrfs_path *path;
7327         struct extent_buffer *leaf;
7328         u32 size = sizeof(*extent_item) + sizeof(*iref);
7329         u64 num_bytes = ins->offset;
7330         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7331                                                  SKINNY_METADATA);
7332
7333         if (!skinny_metadata)
7334                 size += sizeof(*block_info);
7335
7336         path = btrfs_alloc_path();
7337         if (!path) {
7338                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7339                                                    root->nodesize);
7340                 return -ENOMEM;
7341         }
7342
7343         path->leave_spinning = 1;
7344         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7345                                       ins, size);
7346         if (ret) {
7347                 btrfs_free_path(path);
7348                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7349                                                    root->nodesize);
7350                 return ret;
7351         }
7352
7353         leaf = path->nodes[0];
7354         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7355                                      struct btrfs_extent_item);
7356         btrfs_set_extent_refs(leaf, extent_item, 1);
7357         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7358         btrfs_set_extent_flags(leaf, extent_item,
7359                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7360
7361         if (skinny_metadata) {
7362                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7363                 num_bytes = root->nodesize;
7364         } else {
7365                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7366                 btrfs_set_tree_block_key(leaf, block_info, key);
7367                 btrfs_set_tree_block_level(leaf, block_info, level);
7368                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7369         }
7370
7371         if (parent > 0) {
7372                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7373                 btrfs_set_extent_inline_ref_type(leaf, iref,
7374                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7375                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7376         } else {
7377                 btrfs_set_extent_inline_ref_type(leaf, iref,
7378                                                  BTRFS_TREE_BLOCK_REF_KEY);
7379                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7380         }
7381
7382         btrfs_mark_buffer_dirty(leaf);
7383         btrfs_free_path(path);
7384
7385         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7386                                  1);
7387         if (ret) { /* -ENOENT, logic error */
7388                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7389                         ins->objectid, ins->offset);
7390                 BUG();
7391         }
7392
7393         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7394         return ret;
7395 }
7396
7397 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7398                                      struct btrfs_root *root,
7399                                      u64 root_objectid, u64 owner,
7400                                      u64 offset, struct btrfs_key *ins)
7401 {
7402         int ret;
7403
7404         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7405
7406         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7407                                          ins->offset, 0,
7408                                          root_objectid, owner, offset,
7409                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7410         return ret;
7411 }
7412
7413 /*
7414  * this is used by the tree logging recovery code.  It records that
7415  * an extent has been allocated and makes sure to clear the free
7416  * space cache bits as well
7417  */
7418 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7419                                    struct btrfs_root *root,
7420                                    u64 root_objectid, u64 owner, u64 offset,
7421                                    struct btrfs_key *ins)
7422 {
7423         int ret;
7424         struct btrfs_block_group_cache *block_group;
7425
7426         /*
7427          * Mixed block groups will exclude before processing the log so we only
7428          * need to do the exlude dance if this fs isn't mixed.
7429          */
7430         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7431                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7432                 if (ret)
7433                         return ret;
7434         }
7435
7436         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7437         if (!block_group)
7438                 return -EINVAL;
7439
7440         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7441                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7442         BUG_ON(ret); /* logic error */
7443         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7444                                          0, owner, offset, ins, 1);
7445         btrfs_put_block_group(block_group);
7446         return ret;
7447 }
7448
7449 static struct extent_buffer *
7450 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7451                       u64 bytenr, int level)
7452 {
7453         struct extent_buffer *buf;
7454
7455         buf = btrfs_find_create_tree_block(root, bytenr);
7456         if (!buf)
7457                 return ERR_PTR(-ENOMEM);
7458         btrfs_set_header_generation(buf, trans->transid);
7459         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7460         btrfs_tree_lock(buf);
7461         clean_tree_block(trans, root->fs_info, buf);
7462         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7463
7464         btrfs_set_lock_blocking(buf);
7465         btrfs_set_buffer_uptodate(buf);
7466
7467         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7468                 buf->log_index = root->log_transid % 2;
7469                 /*
7470                  * we allow two log transactions at a time, use different
7471                  * EXENT bit to differentiate dirty pages.
7472                  */
7473                 if (buf->log_index == 0)
7474                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7475                                         buf->start + buf->len - 1, GFP_NOFS);
7476                 else
7477                         set_extent_new(&root->dirty_log_pages, buf->start,
7478                                         buf->start + buf->len - 1, GFP_NOFS);
7479         } else {
7480                 buf->log_index = -1;
7481                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7482                          buf->start + buf->len - 1, GFP_NOFS);
7483         }
7484         trans->blocks_used++;
7485         /* this returns a buffer locked for blocking */
7486         return buf;
7487 }
7488
7489 static struct btrfs_block_rsv *
7490 use_block_rsv(struct btrfs_trans_handle *trans,
7491               struct btrfs_root *root, u32 blocksize)
7492 {
7493         struct btrfs_block_rsv *block_rsv;
7494         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7495         int ret;
7496         bool global_updated = false;
7497
7498         block_rsv = get_block_rsv(trans, root);
7499
7500         if (unlikely(block_rsv->size == 0))
7501                 goto try_reserve;
7502 again:
7503         ret = block_rsv_use_bytes(block_rsv, blocksize);
7504         if (!ret)
7505                 return block_rsv;
7506
7507         if (block_rsv->failfast)
7508                 return ERR_PTR(ret);
7509
7510         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7511                 global_updated = true;
7512                 update_global_block_rsv(root->fs_info);
7513                 goto again;
7514         }
7515
7516         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7517                 static DEFINE_RATELIMIT_STATE(_rs,
7518                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7519                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7520                 if (__ratelimit(&_rs))
7521                         WARN(1, KERN_DEBUG
7522                                 "BTRFS: block rsv returned %d\n", ret);
7523         }
7524 try_reserve:
7525         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7526                                      BTRFS_RESERVE_NO_FLUSH);
7527         if (!ret)
7528                 return block_rsv;
7529         /*
7530          * If we couldn't reserve metadata bytes try and use some from
7531          * the global reserve if its space type is the same as the global
7532          * reservation.
7533          */
7534         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7535             block_rsv->space_info == global_rsv->space_info) {
7536                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7537                 if (!ret)
7538                         return global_rsv;
7539         }
7540         return ERR_PTR(ret);
7541 }
7542
7543 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7544                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7545 {
7546         block_rsv_add_bytes(block_rsv, blocksize, 0);
7547         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7548 }
7549
7550 /*
7551  * finds a free extent and does all the dirty work required for allocation
7552  * returns the key for the extent through ins, and a tree buffer for
7553  * the first block of the extent through buf.
7554  *
7555  * returns the tree buffer or an ERR_PTR on error.
7556  */
7557 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7558                                         struct btrfs_root *root,
7559                                         u64 parent, u64 root_objectid,
7560                                         struct btrfs_disk_key *key, int level,
7561                                         u64 hint, u64 empty_size)
7562 {
7563         struct btrfs_key ins;
7564         struct btrfs_block_rsv *block_rsv;
7565         struct extent_buffer *buf;
7566         struct btrfs_delayed_extent_op *extent_op;
7567         u64 flags = 0;
7568         int ret;
7569         u32 blocksize = root->nodesize;
7570         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7571                                                  SKINNY_METADATA);
7572
7573         if (btrfs_test_is_dummy_root(root)) {
7574                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7575                                             level);
7576                 if (!IS_ERR(buf))
7577                         root->alloc_bytenr += blocksize;
7578                 return buf;
7579         }
7580
7581         block_rsv = use_block_rsv(trans, root, blocksize);
7582         if (IS_ERR(block_rsv))
7583                 return ERR_CAST(block_rsv);
7584
7585         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7586                                    empty_size, hint, &ins, 0, 0);
7587         if (ret)
7588                 goto out_unuse;
7589
7590         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
7591         if (IS_ERR(buf)) {
7592                 ret = PTR_ERR(buf);
7593                 goto out_free_reserved;
7594         }
7595
7596         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7597                 if (parent == 0)
7598                         parent = ins.objectid;
7599                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7600         } else
7601                 BUG_ON(parent > 0);
7602
7603         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7604                 extent_op = btrfs_alloc_delayed_extent_op();
7605                 if (!extent_op) {
7606                         ret = -ENOMEM;
7607                         goto out_free_buf;
7608                 }
7609                 if (key)
7610                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7611                 else
7612                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7613                 extent_op->flags_to_set = flags;
7614                 if (skinny_metadata)
7615                         extent_op->update_key = 0;
7616                 else
7617                         extent_op->update_key = 1;
7618                 extent_op->update_flags = 1;
7619                 extent_op->is_data = 0;
7620                 extent_op->level = level;
7621
7622                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7623                                                  ins.objectid, ins.offset,
7624                                                  parent, root_objectid, level,
7625                                                  BTRFS_ADD_DELAYED_EXTENT,
7626                                                  extent_op, 0);
7627                 if (ret)
7628                         goto out_free_delayed;
7629         }
7630         return buf;
7631
7632 out_free_delayed:
7633         btrfs_free_delayed_extent_op(extent_op);
7634 out_free_buf:
7635         free_extent_buffer(buf);
7636 out_free_reserved:
7637         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
7638 out_unuse:
7639         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7640         return ERR_PTR(ret);
7641 }
7642
7643 struct walk_control {
7644         u64 refs[BTRFS_MAX_LEVEL];
7645         u64 flags[BTRFS_MAX_LEVEL];
7646         struct btrfs_key update_progress;
7647         int stage;
7648         int level;
7649         int shared_level;
7650         int update_ref;
7651         int keep_locks;
7652         int reada_slot;
7653         int reada_count;
7654         int for_reloc;
7655 };
7656
7657 #define DROP_REFERENCE  1
7658 #define UPDATE_BACKREF  2
7659
7660 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7661                                      struct btrfs_root *root,
7662                                      struct walk_control *wc,
7663                                      struct btrfs_path *path)
7664 {
7665         u64 bytenr;
7666         u64 generation;
7667         u64 refs;
7668         u64 flags;
7669         u32 nritems;
7670         u32 blocksize;
7671         struct btrfs_key key;
7672         struct extent_buffer *eb;
7673         int ret;
7674         int slot;
7675         int nread = 0;
7676
7677         if (path->slots[wc->level] < wc->reada_slot) {
7678                 wc->reada_count = wc->reada_count * 2 / 3;
7679                 wc->reada_count = max(wc->reada_count, 2);
7680         } else {
7681                 wc->reada_count = wc->reada_count * 3 / 2;
7682                 wc->reada_count = min_t(int, wc->reada_count,
7683                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7684         }
7685
7686         eb = path->nodes[wc->level];
7687         nritems = btrfs_header_nritems(eb);
7688         blocksize = root->nodesize;
7689
7690         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7691                 if (nread >= wc->reada_count)
7692                         break;
7693
7694                 cond_resched();
7695                 bytenr = btrfs_node_blockptr(eb, slot);
7696                 generation = btrfs_node_ptr_generation(eb, slot);
7697
7698                 if (slot == path->slots[wc->level])
7699                         goto reada;
7700
7701                 if (wc->stage == UPDATE_BACKREF &&
7702                     generation <= root->root_key.offset)
7703                         continue;
7704
7705                 /* We don't lock the tree block, it's OK to be racy here */
7706                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7707                                                wc->level - 1, 1, &refs,
7708                                                &flags);
7709                 /* We don't care about errors in readahead. */
7710                 if (ret < 0)
7711                         continue;
7712                 BUG_ON(refs == 0);
7713
7714                 if (wc->stage == DROP_REFERENCE) {
7715                         if (refs == 1)
7716                                 goto reada;
7717
7718                         if (wc->level == 1 &&
7719                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7720                                 continue;
7721                         if (!wc->update_ref ||
7722                             generation <= root->root_key.offset)
7723                                 continue;
7724                         btrfs_node_key_to_cpu(eb, &key, slot);
7725                         ret = btrfs_comp_cpu_keys(&key,
7726                                                   &wc->update_progress);
7727                         if (ret < 0)
7728                                 continue;
7729                 } else {
7730                         if (wc->level == 1 &&
7731                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7732                                 continue;
7733                 }
7734 reada:
7735                 readahead_tree_block(root, bytenr);
7736                 nread++;
7737         }
7738         wc->reada_slot = slot;
7739 }
7740
7741 /*
7742  * TODO: Modify related function to add related node/leaf to dirty_extent_root,
7743  * for later qgroup accounting.
7744  *
7745  * Current, this function does nothing.
7746  */
7747 static int account_leaf_items(struct btrfs_trans_handle *trans,
7748                               struct btrfs_root *root,
7749                               struct extent_buffer *eb)
7750 {
7751         int nr = btrfs_header_nritems(eb);
7752         int i, extent_type;
7753         struct btrfs_key key;
7754         struct btrfs_file_extent_item *fi;
7755         u64 bytenr, num_bytes;
7756
7757         for (i = 0; i < nr; i++) {
7758                 btrfs_item_key_to_cpu(eb, &key, i);
7759
7760                 if (key.type != BTRFS_EXTENT_DATA_KEY)
7761                         continue;
7762
7763                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
7764                 /* filter out non qgroup-accountable extents  */
7765                 extent_type = btrfs_file_extent_type(eb, fi);
7766
7767                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
7768                         continue;
7769
7770                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
7771                 if (!bytenr)
7772                         continue;
7773
7774                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
7775         }
7776         return 0;
7777 }
7778
7779 /*
7780  * Walk up the tree from the bottom, freeing leaves and any interior
7781  * nodes which have had all slots visited. If a node (leaf or
7782  * interior) is freed, the node above it will have it's slot
7783  * incremented. The root node will never be freed.
7784  *
7785  * At the end of this function, we should have a path which has all
7786  * slots incremented to the next position for a search. If we need to
7787  * read a new node it will be NULL and the node above it will have the
7788  * correct slot selected for a later read.
7789  *
7790  * If we increment the root nodes slot counter past the number of
7791  * elements, 1 is returned to signal completion of the search.
7792  */
7793 static int adjust_slots_upwards(struct btrfs_root *root,
7794                                 struct btrfs_path *path, int root_level)
7795 {
7796         int level = 0;
7797         int nr, slot;
7798         struct extent_buffer *eb;
7799
7800         if (root_level == 0)
7801                 return 1;
7802
7803         while (level <= root_level) {
7804                 eb = path->nodes[level];
7805                 nr = btrfs_header_nritems(eb);
7806                 path->slots[level]++;
7807                 slot = path->slots[level];
7808                 if (slot >= nr || level == 0) {
7809                         /*
7810                          * Don't free the root -  we will detect this
7811                          * condition after our loop and return a
7812                          * positive value for caller to stop walking the tree.
7813                          */
7814                         if (level != root_level) {
7815                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7816                                 path->locks[level] = 0;
7817
7818                                 free_extent_buffer(eb);
7819                                 path->nodes[level] = NULL;
7820                                 path->slots[level] = 0;
7821                         }
7822                 } else {
7823                         /*
7824                          * We have a valid slot to walk back down
7825                          * from. Stop here so caller can process these
7826                          * new nodes.
7827                          */
7828                         break;
7829                 }
7830
7831                 level++;
7832         }
7833
7834         eb = path->nodes[root_level];
7835         if (path->slots[root_level] >= btrfs_header_nritems(eb))
7836                 return 1;
7837
7838         return 0;
7839 }
7840
7841 /*
7842  * root_eb is the subtree root and is locked before this function is called.
7843  * TODO: Modify this function to mark all (including complete shared node)
7844  * to dirty_extent_root to allow it get accounted in qgroup.
7845  */
7846 static int account_shared_subtree(struct btrfs_trans_handle *trans,
7847                                   struct btrfs_root *root,
7848                                   struct extent_buffer *root_eb,
7849                                   u64 root_gen,
7850                                   int root_level)
7851 {
7852         int ret = 0;
7853         int level;
7854         struct extent_buffer *eb = root_eb;
7855         struct btrfs_path *path = NULL;
7856
7857         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
7858         BUG_ON(root_eb == NULL);
7859
7860         if (!root->fs_info->quota_enabled)
7861                 return 0;
7862
7863         if (!extent_buffer_uptodate(root_eb)) {
7864                 ret = btrfs_read_buffer(root_eb, root_gen);
7865                 if (ret)
7866                         goto out;
7867         }
7868
7869         if (root_level == 0) {
7870                 ret = account_leaf_items(trans, root, root_eb);
7871                 goto out;
7872         }
7873
7874         path = btrfs_alloc_path();
7875         if (!path)
7876                 return -ENOMEM;
7877
7878         /*
7879          * Walk down the tree.  Missing extent blocks are filled in as
7880          * we go. Metadata is accounted every time we read a new
7881          * extent block.
7882          *
7883          * When we reach a leaf, we account for file extent items in it,
7884          * walk back up the tree (adjusting slot pointers as we go)
7885          * and restart the search process.
7886          */
7887         extent_buffer_get(root_eb); /* For path */
7888         path->nodes[root_level] = root_eb;
7889         path->slots[root_level] = 0;
7890         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
7891 walk_down:
7892         level = root_level;
7893         while (level >= 0) {
7894                 if (path->nodes[level] == NULL) {
7895                         int parent_slot;
7896                         u64 child_gen;
7897                         u64 child_bytenr;
7898
7899                         /* We need to get child blockptr/gen from
7900                          * parent before we can read it. */
7901                         eb = path->nodes[level + 1];
7902                         parent_slot = path->slots[level + 1];
7903                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
7904                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
7905
7906                         eb = read_tree_block(root, child_bytenr, child_gen);
7907                         if (IS_ERR(eb)) {
7908                                 ret = PTR_ERR(eb);
7909                                 goto out;
7910                         } else if (!extent_buffer_uptodate(eb)) {
7911                                 free_extent_buffer(eb);
7912                                 ret = -EIO;
7913                                 goto out;
7914                         }
7915
7916                         path->nodes[level] = eb;
7917                         path->slots[level] = 0;
7918
7919                         btrfs_tree_read_lock(eb);
7920                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
7921                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
7922                 }
7923
7924                 if (level == 0) {
7925                         ret = account_leaf_items(trans, root, path->nodes[level]);
7926                         if (ret)
7927                                 goto out;
7928
7929                         /* Nonzero return here means we completed our search */
7930                         ret = adjust_slots_upwards(root, path, root_level);
7931                         if (ret)
7932                                 break;
7933
7934                         /* Restart search with new slots */
7935                         goto walk_down;
7936                 }
7937
7938                 level--;
7939         }
7940
7941         ret = 0;
7942 out:
7943         btrfs_free_path(path);
7944
7945         return ret;
7946 }
7947
7948 /*
7949  * helper to process tree block while walking down the tree.
7950  *
7951  * when wc->stage == UPDATE_BACKREF, this function updates
7952  * back refs for pointers in the block.
7953  *
7954  * NOTE: return value 1 means we should stop walking down.
7955  */
7956 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7957                                    struct btrfs_root *root,
7958                                    struct btrfs_path *path,
7959                                    struct walk_control *wc, int lookup_info)
7960 {
7961         int level = wc->level;
7962         struct extent_buffer *eb = path->nodes[level];
7963         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7964         int ret;
7965
7966         if (wc->stage == UPDATE_BACKREF &&
7967             btrfs_header_owner(eb) != root->root_key.objectid)
7968                 return 1;
7969
7970         /*
7971          * when reference count of tree block is 1, it won't increase
7972          * again. once full backref flag is set, we never clear it.
7973          */
7974         if (lookup_info &&
7975             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7976              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7977                 BUG_ON(!path->locks[level]);
7978                 ret = btrfs_lookup_extent_info(trans, root,
7979                                                eb->start, level, 1,
7980                                                &wc->refs[level],
7981                                                &wc->flags[level]);
7982                 BUG_ON(ret == -ENOMEM);
7983                 if (ret)
7984                         return ret;
7985                 BUG_ON(wc->refs[level] == 0);
7986         }
7987
7988         if (wc->stage == DROP_REFERENCE) {
7989                 if (wc->refs[level] > 1)
7990                         return 1;
7991
7992                 if (path->locks[level] && !wc->keep_locks) {
7993                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7994                         path->locks[level] = 0;
7995                 }
7996                 return 0;
7997         }
7998
7999         /* wc->stage == UPDATE_BACKREF */
8000         if (!(wc->flags[level] & flag)) {
8001                 BUG_ON(!path->locks[level]);
8002                 ret = btrfs_inc_ref(trans, root, eb, 1);
8003                 BUG_ON(ret); /* -ENOMEM */
8004                 ret = btrfs_dec_ref(trans, root, eb, 0);
8005                 BUG_ON(ret); /* -ENOMEM */
8006                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8007                                                   eb->len, flag,
8008                                                   btrfs_header_level(eb), 0);
8009                 BUG_ON(ret); /* -ENOMEM */
8010                 wc->flags[level] |= flag;
8011         }
8012
8013         /*
8014          * the block is shared by multiple trees, so it's not good to
8015          * keep the tree lock
8016          */
8017         if (path->locks[level] && level > 0) {
8018                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8019                 path->locks[level] = 0;
8020         }
8021         return 0;
8022 }
8023
8024 /*
8025  * helper to process tree block pointer.
8026  *
8027  * when wc->stage == DROP_REFERENCE, this function checks
8028  * reference count of the block pointed to. if the block
8029  * is shared and we need update back refs for the subtree
8030  * rooted at the block, this function changes wc->stage to
8031  * UPDATE_BACKREF. if the block is shared and there is no
8032  * need to update back, this function drops the reference
8033  * to the block.
8034  *
8035  * NOTE: return value 1 means we should stop walking down.
8036  */
8037 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8038                                  struct btrfs_root *root,
8039                                  struct btrfs_path *path,
8040                                  struct walk_control *wc, int *lookup_info)
8041 {
8042         u64 bytenr;
8043         u64 generation;
8044         u64 parent;
8045         u32 blocksize;
8046         struct btrfs_key key;
8047         struct extent_buffer *next;
8048         int level = wc->level;
8049         int reada = 0;
8050         int ret = 0;
8051         bool need_account = false;
8052
8053         generation = btrfs_node_ptr_generation(path->nodes[level],
8054                                                path->slots[level]);
8055         /*
8056          * if the lower level block was created before the snapshot
8057          * was created, we know there is no need to update back refs
8058          * for the subtree
8059          */
8060         if (wc->stage == UPDATE_BACKREF &&
8061             generation <= root->root_key.offset) {
8062                 *lookup_info = 1;
8063                 return 1;
8064         }
8065
8066         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8067         blocksize = root->nodesize;
8068
8069         next = btrfs_find_tree_block(root->fs_info, bytenr);
8070         if (!next) {
8071                 next = btrfs_find_create_tree_block(root, bytenr);
8072                 if (!next)
8073                         return -ENOMEM;
8074                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8075                                                level - 1);
8076                 reada = 1;
8077         }
8078         btrfs_tree_lock(next);
8079         btrfs_set_lock_blocking(next);
8080
8081         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8082                                        &wc->refs[level - 1],
8083                                        &wc->flags[level - 1]);
8084         if (ret < 0) {
8085                 btrfs_tree_unlock(next);
8086                 return ret;
8087         }
8088
8089         if (unlikely(wc->refs[level - 1] == 0)) {
8090                 btrfs_err(root->fs_info, "Missing references.");
8091                 BUG();
8092         }
8093         *lookup_info = 0;
8094
8095         if (wc->stage == DROP_REFERENCE) {
8096                 if (wc->refs[level - 1] > 1) {
8097                         need_account = true;
8098                         if (level == 1 &&
8099                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8100                                 goto skip;
8101
8102                         if (!wc->update_ref ||
8103                             generation <= root->root_key.offset)
8104                                 goto skip;
8105
8106                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8107                                               path->slots[level]);
8108                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8109                         if (ret < 0)
8110                                 goto skip;
8111
8112                         wc->stage = UPDATE_BACKREF;
8113                         wc->shared_level = level - 1;
8114                 }
8115         } else {
8116                 if (level == 1 &&
8117                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8118                         goto skip;
8119         }
8120
8121         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8122                 btrfs_tree_unlock(next);
8123                 free_extent_buffer(next);
8124                 next = NULL;
8125                 *lookup_info = 1;
8126         }
8127
8128         if (!next) {
8129                 if (reada && level == 1)
8130                         reada_walk_down(trans, root, wc, path);
8131                 next = read_tree_block(root, bytenr, generation);
8132                 if (IS_ERR(next)) {
8133                         return PTR_ERR(next);
8134                 } else if (!extent_buffer_uptodate(next)) {
8135                         free_extent_buffer(next);
8136                         return -EIO;
8137                 }
8138                 btrfs_tree_lock(next);
8139                 btrfs_set_lock_blocking(next);
8140         }
8141
8142         level--;
8143         BUG_ON(level != btrfs_header_level(next));
8144         path->nodes[level] = next;
8145         path->slots[level] = 0;
8146         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8147         wc->level = level;
8148         if (wc->level == 1)
8149                 wc->reada_slot = 0;
8150         return 0;
8151 skip:
8152         wc->refs[level - 1] = 0;
8153         wc->flags[level - 1] = 0;
8154         if (wc->stage == DROP_REFERENCE) {
8155                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8156                         parent = path->nodes[level]->start;
8157                 } else {
8158                         BUG_ON(root->root_key.objectid !=
8159                                btrfs_header_owner(path->nodes[level]));
8160                         parent = 0;
8161                 }
8162
8163                 if (need_account) {
8164                         ret = account_shared_subtree(trans, root, next,
8165                                                      generation, level - 1);
8166                         if (ret) {
8167                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8168                                         "%d accounting shared subtree. Quota "
8169                                         "is out of sync, rescan required.\n",
8170                                         root->fs_info->sb->s_id, ret);
8171                         }
8172                 }
8173                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8174                                 root->root_key.objectid, level - 1, 0, 0);
8175                 BUG_ON(ret); /* -ENOMEM */
8176         }
8177         btrfs_tree_unlock(next);
8178         free_extent_buffer(next);
8179         *lookup_info = 1;
8180         return 1;
8181 }
8182
8183 /*
8184  * helper to process tree block while walking up the tree.
8185  *
8186  * when wc->stage == DROP_REFERENCE, this function drops
8187  * reference count on the block.
8188  *
8189  * when wc->stage == UPDATE_BACKREF, this function changes
8190  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8191  * to UPDATE_BACKREF previously while processing the block.
8192  *
8193  * NOTE: return value 1 means we should stop walking up.
8194  */
8195 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8196                                  struct btrfs_root *root,
8197                                  struct btrfs_path *path,
8198                                  struct walk_control *wc)
8199 {
8200         int ret;
8201         int level = wc->level;
8202         struct extent_buffer *eb = path->nodes[level];
8203         u64 parent = 0;
8204
8205         if (wc->stage == UPDATE_BACKREF) {
8206                 BUG_ON(wc->shared_level < level);
8207                 if (level < wc->shared_level)
8208                         goto out;
8209
8210                 ret = find_next_key(path, level + 1, &wc->update_progress);
8211                 if (ret > 0)
8212                         wc->update_ref = 0;
8213
8214                 wc->stage = DROP_REFERENCE;
8215                 wc->shared_level = -1;
8216                 path->slots[level] = 0;
8217
8218                 /*
8219                  * check reference count again if the block isn't locked.
8220                  * we should start walking down the tree again if reference
8221                  * count is one.
8222                  */
8223                 if (!path->locks[level]) {
8224                         BUG_ON(level == 0);
8225                         btrfs_tree_lock(eb);
8226                         btrfs_set_lock_blocking(eb);
8227                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8228
8229                         ret = btrfs_lookup_extent_info(trans, root,
8230                                                        eb->start, level, 1,
8231                                                        &wc->refs[level],
8232                                                        &wc->flags[level]);
8233                         if (ret < 0) {
8234                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8235                                 path->locks[level] = 0;
8236                                 return ret;
8237                         }
8238                         BUG_ON(wc->refs[level] == 0);
8239                         if (wc->refs[level] == 1) {
8240                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8241                                 path->locks[level] = 0;
8242                                 return 1;
8243                         }
8244                 }
8245         }
8246
8247         /* wc->stage == DROP_REFERENCE */
8248         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8249
8250         if (wc->refs[level] == 1) {
8251                 if (level == 0) {
8252                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8253                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8254                         else
8255                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8256                         BUG_ON(ret); /* -ENOMEM */
8257                         ret = account_leaf_items(trans, root, eb);
8258                         if (ret) {
8259                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8260                                         "%d accounting leaf items. Quota "
8261                                         "is out of sync, rescan required.\n",
8262                                         root->fs_info->sb->s_id, ret);
8263                         }
8264                 }
8265                 /* make block locked assertion in clean_tree_block happy */
8266                 if (!path->locks[level] &&
8267                     btrfs_header_generation(eb) == trans->transid) {
8268                         btrfs_tree_lock(eb);
8269                         btrfs_set_lock_blocking(eb);
8270                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8271                 }
8272                 clean_tree_block(trans, root->fs_info, eb);
8273         }
8274
8275         if (eb == root->node) {
8276                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8277                         parent = eb->start;
8278                 else
8279                         BUG_ON(root->root_key.objectid !=
8280                                btrfs_header_owner(eb));
8281         } else {
8282                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8283                         parent = path->nodes[level + 1]->start;
8284                 else
8285                         BUG_ON(root->root_key.objectid !=
8286                                btrfs_header_owner(path->nodes[level + 1]));
8287         }
8288
8289         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8290 out:
8291         wc->refs[level] = 0;
8292         wc->flags[level] = 0;
8293         return 0;
8294 }
8295
8296 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8297                                    struct btrfs_root *root,
8298                                    struct btrfs_path *path,
8299                                    struct walk_control *wc)
8300 {
8301         int level = wc->level;
8302         int lookup_info = 1;
8303         int ret;
8304
8305         while (level >= 0) {
8306                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8307                 if (ret > 0)
8308                         break;
8309
8310                 if (level == 0)
8311                         break;
8312
8313                 if (path->slots[level] >=
8314                     btrfs_header_nritems(path->nodes[level]))
8315                         break;
8316
8317                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8318                 if (ret > 0) {
8319                         path->slots[level]++;
8320                         continue;
8321                 } else if (ret < 0)
8322                         return ret;
8323                 level = wc->level;
8324         }
8325         return 0;
8326 }
8327
8328 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8329                                  struct btrfs_root *root,
8330                                  struct btrfs_path *path,
8331                                  struct walk_control *wc, int max_level)
8332 {
8333         int level = wc->level;
8334         int ret;
8335
8336         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8337         while (level < max_level && path->nodes[level]) {
8338                 wc->level = level;
8339                 if (path->slots[level] + 1 <
8340                     btrfs_header_nritems(path->nodes[level])) {
8341                         path->slots[level]++;
8342                         return 0;
8343                 } else {
8344                         ret = walk_up_proc(trans, root, path, wc);
8345                         if (ret > 0)
8346                                 return 0;
8347
8348                         if (path->locks[level]) {
8349                                 btrfs_tree_unlock_rw(path->nodes[level],
8350                                                      path->locks[level]);
8351                                 path->locks[level] = 0;
8352                         }
8353                         free_extent_buffer(path->nodes[level]);
8354                         path->nodes[level] = NULL;
8355                         level++;
8356                 }
8357         }
8358         return 1;
8359 }
8360
8361 /*
8362  * drop a subvolume tree.
8363  *
8364  * this function traverses the tree freeing any blocks that only
8365  * referenced by the tree.
8366  *
8367  * when a shared tree block is found. this function decreases its
8368  * reference count by one. if update_ref is true, this function
8369  * also make sure backrefs for the shared block and all lower level
8370  * blocks are properly updated.
8371  *
8372  * If called with for_reloc == 0, may exit early with -EAGAIN
8373  */
8374 int btrfs_drop_snapshot(struct btrfs_root *root,
8375                          struct btrfs_block_rsv *block_rsv, int update_ref,
8376                          int for_reloc)
8377 {
8378         struct btrfs_path *path;
8379         struct btrfs_trans_handle *trans;
8380         struct btrfs_root *tree_root = root->fs_info->tree_root;
8381         struct btrfs_root_item *root_item = &root->root_item;
8382         struct walk_control *wc;
8383         struct btrfs_key key;
8384         int err = 0;
8385         int ret;
8386         int level;
8387         bool root_dropped = false;
8388
8389         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8390
8391         path = btrfs_alloc_path();
8392         if (!path) {
8393                 err = -ENOMEM;
8394                 goto out;
8395         }
8396
8397         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8398         if (!wc) {
8399                 btrfs_free_path(path);
8400                 err = -ENOMEM;
8401                 goto out;
8402         }
8403
8404         trans = btrfs_start_transaction(tree_root, 0);
8405         if (IS_ERR(trans)) {
8406                 err = PTR_ERR(trans);
8407                 goto out_free;
8408         }
8409
8410         if (block_rsv)
8411                 trans->block_rsv = block_rsv;
8412
8413         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8414                 level = btrfs_header_level(root->node);
8415                 path->nodes[level] = btrfs_lock_root_node(root);
8416                 btrfs_set_lock_blocking(path->nodes[level]);
8417                 path->slots[level] = 0;
8418                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8419                 memset(&wc->update_progress, 0,
8420                        sizeof(wc->update_progress));
8421         } else {
8422                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8423                 memcpy(&wc->update_progress, &key,
8424                        sizeof(wc->update_progress));
8425
8426                 level = root_item->drop_level;
8427                 BUG_ON(level == 0);
8428                 path->lowest_level = level;
8429                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8430                 path->lowest_level = 0;
8431                 if (ret < 0) {
8432                         err = ret;
8433                         goto out_end_trans;
8434                 }
8435                 WARN_ON(ret > 0);
8436
8437                 /*
8438                  * unlock our path, this is safe because only this
8439                  * function is allowed to delete this snapshot
8440                  */
8441                 btrfs_unlock_up_safe(path, 0);
8442
8443                 level = btrfs_header_level(root->node);
8444                 while (1) {
8445                         btrfs_tree_lock(path->nodes[level]);
8446                         btrfs_set_lock_blocking(path->nodes[level]);
8447                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8448
8449                         ret = btrfs_lookup_extent_info(trans, root,
8450                                                 path->nodes[level]->start,
8451                                                 level, 1, &wc->refs[level],
8452                                                 &wc->flags[level]);
8453                         if (ret < 0) {
8454                                 err = ret;
8455                                 goto out_end_trans;
8456                         }
8457                         BUG_ON(wc->refs[level] == 0);
8458
8459                         if (level == root_item->drop_level)
8460                                 break;
8461
8462                         btrfs_tree_unlock(path->nodes[level]);
8463                         path->locks[level] = 0;
8464                         WARN_ON(wc->refs[level] != 1);
8465                         level--;
8466                 }
8467         }
8468
8469         wc->level = level;
8470         wc->shared_level = -1;
8471         wc->stage = DROP_REFERENCE;
8472         wc->update_ref = update_ref;
8473         wc->keep_locks = 0;
8474         wc->for_reloc = for_reloc;
8475         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8476
8477         while (1) {
8478
8479                 ret = walk_down_tree(trans, root, path, wc);
8480                 if (ret < 0) {
8481                         err = ret;
8482                         break;
8483                 }
8484
8485                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8486                 if (ret < 0) {
8487                         err = ret;
8488                         break;
8489                 }
8490
8491                 if (ret > 0) {
8492                         BUG_ON(wc->stage != DROP_REFERENCE);
8493                         break;
8494                 }
8495
8496                 if (wc->stage == DROP_REFERENCE) {
8497                         level = wc->level;
8498                         btrfs_node_key(path->nodes[level],
8499                                        &root_item->drop_progress,
8500                                        path->slots[level]);
8501                         root_item->drop_level = level;
8502                 }
8503
8504                 BUG_ON(wc->level == 0);
8505                 if (btrfs_should_end_transaction(trans, tree_root) ||
8506                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8507                         ret = btrfs_update_root(trans, tree_root,
8508                                                 &root->root_key,
8509                                                 root_item);
8510                         if (ret) {
8511                                 btrfs_abort_transaction(trans, tree_root, ret);
8512                                 err = ret;
8513                                 goto out_end_trans;
8514                         }
8515
8516                         btrfs_end_transaction_throttle(trans, tree_root);
8517                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8518                                 pr_debug("BTRFS: drop snapshot early exit\n");
8519                                 err = -EAGAIN;
8520                                 goto out_free;
8521                         }
8522
8523                         trans = btrfs_start_transaction(tree_root, 0);
8524                         if (IS_ERR(trans)) {
8525                                 err = PTR_ERR(trans);
8526                                 goto out_free;
8527                         }
8528                         if (block_rsv)
8529                                 trans->block_rsv = block_rsv;
8530                 }
8531         }
8532         btrfs_release_path(path);
8533         if (err)
8534                 goto out_end_trans;
8535
8536         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8537         if (ret) {
8538                 btrfs_abort_transaction(trans, tree_root, ret);
8539                 goto out_end_trans;
8540         }
8541
8542         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8543                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8544                                       NULL, NULL);
8545                 if (ret < 0) {
8546                         btrfs_abort_transaction(trans, tree_root, ret);
8547                         err = ret;
8548                         goto out_end_trans;
8549                 } else if (ret > 0) {
8550                         /* if we fail to delete the orphan item this time
8551                          * around, it'll get picked up the next time.
8552                          *
8553                          * The most common failure here is just -ENOENT.
8554                          */
8555                         btrfs_del_orphan_item(trans, tree_root,
8556                                               root->root_key.objectid);
8557                 }
8558         }
8559
8560         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8561                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
8562         } else {
8563                 free_extent_buffer(root->node);
8564                 free_extent_buffer(root->commit_root);
8565                 btrfs_put_fs_root(root);
8566         }
8567         root_dropped = true;
8568 out_end_trans:
8569         btrfs_end_transaction_throttle(trans, tree_root);
8570 out_free:
8571         kfree(wc);
8572         btrfs_free_path(path);
8573 out:
8574         /*
8575          * So if we need to stop dropping the snapshot for whatever reason we
8576          * need to make sure to add it back to the dead root list so that we
8577          * keep trying to do the work later.  This also cleans up roots if we
8578          * don't have it in the radix (like when we recover after a power fail
8579          * or unmount) so we don't leak memory.
8580          */
8581         if (!for_reloc && root_dropped == false)
8582                 btrfs_add_dead_root(root);
8583         if (err && err != -EAGAIN)
8584                 btrfs_std_error(root->fs_info, err);
8585         return err;
8586 }
8587
8588 /*
8589  * drop subtree rooted at tree block 'node'.
8590  *
8591  * NOTE: this function will unlock and release tree block 'node'
8592  * only used by relocation code
8593  */
8594 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8595                         struct btrfs_root *root,
8596                         struct extent_buffer *node,
8597                         struct extent_buffer *parent)
8598 {
8599         struct btrfs_path *path;
8600         struct walk_control *wc;
8601         int level;
8602         int parent_level;
8603         int ret = 0;
8604         int wret;
8605
8606         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8607
8608         path = btrfs_alloc_path();
8609         if (!path)
8610                 return -ENOMEM;
8611
8612         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8613         if (!wc) {
8614                 btrfs_free_path(path);
8615                 return -ENOMEM;
8616         }
8617
8618         btrfs_assert_tree_locked(parent);
8619         parent_level = btrfs_header_level(parent);
8620         extent_buffer_get(parent);
8621         path->nodes[parent_level] = parent;
8622         path->slots[parent_level] = btrfs_header_nritems(parent);
8623
8624         btrfs_assert_tree_locked(node);
8625         level = btrfs_header_level(node);
8626         path->nodes[level] = node;
8627         path->slots[level] = 0;
8628         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8629
8630         wc->refs[parent_level] = 1;
8631         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8632         wc->level = level;
8633         wc->shared_level = -1;
8634         wc->stage = DROP_REFERENCE;
8635         wc->update_ref = 0;
8636         wc->keep_locks = 1;
8637         wc->for_reloc = 1;
8638         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8639
8640         while (1) {
8641                 wret = walk_down_tree(trans, root, path, wc);
8642                 if (wret < 0) {
8643                         ret = wret;
8644                         break;
8645                 }
8646
8647                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8648                 if (wret < 0)
8649                         ret = wret;
8650                 if (wret != 0)
8651                         break;
8652         }
8653
8654         kfree(wc);
8655         btrfs_free_path(path);
8656         return ret;
8657 }
8658
8659 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8660 {
8661         u64 num_devices;
8662         u64 stripped;
8663
8664         /*
8665          * if restripe for this chunk_type is on pick target profile and
8666          * return, otherwise do the usual balance
8667          */
8668         stripped = get_restripe_target(root->fs_info, flags);
8669         if (stripped)
8670                 return extended_to_chunk(stripped);
8671
8672         num_devices = root->fs_info->fs_devices->rw_devices;
8673
8674         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8675                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8676                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8677
8678         if (num_devices == 1) {
8679                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8680                 stripped = flags & ~stripped;
8681
8682                 /* turn raid0 into single device chunks */
8683                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8684                         return stripped;
8685
8686                 /* turn mirroring into duplication */
8687                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8688                              BTRFS_BLOCK_GROUP_RAID10))
8689                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8690         } else {
8691                 /* they already had raid on here, just return */
8692                 if (flags & stripped)
8693                         return flags;
8694
8695                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8696                 stripped = flags & ~stripped;
8697
8698                 /* switch duplicated blocks with raid1 */
8699                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8700                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8701
8702                 /* this is drive concat, leave it alone */
8703         }
8704
8705         return flags;
8706 }
8707
8708 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8709 {
8710         struct btrfs_space_info *sinfo = cache->space_info;
8711         u64 num_bytes;
8712         u64 min_allocable_bytes;
8713         int ret = -ENOSPC;
8714
8715
8716         /*
8717          * We need some metadata space and system metadata space for
8718          * allocating chunks in some corner cases until we force to set
8719          * it to be readonly.
8720          */
8721         if ((sinfo->flags &
8722              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8723             !force)
8724                 min_allocable_bytes = 1 * 1024 * 1024;
8725         else
8726                 min_allocable_bytes = 0;
8727
8728         spin_lock(&sinfo->lock);
8729         spin_lock(&cache->lock);
8730
8731         if (cache->ro) {
8732                 ret = 0;
8733                 goto out;
8734         }
8735
8736         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8737                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8738
8739         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8740             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8741             min_allocable_bytes <= sinfo->total_bytes) {
8742                 sinfo->bytes_readonly += num_bytes;
8743                 cache->ro = 1;
8744                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
8745                 ret = 0;
8746         }
8747 out:
8748         spin_unlock(&cache->lock);
8749         spin_unlock(&sinfo->lock);
8750         return ret;
8751 }
8752
8753 int btrfs_set_block_group_ro(struct btrfs_root *root,
8754                              struct btrfs_block_group_cache *cache)
8755
8756 {
8757         struct btrfs_trans_handle *trans;
8758         u64 alloc_flags;
8759         int ret;
8760
8761         BUG_ON(cache->ro);
8762
8763 again:
8764         trans = btrfs_join_transaction(root);
8765         if (IS_ERR(trans))
8766                 return PTR_ERR(trans);
8767
8768         /*
8769          * we're not allowed to set block groups readonly after the dirty
8770          * block groups cache has started writing.  If it already started,
8771          * back off and let this transaction commit
8772          */
8773         mutex_lock(&root->fs_info->ro_block_group_mutex);
8774         if (trans->transaction->dirty_bg_run) {
8775                 u64 transid = trans->transid;
8776
8777                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
8778                 btrfs_end_transaction(trans, root);
8779
8780                 ret = btrfs_wait_for_commit(root, transid);
8781                 if (ret)
8782                         return ret;
8783                 goto again;
8784         }
8785
8786         /*
8787          * if we are changing raid levels, try to allocate a corresponding
8788          * block group with the new raid level.
8789          */
8790         alloc_flags = update_block_group_flags(root, cache->flags);
8791         if (alloc_flags != cache->flags) {
8792                 ret = do_chunk_alloc(trans, root, alloc_flags,
8793                                      CHUNK_ALLOC_FORCE);
8794                 /*
8795                  * ENOSPC is allowed here, we may have enough space
8796                  * already allocated at the new raid level to
8797                  * carry on
8798                  */
8799                 if (ret == -ENOSPC)
8800                         ret = 0;
8801                 if (ret < 0)
8802                         goto out;
8803         }
8804
8805         ret = set_block_group_ro(cache, 0);
8806         if (!ret)
8807                 goto out;
8808         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8809         ret = do_chunk_alloc(trans, root, alloc_flags,
8810                              CHUNK_ALLOC_FORCE);
8811         if (ret < 0)
8812                 goto out;
8813         ret = set_block_group_ro(cache, 0);
8814 out:
8815         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
8816                 alloc_flags = update_block_group_flags(root, cache->flags);
8817                 lock_chunks(root->fs_info->chunk_root);
8818                 check_system_chunk(trans, root, alloc_flags);
8819                 unlock_chunks(root->fs_info->chunk_root);
8820         }
8821         mutex_unlock(&root->fs_info->ro_block_group_mutex);
8822
8823         btrfs_end_transaction(trans, root);
8824         return ret;
8825 }
8826
8827 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8828                             struct btrfs_root *root, u64 type)
8829 {
8830         u64 alloc_flags = get_alloc_profile(root, type);
8831         return do_chunk_alloc(trans, root, alloc_flags,
8832                               CHUNK_ALLOC_FORCE);
8833 }
8834
8835 /*
8836  * helper to account the unused space of all the readonly block group in the
8837  * space_info. takes mirrors into account.
8838  */
8839 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8840 {
8841         struct btrfs_block_group_cache *block_group;
8842         u64 free_bytes = 0;
8843         int factor;
8844
8845         /* It's df, we don't care if it's racey */
8846         if (list_empty(&sinfo->ro_bgs))
8847                 return 0;
8848
8849         spin_lock(&sinfo->lock);
8850         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
8851                 spin_lock(&block_group->lock);
8852
8853                 if (!block_group->ro) {
8854                         spin_unlock(&block_group->lock);
8855                         continue;
8856                 }
8857
8858                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8859                                           BTRFS_BLOCK_GROUP_RAID10 |
8860                                           BTRFS_BLOCK_GROUP_DUP))
8861                         factor = 2;
8862                 else
8863                         factor = 1;
8864
8865                 free_bytes += (block_group->key.offset -
8866                                btrfs_block_group_used(&block_group->item)) *
8867                                factor;
8868
8869                 spin_unlock(&block_group->lock);
8870         }
8871         spin_unlock(&sinfo->lock);
8872
8873         return free_bytes;
8874 }
8875
8876 void btrfs_set_block_group_rw(struct btrfs_root *root,
8877                               struct btrfs_block_group_cache *cache)
8878 {
8879         struct btrfs_space_info *sinfo = cache->space_info;
8880         u64 num_bytes;
8881
8882         BUG_ON(!cache->ro);
8883
8884         spin_lock(&sinfo->lock);
8885         spin_lock(&cache->lock);
8886         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8887                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8888         sinfo->bytes_readonly -= num_bytes;
8889         cache->ro = 0;
8890         list_del_init(&cache->ro_list);
8891         spin_unlock(&cache->lock);
8892         spin_unlock(&sinfo->lock);
8893 }
8894
8895 /*
8896  * checks to see if its even possible to relocate this block group.
8897  *
8898  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8899  * ok to go ahead and try.
8900  */
8901 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8902 {
8903         struct btrfs_block_group_cache *block_group;
8904         struct btrfs_space_info *space_info;
8905         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8906         struct btrfs_device *device;
8907         struct btrfs_trans_handle *trans;
8908         u64 min_free;
8909         u64 dev_min = 1;
8910         u64 dev_nr = 0;
8911         u64 target;
8912         int index;
8913         int full = 0;
8914         int ret = 0;
8915
8916         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8917
8918         /* odd, couldn't find the block group, leave it alone */
8919         if (!block_group)
8920                 return -1;
8921
8922         min_free = btrfs_block_group_used(&block_group->item);
8923
8924         /* no bytes used, we're good */
8925         if (!min_free)
8926                 goto out;
8927
8928         space_info = block_group->space_info;
8929         spin_lock(&space_info->lock);
8930
8931         full = space_info->full;
8932
8933         /*
8934          * if this is the last block group we have in this space, we can't
8935          * relocate it unless we're able to allocate a new chunk below.
8936          *
8937          * Otherwise, we need to make sure we have room in the space to handle
8938          * all of the extents from this block group.  If we can, we're good
8939          */
8940         if ((space_info->total_bytes != block_group->key.offset) &&
8941             (space_info->bytes_used + space_info->bytes_reserved +
8942              space_info->bytes_pinned + space_info->bytes_readonly +
8943              min_free < space_info->total_bytes)) {
8944                 spin_unlock(&space_info->lock);
8945                 goto out;
8946         }
8947         spin_unlock(&space_info->lock);
8948
8949         /*
8950          * ok we don't have enough space, but maybe we have free space on our
8951          * devices to allocate new chunks for relocation, so loop through our
8952          * alloc devices and guess if we have enough space.  if this block
8953          * group is going to be restriped, run checks against the target
8954          * profile instead of the current one.
8955          */
8956         ret = -1;
8957
8958         /*
8959          * index:
8960          *      0: raid10
8961          *      1: raid1
8962          *      2: dup
8963          *      3: raid0
8964          *      4: single
8965          */
8966         target = get_restripe_target(root->fs_info, block_group->flags);
8967         if (target) {
8968                 index = __get_raid_index(extended_to_chunk(target));
8969         } else {
8970                 /*
8971                  * this is just a balance, so if we were marked as full
8972                  * we know there is no space for a new chunk
8973                  */
8974                 if (full)
8975                         goto out;
8976
8977                 index = get_block_group_index(block_group);
8978         }
8979
8980         if (index == BTRFS_RAID_RAID10) {
8981                 dev_min = 4;
8982                 /* Divide by 2 */
8983                 min_free >>= 1;
8984         } else if (index == BTRFS_RAID_RAID1) {
8985                 dev_min = 2;
8986         } else if (index == BTRFS_RAID_DUP) {
8987                 /* Multiply by 2 */
8988                 min_free <<= 1;
8989         } else if (index == BTRFS_RAID_RAID0) {
8990                 dev_min = fs_devices->rw_devices;
8991                 min_free = div64_u64(min_free, dev_min);
8992         }
8993
8994         /* We need to do this so that we can look at pending chunks */
8995         trans = btrfs_join_transaction(root);
8996         if (IS_ERR(trans)) {
8997                 ret = PTR_ERR(trans);
8998                 goto out;
8999         }
9000
9001         mutex_lock(&root->fs_info->chunk_mutex);
9002         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9003                 u64 dev_offset;
9004
9005                 /*
9006                  * check to make sure we can actually find a chunk with enough
9007                  * space to fit our block group in.
9008                  */
9009                 if (device->total_bytes > device->bytes_used + min_free &&
9010                     !device->is_tgtdev_for_dev_replace) {
9011                         ret = find_free_dev_extent(trans, device, min_free,
9012                                                    &dev_offset, NULL);
9013                         if (!ret)
9014                                 dev_nr++;
9015
9016                         if (dev_nr >= dev_min)
9017                                 break;
9018
9019                         ret = -1;
9020                 }
9021         }
9022         mutex_unlock(&root->fs_info->chunk_mutex);
9023         btrfs_end_transaction(trans, root);
9024 out:
9025         btrfs_put_block_group(block_group);
9026         return ret;
9027 }
9028
9029 static int find_first_block_group(struct btrfs_root *root,
9030                 struct btrfs_path *path, struct btrfs_key *key)
9031 {
9032         int ret = 0;
9033         struct btrfs_key found_key;
9034         struct extent_buffer *leaf;
9035         int slot;
9036
9037         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9038         if (ret < 0)
9039                 goto out;
9040
9041         while (1) {
9042                 slot = path->slots[0];
9043                 leaf = path->nodes[0];
9044                 if (slot >= btrfs_header_nritems(leaf)) {
9045                         ret = btrfs_next_leaf(root, path);
9046                         if (ret == 0)
9047                                 continue;
9048                         if (ret < 0)
9049                                 goto out;
9050                         break;
9051                 }
9052                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9053
9054                 if (found_key.objectid >= key->objectid &&
9055                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9056                         ret = 0;
9057                         goto out;
9058                 }
9059                 path->slots[0]++;
9060         }
9061 out:
9062         return ret;
9063 }
9064
9065 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9066 {
9067         struct btrfs_block_group_cache *block_group;
9068         u64 last = 0;
9069
9070         while (1) {
9071                 struct inode *inode;
9072
9073                 block_group = btrfs_lookup_first_block_group(info, last);
9074                 while (block_group) {
9075                         spin_lock(&block_group->lock);
9076                         if (block_group->iref)
9077                                 break;
9078                         spin_unlock(&block_group->lock);
9079                         block_group = next_block_group(info->tree_root,
9080                                                        block_group);
9081                 }
9082                 if (!block_group) {
9083                         if (last == 0)
9084                                 break;
9085                         last = 0;
9086                         continue;
9087                 }
9088
9089                 inode = block_group->inode;
9090                 block_group->iref = 0;
9091                 block_group->inode = NULL;
9092                 spin_unlock(&block_group->lock);
9093                 iput(inode);
9094                 last = block_group->key.objectid + block_group->key.offset;
9095                 btrfs_put_block_group(block_group);
9096         }
9097 }
9098
9099 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9100 {
9101         struct btrfs_block_group_cache *block_group;
9102         struct btrfs_space_info *space_info;
9103         struct btrfs_caching_control *caching_ctl;
9104         struct rb_node *n;
9105
9106         down_write(&info->commit_root_sem);
9107         while (!list_empty(&info->caching_block_groups)) {
9108                 caching_ctl = list_entry(info->caching_block_groups.next,
9109                                          struct btrfs_caching_control, list);
9110                 list_del(&caching_ctl->list);
9111                 put_caching_control(caching_ctl);
9112         }
9113         up_write(&info->commit_root_sem);
9114
9115         spin_lock(&info->unused_bgs_lock);
9116         while (!list_empty(&info->unused_bgs)) {
9117                 block_group = list_first_entry(&info->unused_bgs,
9118                                                struct btrfs_block_group_cache,
9119                                                bg_list);
9120                 list_del_init(&block_group->bg_list);
9121                 btrfs_put_block_group(block_group);
9122         }
9123         spin_unlock(&info->unused_bgs_lock);
9124
9125         spin_lock(&info->block_group_cache_lock);
9126         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9127                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9128                                        cache_node);
9129                 rb_erase(&block_group->cache_node,
9130                          &info->block_group_cache_tree);
9131                 RB_CLEAR_NODE(&block_group->cache_node);
9132                 spin_unlock(&info->block_group_cache_lock);
9133
9134                 down_write(&block_group->space_info->groups_sem);
9135                 list_del(&block_group->list);
9136                 up_write(&block_group->space_info->groups_sem);
9137
9138                 if (block_group->cached == BTRFS_CACHE_STARTED)
9139                         wait_block_group_cache_done(block_group);
9140
9141                 /*
9142                  * We haven't cached this block group, which means we could
9143                  * possibly have excluded extents on this block group.
9144                  */
9145                 if (block_group->cached == BTRFS_CACHE_NO ||
9146                     block_group->cached == BTRFS_CACHE_ERROR)
9147                         free_excluded_extents(info->extent_root, block_group);
9148
9149                 btrfs_remove_free_space_cache(block_group);
9150                 btrfs_put_block_group(block_group);
9151
9152                 spin_lock(&info->block_group_cache_lock);
9153         }
9154         spin_unlock(&info->block_group_cache_lock);
9155
9156         /* now that all the block groups are freed, go through and
9157          * free all the space_info structs.  This is only called during
9158          * the final stages of unmount, and so we know nobody is
9159          * using them.  We call synchronize_rcu() once before we start,
9160          * just to be on the safe side.
9161          */
9162         synchronize_rcu();
9163
9164         release_global_block_rsv(info);
9165
9166         while (!list_empty(&info->space_info)) {
9167                 int i;
9168
9169                 space_info = list_entry(info->space_info.next,
9170                                         struct btrfs_space_info,
9171                                         list);
9172                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9173                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9174                             space_info->bytes_reserved > 0 ||
9175                             space_info->bytes_may_use > 0)) {
9176                                 dump_space_info(space_info, 0, 0);
9177                         }
9178                 }
9179                 list_del(&space_info->list);
9180                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9181                         struct kobject *kobj;
9182                         kobj = space_info->block_group_kobjs[i];
9183                         space_info->block_group_kobjs[i] = NULL;
9184                         if (kobj) {
9185                                 kobject_del(kobj);
9186                                 kobject_put(kobj);
9187                         }
9188                 }
9189                 kobject_del(&space_info->kobj);
9190                 kobject_put(&space_info->kobj);
9191         }
9192         return 0;
9193 }
9194
9195 static void __link_block_group(struct btrfs_space_info *space_info,
9196                                struct btrfs_block_group_cache *cache)
9197 {
9198         int index = get_block_group_index(cache);
9199         bool first = false;
9200
9201         down_write(&space_info->groups_sem);
9202         if (list_empty(&space_info->block_groups[index]))
9203                 first = true;
9204         list_add_tail(&cache->list, &space_info->block_groups[index]);
9205         up_write(&space_info->groups_sem);
9206
9207         if (first) {
9208                 struct raid_kobject *rkobj;
9209                 int ret;
9210
9211                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9212                 if (!rkobj)
9213                         goto out_err;
9214                 rkobj->raid_type = index;
9215                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9216                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9217                                   "%s", get_raid_name(index));
9218                 if (ret) {
9219                         kobject_put(&rkobj->kobj);
9220                         goto out_err;
9221                 }
9222                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9223         }
9224
9225         return;
9226 out_err:
9227         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9228 }
9229
9230 static struct btrfs_block_group_cache *
9231 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9232 {
9233         struct btrfs_block_group_cache *cache;
9234
9235         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9236         if (!cache)
9237                 return NULL;
9238
9239         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9240                                         GFP_NOFS);
9241         if (!cache->free_space_ctl) {
9242                 kfree(cache);
9243                 return NULL;
9244         }
9245
9246         cache->key.objectid = start;
9247         cache->key.offset = size;
9248         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9249
9250         cache->sectorsize = root->sectorsize;
9251         cache->fs_info = root->fs_info;
9252         cache->full_stripe_len = btrfs_full_stripe_len(root,
9253                                                &root->fs_info->mapping_tree,
9254                                                start);
9255         atomic_set(&cache->count, 1);
9256         spin_lock_init(&cache->lock);
9257         init_rwsem(&cache->data_rwsem);
9258         INIT_LIST_HEAD(&cache->list);
9259         INIT_LIST_HEAD(&cache->cluster_list);
9260         INIT_LIST_HEAD(&cache->bg_list);
9261         INIT_LIST_HEAD(&cache->ro_list);
9262         INIT_LIST_HEAD(&cache->dirty_list);
9263         INIT_LIST_HEAD(&cache->io_list);
9264         btrfs_init_free_space_ctl(cache);
9265         atomic_set(&cache->trimming, 0);
9266
9267         return cache;
9268 }
9269
9270 int btrfs_read_block_groups(struct btrfs_root *root)
9271 {
9272         struct btrfs_path *path;
9273         int ret;
9274         struct btrfs_block_group_cache *cache;
9275         struct btrfs_fs_info *info = root->fs_info;
9276         struct btrfs_space_info *space_info;
9277         struct btrfs_key key;
9278         struct btrfs_key found_key;
9279         struct extent_buffer *leaf;
9280         int need_clear = 0;
9281         u64 cache_gen;
9282
9283         root = info->extent_root;
9284         key.objectid = 0;
9285         key.offset = 0;
9286         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9287         path = btrfs_alloc_path();
9288         if (!path)
9289                 return -ENOMEM;
9290         path->reada = 1;
9291
9292         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9293         if (btrfs_test_opt(root, SPACE_CACHE) &&
9294             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9295                 need_clear = 1;
9296         if (btrfs_test_opt(root, CLEAR_CACHE))
9297                 need_clear = 1;
9298
9299         while (1) {
9300                 ret = find_first_block_group(root, path, &key);
9301                 if (ret > 0)
9302                         break;
9303                 if (ret != 0)
9304                         goto error;
9305
9306                 leaf = path->nodes[0];
9307                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9308
9309                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9310                                                        found_key.offset);
9311                 if (!cache) {
9312                         ret = -ENOMEM;
9313                         goto error;
9314                 }
9315
9316                 if (need_clear) {
9317                         /*
9318                          * When we mount with old space cache, we need to
9319                          * set BTRFS_DC_CLEAR and set dirty flag.
9320                          *
9321                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9322                          *    truncate the old free space cache inode and
9323                          *    setup a new one.
9324                          * b) Setting 'dirty flag' makes sure that we flush
9325                          *    the new space cache info onto disk.
9326                          */
9327                         if (btrfs_test_opt(root, SPACE_CACHE))
9328                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9329                 }
9330
9331                 read_extent_buffer(leaf, &cache->item,
9332                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9333                                    sizeof(cache->item));
9334                 cache->flags = btrfs_block_group_flags(&cache->item);
9335
9336                 key.objectid = found_key.objectid + found_key.offset;
9337                 btrfs_release_path(path);
9338
9339                 /*
9340                  * We need to exclude the super stripes now so that the space
9341                  * info has super bytes accounted for, otherwise we'll think
9342                  * we have more space than we actually do.
9343                  */
9344                 ret = exclude_super_stripes(root, cache);
9345                 if (ret) {
9346                         /*
9347                          * We may have excluded something, so call this just in
9348                          * case.
9349                          */
9350                         free_excluded_extents(root, cache);
9351                         btrfs_put_block_group(cache);
9352                         goto error;
9353                 }
9354
9355                 /*
9356                  * check for two cases, either we are full, and therefore
9357                  * don't need to bother with the caching work since we won't
9358                  * find any space, or we are empty, and we can just add all
9359                  * the space in and be done with it.  This saves us _alot_ of
9360                  * time, particularly in the full case.
9361                  */
9362                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9363                         cache->last_byte_to_unpin = (u64)-1;
9364                         cache->cached = BTRFS_CACHE_FINISHED;
9365                         free_excluded_extents(root, cache);
9366                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9367                         cache->last_byte_to_unpin = (u64)-1;
9368                         cache->cached = BTRFS_CACHE_FINISHED;
9369                         add_new_free_space(cache, root->fs_info,
9370                                            found_key.objectid,
9371                                            found_key.objectid +
9372                                            found_key.offset);
9373                         free_excluded_extents(root, cache);
9374                 }
9375
9376                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9377                 if (ret) {
9378                         btrfs_remove_free_space_cache(cache);
9379                         btrfs_put_block_group(cache);
9380                         goto error;
9381                 }
9382
9383                 ret = update_space_info(info, cache->flags, found_key.offset,
9384                                         btrfs_block_group_used(&cache->item),
9385                                         &space_info);
9386                 if (ret) {
9387                         btrfs_remove_free_space_cache(cache);
9388                         spin_lock(&info->block_group_cache_lock);
9389                         rb_erase(&cache->cache_node,
9390                                  &info->block_group_cache_tree);
9391                         RB_CLEAR_NODE(&cache->cache_node);
9392                         spin_unlock(&info->block_group_cache_lock);
9393                         btrfs_put_block_group(cache);
9394                         goto error;
9395                 }
9396
9397                 cache->space_info = space_info;
9398                 spin_lock(&cache->space_info->lock);
9399                 cache->space_info->bytes_readonly += cache->bytes_super;
9400                 spin_unlock(&cache->space_info->lock);
9401
9402                 __link_block_group(space_info, cache);
9403
9404                 set_avail_alloc_bits(root->fs_info, cache->flags);
9405                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9406                         set_block_group_ro(cache, 1);
9407                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9408                         spin_lock(&info->unused_bgs_lock);
9409                         /* Should always be true but just in case. */
9410                         if (list_empty(&cache->bg_list)) {
9411                                 btrfs_get_block_group(cache);
9412                                 list_add_tail(&cache->bg_list,
9413                                               &info->unused_bgs);
9414                         }
9415                         spin_unlock(&info->unused_bgs_lock);
9416                 }
9417         }
9418
9419         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9420                 if (!(get_alloc_profile(root, space_info->flags) &
9421                       (BTRFS_BLOCK_GROUP_RAID10 |
9422                        BTRFS_BLOCK_GROUP_RAID1 |
9423                        BTRFS_BLOCK_GROUP_RAID5 |
9424                        BTRFS_BLOCK_GROUP_RAID6 |
9425                        BTRFS_BLOCK_GROUP_DUP)))
9426                         continue;
9427                 /*
9428                  * avoid allocating from un-mirrored block group if there are
9429                  * mirrored block groups.
9430                  */
9431                 list_for_each_entry(cache,
9432                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9433                                 list)
9434                         set_block_group_ro(cache, 1);
9435                 list_for_each_entry(cache,
9436                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9437                                 list)
9438                         set_block_group_ro(cache, 1);
9439         }
9440
9441         init_global_block_rsv(info);
9442         ret = 0;
9443 error:
9444         btrfs_free_path(path);
9445         return ret;
9446 }
9447
9448 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9449                                        struct btrfs_root *root)
9450 {
9451         struct btrfs_block_group_cache *block_group, *tmp;
9452         struct btrfs_root *extent_root = root->fs_info->extent_root;
9453         struct btrfs_block_group_item item;
9454         struct btrfs_key key;
9455         int ret = 0;
9456
9457         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9458                 if (ret)
9459                         goto next;
9460
9461                 spin_lock(&block_group->lock);
9462                 memcpy(&item, &block_group->item, sizeof(item));
9463                 memcpy(&key, &block_group->key, sizeof(key));
9464                 spin_unlock(&block_group->lock);
9465
9466                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9467                                         sizeof(item));
9468                 if (ret)
9469                         btrfs_abort_transaction(trans, extent_root, ret);
9470                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9471                                                key.objectid, key.offset);
9472                 if (ret)
9473                         btrfs_abort_transaction(trans, extent_root, ret);
9474 next:
9475                 list_del_init(&block_group->bg_list);
9476         }
9477 }
9478
9479 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9480                            struct btrfs_root *root, u64 bytes_used,
9481                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9482                            u64 size)
9483 {
9484         int ret;
9485         struct btrfs_root *extent_root;
9486         struct btrfs_block_group_cache *cache;
9487
9488         extent_root = root->fs_info->extent_root;
9489
9490         btrfs_set_log_full_commit(root->fs_info, trans);
9491
9492         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9493         if (!cache)
9494                 return -ENOMEM;
9495
9496         btrfs_set_block_group_used(&cache->item, bytes_used);
9497         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9498         btrfs_set_block_group_flags(&cache->item, type);
9499
9500         cache->flags = type;
9501         cache->last_byte_to_unpin = (u64)-1;
9502         cache->cached = BTRFS_CACHE_FINISHED;
9503         ret = exclude_super_stripes(root, cache);
9504         if (ret) {
9505                 /*
9506                  * We may have excluded something, so call this just in
9507                  * case.
9508                  */
9509                 free_excluded_extents(root, cache);
9510                 btrfs_put_block_group(cache);
9511                 return ret;
9512         }
9513
9514         add_new_free_space(cache, root->fs_info, chunk_offset,
9515                            chunk_offset + size);
9516
9517         free_excluded_extents(root, cache);
9518
9519         /*
9520          * Call to ensure the corresponding space_info object is created and
9521          * assigned to our block group, but don't update its counters just yet.
9522          * We want our bg to be added to the rbtree with its ->space_info set.
9523          */
9524         ret = update_space_info(root->fs_info, cache->flags, 0, 0,
9525                                 &cache->space_info);
9526         if (ret) {
9527                 btrfs_remove_free_space_cache(cache);
9528                 btrfs_put_block_group(cache);
9529                 return ret;
9530         }
9531
9532         ret = btrfs_add_block_group_cache(root->fs_info, cache);
9533         if (ret) {
9534                 btrfs_remove_free_space_cache(cache);
9535                 btrfs_put_block_group(cache);
9536                 return ret;
9537         }
9538
9539         /*
9540          * Now that our block group has its ->space_info set and is inserted in
9541          * the rbtree, update the space info's counters.
9542          */
9543         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
9544                                 &cache->space_info);
9545         if (ret) {
9546                 btrfs_remove_free_space_cache(cache);
9547                 spin_lock(&root->fs_info->block_group_cache_lock);
9548                 rb_erase(&cache->cache_node,
9549                          &root->fs_info->block_group_cache_tree);
9550                 RB_CLEAR_NODE(&cache->cache_node);
9551                 spin_unlock(&root->fs_info->block_group_cache_lock);
9552                 btrfs_put_block_group(cache);
9553                 return ret;
9554         }
9555         update_global_block_rsv(root->fs_info);
9556
9557         spin_lock(&cache->space_info->lock);
9558         cache->space_info->bytes_readonly += cache->bytes_super;
9559         spin_unlock(&cache->space_info->lock);
9560
9561         __link_block_group(cache->space_info, cache);
9562
9563         list_add_tail(&cache->bg_list, &trans->new_bgs);
9564
9565         set_avail_alloc_bits(extent_root->fs_info, type);
9566
9567         return 0;
9568 }
9569
9570 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9571 {
9572         u64 extra_flags = chunk_to_extended(flags) &
9573                                 BTRFS_EXTENDED_PROFILE_MASK;
9574
9575         write_seqlock(&fs_info->profiles_lock);
9576         if (flags & BTRFS_BLOCK_GROUP_DATA)
9577                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9578         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9579                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9580         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9581                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9582         write_sequnlock(&fs_info->profiles_lock);
9583 }
9584
9585 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9586                              struct btrfs_root *root, u64 group_start,
9587                              struct extent_map *em)
9588 {
9589         struct btrfs_path *path;
9590         struct btrfs_block_group_cache *block_group;
9591         struct btrfs_free_cluster *cluster;
9592         struct btrfs_root *tree_root = root->fs_info->tree_root;
9593         struct btrfs_key key;
9594         struct inode *inode;
9595         struct kobject *kobj = NULL;
9596         int ret;
9597         int index;
9598         int factor;
9599         struct btrfs_caching_control *caching_ctl = NULL;
9600         bool remove_em;
9601
9602         root = root->fs_info->extent_root;
9603
9604         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9605         BUG_ON(!block_group);
9606         BUG_ON(!block_group->ro);
9607
9608         /*
9609          * Free the reserved super bytes from this block group before
9610          * remove it.
9611          */
9612         free_excluded_extents(root, block_group);
9613
9614         memcpy(&key, &block_group->key, sizeof(key));
9615         index = get_block_group_index(block_group);
9616         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9617                                   BTRFS_BLOCK_GROUP_RAID1 |
9618                                   BTRFS_BLOCK_GROUP_RAID10))
9619                 factor = 2;
9620         else
9621                 factor = 1;
9622
9623         /* make sure this block group isn't part of an allocation cluster */
9624         cluster = &root->fs_info->data_alloc_cluster;
9625         spin_lock(&cluster->refill_lock);
9626         btrfs_return_cluster_to_free_space(block_group, cluster);
9627         spin_unlock(&cluster->refill_lock);
9628
9629         /*
9630          * make sure this block group isn't part of a metadata
9631          * allocation cluster
9632          */
9633         cluster = &root->fs_info->meta_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         path = btrfs_alloc_path();
9639         if (!path) {
9640                 ret = -ENOMEM;
9641                 goto out;
9642         }
9643
9644         /*
9645          * get the inode first so any iput calls done for the io_list
9646          * aren't the final iput (no unlinks allowed now)
9647          */
9648         inode = lookup_free_space_inode(tree_root, block_group, path);
9649
9650         mutex_lock(&trans->transaction->cache_write_mutex);
9651         /*
9652          * make sure our free spache cache IO is done before remove the
9653          * free space inode
9654          */
9655         spin_lock(&trans->transaction->dirty_bgs_lock);
9656         if (!list_empty(&block_group->io_list)) {
9657                 list_del_init(&block_group->io_list);
9658
9659                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
9660
9661                 spin_unlock(&trans->transaction->dirty_bgs_lock);
9662                 btrfs_wait_cache_io(root, trans, block_group,
9663                                     &block_group->io_ctl, path,
9664                                     block_group->key.objectid);
9665                 btrfs_put_block_group(block_group);
9666                 spin_lock(&trans->transaction->dirty_bgs_lock);
9667         }
9668
9669         if (!list_empty(&block_group->dirty_list)) {
9670                 list_del_init(&block_group->dirty_list);
9671                 btrfs_put_block_group(block_group);
9672         }
9673         spin_unlock(&trans->transaction->dirty_bgs_lock);
9674         mutex_unlock(&trans->transaction->cache_write_mutex);
9675
9676         if (!IS_ERR(inode)) {
9677                 ret = btrfs_orphan_add(trans, inode);
9678                 if (ret) {
9679                         btrfs_add_delayed_iput(inode);
9680                         goto out;
9681                 }
9682                 clear_nlink(inode);
9683                 /* One for the block groups ref */
9684                 spin_lock(&block_group->lock);
9685                 if (block_group->iref) {
9686                         block_group->iref = 0;
9687                         block_group->inode = NULL;
9688                         spin_unlock(&block_group->lock);
9689                         iput(inode);
9690                 } else {
9691                         spin_unlock(&block_group->lock);
9692                 }
9693                 /* One for our lookup ref */
9694                 btrfs_add_delayed_iput(inode);
9695         }
9696
9697         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9698         key.offset = block_group->key.objectid;
9699         key.type = 0;
9700
9701         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9702         if (ret < 0)
9703                 goto out;
9704         if (ret > 0)
9705                 btrfs_release_path(path);
9706         if (ret == 0) {
9707                 ret = btrfs_del_item(trans, tree_root, path);
9708                 if (ret)
9709                         goto out;
9710                 btrfs_release_path(path);
9711         }
9712
9713         spin_lock(&root->fs_info->block_group_cache_lock);
9714         rb_erase(&block_group->cache_node,
9715                  &root->fs_info->block_group_cache_tree);
9716         RB_CLEAR_NODE(&block_group->cache_node);
9717
9718         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9719                 root->fs_info->first_logical_byte = (u64)-1;
9720         spin_unlock(&root->fs_info->block_group_cache_lock);
9721
9722         down_write(&block_group->space_info->groups_sem);
9723         /*
9724          * we must use list_del_init so people can check to see if they
9725          * are still on the list after taking the semaphore
9726          */
9727         list_del_init(&block_group->list);
9728         if (list_empty(&block_group->space_info->block_groups[index])) {
9729                 kobj = block_group->space_info->block_group_kobjs[index];
9730                 block_group->space_info->block_group_kobjs[index] = NULL;
9731                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9732         }
9733         up_write(&block_group->space_info->groups_sem);
9734         if (kobj) {
9735                 kobject_del(kobj);
9736                 kobject_put(kobj);
9737         }
9738
9739         if (block_group->has_caching_ctl)
9740                 caching_ctl = get_caching_control(block_group);
9741         if (block_group->cached == BTRFS_CACHE_STARTED)
9742                 wait_block_group_cache_done(block_group);
9743         if (block_group->has_caching_ctl) {
9744                 down_write(&root->fs_info->commit_root_sem);
9745                 if (!caching_ctl) {
9746                         struct btrfs_caching_control *ctl;
9747
9748                         list_for_each_entry(ctl,
9749                                     &root->fs_info->caching_block_groups, list)
9750                                 if (ctl->block_group == block_group) {
9751                                         caching_ctl = ctl;
9752                                         atomic_inc(&caching_ctl->count);
9753                                         break;
9754                                 }
9755                 }
9756                 if (caching_ctl)
9757                         list_del_init(&caching_ctl->list);
9758                 up_write(&root->fs_info->commit_root_sem);
9759                 if (caching_ctl) {
9760                         /* Once for the caching bgs list and once for us. */
9761                         put_caching_control(caching_ctl);
9762                         put_caching_control(caching_ctl);
9763                 }
9764         }
9765
9766         spin_lock(&trans->transaction->dirty_bgs_lock);
9767         if (!list_empty(&block_group->dirty_list)) {
9768                 WARN_ON(1);
9769         }
9770         if (!list_empty(&block_group->io_list)) {
9771                 WARN_ON(1);
9772         }
9773         spin_unlock(&trans->transaction->dirty_bgs_lock);
9774         btrfs_remove_free_space_cache(block_group);
9775
9776         spin_lock(&block_group->space_info->lock);
9777         list_del_init(&block_group->ro_list);
9778
9779         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
9780                 WARN_ON(block_group->space_info->total_bytes
9781                         < block_group->key.offset);
9782                 WARN_ON(block_group->space_info->bytes_readonly
9783                         < block_group->key.offset);
9784                 WARN_ON(block_group->space_info->disk_total
9785                         < block_group->key.offset * factor);
9786         }
9787         block_group->space_info->total_bytes -= block_group->key.offset;
9788         block_group->space_info->bytes_readonly -= block_group->key.offset;
9789         block_group->space_info->disk_total -= block_group->key.offset * factor;
9790
9791         spin_unlock(&block_group->space_info->lock);
9792
9793         memcpy(&key, &block_group->key, sizeof(key));
9794
9795         lock_chunks(root);
9796         if (!list_empty(&em->list)) {
9797                 /* We're in the transaction->pending_chunks list. */
9798                 free_extent_map(em);
9799         }
9800         spin_lock(&block_group->lock);
9801         block_group->removed = 1;
9802         /*
9803          * At this point trimming can't start on this block group, because we
9804          * removed the block group from the tree fs_info->block_group_cache_tree
9805          * so no one can't find it anymore and even if someone already got this
9806          * block group before we removed it from the rbtree, they have already
9807          * incremented block_group->trimming - if they didn't, they won't find
9808          * any free space entries because we already removed them all when we
9809          * called btrfs_remove_free_space_cache().
9810          *
9811          * And we must not remove the extent map from the fs_info->mapping_tree
9812          * to prevent the same logical address range and physical device space
9813          * ranges from being reused for a new block group. This is because our
9814          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
9815          * completely transactionless, so while it is trimming a range the
9816          * currently running transaction might finish and a new one start,
9817          * allowing for new block groups to be created that can reuse the same
9818          * physical device locations unless we take this special care.
9819          */
9820         remove_em = (atomic_read(&block_group->trimming) == 0);
9821         /*
9822          * Make sure a trimmer task always sees the em in the pinned_chunks list
9823          * if it sees block_group->removed == 1 (needs to lock block_group->lock
9824          * before checking block_group->removed).
9825          */
9826         if (!remove_em) {
9827                 /*
9828                  * Our em might be in trans->transaction->pending_chunks which
9829                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
9830                  * and so is the fs_info->pinned_chunks list.
9831                  *
9832                  * So at this point we must be holding the chunk_mutex to avoid
9833                  * any races with chunk allocation (more specifically at
9834                  * volumes.c:contains_pending_extent()), to ensure it always
9835                  * sees the em, either in the pending_chunks list or in the
9836                  * pinned_chunks list.
9837                  */
9838                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
9839         }
9840         spin_unlock(&block_group->lock);
9841
9842         if (remove_em) {
9843                 struct extent_map_tree *em_tree;
9844
9845                 em_tree = &root->fs_info->mapping_tree.map_tree;
9846                 write_lock(&em_tree->lock);
9847                 /*
9848                  * The em might be in the pending_chunks list, so make sure the
9849                  * chunk mutex is locked, since remove_extent_mapping() will
9850                  * delete us from that list.
9851                  */
9852                 remove_extent_mapping(em_tree, em);
9853                 write_unlock(&em_tree->lock);
9854                 /* once for the tree */
9855                 free_extent_map(em);
9856         }
9857
9858         unlock_chunks(root);
9859
9860         btrfs_put_block_group(block_group);
9861         btrfs_put_block_group(block_group);
9862
9863         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9864         if (ret > 0)
9865                 ret = -EIO;
9866         if (ret < 0)
9867                 goto out;
9868
9869         ret = btrfs_del_item(trans, root, path);
9870 out:
9871         btrfs_free_path(path);
9872         return ret;
9873 }
9874
9875 /*
9876  * Process the unused_bgs list and remove any that don't have any allocated
9877  * space inside of them.
9878  */
9879 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
9880 {
9881         struct btrfs_block_group_cache *block_group;
9882         struct btrfs_space_info *space_info;
9883         struct btrfs_root *root = fs_info->extent_root;
9884         struct btrfs_trans_handle *trans;
9885         int ret = 0;
9886
9887         if (!fs_info->open)
9888                 return;
9889
9890         spin_lock(&fs_info->unused_bgs_lock);
9891         while (!list_empty(&fs_info->unused_bgs)) {
9892                 u64 start, end;
9893
9894                 block_group = list_first_entry(&fs_info->unused_bgs,
9895                                                struct btrfs_block_group_cache,
9896                                                bg_list);
9897                 space_info = block_group->space_info;
9898                 list_del_init(&block_group->bg_list);
9899                 if (ret || btrfs_mixed_space_info(space_info)) {
9900                         btrfs_put_block_group(block_group);
9901                         continue;
9902                 }
9903                 spin_unlock(&fs_info->unused_bgs_lock);
9904
9905                 mutex_lock(&root->fs_info->delete_unused_bgs_mutex);
9906
9907                 /* Don't want to race with allocators so take the groups_sem */
9908                 down_write(&space_info->groups_sem);
9909                 spin_lock(&block_group->lock);
9910                 if (block_group->reserved ||
9911                     btrfs_block_group_used(&block_group->item) ||
9912                     block_group->ro) {
9913                         /*
9914                          * We want to bail if we made new allocations or have
9915                          * outstanding allocations in this block group.  We do
9916                          * the ro check in case balance is currently acting on
9917                          * this block group.
9918                          */
9919                         spin_unlock(&block_group->lock);
9920                         up_write(&space_info->groups_sem);
9921                         goto next;
9922                 }
9923                 spin_unlock(&block_group->lock);
9924
9925                 /* We don't want to force the issue, only flip if it's ok. */
9926                 ret = set_block_group_ro(block_group, 0);
9927                 up_write(&space_info->groups_sem);
9928                 if (ret < 0) {
9929                         ret = 0;
9930                         goto next;
9931                 }
9932
9933                 /*
9934                  * Want to do this before we do anything else so we can recover
9935                  * properly if we fail to join the transaction.
9936                  */
9937                 /* 1 for btrfs_orphan_reserve_metadata() */
9938                 trans = btrfs_start_transaction(root, 1);
9939                 if (IS_ERR(trans)) {
9940                         btrfs_set_block_group_rw(root, block_group);
9941                         ret = PTR_ERR(trans);
9942                         goto next;
9943                 }
9944
9945                 /*
9946                  * We could have pending pinned extents for this block group,
9947                  * just delete them, we don't care about them anymore.
9948                  */
9949                 start = block_group->key.objectid;
9950                 end = start + block_group->key.offset - 1;
9951                 /*
9952                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
9953                  * btrfs_finish_extent_commit(). If we are at transaction N,
9954                  * another task might be running finish_extent_commit() for the
9955                  * previous transaction N - 1, and have seen a range belonging
9956                  * to the block group in freed_extents[] before we were able to
9957                  * clear the whole block group range from freed_extents[]. This
9958                  * means that task can lookup for the block group after we
9959                  * unpinned it from freed_extents[] and removed it, leading to
9960                  * a BUG_ON() at btrfs_unpin_extent_range().
9961                  */
9962                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
9963                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
9964                                   EXTENT_DIRTY, GFP_NOFS);
9965                 if (ret) {
9966                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9967                         btrfs_set_block_group_rw(root, block_group);
9968                         goto end_trans;
9969                 }
9970                 ret = clear_extent_bits(&fs_info->freed_extents[1], 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                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9978
9979                 /* Reset pinned so btrfs_put_block_group doesn't complain */
9980                 spin_lock(&space_info->lock);
9981                 spin_lock(&block_group->lock);
9982
9983                 space_info->bytes_pinned -= block_group->pinned;
9984                 space_info->bytes_readonly += block_group->pinned;
9985                 percpu_counter_add(&space_info->total_bytes_pinned,
9986                                    -block_group->pinned);
9987                 block_group->pinned = 0;
9988
9989                 spin_unlock(&block_group->lock);
9990                 spin_unlock(&space_info->lock);
9991
9992                 /*
9993                  * Btrfs_remove_chunk will abort the transaction if things go
9994                  * horribly wrong.
9995                  */
9996                 ret = btrfs_remove_chunk(trans, root,
9997                                          block_group->key.objectid);
9998 end_trans:
9999                 btrfs_end_transaction(trans, root);
10000 next:
10001                 mutex_unlock(&root->fs_info->delete_unused_bgs_mutex);
10002                 btrfs_put_block_group(block_group);
10003                 spin_lock(&fs_info->unused_bgs_lock);
10004         }
10005         spin_unlock(&fs_info->unused_bgs_lock);
10006 }
10007
10008 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10009 {
10010         struct btrfs_space_info *space_info;
10011         struct btrfs_super_block *disk_super;
10012         u64 features;
10013         u64 flags;
10014         int mixed = 0;
10015         int ret;
10016
10017         disk_super = fs_info->super_copy;
10018         if (!btrfs_super_root(disk_super))
10019                 return 1;
10020
10021         features = btrfs_super_incompat_flags(disk_super);
10022         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10023                 mixed = 1;
10024
10025         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10026         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10027         if (ret)
10028                 goto out;
10029
10030         if (mixed) {
10031                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10032                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10033         } else {
10034                 flags = BTRFS_BLOCK_GROUP_METADATA;
10035                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10036                 if (ret)
10037                         goto out;
10038
10039                 flags = BTRFS_BLOCK_GROUP_DATA;
10040                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10041         }
10042 out:
10043         return ret;
10044 }
10045
10046 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10047 {
10048         return unpin_extent_range(root, start, end, false);
10049 }
10050
10051 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10052 {
10053         struct btrfs_fs_info *fs_info = root->fs_info;
10054         struct btrfs_block_group_cache *cache = NULL;
10055         u64 group_trimmed;
10056         u64 start;
10057         u64 end;
10058         u64 trimmed = 0;
10059         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10060         int ret = 0;
10061
10062         /*
10063          * try to trim all FS space, our block group may start from non-zero.
10064          */
10065         if (range->len == total_bytes)
10066                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10067         else
10068                 cache = btrfs_lookup_block_group(fs_info, range->start);
10069
10070         while (cache) {
10071                 if (cache->key.objectid >= (range->start + range->len)) {
10072                         btrfs_put_block_group(cache);
10073                         break;
10074                 }
10075
10076                 start = max(range->start, cache->key.objectid);
10077                 end = min(range->start + range->len,
10078                                 cache->key.objectid + cache->key.offset);
10079
10080                 if (end - start >= range->minlen) {
10081                         if (!block_group_cache_done(cache)) {
10082                                 ret = cache_block_group(cache, 0);
10083                                 if (ret) {
10084                                         btrfs_put_block_group(cache);
10085                                         break;
10086                                 }
10087                                 ret = wait_block_group_cache_done(cache);
10088                                 if (ret) {
10089                                         btrfs_put_block_group(cache);
10090                                         break;
10091                                 }
10092                         }
10093                         ret = btrfs_trim_block_group(cache,
10094                                                      &group_trimmed,
10095                                                      start,
10096                                                      end,
10097                                                      range->minlen);
10098
10099                         trimmed += group_trimmed;
10100                         if (ret) {
10101                                 btrfs_put_block_group(cache);
10102                                 break;
10103                         }
10104                 }
10105
10106                 cache = next_block_group(fs_info->tree_root, cache);
10107         }
10108
10109         range->len = trimmed;
10110         return ret;
10111 }
10112
10113 /*
10114  * btrfs_{start,end}_write_no_snapshoting() are similar to
10115  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10116  * data into the page cache through nocow before the subvolume is snapshoted,
10117  * but flush the data into disk after the snapshot creation, or to prevent
10118  * operations while snapshoting is ongoing and that cause the snapshot to be
10119  * inconsistent (writes followed by expanding truncates for example).
10120  */
10121 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10122 {
10123         percpu_counter_dec(&root->subv_writers->counter);
10124         /*
10125          * Make sure counter is updated before we wake up
10126          * waiters.
10127          */
10128         smp_mb();
10129         if (waitqueue_active(&root->subv_writers->wait))
10130                 wake_up(&root->subv_writers->wait);
10131 }
10132
10133 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10134 {
10135         if (atomic_read(&root->will_be_snapshoted))
10136                 return 0;
10137
10138         percpu_counter_inc(&root->subv_writers->counter);
10139         /*
10140          * Make sure counter is updated before we check for snapshot creation.
10141          */
10142         smp_mb();
10143         if (atomic_read(&root->will_be_snapshoted)) {
10144                 btrfs_end_write_no_snapshoting(root);
10145                 return 0;
10146         }
10147         return 1;
10148 }