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 "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "raid56.h"
  35 #include "locking.h"
  36 #include "free-space-cache.h"
  37 #include "math.h"
  38 #include "sysfs.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_root *root,
  78                               u64 bytenr, u64 num_bytes, int alloc);
  79 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  80                                 struct btrfs_root *root,
  81                                 u64 bytenr, u64 num_bytes, u64 parent,
  82                                 u64 root_objectid, u64 owner_objectid,
  83                                 u64 owner_offset, int refs_to_drop,
  84                                 struct btrfs_delayed_extent_op *extra_op);
  85 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  86                                     struct extent_buffer *leaf,
  87                                     struct btrfs_extent_item *ei);
  88 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  89                                       struct btrfs_root *root,
  90                                       u64 parent, u64 root_objectid,
  91                                       u64 flags, u64 owner, u64 offset,
  92                                       struct btrfs_key *ins, int ref_mod);
  93 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  94                                      struct btrfs_root *root,
  95                                      u64 parent, u64 root_objectid,
  96                                      u64 flags, struct btrfs_disk_key *key,
  97                                      int level, struct btrfs_key *ins);
  98 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  99                           struct btrfs_root *extent_root, u64 flags,
 100                           int force);
 101 static int find_next_key(struct btrfs_path *path, int level,
 102                          struct btrfs_key *key);
 103 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 104                             int dump_block_groups);
 105 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 106                                        u64 num_bytes, int reserve);
 107 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 108                                u64 num_bytes);
 109 int btrfs_pin_extent(struct btrfs_root *root,
 110                      u64 bytenr, u64 num_bytes, int reserved);
 111
 112 static noinline int
 113 block_group_cache_done(struct btrfs_block_group_cache *cache)
 114 {
 115         smp_mb();
 116         return cache->cached == BTRFS_CACHE_FINISHED ||
 117                 cache->cached == BTRFS_CACHE_ERROR;
 118 }
 119
 120 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 121 {
 122         return (cache->flags & bits) == bits;
 123 }
 124
 125 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 126 {
 127         atomic_inc(&cache->count);
 128 }
 129
 130 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 131 {
 132         if (atomic_dec_and_test(&cache->count)) {
 133                 WARN_ON(cache->pinned > 0);
 134                 WARN_ON(cache->reserved > 0);
 135                 kfree(cache->free_space_ctl);
 136                 kfree(cache);
 137         }
 138 }
 139
 140 /*
 141  * this adds the block group to the fs_info rb tree for the block group
 142  * cache
 143  */
 144 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 145                                 struct btrfs_block_group_cache *block_group)
 146 {
 147         struct rb_node **p;
 148         struct rb_node *parent = NULL;
 149         struct btrfs_block_group_cache *cache;
 150
 151         spin_lock(&info->block_group_cache_lock);
 152         p = &info->block_group_cache_tree.rb_node;
 153
 154         while (*p) {
 155                 parent = *p;
 156                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 157                                  cache_node);
 158                 if (block_group->key.objectid < cache->key.objectid) {
 159                         p = &(*p)->rb_left;
 160                 } else if (block_group->key.objectid > cache->key.objectid) {
 161                         p = &(*p)->rb_right;
 162                 } else {
 163                         spin_unlock(&info->block_group_cache_lock);
 164                         return -EEXIST;
 165                 }
 166         }
 167
 168         rb_link_node(&block_group->cache_node, parent, p);
 169         rb_insert_color(&block_group->cache_node,
 170                         &info->block_group_cache_tree);
 171
 172         if (info->first_logical_byte > block_group->key.objectid)
 173                 info->first_logical_byte = block_group->key.objectid;
 174
 175         spin_unlock(&info->block_group_cache_lock);
 176
 177         return 0;
 178 }
 179
 180 /*
 181  * This will return the block group at or after bytenr if contains is 0, else
 182  * it will return the block group that contains the bytenr
 183  */
 184 static struct btrfs_block_group_cache *
 185 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 186                               int contains)
 187 {
 188         struct btrfs_block_group_cache *cache, *ret = NULL;
 189         struct rb_node *n;
 190         u64 end, start;
 191
 192         spin_lock(&info->block_group_cache_lock);
 193         n = info->block_group_cache_tree.rb_node;
 194
 195         while (n) {
 196                 cache = rb_entry(n, struct btrfs_block_group_cache,
 197                                  cache_node);
 198                 end = cache->key.objectid + cache->key.offset - 1;
 199                 start = cache->key.objectid;
 200
 201                 if (bytenr < start) {
 202                         if (!contains && (!ret || start < ret->key.objectid))
 203                                 ret = cache;
 204                         n = n->rb_left;
 205                 } else if (bytenr > start) {
 206                         if (contains && bytenr <= end) {
 207                                 ret = cache;
 208                                 break;
 209                         }
 210                         n = n->rb_right;
 211                 } else {
 212                         ret = cache;
 213                         break;
 214                 }
 215         }
 216         if (ret) {
 217                 btrfs_get_block_group(ret);
 218                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 219                         info->first_logical_byte = ret->key.objectid;
 220         }
 221         spin_unlock(&info->block_group_cache_lock);
 222
 223         return ret;
 224 }
 225
 226 static int add_excluded_extent(struct btrfs_root *root,
 227                                u64 start, u64 num_bytes)
 228 {
 229         u64 end = start + num_bytes - 1;
 230         set_extent_bits(&root->fs_info->freed_extents[0],
 231                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 232         set_extent_bits(&root->fs_info->freed_extents[1],
 233                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 234         return 0;
 235 }
 236
 237 static void free_excluded_extents(struct btrfs_root *root,
 238                                   struct btrfs_block_group_cache *cache)
 239 {
 240         u64 start, end;
 241
 242         start = cache->key.objectid;
 243         end = start + cache->key.offset - 1;
 244
 245         clear_extent_bits(&root->fs_info->freed_extents[0],
 246                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 247         clear_extent_bits(&root->fs_info->freed_extents[1],
 248                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 249 }
 250
 251 static int exclude_super_stripes(struct btrfs_root *root,
 252                                  struct btrfs_block_group_cache *cache)
 253 {
 254         u64 bytenr;
 255         u64 *logical;
 256         int stripe_len;
 257         int i, nr, ret;
 258
 259         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 260                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 261                 cache->bytes_super += stripe_len;
 262                 ret = add_excluded_extent(root, cache->key.objectid,
 263                                           stripe_len);
 264                 if (ret)
 265                         return ret;
 266         }
 267
 268         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 269                 bytenr = btrfs_sb_offset(i);
 270                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 271                                        cache->key.objectid, bytenr,
 272                                        0, &logical, &nr, &stripe_len);
 273                 if (ret)
 274                         return ret;
 275
 276                 while (nr--) {
 277                         u64 start, len;
 278
 279                         if (logical[nr] > cache->key.objectid +
 280                             cache->key.offset)
 281                                 continue;
 282
 283                         if (logical[nr] + stripe_len <= cache->key.objectid)
 284                                 continue;
 285
 286                         start = logical[nr];
 287                         if (start < cache->key.objectid) {
 288                                 start = cache->key.objectid;
 289                                 len = (logical[nr] + stripe_len) - start;
 290                         } else {
 291                                 len = min_t(u64, stripe_len,
 292                                             cache->key.objectid +
 293                                             cache->key.offset - start);
 294                         }
 295
 296                         cache->bytes_super += len;
 297                         ret = add_excluded_extent(root, start, len);
 298                         if (ret) {
 299                                 kfree(logical);
 300                                 return ret;
 301                         }
 302                 }
 303
 304                 kfree(logical);
 305         }
 306         return 0;
 307 }
 308
 309 static struct btrfs_caching_control *
 310 get_caching_control(struct btrfs_block_group_cache *cache)
 311 {
 312         struct btrfs_caching_control *ctl;
 313
 314         spin_lock(&cache->lock);
 315         if (cache->cached != BTRFS_CACHE_STARTED) {
 316                 spin_unlock(&cache->lock);
 317                 return NULL;
 318         }
 319
 320         /* We're loading it the fast way, so we don't have a caching_ctl. */
 321         if (!cache->caching_ctl) {
 322                 spin_unlock(&cache->lock);
 323                 return NULL;
 324         }
 325
 326         ctl = cache->caching_ctl;
 327         atomic_inc(&ctl->count);
 328         spin_unlock(&cache->lock);
 329         return ctl;
 330 }
 331
 332 static void put_caching_control(struct btrfs_caching_control *ctl)
 333 {
 334         if (atomic_dec_and_test(&ctl->count))
 335                 kfree(ctl);
 336 }
 337
 338 /*
 339  * this is only called by cache_block_group, since we could have freed extents
 340  * we need to check the pinned_extents for any extents that can't be used yet
 341  * since their free space will be released as soon as the transaction commits.
 342  */
 343 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 344                               struct btrfs_fs_info *info, u64 start, u64 end)
 345 {
 346         u64 extent_start, extent_end, size, total_added = 0;
 347         int ret;
 348
 349         while (start < end) {
 350                 ret = find_first_extent_bit(info->pinned_extents, start,
 351                                             &extent_start, &extent_end,
 352                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 353                                             NULL);
 354                 if (ret)
 355                         break;
 356
 357                 if (extent_start <= start) {
 358                         start = extent_end + 1;
 359                 } else if (extent_start > start && extent_start < end) {
 360                         size = extent_start - start;
 361                         total_added += size;
 362                         ret = btrfs_add_free_space(block_group, start,
 363                                                    size);
 364                         BUG_ON(ret); /* -ENOMEM or logic error */
 365                         start = extent_end + 1;
 366                 } else {
 367                         break;
 368                 }
 369         }
 370
 371         if (start < end) {
 372                 size = end - start;
 373                 total_added += size;
 374                 ret = btrfs_add_free_space(block_group, start, size);
 375                 BUG_ON(ret); /* -ENOMEM or logic error */
 376         }
 377
 378         return total_added;
 379 }
 380
 381 static noinline void caching_thread(struct btrfs_work *work)
 382 {
 383         struct btrfs_block_group_cache *block_group;
 384         struct btrfs_fs_info *fs_info;
 385         struct btrfs_caching_control *caching_ctl;
 386         struct btrfs_root *extent_root;
 387         struct btrfs_path *path;
 388         struct extent_buffer *leaf;
 389         struct btrfs_key key;
 390         u64 total_found = 0;
 391         u64 last = 0;
 392         u32 nritems;
 393         int ret = -ENOMEM;
 394
 395         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 396         block_group = caching_ctl->block_group;
 397         fs_info = block_group->fs_info;
 398         extent_root = fs_info->extent_root;
 399
 400         path = btrfs_alloc_path();
 401         if (!path)
 402                 goto out;
 403
 404         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 405
 406         /*
 407          * We don't want to deadlock with somebody trying to allocate a new
 408          * extent for the extent root while also trying to search the extent
 409          * root to add free space.  So we skip locking and search the commit
 410          * root, since its read-only
 411          */
 412         path->skip_locking = 1;
 413         path->search_commit_root = 1;
 414         path->reada = 1;
 415
 416         key.objectid = last;
 417         key.offset = 0;
 418         key.type = BTRFS_EXTENT_ITEM_KEY;
 419 again:
 420         mutex_lock(&caching_ctl->mutex);
 421         /* need to make sure the commit_root doesn't disappear */
 422         down_read(&fs_info->commit_root_sem);
 423
 424 next:
 425         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 426         if (ret < 0)
 427                 goto err;
 428
 429         leaf = path->nodes[0];
 430         nritems = btrfs_header_nritems(leaf);
 431
 432         while (1) {
 433                 if (btrfs_fs_closing(fs_info) > 1) {
 434                         last = (u64)-1;
 435                         break;
 436                 }
 437
 438                 if (path->slots[0] < nritems) {
 439                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 440                 } else {
 441                         ret = find_next_key(path, 0, &key);
 442                         if (ret)
 443                                 break;
 444
 445                         if (need_resched() ||
 446                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 447                                 caching_ctl->progress = last;
 448                                 btrfs_release_path(path);
 449                                 up_read(&fs_info->commit_root_sem);
 450                                 mutex_unlock(&caching_ctl->mutex);
 451                                 cond_resched();
 452                                 goto again;
 453                         }
 454
 455                         ret = btrfs_next_leaf(extent_root, path);
 456                         if (ret < 0)
 457                                 goto err;
 458                         if (ret)
 459                                 break;
 460                         leaf = path->nodes[0];
 461                         nritems = btrfs_header_nritems(leaf);
 462                         continue;
 463                 }
 464
 465                 if (key.objectid < last) {
 466                         key.objectid = last;
 467                         key.offset = 0;
 468                         key.type = BTRFS_EXTENT_ITEM_KEY;
 469
 470                         caching_ctl->progress = last;
 471                         btrfs_release_path(path);
 472                         goto next;
 473                 }
 474
 475                 if (key.objectid < block_group->key.objectid) {
 476                         path->slots[0]++;
 477                         continue;
 478                 }
 479
 480                 if (key.objectid >= block_group->key.objectid +
 481                     block_group->key.offset)
 482                         break;
 483
 484                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 485                     key.type == BTRFS_METADATA_ITEM_KEY) {
 486                         total_found += add_new_free_space(block_group,
 487                                                           fs_info, last,
 488                                                           key.objectid);
 489                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 490                                 last = key.objectid +
 491                                         fs_info->tree_root->leafsize;
 492                         else
 493                                 last = key.objectid + key.offset;
 494
 495                         if (total_found > (1024 * 1024 * 2)) {
 496                                 total_found = 0;
 497                                 wake_up(&caching_ctl->wait);
 498                         }
 499                 }
 500                 path->slots[0]++;
 501         }
 502         ret = 0;
 503
 504         total_found += add_new_free_space(block_group, fs_info, last,
 505                                           block_group->key.objectid +
 506                                           block_group->key.offset);
 507         caching_ctl->progress = (u64)-1;
 508
 509         spin_lock(&block_group->lock);
 510         block_group->caching_ctl = NULL;
 511         block_group->cached = BTRFS_CACHE_FINISHED;
 512         spin_unlock(&block_group->lock);
 513
 514 err:
 515         btrfs_free_path(path);
 516         up_read(&fs_info->commit_root_sem);
 517
 518         free_excluded_extents(extent_root, block_group);
 519
 520         mutex_unlock(&caching_ctl->mutex);
 521 out:
 522         if (ret) {
 523                 spin_lock(&block_group->lock);
 524                 block_group->caching_ctl = NULL;
 525                 block_group->cached = BTRFS_CACHE_ERROR;
 526                 spin_unlock(&block_group->lock);
 527         }
 528         wake_up(&caching_ctl->wait);
 529
 530         put_caching_control(caching_ctl);
 531         btrfs_put_block_group(block_group);
 532 }
 533
 534 static int cache_block_group(struct btrfs_block_group_cache *cache,
 535                              int load_cache_only)
 536 {
 537         DEFINE_WAIT(wait);
 538         struct btrfs_fs_info *fs_info = cache->fs_info;
 539         struct btrfs_caching_control *caching_ctl;
 540         int ret = 0;
 541
 542         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 543         if (!caching_ctl)
 544                 return -ENOMEM;
 545
 546         INIT_LIST_HEAD(&caching_ctl->list);
 547         mutex_init(&caching_ctl->mutex);
 548         init_waitqueue_head(&caching_ctl->wait);
 549         caching_ctl->block_group = cache;
 550         caching_ctl->progress = cache->key.objectid;
 551         atomic_set(&caching_ctl->count, 1);
 552         btrfs_init_work(&caching_ctl->work, caching_thread, NULL, NULL);
 553
 554         spin_lock(&cache->lock);
 555         /*
 556          * This should be a rare occasion, but this could happen I think in the
 557          * case where one thread starts to load the space cache info, and then
 558          * some other thread starts a transaction commit which tries to do an
 559          * allocation while the other thread is still loading the space cache
 560          * info.  The previous loop should have kept us from choosing this block
 561          * group, but if we've moved to the state where we will wait on caching
 562          * block groups we need to first check if we're doing a fast load here,
 563          * so we can wait for it to finish, otherwise we could end up allocating
 564          * from a block group who's cache gets evicted for one reason or
 565          * another.
 566          */
 567         while (cache->cached == BTRFS_CACHE_FAST) {
 568                 struct btrfs_caching_control *ctl;
 569
 570                 ctl = cache->caching_ctl;
 571                 atomic_inc(&ctl->count);
 572                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 573                 spin_unlock(&cache->lock);
 574
 575                 schedule();
 576
 577                 finish_wait(&ctl->wait, &wait);
 578                 put_caching_control(ctl);
 579                 spin_lock(&cache->lock);
 580         }
 581
 582         if (cache->cached != BTRFS_CACHE_NO) {
 583                 spin_unlock(&cache->lock);
 584                 kfree(caching_ctl);
 585                 return 0;
 586         }
 587         WARN_ON(cache->caching_ctl);
 588         cache->caching_ctl = caching_ctl;
 589         cache->cached = BTRFS_CACHE_FAST;
 590         spin_unlock(&cache->lock);
 591
 592         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 593                 ret = load_free_space_cache(fs_info, cache);
 594
 595                 spin_lock(&cache->lock);
 596                 if (ret == 1) {
 597                         cache->caching_ctl = NULL;
 598                         cache->cached = BTRFS_CACHE_FINISHED;
 599                         cache->last_byte_to_unpin = (u64)-1;
 600                 } 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                         }
 607                 }
 608                 spin_unlock(&cache->lock);
 609                 wake_up(&caching_ctl->wait);
 610                 if (ret == 1) {
 611                         put_caching_control(caching_ctl);
 612                         free_excluded_extents(fs_info->extent_root, cache);
 613                         return 0;
 614                 }
 615         } else {
 616                 /*
 617                  * We are not going to do the fast caching, set cached to the
 618                  * appropriate value and wakeup any waiters.
 619                  */
 620                 spin_lock(&cache->lock);
 621                 if (load_cache_only) {
 622                         cache->caching_ctl = NULL;
 623                         cache->cached = BTRFS_CACHE_NO;
 624                 } else {
 625                         cache->cached = BTRFS_CACHE_STARTED;
 626                 }
 627                 spin_unlock(&cache->lock);
 628                 wake_up(&caching_ctl->wait);
 629         }
 630
 631         if (load_cache_only) {
 632                 put_caching_control(caching_ctl);
 633                 return 0;
 634         }
 635
 636         down_write(&fs_info->commit_root_sem);
 637         atomic_inc(&caching_ctl->count);
 638         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 639         up_write(&fs_info->commit_root_sem);
 640
 641         btrfs_get_block_group(cache);
 642
 643         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 644
 645         return ret;
 646 }
 647
 648 /*
 649  * return the block group that starts at or after bytenr
 650  */
 651 static struct btrfs_block_group_cache *
 652 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 653 {
 654         struct btrfs_block_group_cache *cache;
 655
 656         cache = block_group_cache_tree_search(info, bytenr, 0);
 657
 658         return cache;
 659 }
 660
 661 /*
 662  * return the block group that contains the given bytenr
 663  */
 664 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 665                                                  struct btrfs_fs_info *info,
 666                                                  u64 bytenr)
 667 {
 668         struct btrfs_block_group_cache *cache;
 669
 670         cache = block_group_cache_tree_search(info, bytenr, 1);
 671
 672         return cache;
 673 }
 674
 675 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 676                                                   u64 flags)
 677 {
 678         struct list_head *head = &info->space_info;
 679         struct btrfs_space_info *found;
 680
 681         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 682
 683         rcu_read_lock();
 684         list_for_each_entry_rcu(found, head, list) {
 685                 if (found->flags & flags) {
 686                         rcu_read_unlock();
 687                         return found;
 688                 }
 689         }
 690         rcu_read_unlock();
 691         return NULL;
 692 }
 693
 694 /*
 695  * after adding space to the filesystem, we need to clear the full flags
 696  * on all the space infos.
 697  */
 698 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 699 {
 700         struct list_head *head = &info->space_info;
 701         struct btrfs_space_info *found;
 702
 703         rcu_read_lock();
 704         list_for_each_entry_rcu(found, head, list)
 705                 found->full = 0;
 706         rcu_read_unlock();
 707 }
 708
 709 /* simple helper to search for an existing extent at a given offset */
 710 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 711 {
 712         int ret;
 713         struct btrfs_key key;
 714         struct btrfs_path *path;
 715
 716         path = btrfs_alloc_path();
 717         if (!path)
 718                 return -ENOMEM;
 719
 720         key.objectid = start;
 721         key.offset = len;
 722         key.type = BTRFS_EXTENT_ITEM_KEY;
 723         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 724                                 0, 0);
 725         if (ret > 0) {
 726                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 727                 if (key.objectid == start &&
 728                     key.type == BTRFS_METADATA_ITEM_KEY)
 729                         ret = 0;
 730         }
 731         btrfs_free_path(path);
 732         return ret;
 733 }
 734
 735 /*
 736  * helper function to lookup reference count and flags of a tree block.
 737  *
 738  * the head node for delayed ref is used to store the sum of all the
 739  * reference count modifications queued up in the rbtree. the head
 740  * node may also store the extent flags to set. This way you can check
 741  * to see what the reference count and extent flags would be if all of
 742  * the delayed refs are not processed.
 743  */
 744 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 745                              struct btrfs_root *root, u64 bytenr,
 746                              u64 offset, int metadata, u64 *refs, u64 *flags)
 747 {
 748         struct btrfs_delayed_ref_head *head;
 749         struct btrfs_delayed_ref_root *delayed_refs;
 750         struct btrfs_path *path;
 751         struct btrfs_extent_item *ei;
 752         struct extent_buffer *leaf;
 753         struct btrfs_key key;
 754         u32 item_size;
 755         u64 num_refs;
 756         u64 extent_flags;
 757         int ret;
 758
 759         /*
 760          * If we don't have skinny metadata, don't bother doing anything
 761          * different
 762          */
 763         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 764                 offset = root->leafsize;
 765                 metadata = 0;
 766         }
 767
 768         path = btrfs_alloc_path();
 769         if (!path)
 770                 return -ENOMEM;
 771
 772         if (!trans) {
 773                 path->skip_locking = 1;
 774                 path->search_commit_root = 1;
 775         }
 776
 777 search_again:
 778         key.objectid = bytenr;
 779         key.offset = offset;
 780         if (metadata)
 781                 key.type = BTRFS_METADATA_ITEM_KEY;
 782         else
 783                 key.type = BTRFS_EXTENT_ITEM_KEY;
 784
 785 again:
 786         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 787                                 &key, path, 0, 0);
 788         if (ret < 0)
 789                 goto out_free;
 790
 791         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 792                 if (path->slots[0]) {
 793                         path->slots[0]--;
 794                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 795                                               path->slots[0]);
 796                         if (key.objectid == bytenr &&
 797                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 798                             key.offset == root->leafsize)
 799                                 ret = 0;
 800                 }
 801                 if (ret) {
 802                         key.objectid = bytenr;
 803                         key.type = BTRFS_EXTENT_ITEM_KEY;
 804                         key.offset = root->leafsize;
 805                         btrfs_release_path(path);
 806                         goto again;
 807                 }
 808         }
 809
 810         if (ret == 0) {
 811                 leaf = path->nodes[0];
 812                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 813                 if (item_size >= sizeof(*ei)) {
 814                         ei = btrfs_item_ptr(leaf, path->slots[0],
 815                                             struct btrfs_extent_item);
 816                         num_refs = btrfs_extent_refs(leaf, ei);
 817                         extent_flags = btrfs_extent_flags(leaf, ei);
 818                 } else {
 819 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 820                         struct btrfs_extent_item_v0 *ei0;
 821                         BUG_ON(item_size != sizeof(*ei0));
 822                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 823                                              struct btrfs_extent_item_v0);
 824                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 825                         /* FIXME: this isn't correct for data */
 826                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 827 #else
 828                         BUG();
 829 #endif
 830                 }
 831                 BUG_ON(num_refs == 0);
 832         } else {
 833                 num_refs = 0;
 834                 extent_flags = 0;
 835                 ret = 0;
 836         }
 837
 838         if (!trans)
 839                 goto out;
 840
 841         delayed_refs = &trans->transaction->delayed_refs;
 842         spin_lock(&delayed_refs->lock);
 843         head = btrfs_find_delayed_ref_head(trans, bytenr);
 844         if (head) {
 845                 if (!mutex_trylock(&head->mutex)) {
 846                         atomic_inc(&head->node.refs);
 847                         spin_unlock(&delayed_refs->lock);
 848
 849                         btrfs_release_path(path);
 850
 851                         /*
 852                          * Mutex was contended, block until it's released and try
 853                          * again
 854                          */
 855                         mutex_lock(&head->mutex);
 856                         mutex_unlock(&head->mutex);
 857                         btrfs_put_delayed_ref(&head->node);
 858                         goto search_again;
 859                 }
 860                 spin_lock(&head->lock);
 861                 if (head->extent_op && head->extent_op->update_flags)
 862                         extent_flags |= head->extent_op->flags_to_set;
 863                 else
 864                         BUG_ON(num_refs == 0);
 865
 866                 num_refs += head->node.ref_mod;
 867                 spin_unlock(&head->lock);
 868                 mutex_unlock(&head->mutex);
 869         }
 870         spin_unlock(&delayed_refs->lock);
 871 out:
 872         WARN_ON(num_refs == 0);
 873         if (refs)
 874                 *refs = num_refs;
 875         if (flags)
 876                 *flags = extent_flags;
 877 out_free:
 878         btrfs_free_path(path);
 879         return ret;
 880 }
 881
 882 /*
 883  * Back reference rules.  Back refs have three main goals:
 884  *
 885  * 1) differentiate between all holders of references to an extent so that
 886  *    when a reference is dropped we can make sure it was a valid reference
 887  *    before freeing the extent.
 888  *
 889  * 2) Provide enough information to quickly find the holders of an extent
 890  *    if we notice a given block is corrupted or bad.
 891  *
 892  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 893  *    maintenance.  This is actually the same as #2, but with a slightly
 894  *    different use case.
 895  *
 896  * There are two kinds of back refs. The implicit back refs is optimized
 897  * for pointers in non-shared tree blocks. For a given pointer in a block,
 898  * back refs of this kind provide information about the block's owner tree
 899  * and the pointer's key. These information allow us to find the block by
 900  * b-tree searching. The full back refs is for pointers in tree blocks not
 901  * referenced by their owner trees. The location of tree block is recorded
 902  * in the back refs. Actually the full back refs is generic, and can be
 903  * used in all cases the implicit back refs is used. The major shortcoming
 904  * of the full back refs is its overhead. Every time a tree block gets
 905  * COWed, we have to update back refs entry for all pointers in it.
 906  *
 907  * For a newly allocated tree block, we use implicit back refs for
 908  * pointers in it. This means most tree related operations only involve
 909  * implicit back refs. For a tree block created in old transaction, the
 910  * only way to drop a reference to it is COW it. So we can detect the
 911  * event that tree block loses its owner tree's reference and do the
 912  * back refs conversion.
 913  *
 914  * When a tree block is COW'd through a tree, there are four cases:
 915  *
 916  * The reference count of the block is one and the tree is the block's
 917  * owner tree. Nothing to do in this case.
 918  *
 919  * The reference count of the block is one and the tree is not the
 920  * block's owner tree. In this case, full back refs is used for pointers
 921  * in the block. Remove these full back refs, add implicit back refs for
 922  * every pointers in the new block.
 923  *
 924  * The reference count of the block is greater than one and the tree is
 925  * the block's owner tree. In this case, implicit back refs is used for
 926  * pointers in the block. Add full back refs for every pointers in the
 927  * block, increase lower level extents' reference counts. The original
 928  * implicit back refs are entailed to the new block.
 929  *
 930  * The reference count of the block is greater than one and the tree is
 931  * not the block's owner tree. Add implicit back refs for every pointer in
 932  * the new block, increase lower level extents' reference count.
 933  *
 934  * Back Reference Key composing:
 935  *
 936  * The key objectid corresponds to the first byte in the extent,
 937  * The key type is used to differentiate between types of back refs.
 938  * There are different meanings of the key offset for different types
 939  * of back refs.
 940  *
 941  * File extents can be referenced by:
 942  *
 943  * - multiple snapshots, subvolumes, or different generations in one subvol
 944  * - different files inside a single subvolume
 945  * - different offsets inside a file (bookend extents in file.c)
 946  *
 947  * The extent ref structure for the implicit back refs has fields for:
 948  *
 949  * - Objectid of the subvolume root
 950  * - objectid of the file holding the reference
 951  * - original offset in the file
 952  * - how many bookend extents
 953  *
 954  * The key offset for the implicit back refs is hash of the first
 955  * three fields.
 956  *
 957  * The extent ref structure for the full back refs has field for:
 958  *
 959  * - number of pointers in the tree leaf
 960  *
 961  * The key offset for the implicit back refs is the first byte of
 962  * the tree leaf
 963  *
 964  * When a file extent is allocated, The implicit back refs is used.
 965  * the fields are filled in:
 966  *
 967  *     (root_key.objectid, inode objectid, offset in file, 1)
 968  *
 969  * When a file extent is removed file truncation, we find the
 970  * corresponding implicit back refs and check the following fields:
 971  *
 972  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 973  *
 974  * Btree extents can be referenced by:
 975  *
 976  * - Different subvolumes
 977  *
 978  * Both the implicit back refs and the full back refs for tree blocks
 979  * only consist of key. The key offset for the implicit back refs is
 980  * objectid of block's owner tree. The key offset for the full back refs
 981  * is the first byte of parent block.
 982  *
 983  * When implicit back refs is used, information about the lowest key and
 984  * level of the tree block are required. These information are stored in
 985  * tree block info structure.
 986  */
 987
 988 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 989 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 990                                   struct btrfs_root *root,
 991                                   struct btrfs_path *path,
 992                                   u64 owner, u32 extra_size)
 993 {
 994         struct btrfs_extent_item *item;
 995         struct btrfs_extent_item_v0 *ei0;
 996         struct btrfs_extent_ref_v0 *ref0;
 997         struct btrfs_tree_block_info *bi;
 998         struct extent_buffer *leaf;
 999         struct btrfs_key key;
1000         struct btrfs_key found_key;
1001         u32 new_size = sizeof(*item);
1002         u64 refs;
1003         int ret;
1004
1005         leaf = path->nodes[0];
1006         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1007
1008         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1009         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1010                              struct btrfs_extent_item_v0);
1011         refs = btrfs_extent_refs_v0(leaf, ei0);
1012
1013         if (owner == (u64)-1) {
1014                 while (1) {
1015                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1016                                 ret = btrfs_next_leaf(root, path);
1017                                 if (ret < 0)
1018                                         return ret;
1019                                 BUG_ON(ret > 0); /* Corruption */
1020                                 leaf = path->nodes[0];
1021                         }
1022                         btrfs_item_key_to_cpu(leaf, &found_key,
1023                                               path->slots[0]);
1024                         BUG_ON(key.objectid != found_key.objectid);
1025                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1026                                 path->slots[0]++;
1027                                 continue;
1028                         }
1029                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1030                                               struct btrfs_extent_ref_v0);
1031                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1032                         break;
1033                 }
1034         }
1035         btrfs_release_path(path);
1036
1037         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1038                 new_size += sizeof(*bi);
1039
1040         new_size -= sizeof(*ei0);
1041         ret = btrfs_search_slot(trans, root, &key, path,
1042                                 new_size + extra_size, 1);
1043         if (ret < 0)
1044                 return ret;
1045         BUG_ON(ret); /* Corruption */
1046
1047         btrfs_extend_item(root, path, new_size);
1048
1049         leaf = path->nodes[0];
1050         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1051         btrfs_set_extent_refs(leaf, item, refs);
1052         /* FIXME: get real generation */
1053         btrfs_set_extent_generation(leaf, item, 0);
1054         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1055                 btrfs_set_extent_flags(leaf, item,
1056                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1057                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1058                 bi = (struct btrfs_tree_block_info *)(item + 1);
1059                 /* FIXME: get first key of the block */
1060                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1061                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1062         } else {
1063                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1064         }
1065         btrfs_mark_buffer_dirty(leaf);
1066         return 0;
1067 }
1068 #endif
1069
1070 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1071 {
1072         u32 high_crc = ~(u32)0;
1073         u32 low_crc = ~(u32)0;
1074         __le64 lenum;
1075
1076         lenum = cpu_to_le64(root_objectid);
1077         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1078         lenum = cpu_to_le64(owner);
1079         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1080         lenum = cpu_to_le64(offset);
1081         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1082
1083         return ((u64)high_crc << 31) ^ (u64)low_crc;
1084 }
1085
1086 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1087                                      struct btrfs_extent_data_ref *ref)
1088 {
1089         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1090                                     btrfs_extent_data_ref_objectid(leaf, ref),
1091                                     btrfs_extent_data_ref_offset(leaf, ref));
1092 }
1093
1094 static int match_extent_data_ref(struct extent_buffer *leaf,
1095                                  struct btrfs_extent_data_ref *ref,
1096                                  u64 root_objectid, u64 owner, u64 offset)
1097 {
1098         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1099             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1100             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1101                 return 0;
1102         return 1;
1103 }
1104
1105 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1106                                            struct btrfs_root *root,
1107                                            struct btrfs_path *path,
1108                                            u64 bytenr, u64 parent,
1109                                            u64 root_objectid,
1110                                            u64 owner, u64 offset)
1111 {
1112         struct btrfs_key key;
1113         struct btrfs_extent_data_ref *ref;
1114         struct extent_buffer *leaf;
1115         u32 nritems;
1116         int ret;
1117         int recow;
1118         int err = -ENOENT;
1119
1120         key.objectid = bytenr;
1121         if (parent) {
1122                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1123                 key.offset = parent;
1124         } else {
1125                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1126                 key.offset = hash_extent_data_ref(root_objectid,
1127                                                   owner, offset);
1128         }
1129 again:
1130         recow = 0;
1131         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1132         if (ret < 0) {
1133                 err = ret;
1134                 goto fail;
1135         }
1136
1137         if (parent) {
1138                 if (!ret)
1139                         return 0;
1140 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1141                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1142                 btrfs_release_path(path);
1143                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1144                 if (ret < 0) {
1145                         err = ret;
1146                         goto fail;
1147                 }
1148                 if (!ret)
1149                         return 0;
1150 #endif
1151                 goto fail;
1152         }
1153
1154         leaf = path->nodes[0];
1155         nritems = btrfs_header_nritems(leaf);
1156         while (1) {
1157                 if (path->slots[0] >= nritems) {
1158                         ret = btrfs_next_leaf(root, path);
1159                         if (ret < 0)
1160                                 err = ret;
1161                         if (ret)
1162                                 goto fail;
1163
1164                         leaf = path->nodes[0];
1165                         nritems = btrfs_header_nritems(leaf);
1166                         recow = 1;
1167                 }
1168
1169                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1170                 if (key.objectid != bytenr ||
1171                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1172                         goto fail;
1173
1174                 ref = btrfs_item_ptr(leaf, path->slots[0],
1175                                      struct btrfs_extent_data_ref);
1176
1177                 if (match_extent_data_ref(leaf, ref, root_objectid,
1178                                           owner, offset)) {
1179                         if (recow) {
1180                                 btrfs_release_path(path);
1181                                 goto again;
1182                         }
1183                         err = 0;
1184                         break;
1185                 }
1186                 path->slots[0]++;
1187         }
1188 fail:
1189         return err;
1190 }
1191
1192 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1193                                            struct btrfs_root *root,
1194                                            struct btrfs_path *path,
1195                                            u64 bytenr, u64 parent,
1196                                            u64 root_objectid, u64 owner,
1197                                            u64 offset, int refs_to_add)
1198 {
1199         struct btrfs_key key;
1200         struct extent_buffer *leaf;
1201         u32 size;
1202         u32 num_refs;
1203         int ret;
1204
1205         key.objectid = bytenr;
1206         if (parent) {
1207                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1208                 key.offset = parent;
1209                 size = sizeof(struct btrfs_shared_data_ref);
1210         } else {
1211                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1212                 key.offset = hash_extent_data_ref(root_objectid,
1213                                                   owner, offset);
1214                 size = sizeof(struct btrfs_extent_data_ref);
1215         }
1216
1217         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1218         if (ret && ret != -EEXIST)
1219                 goto fail;
1220
1221         leaf = path->nodes[0];
1222         if (parent) {
1223                 struct btrfs_shared_data_ref *ref;
1224                 ref = btrfs_item_ptr(leaf, path->slots[0],
1225                                      struct btrfs_shared_data_ref);
1226                 if (ret == 0) {
1227                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1228                 } else {
1229                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1230                         num_refs += refs_to_add;
1231                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1232                 }
1233         } else {
1234                 struct btrfs_extent_data_ref *ref;
1235                 while (ret == -EEXIST) {
1236                         ref = btrfs_item_ptr(leaf, path->slots[0],
1237                                              struct btrfs_extent_data_ref);
1238                         if (match_extent_data_ref(leaf, ref, root_objectid,
1239                                                   owner, offset))
1240                                 break;
1241                         btrfs_release_path(path);
1242                         key.offset++;
1243                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1244                                                       size);
1245                         if (ret && ret != -EEXIST)
1246                                 goto fail;
1247
1248                         leaf = path->nodes[0];
1249                 }
1250                 ref = btrfs_item_ptr(leaf, path->slots[0],
1251                                      struct btrfs_extent_data_ref);
1252                 if (ret == 0) {
1253                         btrfs_set_extent_data_ref_root(leaf, ref,
1254                                                        root_objectid);
1255                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1256                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1257                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1258                 } else {
1259                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1260                         num_refs += refs_to_add;
1261                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1262                 }
1263         }
1264         btrfs_mark_buffer_dirty(leaf);
1265         ret = 0;
1266 fail:
1267         btrfs_release_path(path);
1268         return ret;
1269 }
1270
1271 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1272                                            struct btrfs_root *root,
1273                                            struct btrfs_path *path,
1274                                            int refs_to_drop)
1275 {
1276         struct btrfs_key key;
1277         struct btrfs_extent_data_ref *ref1 = NULL;
1278         struct btrfs_shared_data_ref *ref2 = NULL;
1279         struct extent_buffer *leaf;
1280         u32 num_refs = 0;
1281         int ret = 0;
1282
1283         leaf = path->nodes[0];
1284         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1285
1286         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1287                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1288                                       struct btrfs_extent_data_ref);
1289                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1290         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1291                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1292                                       struct btrfs_shared_data_ref);
1293                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1294 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1295         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1296                 struct btrfs_extent_ref_v0 *ref0;
1297                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1298                                       struct btrfs_extent_ref_v0);
1299                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1300 #endif
1301         } else {
1302                 BUG();
1303         }
1304
1305         BUG_ON(num_refs < refs_to_drop);
1306         num_refs -= refs_to_drop;
1307
1308         if (num_refs == 0) {
1309                 ret = btrfs_del_item(trans, root, path);
1310         } else {
1311                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1312                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1313                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1314                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1315 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1316                 else {
1317                         struct btrfs_extent_ref_v0 *ref0;
1318                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1319                                         struct btrfs_extent_ref_v0);
1320                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1321                 }
1322 #endif
1323                 btrfs_mark_buffer_dirty(leaf);
1324         }
1325         return ret;
1326 }
1327
1328 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1329                                           struct btrfs_path *path,
1330                                           struct btrfs_extent_inline_ref *iref)
1331 {
1332         struct btrfs_key key;
1333         struct extent_buffer *leaf;
1334         struct btrfs_extent_data_ref *ref1;
1335         struct btrfs_shared_data_ref *ref2;
1336         u32 num_refs = 0;
1337
1338         leaf = path->nodes[0];
1339         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1340         if (iref) {
1341                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1342                     BTRFS_EXTENT_DATA_REF_KEY) {
1343                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1344                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1345                 } else {
1346                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1347                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1348                 }
1349         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1350                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1351                                       struct btrfs_extent_data_ref);
1352                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1353         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1354                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1355                                       struct btrfs_shared_data_ref);
1356                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1357 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1358         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1359                 struct btrfs_extent_ref_v0 *ref0;
1360                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1361                                       struct btrfs_extent_ref_v0);
1362                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1363 #endif
1364         } else {
1365                 WARN_ON(1);
1366         }
1367         return num_refs;
1368 }
1369
1370 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1371                                           struct btrfs_root *root,
1372                                           struct btrfs_path *path,
1373                                           u64 bytenr, u64 parent,
1374                                           u64 root_objectid)
1375 {
1376         struct btrfs_key key;
1377         int ret;
1378
1379         key.objectid = bytenr;
1380         if (parent) {
1381                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1382                 key.offset = parent;
1383         } else {
1384                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1385                 key.offset = root_objectid;
1386         }
1387
1388         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1389         if (ret > 0)
1390                 ret = -ENOENT;
1391 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1392         if (ret == -ENOENT && parent) {
1393                 btrfs_release_path(path);
1394                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1395                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1396                 if (ret > 0)
1397                         ret = -ENOENT;
1398         }
1399 #endif
1400         return ret;
1401 }
1402
1403 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1404                                           struct btrfs_root *root,
1405                                           struct btrfs_path *path,
1406                                           u64 bytenr, u64 parent,
1407                                           u64 root_objectid)
1408 {
1409         struct btrfs_key key;
1410         int ret;
1411
1412         key.objectid = bytenr;
1413         if (parent) {
1414                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1415                 key.offset = parent;
1416         } else {
1417                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1418                 key.offset = root_objectid;
1419         }
1420
1421         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1422         btrfs_release_path(path);
1423         return ret;
1424 }
1425
1426 static inline int extent_ref_type(u64 parent, u64 owner)
1427 {
1428         int type;
1429         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1430                 if (parent > 0)
1431                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1432                 else
1433                         type = BTRFS_TREE_BLOCK_REF_KEY;
1434         } else {
1435                 if (parent > 0)
1436                         type = BTRFS_SHARED_DATA_REF_KEY;
1437                 else
1438                         type = BTRFS_EXTENT_DATA_REF_KEY;
1439         }
1440         return type;
1441 }
1442
1443 static int find_next_key(struct btrfs_path *path, int level,
1444                          struct btrfs_key *key)
1445
1446 {
1447         for (; level < BTRFS_MAX_LEVEL; level++) {
1448                 if (!path->nodes[level])
1449                         break;
1450                 if (path->slots[level] + 1 >=
1451                     btrfs_header_nritems(path->nodes[level]))
1452                         continue;
1453                 if (level == 0)
1454                         btrfs_item_key_to_cpu(path->nodes[level], key,
1455                                               path->slots[level] + 1);
1456                 else
1457                         btrfs_node_key_to_cpu(path->nodes[level], key,
1458                                               path->slots[level] + 1);
1459                 return 0;
1460         }
1461         return 1;
1462 }
1463
1464 /*
1465  * look for inline back ref. if back ref is found, *ref_ret is set
1466  * to the address of inline back ref, and 0 is returned.
1467  *
1468  * if back ref isn't found, *ref_ret is set to the address where it
1469  * should be inserted, and -ENOENT is returned.
1470  *
1471  * if insert is true and there are too many inline back refs, the path
1472  * points to the extent item, and -EAGAIN is returned.
1473  *
1474  * NOTE: inline back refs are ordered in the same way that back ref
1475  *       items in the tree are ordered.
1476  */
1477 static noinline_for_stack
1478 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1479                                  struct btrfs_root *root,
1480                                  struct btrfs_path *path,
1481                                  struct btrfs_extent_inline_ref **ref_ret,
1482                                  u64 bytenr, u64 num_bytes,
1483                                  u64 parent, u64 root_objectid,
1484                                  u64 owner, u64 offset, int insert)
1485 {
1486         struct btrfs_key key;
1487         struct extent_buffer *leaf;
1488         struct btrfs_extent_item *ei;
1489         struct btrfs_extent_inline_ref *iref;
1490         u64 flags;
1491         u64 item_size;
1492         unsigned long ptr;
1493         unsigned long end;
1494         int extra_size;
1495         int type;
1496         int want;
1497         int ret;
1498         int err = 0;
1499         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1500                                                  SKINNY_METADATA);
1501
1502         key.objectid = bytenr;
1503         key.type = BTRFS_EXTENT_ITEM_KEY;
1504         key.offset = num_bytes;
1505
1506         want = extent_ref_type(parent, owner);
1507         if (insert) {
1508                 extra_size = btrfs_extent_inline_ref_size(want);
1509                 path->keep_locks = 1;
1510         } else
1511                 extra_size = -1;
1512
1513         /*
1514          * Owner is our parent level, so we can just add one to get the level
1515          * for the block we are interested in.
1516          */
1517         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1518                 key.type = BTRFS_METADATA_ITEM_KEY;
1519                 key.offset = owner;
1520         }
1521
1522 again:
1523         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1524         if (ret < 0) {
1525                 err = ret;
1526                 goto out;
1527         }
1528
1529         /*
1530          * We may be a newly converted file system which still has the old fat
1531          * extent entries for metadata, so try and see if we have one of those.
1532          */
1533         if (ret > 0 && skinny_metadata) {
1534                 skinny_metadata = false;
1535                 if (path->slots[0]) {
1536                         path->slots[0]--;
1537                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1538                                               path->slots[0]);
1539                         if (key.objectid == bytenr &&
1540                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1541                             key.offset == num_bytes)
1542                                 ret = 0;
1543                 }
1544                 if (ret) {
1545                         key.objectid = bytenr;
1546                         key.type = BTRFS_EXTENT_ITEM_KEY;
1547                         key.offset = num_bytes;
1548                         btrfs_release_path(path);
1549                         goto again;
1550                 }
1551         }
1552
1553         if (ret && !insert) {
1554                 err = -ENOENT;
1555                 goto out;
1556         } else if (WARN_ON(ret)) {
1557                 err = -EIO;
1558                 goto out;
1559         }
1560
1561         leaf = path->nodes[0];
1562         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1563 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1564         if (item_size < sizeof(*ei)) {
1565                 if (!insert) {
1566                         err = -ENOENT;
1567                         goto out;
1568                 }
1569                 ret = convert_extent_item_v0(trans, root, path, owner,
1570                                              extra_size);
1571                 if (ret < 0) {
1572                         err = ret;
1573                         goto out;
1574                 }
1575                 leaf = path->nodes[0];
1576                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1577         }
1578 #endif
1579         BUG_ON(item_size < sizeof(*ei));
1580
1581         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1582         flags = btrfs_extent_flags(leaf, ei);
1583
1584         ptr = (unsigned long)(ei + 1);
1585         end = (unsigned long)ei + item_size;
1586
1587         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1588                 ptr += sizeof(struct btrfs_tree_block_info);
1589                 BUG_ON(ptr > end);
1590         }
1591
1592         err = -ENOENT;
1593         while (1) {
1594                 if (ptr >= end) {
1595                         WARN_ON(ptr > end);
1596                         break;
1597                 }
1598                 iref = (struct btrfs_extent_inline_ref *)ptr;
1599                 type = btrfs_extent_inline_ref_type(leaf, iref);
1600                 if (want < type)
1601                         break;
1602                 if (want > type) {
1603                         ptr += btrfs_extent_inline_ref_size(type);
1604                         continue;
1605                 }
1606
1607                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1608                         struct btrfs_extent_data_ref *dref;
1609                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1610                         if (match_extent_data_ref(leaf, dref, root_objectid,
1611                                                   owner, offset)) {
1612                                 err = 0;
1613                                 break;
1614                         }
1615                         if (hash_extent_data_ref_item(leaf, dref) <
1616                             hash_extent_data_ref(root_objectid, owner, offset))
1617                                 break;
1618                 } else {
1619                         u64 ref_offset;
1620                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1621                         if (parent > 0) {
1622                                 if (parent == ref_offset) {
1623                                         err = 0;
1624                                         break;
1625                                 }
1626                                 if (ref_offset < parent)
1627                                         break;
1628                         } else {
1629                                 if (root_objectid == ref_offset) {
1630                                         err = 0;
1631                                         break;
1632                                 }
1633                                 if (ref_offset < root_objectid)
1634                                         break;
1635                         }
1636                 }
1637                 ptr += btrfs_extent_inline_ref_size(type);
1638         }
1639         if (err == -ENOENT && insert) {
1640                 if (item_size + extra_size >=
1641                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1642                         err = -EAGAIN;
1643                         goto out;
1644                 }
1645                 /*
1646                  * To add new inline back ref, we have to make sure
1647                  * there is no corresponding back ref item.
1648                  * For simplicity, we just do not add new inline back
1649                  * ref if there is any kind of item for this block
1650                  */
1651                 if (find_next_key(path, 0, &key) == 0 &&
1652                     key.objectid == bytenr &&
1653                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1654                         err = -EAGAIN;
1655                         goto out;
1656                 }
1657         }
1658         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1659 out:
1660         if (insert) {
1661                 path->keep_locks = 0;
1662                 btrfs_unlock_up_safe(path, 1);
1663         }
1664         return err;
1665 }
1666
1667 /*
1668  * helper to add new inline back ref
1669  */
1670 static noinline_for_stack
1671 void setup_inline_extent_backref(struct btrfs_root *root,
1672                                  struct btrfs_path *path,
1673                                  struct btrfs_extent_inline_ref *iref,
1674                                  u64 parent, u64 root_objectid,
1675                                  u64 owner, u64 offset, int refs_to_add,
1676                                  struct btrfs_delayed_extent_op *extent_op)
1677 {
1678         struct extent_buffer *leaf;
1679         struct btrfs_extent_item *ei;
1680         unsigned long ptr;
1681         unsigned long end;
1682         unsigned long item_offset;
1683         u64 refs;
1684         int size;
1685         int type;
1686
1687         leaf = path->nodes[0];
1688         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1689         item_offset = (unsigned long)iref - (unsigned long)ei;
1690
1691         type = extent_ref_type(parent, owner);
1692         size = btrfs_extent_inline_ref_size(type);
1693
1694         btrfs_extend_item(root, path, size);
1695
1696         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1697         refs = btrfs_extent_refs(leaf, ei);
1698         refs += refs_to_add;
1699         btrfs_set_extent_refs(leaf, ei, refs);
1700         if (extent_op)
1701                 __run_delayed_extent_op(extent_op, leaf, ei);
1702
1703         ptr = (unsigned long)ei + item_offset;
1704         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1705         if (ptr < end - size)
1706                 memmove_extent_buffer(leaf, ptr + size, ptr,
1707                                       end - size - ptr);
1708
1709         iref = (struct btrfs_extent_inline_ref *)ptr;
1710         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1711         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1712                 struct btrfs_extent_data_ref *dref;
1713                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1714                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1715                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1716                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1717                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1718         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1719                 struct btrfs_shared_data_ref *sref;
1720                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1721                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1722                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1723         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1724                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1725         } else {
1726                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1727         }
1728         btrfs_mark_buffer_dirty(leaf);
1729 }
1730
1731 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1732                                  struct btrfs_root *root,
1733                                  struct btrfs_path *path,
1734                                  struct btrfs_extent_inline_ref **ref_ret,
1735                                  u64 bytenr, u64 num_bytes, u64 parent,
1736                                  u64 root_objectid, u64 owner, u64 offset)
1737 {
1738         int ret;
1739
1740         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1741                                            bytenr, num_bytes, parent,
1742                                            root_objectid, owner, offset, 0);
1743         if (ret != -ENOENT)
1744                 return ret;
1745
1746         btrfs_release_path(path);
1747         *ref_ret = NULL;
1748
1749         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1750                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1751                                             root_objectid);
1752         } else {
1753                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1754                                              root_objectid, owner, offset);
1755         }
1756         return ret;
1757 }
1758
1759 /*
1760  * helper to update/remove inline back ref
1761  */
1762 static noinline_for_stack
1763 void update_inline_extent_backref(struct btrfs_root *root,
1764                                   struct btrfs_path *path,
1765                                   struct btrfs_extent_inline_ref *iref,
1766                                   int refs_to_mod,
1767                                   struct btrfs_delayed_extent_op *extent_op)
1768 {
1769         struct extent_buffer *leaf;
1770         struct btrfs_extent_item *ei;
1771         struct btrfs_extent_data_ref *dref = NULL;
1772         struct btrfs_shared_data_ref *sref = NULL;
1773         unsigned long ptr;
1774         unsigned long end;
1775         u32 item_size;
1776         int size;
1777         int type;
1778         u64 refs;
1779
1780         leaf = path->nodes[0];
1781         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1782         refs = btrfs_extent_refs(leaf, ei);
1783         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1784         refs += refs_to_mod;
1785         btrfs_set_extent_refs(leaf, ei, refs);
1786         if (extent_op)
1787                 __run_delayed_extent_op(extent_op, leaf, ei);
1788
1789         type = btrfs_extent_inline_ref_type(leaf, iref);
1790
1791         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1792                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1793                 refs = btrfs_extent_data_ref_count(leaf, dref);
1794         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1795                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1796                 refs = btrfs_shared_data_ref_count(leaf, sref);
1797         } else {
1798                 refs = 1;
1799                 BUG_ON(refs_to_mod != -1);
1800         }
1801
1802         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1803         refs += refs_to_mod;
1804
1805         if (refs > 0) {
1806                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1807                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1808                 else
1809                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1810         } else {
1811                 size =  btrfs_extent_inline_ref_size(type);
1812                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1813                 ptr = (unsigned long)iref;
1814                 end = (unsigned long)ei + item_size;
1815                 if (ptr + size < end)
1816                         memmove_extent_buffer(leaf, ptr, ptr + size,
1817                                               end - ptr - size);
1818                 item_size -= size;
1819                 btrfs_truncate_item(root, path, item_size, 1);
1820         }
1821         btrfs_mark_buffer_dirty(leaf);
1822 }
1823
1824 static noinline_for_stack
1825 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1826                                  struct btrfs_root *root,
1827                                  struct btrfs_path *path,
1828                                  u64 bytenr, u64 num_bytes, u64 parent,
1829                                  u64 root_objectid, u64 owner,
1830                                  u64 offset, int refs_to_add,
1831                                  struct btrfs_delayed_extent_op *extent_op)
1832 {
1833         struct btrfs_extent_inline_ref *iref;
1834         int ret;
1835
1836         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1837                                            bytenr, num_bytes, parent,
1838                                            root_objectid, owner, offset, 1);
1839         if (ret == 0) {
1840                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1841                 update_inline_extent_backref(root, path, iref,
1842                                              refs_to_add, extent_op);
1843         } else if (ret == -ENOENT) {
1844                 setup_inline_extent_backref(root, path, iref, parent,
1845                                             root_objectid, owner, offset,
1846                                             refs_to_add, extent_op);
1847                 ret = 0;
1848         }
1849         return ret;
1850 }
1851
1852 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1853                                  struct btrfs_root *root,
1854                                  struct btrfs_path *path,
1855                                  u64 bytenr, u64 parent, u64 root_objectid,
1856                                  u64 owner, u64 offset, int refs_to_add)
1857 {
1858         int ret;
1859         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1860                 BUG_ON(refs_to_add != 1);
1861                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1862                                             parent, root_objectid);
1863         } else {
1864                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1865                                              parent, root_objectid,
1866                                              owner, offset, refs_to_add);
1867         }
1868         return ret;
1869 }
1870
1871 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1872                                  struct btrfs_root *root,
1873                                  struct btrfs_path *path,
1874                                  struct btrfs_extent_inline_ref *iref,
1875                                  int refs_to_drop, int is_data)
1876 {
1877         int ret = 0;
1878
1879         BUG_ON(!is_data && refs_to_drop != 1);
1880         if (iref) {
1881                 update_inline_extent_backref(root, path, iref,
1882                                              -refs_to_drop, NULL);
1883         } else if (is_data) {
1884                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1885         } else {
1886                 ret = btrfs_del_item(trans, root, path);
1887         }
1888         return ret;
1889 }
1890
1891 static int btrfs_issue_discard(struct block_device *bdev,
1892                                 u64 start, u64 len)
1893 {
1894         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1895 }
1896
1897 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1898                                 u64 num_bytes, u64 *actual_bytes)
1899 {
1900         int ret;
1901         u64 discarded_bytes = 0;
1902         struct btrfs_bio *bbio = NULL;
1903
1904
1905         /* Tell the block device(s) that the sectors can be discarded */
1906         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1907                               bytenr, &num_bytes, &bbio, 0);
1908         /* Error condition is -ENOMEM */
1909         if (!ret) {
1910                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1911                 int i;
1912
1913
1914                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1915                         if (!stripe->dev->can_discard)
1916                                 continue;
1917
1918                         ret = btrfs_issue_discard(stripe->dev->bdev,
1919                                                   stripe->physical,
1920                                                   stripe->length);
1921                         if (!ret)
1922                                 discarded_bytes += stripe->length;
1923                         else if (ret != -EOPNOTSUPP)
1924                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1925
1926                         /*
1927                          * Just in case we get back EOPNOTSUPP for some reason,
1928                          * just ignore the return value so we don't screw up
1929                          * people calling discard_extent.
1930                          */
1931                         ret = 0;
1932                 }
1933                 kfree(bbio);
1934         }
1935
1936         if (actual_bytes)
1937                 *actual_bytes = discarded_bytes;
1938
1939
1940         if (ret == -EOPNOTSUPP)
1941                 ret = 0;
1942         return ret;
1943 }
1944
1945 /* Can return -ENOMEM */
1946 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1947                          struct btrfs_root *root,
1948                          u64 bytenr, u64 num_bytes, u64 parent,
1949                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1950 {
1951         int ret;
1952         struct btrfs_fs_info *fs_info = root->fs_info;
1953
1954         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1955                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1956
1957         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1958                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1959                                         num_bytes,
1960                                         parent, root_objectid, (int)owner,
1961                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1962         } else {
1963                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1964                                         num_bytes,
1965                                         parent, root_objectid, owner, offset,
1966                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1967         }
1968         return ret;
1969 }
1970
1971 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1972                                   struct btrfs_root *root,
1973                                   u64 bytenr, u64 num_bytes,
1974                                   u64 parent, u64 root_objectid,
1975                                   u64 owner, u64 offset, int refs_to_add,
1976                                   struct btrfs_delayed_extent_op *extent_op)
1977 {
1978         struct btrfs_path *path;
1979         struct extent_buffer *leaf;
1980         struct btrfs_extent_item *item;
1981         u64 refs;
1982         int ret;
1983
1984         path = btrfs_alloc_path();
1985         if (!path)
1986                 return -ENOMEM;
1987
1988         path->reada = 1;
1989         path->leave_spinning = 1;
1990         /* this will setup the path even if it fails to insert the back ref */
1991         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1992                                            path, bytenr, num_bytes, parent,
1993                                            root_objectid, owner, offset,
1994                                            refs_to_add, extent_op);
1995         if (ret != -EAGAIN)
1996                 goto out;
1997
1998         leaf = path->nodes[0];
1999         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2000         refs = btrfs_extent_refs(leaf, item);
2001         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2002         if (extent_op)
2003                 __run_delayed_extent_op(extent_op, leaf, item);
2004
2005         btrfs_mark_buffer_dirty(leaf);
2006         btrfs_release_path(path);
2007
2008         path->reada = 1;
2009         path->leave_spinning = 1;
2010
2011         /* now insert the actual backref */
2012         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2013                                     path, bytenr, parent, root_objectid,
2014                                     owner, offset, refs_to_add);
2015         if (ret)
2016                 btrfs_abort_transaction(trans, root, ret);
2017 out:
2018         btrfs_free_path(path);
2019         return ret;
2020 }
2021
2022 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2023                                 struct btrfs_root *root,
2024                                 struct btrfs_delayed_ref_node *node,
2025                                 struct btrfs_delayed_extent_op *extent_op,
2026                                 int insert_reserved)
2027 {
2028         int ret = 0;
2029         struct btrfs_delayed_data_ref *ref;
2030         struct btrfs_key ins;
2031         u64 parent = 0;
2032         u64 ref_root = 0;
2033         u64 flags = 0;
2034
2035         ins.objectid = node->bytenr;
2036         ins.offset = node->num_bytes;
2037         ins.type = BTRFS_EXTENT_ITEM_KEY;
2038
2039         ref = btrfs_delayed_node_to_data_ref(node);
2040         trace_run_delayed_data_ref(node, ref, node->action);
2041
2042         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2043                 parent = ref->parent;
2044         else
2045                 ref_root = ref->root;
2046
2047         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2048                 if (extent_op)
2049                         flags |= extent_op->flags_to_set;
2050                 ret = alloc_reserved_file_extent(trans, root,
2051                                                  parent, ref_root, flags,
2052                                                  ref->objectid, ref->offset,
2053                                                  &ins, node->ref_mod);
2054         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2055                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2056                                              node->num_bytes, parent,
2057                                              ref_root, ref->objectid,
2058                                              ref->offset, node->ref_mod,
2059                                              extent_op);
2060         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2061                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2062                                           node->num_bytes, parent,
2063                                           ref_root, ref->objectid,
2064                                           ref->offset, node->ref_mod,
2065                                           extent_op);
2066         } else {
2067                 BUG();
2068         }
2069         return ret;
2070 }
2071
2072 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2073                                     struct extent_buffer *leaf,
2074                                     struct btrfs_extent_item *ei)
2075 {
2076         u64 flags = btrfs_extent_flags(leaf, ei);
2077         if (extent_op->update_flags) {
2078                 flags |= extent_op->flags_to_set;
2079                 btrfs_set_extent_flags(leaf, ei, flags);
2080         }
2081
2082         if (extent_op->update_key) {
2083                 struct btrfs_tree_block_info *bi;
2084                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2085                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2086                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2087         }
2088 }
2089
2090 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2091                                  struct btrfs_root *root,
2092                                  struct btrfs_delayed_ref_node *node,
2093                                  struct btrfs_delayed_extent_op *extent_op)
2094 {
2095         struct btrfs_key key;
2096         struct btrfs_path *path;
2097         struct btrfs_extent_item *ei;
2098         struct extent_buffer *leaf;
2099         u32 item_size;
2100         int ret;
2101         int err = 0;
2102         int metadata = !extent_op->is_data;
2103
2104         if (trans->aborted)
2105                 return 0;
2106
2107         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2108                 metadata = 0;
2109
2110         path = btrfs_alloc_path();
2111         if (!path)
2112                 return -ENOMEM;
2113
2114         key.objectid = node->bytenr;
2115
2116         if (metadata) {
2117                 key.type = BTRFS_METADATA_ITEM_KEY;
2118                 key.offset = extent_op->level;
2119         } else {
2120                 key.type = BTRFS_EXTENT_ITEM_KEY;
2121                 key.offset = node->num_bytes;
2122         }
2123
2124 again:
2125         path->reada = 1;
2126         path->leave_spinning = 1;
2127         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2128                                 path, 0, 1);
2129         if (ret < 0) {
2130                 err = ret;
2131                 goto out;
2132         }
2133         if (ret > 0) {
2134                 if (metadata) {
2135                         if (path->slots[0] > 0) {
2136                                 path->slots[0]--;
2137                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2138                                                       path->slots[0]);
2139                                 if (key.objectid == node->bytenr &&
2140                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2141                                     key.offset == node->num_bytes)
2142                                         ret = 0;
2143                         }
2144                         if (ret > 0) {
2145                                 btrfs_release_path(path);
2146                                 metadata = 0;
2147
2148                                 key.objectid = node->bytenr;
2149                                 key.offset = node->num_bytes;
2150                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2151                                 goto again;
2152                         }
2153                 } else {
2154                         err = -EIO;
2155                         goto out;
2156                 }
2157         }
2158
2159         leaf = path->nodes[0];
2160         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2161 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2162         if (item_size < sizeof(*ei)) {
2163                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2164                                              path, (u64)-1, 0);
2165                 if (ret < 0) {
2166                         err = ret;
2167                         goto out;
2168                 }
2169                 leaf = path->nodes[0];
2170                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2171         }
2172 #endif
2173         BUG_ON(item_size < sizeof(*ei));
2174         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2175         __run_delayed_extent_op(extent_op, leaf, ei);
2176
2177         btrfs_mark_buffer_dirty(leaf);
2178 out:
2179         btrfs_free_path(path);
2180         return err;
2181 }
2182
2183 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2184                                 struct btrfs_root *root,
2185                                 struct btrfs_delayed_ref_node *node,
2186                                 struct btrfs_delayed_extent_op *extent_op,
2187                                 int insert_reserved)
2188 {
2189         int ret = 0;
2190         struct btrfs_delayed_tree_ref *ref;
2191         struct btrfs_key ins;
2192         u64 parent = 0;
2193         u64 ref_root = 0;
2194         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2195                                                  SKINNY_METADATA);
2196
2197         ref = btrfs_delayed_node_to_tree_ref(node);
2198         trace_run_delayed_tree_ref(node, ref, node->action);
2199
2200         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2201                 parent = ref->parent;
2202         else
2203                 ref_root = ref->root;
2204
2205         ins.objectid = node->bytenr;
2206         if (skinny_metadata) {
2207                 ins.offset = ref->level;
2208                 ins.type = BTRFS_METADATA_ITEM_KEY;
2209         } else {
2210                 ins.offset = node->num_bytes;
2211                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2212         }
2213
2214         BUG_ON(node->ref_mod != 1);
2215         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2216                 BUG_ON(!extent_op || !extent_op->update_flags);
2217                 ret = alloc_reserved_tree_block(trans, root,
2218                                                 parent, ref_root,
2219                                                 extent_op->flags_to_set,
2220                                                 &extent_op->key,
2221                                                 ref->level, &ins);
2222         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2223                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2224                                              node->num_bytes, parent, ref_root,
2225                                              ref->level, 0, 1, extent_op);
2226         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2227                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2228                                           node->num_bytes, parent, ref_root,
2229                                           ref->level, 0, 1, extent_op);
2230         } else {
2231                 BUG();
2232         }
2233         return ret;
2234 }
2235
2236 /* helper function to actually process a single delayed ref entry */
2237 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2238                                struct btrfs_root *root,
2239                                struct btrfs_delayed_ref_node *node,
2240                                struct btrfs_delayed_extent_op *extent_op,
2241                                int insert_reserved)
2242 {
2243         int ret = 0;
2244
2245         if (trans->aborted) {
2246                 if (insert_reserved)
2247                         btrfs_pin_extent(root, node->bytenr,
2248                                          node->num_bytes, 1);
2249                 return 0;
2250         }
2251
2252         if (btrfs_delayed_ref_is_head(node)) {
2253                 struct btrfs_delayed_ref_head *head;
2254                 /*
2255                  * we've hit the end of the chain and we were supposed
2256                  * to insert this extent into the tree.  But, it got
2257                  * deleted before we ever needed to insert it, so all
2258                  * we have to do is clean up the accounting
2259                  */
2260                 BUG_ON(extent_op);
2261                 head = btrfs_delayed_node_to_head(node);
2262                 trace_run_delayed_ref_head(node, head, node->action);
2263
2264                 if (insert_reserved) {
2265                         btrfs_pin_extent(root, node->bytenr,
2266                                          node->num_bytes, 1);
2267                         if (head->is_data) {
2268                                 ret = btrfs_del_csums(trans, root,
2269                                                       node->bytenr,
2270                                                       node->num_bytes);
2271                         }
2272                 }
2273                 return ret;
2274         }
2275
2276         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2277             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2278                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2279                                            insert_reserved);
2280         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2281                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2282                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2283                                            insert_reserved);
2284         else
2285                 BUG();
2286         return ret;
2287 }
2288
2289 static noinline struct btrfs_delayed_ref_node *
2290 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2291 {
2292         struct rb_node *node;
2293         struct btrfs_delayed_ref_node *ref, *last = NULL;;
2294
2295         /*
2296          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2297          * this prevents ref count from going down to zero when
2298          * there still are pending delayed ref.
2299          */
2300         node = rb_first(&head->ref_root);
2301         while (node) {
2302                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2303                                 rb_node);
2304                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2305                         return ref;
2306                 else if (last == NULL)
2307                         last = ref;
2308                 node = rb_next(node);
2309         }
2310         return last;
2311 }
2312
2313 /*
2314  * Returns 0 on success or if called with an already aborted transaction.
2315  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2316  */
2317 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2318                                              struct btrfs_root *root,
2319                                              unsigned long nr)
2320 {
2321         struct btrfs_delayed_ref_root *delayed_refs;
2322         struct btrfs_delayed_ref_node *ref;
2323         struct btrfs_delayed_ref_head *locked_ref = NULL;
2324         struct btrfs_delayed_extent_op *extent_op;
2325         struct btrfs_fs_info *fs_info = root->fs_info;
2326         ktime_t start = ktime_get();
2327         int ret;
2328         unsigned long count = 0;
2329         unsigned long actual_count = 0;
2330         int must_insert_reserved = 0;
2331
2332         delayed_refs = &trans->transaction->delayed_refs;
2333         while (1) {
2334                 if (!locked_ref) {
2335                         if (count >= nr)
2336                                 break;
2337
2338                         spin_lock(&delayed_refs->lock);
2339                         locked_ref = btrfs_select_ref_head(trans);
2340                         if (!locked_ref) {
2341                                 spin_unlock(&delayed_refs->lock);
2342                                 break;
2343                         }
2344
2345                         /* grab the lock that says we are going to process
2346                          * all the refs for this head */
2347                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2348                         spin_unlock(&delayed_refs->lock);
2349                         /*
2350                          * we may have dropped the spin lock to get the head
2351                          * mutex lock, and that might have given someone else
2352                          * time to free the head.  If that's true, it has been
2353                          * removed from our list and we can move on.
2354                          */
2355                         if (ret == -EAGAIN) {
2356                                 locked_ref = NULL;
2357                                 count++;
2358                                 continue;
2359                         }
2360                 }
2361
2362                 /*
2363                  * We need to try and merge add/drops of the same ref since we
2364                  * can run into issues with relocate dropping the implicit ref
2365                  * and then it being added back again before the drop can
2366                  * finish.  If we merged anything we need to re-loop so we can
2367                  * get a good ref.
2368                  */
2369                 spin_lock(&locked_ref->lock);
2370                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2371                                          locked_ref);
2372
2373                 /*
2374                  * locked_ref is the head node, so we have to go one
2375                  * node back for any delayed ref updates
2376                  */
2377                 ref = select_delayed_ref(locked_ref);
2378
2379                 if (ref && ref->seq &&
2380                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2381                         spin_unlock(&locked_ref->lock);
2382                         btrfs_delayed_ref_unlock(locked_ref);
2383                         spin_lock(&delayed_refs->lock);
2384                         locked_ref->processing = 0;
2385                         delayed_refs->num_heads_ready++;
2386                         spin_unlock(&delayed_refs->lock);
2387                         locked_ref = NULL;
2388                         cond_resched();
2389                         count++;
2390                         continue;
2391                 }
2392
2393                 /*
2394                  * record the must insert reserved flag before we
2395                  * drop the spin lock.
2396                  */
2397                 must_insert_reserved = locked_ref->must_insert_reserved;
2398                 locked_ref->must_insert_reserved = 0;
2399
2400                 extent_op = locked_ref->extent_op;
2401                 locked_ref->extent_op = NULL;
2402
2403                 if (!ref) {
2404
2405
2406                         /* All delayed refs have been processed, Go ahead
2407                          * and send the head node to run_one_delayed_ref,
2408                          * so that any accounting fixes can happen
2409                          */
2410                         ref = &locked_ref->node;
2411
2412                         if (extent_op && must_insert_reserved) {
2413                                 btrfs_free_delayed_extent_op(extent_op);
2414                                 extent_op = NULL;
2415                         }
2416
2417                         if (extent_op) {
2418                                 spin_unlock(&locked_ref->lock);
2419                                 ret = run_delayed_extent_op(trans, root,
2420                                                             ref, extent_op);
2421                                 btrfs_free_delayed_extent_op(extent_op);
2422
2423                                 if (ret) {
2424                                         /*
2425                                          * Need to reset must_insert_reserved if
2426                                          * there was an error so the abort stuff
2427                                          * can cleanup the reserved space
2428                                          * properly.
2429                                          */
2430                                         if (must_insert_reserved)
2431                                                 locked_ref->must_insert_reserved = 1;
2432                                         locked_ref->processing = 0;
2433                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2434                                         btrfs_delayed_ref_unlock(locked_ref);
2435                                         return ret;
2436                                 }
2437                                 continue;
2438                         }
2439
2440                         /*
2441                          * Need to drop our head ref lock and re-aqcuire the
2442                          * delayed ref lock and then re-check to make sure
2443                          * nobody got added.
2444                          */
2445                         spin_unlock(&locked_ref->lock);
2446                         spin_lock(&delayed_refs->lock);
2447                         spin_lock(&locked_ref->lock);
2448                         if (rb_first(&locked_ref->ref_root) ||
2449                             locked_ref->extent_op) {
2450                                 spin_unlock(&locked_ref->lock);
2451                                 spin_unlock(&delayed_refs->lock);
2452                                 continue;
2453                         }
2454                         ref->in_tree = 0;
2455                         delayed_refs->num_heads--;
2456                         rb_erase(&locked_ref->href_node,
2457                                  &delayed_refs->href_root);
2458                         spin_unlock(&delayed_refs->lock);
2459                 } else {
2460                         actual_count++;
2461                         ref->in_tree = 0;
2462                         rb_erase(&ref->rb_node, &locked_ref->ref_root);
2463                 }
2464                 atomic_dec(&delayed_refs->num_entries);
2465
2466                 if (!btrfs_delayed_ref_is_head(ref)) {
2467                         /*
2468                          * when we play the delayed ref, also correct the
2469                          * ref_mod on head
2470                          */
2471                         switch (ref->action) {
2472                         case BTRFS_ADD_DELAYED_REF:
2473                         case BTRFS_ADD_DELAYED_EXTENT:
2474                                 locked_ref->node.ref_mod -= ref->ref_mod;
2475                                 break;
2476                         case BTRFS_DROP_DELAYED_REF:
2477                                 locked_ref->node.ref_mod += ref->ref_mod;
2478                                 break;
2479                         default:
2480                                 WARN_ON(1);
2481                         }
2482                 }
2483                 spin_unlock(&locked_ref->lock);
2484
2485                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2486                                           must_insert_reserved);
2487
2488                 btrfs_free_delayed_extent_op(extent_op);
2489                 if (ret) {
2490                         locked_ref->processing = 0;
2491                         btrfs_delayed_ref_unlock(locked_ref);
2492                         btrfs_put_delayed_ref(ref);
2493                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2494                         return ret;
2495                 }
2496
2497                 /*
2498                  * If this node is a head, that means all the refs in this head
2499                  * have been dealt with, and we will pick the next head to deal
2500                  * with, so we must unlock the head and drop it from the cluster
2501                  * list before we release it.
2502                  */
2503                 if (btrfs_delayed_ref_is_head(ref)) {
2504                         btrfs_delayed_ref_unlock(locked_ref);
2505                         locked_ref = NULL;
2506                 }
2507                 btrfs_put_delayed_ref(ref);
2508                 count++;
2509                 cond_resched();
2510         }
2511
2512         /*
2513          * We don't want to include ref heads since we can have empty ref heads
2514          * and those will drastically skew our runtime down since we just do
2515          * accounting, no actual extent tree updates.
2516          */
2517         if (actual_count > 0) {
2518                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2519                 u64 avg;
2520
2521                 /*
2522                  * We weigh the current average higher than our current runtime
2523                  * to avoid large swings in the average.
2524                  */
2525                 spin_lock(&delayed_refs->lock);
2526                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2527                 avg = div64_u64(avg, 4);
2528                 fs_info->avg_delayed_ref_runtime = avg;
2529                 spin_unlock(&delayed_refs->lock);
2530         }
2531         return 0;
2532 }
2533
2534 #ifdef SCRAMBLE_DELAYED_REFS
2535 /*
2536  * Normally delayed refs get processed in ascending bytenr order. This
2537  * correlates in most cases to the order added. To expose dependencies on this
2538  * order, we start to process the tree in the middle instead of the beginning
2539  */
2540 static u64 find_middle(struct rb_root *root)
2541 {
2542         struct rb_node *n = root->rb_node;
2543         struct btrfs_delayed_ref_node *entry;
2544         int alt = 1;
2545         u64 middle;
2546         u64 first = 0, last = 0;
2547
2548         n = rb_first(root);
2549         if (n) {
2550                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2551                 first = entry->bytenr;
2552         }
2553         n = rb_last(root);
2554         if (n) {
2555                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2556                 last = entry->bytenr;
2557         }
2558         n = root->rb_node;
2559
2560         while (n) {
2561                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2562                 WARN_ON(!entry->in_tree);
2563
2564                 middle = entry->bytenr;
2565
2566                 if (alt)
2567                         n = n->rb_left;
2568                 else
2569                         n = n->rb_right;
2570
2571                 alt = 1 - alt;
2572         }
2573         return middle;
2574 }
2575 #endif
2576
2577 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2578                                          struct btrfs_fs_info *fs_info)
2579 {
2580         struct qgroup_update *qgroup_update;
2581         int ret = 0;
2582
2583         if (list_empty(&trans->qgroup_ref_list) !=
2584             !trans->delayed_ref_elem.seq) {
2585                 /* list without seq or seq without list */
2586                 btrfs_err(fs_info,
2587                         "qgroup accounting update error, list is%s empty, seq is %#x.%x",
2588                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2589                         (u32)(trans->delayed_ref_elem.seq >> 32),
2590                         (u32)trans->delayed_ref_elem.seq);
2591                 BUG();
2592         }
2593
2594         if (!trans->delayed_ref_elem.seq)
2595                 return 0;
2596
2597         while (!list_empty(&trans->qgroup_ref_list)) {
2598                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2599                                                  struct qgroup_update, list);
2600                 list_del(&qgroup_update->list);
2601                 if (!ret)
2602                         ret = btrfs_qgroup_account_ref(
2603                                         trans, fs_info, qgroup_update->node,
2604                                         qgroup_update->extent_op);
2605                 kfree(qgroup_update);
2606         }
2607
2608         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2609
2610         return ret;
2611 }
2612
2613 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2614 {
2615         u64 num_bytes;
2616
2617         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2618                              sizeof(struct btrfs_extent_inline_ref));
2619         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2620                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2621
2622         /*
2623          * We don't ever fill up leaves all the way so multiply by 2 just to be
2624          * closer to what we're really going to want to ouse.
2625          */
2626         return div64_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2627 }
2628
2629 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2630                                        struct btrfs_root *root)
2631 {
2632         struct btrfs_block_rsv *global_rsv;
2633         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2634         u64 num_bytes;
2635         int ret = 0;
2636
2637         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2638         num_heads = heads_to_leaves(root, num_heads);
2639         if (num_heads > 1)
2640                 num_bytes += (num_heads - 1) * root->leafsize;
2641         num_bytes <<= 1;
2642         global_rsv = &root->fs_info->global_block_rsv;
2643
2644         /*
2645          * If we can't allocate any more chunks lets make sure we have _lots_ of
2646          * wiggle room since running delayed refs can create more delayed refs.
2647          */
2648         if (global_rsv->space_info->full)
2649                 num_bytes <<= 1;
2650
2651         spin_lock(&global_rsv->lock);
2652         if (global_rsv->reserved <= num_bytes)
2653                 ret = 1;
2654         spin_unlock(&global_rsv->lock);
2655         return ret;
2656 }
2657
2658 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2659                                        struct btrfs_root *root)
2660 {
2661         struct btrfs_fs_info *fs_info = root->fs_info;
2662         u64 num_entries =
2663                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2664         u64 avg_runtime;
2665
2666         smp_mb();
2667         avg_runtime = fs_info->avg_delayed_ref_runtime;
2668         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2669                 return 1;
2670
2671         return btrfs_check_space_for_delayed_refs(trans, root);
2672 }
2673
2674 /*
2675  * this starts processing the delayed reference count updates and
2676  * extent insertions we have queued up so far.  count can be
2677  * 0, which means to process everything in the tree at the start
2678  * of the run (but not newly added entries), or it can be some target
2679  * number you'd like to process.
2680  *
2681  * Returns 0 on success or if called with an aborted transaction
2682  * Returns <0 on error and aborts the transaction
2683  */
2684 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2685                            struct btrfs_root *root, unsigned long count)
2686 {
2687         struct rb_node *node;
2688         struct btrfs_delayed_ref_root *delayed_refs;
2689         struct btrfs_delayed_ref_head *head;
2690         int ret;
2691         int run_all = count == (unsigned long)-1;
2692         int run_most = 0;
2693
2694         /* We'll clean this up in btrfs_cleanup_transaction */
2695         if (trans->aborted)
2696                 return 0;
2697
2698         if (root == root->fs_info->extent_root)
2699                 root = root->fs_info->tree_root;
2700
2701         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2702
2703         delayed_refs = &trans->transaction->delayed_refs;
2704         if (count == 0) {
2705                 count = atomic_read(&delayed_refs->num_entries) * 2;
2706                 run_most = 1;
2707         }
2708
2709 again:
2710 #ifdef SCRAMBLE_DELAYED_REFS
2711         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2712 #endif
2713         ret = __btrfs_run_delayed_refs(trans, root, count);
2714         if (ret < 0) {
2715                 btrfs_abort_transaction(trans, root, ret);
2716                 return ret;
2717         }
2718
2719         if (run_all) {
2720                 if (!list_empty(&trans->new_bgs))
2721                         btrfs_create_pending_block_groups(trans, root);
2722
2723                 spin_lock(&delayed_refs->lock);
2724                 node = rb_first(&delayed_refs->href_root);
2725                 if (!node) {
2726                         spin_unlock(&delayed_refs->lock);
2727                         goto out;
2728                 }
2729                 count = (unsigned long)-1;
2730
2731                 while (node) {
2732                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2733                                         href_node);
2734                         if (btrfs_delayed_ref_is_head(&head->node)) {
2735                                 struct btrfs_delayed_ref_node *ref;
2736
2737                                 ref = &head->node;
2738                                 atomic_inc(&ref->refs);
2739
2740                                 spin_unlock(&delayed_refs->lock);
2741                                 /*
2742                                  * Mutex was contended, block until it's
2743                                  * released and try again
2744                                  */
2745                                 mutex_lock(&head->mutex);
2746                                 mutex_unlock(&head->mutex);
2747
2748                                 btrfs_put_delayed_ref(ref);
2749                                 cond_resched();
2750                                 goto again;
2751                         } else {
2752                                 WARN_ON(1);
2753                         }
2754                         node = rb_next(node);
2755                 }
2756                 spin_unlock(&delayed_refs->lock);
2757                 cond_resched();
2758                 goto again;
2759         }
2760 out:
2761         assert_qgroups_uptodate(trans);
2762         return 0;
2763 }
2764
2765 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2766                                 struct btrfs_root *root,
2767                                 u64 bytenr, u64 num_bytes, u64 flags,
2768                                 int level, int is_data)
2769 {
2770         struct btrfs_delayed_extent_op *extent_op;
2771         int ret;
2772
2773         extent_op = btrfs_alloc_delayed_extent_op();
2774         if (!extent_op)
2775                 return -ENOMEM;
2776
2777         extent_op->flags_to_set = flags;
2778         extent_op->update_flags = 1;
2779         extent_op->update_key = 0;
2780         extent_op->is_data = is_data ? 1 : 0;
2781         extent_op->level = level;
2782
2783         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2784                                           num_bytes, extent_op);
2785         if (ret)
2786                 btrfs_free_delayed_extent_op(extent_op);
2787         return ret;
2788 }
2789
2790 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2791                                       struct btrfs_root *root,
2792                                       struct btrfs_path *path,
2793                                       u64 objectid, u64 offset, u64 bytenr)
2794 {
2795         struct btrfs_delayed_ref_head *head;
2796         struct btrfs_delayed_ref_node *ref;
2797         struct btrfs_delayed_data_ref *data_ref;
2798         struct btrfs_delayed_ref_root *delayed_refs;
2799         struct rb_node *node;
2800         int ret = 0;
2801
2802         delayed_refs = &trans->transaction->delayed_refs;
2803         spin_lock(&delayed_refs->lock);
2804         head = btrfs_find_delayed_ref_head(trans, bytenr);
2805         if (!head) {
2806                 spin_unlock(&delayed_refs->lock);
2807                 return 0;
2808         }
2809
2810         if (!mutex_trylock(&head->mutex)) {
2811                 atomic_inc(&head->node.refs);
2812                 spin_unlock(&delayed_refs->lock);
2813
2814                 btrfs_release_path(path);
2815
2816                 /*
2817                  * Mutex was contended, block until it's released and let
2818                  * caller try again
2819                  */
2820                 mutex_lock(&head->mutex);
2821                 mutex_unlock(&head->mutex);
2822                 btrfs_put_delayed_ref(&head->node);
2823                 return -EAGAIN;
2824         }
2825         spin_unlock(&delayed_refs->lock);
2826
2827         spin_lock(&head->lock);
2828         node = rb_first(&head->ref_root);
2829         while (node) {
2830                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2831                 node = rb_next(node);
2832
2833                 /* If it's a shared ref we know a cross reference exists */
2834                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2835                         ret = 1;
2836                         break;
2837                 }
2838
2839                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2840
2841                 /*
2842                  * If our ref doesn't match the one we're currently looking at
2843                  * then we have a cross reference.
2844                  */
2845                 if (data_ref->root != root->root_key.objectid ||
2846                     data_ref->objectid != objectid ||
2847                     data_ref->offset != offset) {
2848                         ret = 1;
2849                         break;
2850                 }
2851         }
2852         spin_unlock(&head->lock);
2853         mutex_unlock(&head->mutex);
2854         return ret;
2855 }
2856
2857 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2858                                         struct btrfs_root *root,
2859                                         struct btrfs_path *path,
2860                                         u64 objectid, u64 offset, u64 bytenr)
2861 {
2862         struct btrfs_root *extent_root = root->fs_info->extent_root;
2863         struct extent_buffer *leaf;
2864         struct btrfs_extent_data_ref *ref;
2865         struct btrfs_extent_inline_ref *iref;
2866         struct btrfs_extent_item *ei;
2867         struct btrfs_key key;
2868         u32 item_size;
2869         int ret;
2870
2871         key.objectid = bytenr;
2872         key.offset = (u64)-1;
2873         key.type = BTRFS_EXTENT_ITEM_KEY;
2874
2875         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2876         if (ret < 0)
2877                 goto out;
2878         BUG_ON(ret == 0); /* Corruption */
2879
2880         ret = -ENOENT;
2881         if (path->slots[0] == 0)
2882                 goto out;
2883
2884         path->slots[0]--;
2885         leaf = path->nodes[0];
2886         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2887
2888         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2889                 goto out;
2890
2891         ret = 1;
2892         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2893 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2894         if (item_size < sizeof(*ei)) {
2895                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2896                 goto out;
2897         }
2898 #endif
2899         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2900
2901         if (item_size != sizeof(*ei) +
2902             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2903                 goto out;
2904
2905         if (btrfs_extent_generation(leaf, ei) <=
2906             btrfs_root_last_snapshot(&root->root_item))
2907                 goto out;
2908
2909         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2910         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2911             BTRFS_EXTENT_DATA_REF_KEY)
2912                 goto out;
2913
2914         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2915         if (btrfs_extent_refs(leaf, ei) !=
2916             btrfs_extent_data_ref_count(leaf, ref) ||
2917             btrfs_extent_data_ref_root(leaf, ref) !=
2918             root->root_key.objectid ||
2919             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2920             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2921                 goto out;
2922
2923         ret = 0;
2924 out:
2925         return ret;
2926 }
2927
2928 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2929                           struct btrfs_root *root,
2930                           u64 objectid, u64 offset, u64 bytenr)
2931 {
2932         struct btrfs_path *path;
2933         int ret;
2934         int ret2;
2935
2936         path = btrfs_alloc_path();
2937         if (!path)
2938                 return -ENOENT;
2939
2940         do {
2941                 ret = check_committed_ref(trans, root, path, objectid,
2942                                           offset, bytenr);
2943                 if (ret && ret != -ENOENT)
2944                         goto out;
2945
2946                 ret2 = check_delayed_ref(trans, root, path, objectid,
2947                                          offset, bytenr);
2948         } while (ret2 == -EAGAIN);
2949
2950         if (ret2 && ret2 != -ENOENT) {
2951                 ret = ret2;
2952                 goto out;
2953         }
2954
2955         if (ret != -ENOENT || ret2 != -ENOENT)
2956                 ret = 0;
2957 out:
2958         btrfs_free_path(path);
2959         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2960                 WARN_ON(ret > 0);
2961         return ret;
2962 }
2963
2964 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2965                            struct btrfs_root *root,
2966                            struct extent_buffer *buf,
2967                            int full_backref, int inc, int for_cow)
2968 {
2969         u64 bytenr;
2970         u64 num_bytes;
2971         u64 parent;
2972         u64 ref_root;
2973         u32 nritems;
2974         struct btrfs_key key;
2975         struct btrfs_file_extent_item *fi;
2976         int i;
2977         int level;
2978         int ret = 0;
2979         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2980                             u64, u64, u64, u64, u64, u64, int);
2981
2982         ref_root = btrfs_header_owner(buf);
2983         nritems = btrfs_header_nritems(buf);
2984         level = btrfs_header_level(buf);
2985
2986         if (!root->ref_cows && level == 0)
2987                 return 0;
2988
2989         if (inc)
2990                 process_func = btrfs_inc_extent_ref;
2991         else
2992                 process_func = btrfs_free_extent;
2993
2994         if (full_backref)
2995                 parent = buf->start;
2996         else
2997                 parent = 0;
2998
2999         for (i = 0; i < nritems; i++) {
3000                 if (level == 0) {
3001                         btrfs_item_key_to_cpu(buf, &key, i);
3002                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
3003                                 continue;
3004                         fi = btrfs_item_ptr(buf, i,
3005                                             struct btrfs_file_extent_item);
3006                         if (btrfs_file_extent_type(buf, fi) ==
3007                             BTRFS_FILE_EXTENT_INLINE)
3008                                 continue;
3009                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3010                         if (bytenr == 0)
3011                                 continue;
3012
3013                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3014                         key.offset -= btrfs_file_extent_offset(buf, fi);
3015                         ret = process_func(trans, root, bytenr, num_bytes,
3016                                            parent, ref_root, key.objectid,
3017                                            key.offset, for_cow);
3018                         if (ret)
3019                                 goto fail;
3020                 } else {
3021                         bytenr = btrfs_node_blockptr(buf, i);
3022                         num_bytes = btrfs_level_size(root, level - 1);
3023                         ret = process_func(trans, root, bytenr, num_bytes,
3024                                            parent, ref_root, level - 1, 0,
3025                                            for_cow);
3026                         if (ret)
3027                                 goto fail;
3028                 }
3029         }
3030         return 0;
3031 fail:
3032         return ret;
3033 }
3034
3035 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3036                   struct extent_buffer *buf, int full_backref, int for_cow)
3037 {
3038         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
3039 }
3040
3041 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3042                   struct extent_buffer *buf, int full_backref, int for_cow)
3043 {
3044         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
3045 }
3046
3047 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3048                                  struct btrfs_root *root,
3049                                  struct btrfs_path *path,
3050                                  struct btrfs_block_group_cache *cache)
3051 {
3052         int ret;
3053         struct btrfs_root *extent_root = root->fs_info->extent_root;
3054         unsigned long bi;
3055         struct extent_buffer *leaf;
3056
3057         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3058         if (ret < 0)
3059                 goto fail;
3060         BUG_ON(ret); /* Corruption */
3061
3062         leaf = path->nodes[0];
3063         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3064         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3065         btrfs_mark_buffer_dirty(leaf);
3066         btrfs_release_path(path);
3067 fail:
3068         if (ret) {
3069                 btrfs_abort_transaction(trans, root, ret);
3070                 return ret;
3071         }
3072         return 0;
3073
3074 }
3075
3076 static struct btrfs_block_group_cache *
3077 next_block_group(struct btrfs_root *root,
3078                  struct btrfs_block_group_cache *cache)
3079 {
3080         struct rb_node *node;
3081         spin_lock(&root->fs_info->block_group_cache_lock);
3082         node = rb_next(&cache->cache_node);
3083         btrfs_put_block_group(cache);
3084         if (node) {
3085                 cache = rb_entry(node, struct btrfs_block_group_cache,
3086                                  cache_node);
3087                 btrfs_get_block_group(cache);
3088         } else
3089                 cache = NULL;
3090         spin_unlock(&root->fs_info->block_group_cache_lock);
3091         return cache;
3092 }
3093
3094 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3095                             struct btrfs_trans_handle *trans,
3096                             struct btrfs_path *path)
3097 {
3098         struct btrfs_root *root = block_group->fs_info->tree_root;
3099         struct inode *inode = NULL;
3100         u64 alloc_hint = 0;
3101         int dcs = BTRFS_DC_ERROR;
3102         int num_pages = 0;
3103         int retries = 0;
3104         int ret = 0;
3105
3106         /*
3107          * If this block group is smaller than 100 megs don't bother caching the
3108          * block group.
3109          */
3110         if (block_group->key.offset < (100 * 1024 * 1024)) {
3111                 spin_lock(&block_group->lock);
3112                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3113                 spin_unlock(&block_group->lock);
3114                 return 0;
3115         }
3116
3117 again:
3118         inode = lookup_free_space_inode(root, block_group, path);
3119         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3120                 ret = PTR_ERR(inode);
3121                 btrfs_release_path(path);
3122                 goto out;
3123         }
3124
3125         if (IS_ERR(inode)) {
3126                 BUG_ON(retries);
3127                 retries++;
3128
3129                 if (block_group->ro)
3130                         goto out_free;
3131
3132                 ret = create_free_space_inode(root, trans, block_group, path);
3133                 if (ret)
3134                         goto out_free;
3135                 goto again;
3136         }
3137
3138         /* We've already setup this transaction, go ahead and exit */
3139         if (block_group->cache_generation == trans->transid &&
3140             i_size_read(inode)) {
3141                 dcs = BTRFS_DC_SETUP;
3142                 goto out_put;
3143         }
3144
3145         /*
3146          * We want to set the generation to 0, that way if anything goes wrong
3147          * from here on out we know not to trust this cache when we load up next
3148          * time.
3149          */
3150         BTRFS_I(inode)->generation = 0;
3151         ret = btrfs_update_inode(trans, root, inode);
3152         WARN_ON(ret);
3153
3154         if (i_size_read(inode) > 0) {
3155                 ret = btrfs_check_trunc_cache_free_space(root,
3156                                         &root->fs_info->global_block_rsv);
3157                 if (ret)
3158                         goto out_put;
3159
3160                 ret = btrfs_truncate_free_space_cache(root, trans, inode);
3161                 if (ret)
3162                         goto out_put;
3163         }
3164
3165         spin_lock(&block_group->lock);
3166         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3167             !btrfs_test_opt(root, SPACE_CACHE)) {
3168                 /*
3169                  * don't bother trying to write stuff out _if_
3170                  * a) we're not cached,
3171                  * b) we're with nospace_cache mount option.
3172                  */
3173                 dcs = BTRFS_DC_WRITTEN;
3174                 spin_unlock(&block_group->lock);
3175                 goto out_put;
3176         }
3177         spin_unlock(&block_group->lock);
3178
3179         /*
3180          * Try to preallocate enough space based on how big the block group is.
3181          * Keep in mind this has to include any pinned space which could end up
3182          * taking up quite a bit since it's not folded into the other space
3183          * cache.
3184          */
3185         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3186         if (!num_pages)
3187                 num_pages = 1;
3188
3189         num_pages *= 16;
3190         num_pages *= PAGE_CACHE_SIZE;
3191
3192         ret = btrfs_check_data_free_space(inode, num_pages);
3193         if (ret)
3194                 goto out_put;
3195
3196         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3197                                               num_pages, num_pages,
3198                                               &alloc_hint);
3199         if (!ret)
3200                 dcs = BTRFS_DC_SETUP;
3201         btrfs_free_reserved_data_space(inode, num_pages);
3202
3203 out_put:
3204         iput(inode);
3205 out_free:
3206         btrfs_release_path(path);
3207 out:
3208         spin_lock(&block_group->lock);
3209         if (!ret && dcs == BTRFS_DC_SETUP)
3210                 block_group->cache_generation = trans->transid;
3211         block_group->disk_cache_state = dcs;
3212         spin_unlock(&block_group->lock);
3213
3214         return ret;
3215 }
3216
3217 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3218                                    struct btrfs_root *root)
3219 {
3220         struct btrfs_block_group_cache *cache;
3221         int err = 0;
3222         struct btrfs_path *path;
3223         u64 last = 0;
3224
3225         path = btrfs_alloc_path();
3226         if (!path)
3227                 return -ENOMEM;
3228
3229 again:
3230         while (1) {
3231                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3232                 while (cache) {
3233                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3234                                 break;
3235                         cache = next_block_group(root, cache);
3236                 }
3237                 if (!cache) {
3238                         if (last == 0)
3239                                 break;
3240                         last = 0;
3241                         continue;
3242                 }
3243                 err = cache_save_setup(cache, trans, path);
3244                 last = cache->key.objectid + cache->key.offset;
3245                 btrfs_put_block_group(cache);
3246         }
3247
3248         while (1) {
3249                 if (last == 0) {
3250                         err = btrfs_run_delayed_refs(trans, root,
3251                                                      (unsigned long)-1);
3252                         if (err) /* File system offline */
3253                                 goto out;
3254                 }
3255
3256                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3257                 while (cache) {
3258                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3259                                 btrfs_put_block_group(cache);
3260                                 goto again;
3261                         }
3262
3263                         if (cache->dirty)
3264                                 break;
3265                         cache = next_block_group(root, cache);
3266                 }
3267                 if (!cache) {
3268                         if (last == 0)
3269                                 break;
3270                         last = 0;
3271                         continue;
3272                 }
3273
3274                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3275                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3276                 cache->dirty = 0;
3277                 last = cache->key.objectid + cache->key.offset;
3278
3279                 err = write_one_cache_group(trans, root, path, cache);
3280                 btrfs_put_block_group(cache);
3281                 if (err) /* File system offline */
3282                         goto out;
3283         }
3284
3285         while (1) {
3286                 /*
3287                  * I don't think this is needed since we're just marking our
3288                  * preallocated extent as written, but just in case it can't
3289                  * hurt.
3290                  */
3291                 if (last == 0) {
3292                         err = btrfs_run_delayed_refs(trans, root,
3293                                                      (unsigned long)-1);
3294                         if (err) /* File system offline */
3295                                 goto out;
3296                 }
3297
3298                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3299                 while (cache) {
3300                         /*
3301                          * Really this shouldn't happen, but it could if we
3302                          * couldn't write the entire preallocated extent and
3303                          * splitting the extent resulted in a new block.
3304                          */
3305                         if (cache->dirty) {
3306                                 btrfs_put_block_group(cache);
3307                                 goto again;
3308                         }
3309                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3310                                 break;
3311                         cache = next_block_group(root, cache);
3312                 }
3313                 if (!cache) {
3314                         if (last == 0)
3315                                 break;
3316                         last = 0;
3317                         continue;
3318                 }
3319
3320                 err = btrfs_write_out_cache(root, trans, cache, path);
3321
3322                 /*
3323                  * If we didn't have an error then the cache state is still
3324                  * NEED_WRITE, so we can set it to WRITTEN.
3325                  */
3326                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3327                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3328                 last = cache->key.objectid + cache->key.offset;
3329                 btrfs_put_block_group(cache);
3330         }
3331 out:
3332
3333         btrfs_free_path(path);
3334         return err;
3335 }
3336
3337 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3338 {
3339         struct btrfs_block_group_cache *block_group;
3340         int readonly = 0;
3341
3342         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3343         if (!block_group || block_group->ro)
3344                 readonly = 1;
3345         if (block_group)
3346                 btrfs_put_block_group(block_group);
3347         return readonly;
3348 }
3349
3350 static const char *alloc_name(u64 flags)
3351 {
3352         switch (flags) {
3353         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3354                 return "mixed";
3355         case BTRFS_BLOCK_GROUP_METADATA:
3356                 return "metadata";
3357         case BTRFS_BLOCK_GROUP_DATA:
3358                 return "data";
3359         case BTRFS_BLOCK_GROUP_SYSTEM:
3360                 return "system";
3361         default:
3362                 WARN_ON(1);
3363                 return "invalid-combination";
3364         };
3365 }
3366
3367 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3368                              u64 total_bytes, u64 bytes_used,
3369                              struct btrfs_space_info **space_info)
3370 {
3371         struct btrfs_space_info *found;
3372         int i;
3373         int factor;
3374         int ret;
3375
3376         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3377                      BTRFS_BLOCK_GROUP_RAID10))
3378                 factor = 2;
3379         else
3380                 factor = 1;
3381
3382         found = __find_space_info(info, flags);
3383         if (found) {
3384                 spin_lock(&found->lock);
3385                 found->total_bytes += total_bytes;
3386                 found->disk_total += total_bytes * factor;
3387                 found->bytes_used += bytes_used;
3388                 found->disk_used += bytes_used * factor;
3389                 found->full = 0;
3390                 spin_unlock(&found->lock);
3391                 *space_info = found;
3392                 return 0;
3393         }
3394         found = kzalloc(sizeof(*found), GFP_NOFS);
3395         if (!found)
3396                 return -ENOMEM;
3397
3398         ret = percpu_counter_init(&found->total_bytes_pinned, 0);
3399         if (ret) {
3400                 kfree(found);
3401                 return ret;
3402         }
3403
3404         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
3405                 INIT_LIST_HEAD(&found->block_groups[i]);
3406                 kobject_init(&found->block_group_kobjs[i], &btrfs_raid_ktype);
3407         }
3408         init_rwsem(&found->groups_sem);
3409         spin_lock_init(&found->lock);
3410         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3411         found->total_bytes = total_bytes;
3412         found->disk_total = total_bytes * factor;
3413         found->bytes_used = bytes_used;
3414         found->disk_used = bytes_used * factor;
3415         found->bytes_pinned = 0;
3416         found->bytes_reserved = 0;
3417         found->bytes_readonly = 0;
3418         found->bytes_may_use = 0;
3419         found->full = 0;
3420         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3421         found->chunk_alloc = 0;
3422         found->flush = 0;
3423         init_waitqueue_head(&found->wait);
3424
3425         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3426                                     info->space_info_kobj, "%s",
3427                                     alloc_name(found->flags));
3428         if (ret) {
3429                 kfree(found);
3430                 return ret;
3431         }
3432
3433         *space_info = found;
3434         list_add_rcu(&found->list, &info->space_info);
3435         if (flags & BTRFS_BLOCK_GROUP_DATA)
3436                 info->data_sinfo = found;
3437
3438         return ret;
3439 }
3440
3441 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3442 {
3443         u64 extra_flags = chunk_to_extended(flags) &
3444                                 BTRFS_EXTENDED_PROFILE_MASK;
3445
3446         write_seqlock(&fs_info->profiles_lock);
3447         if (flags & BTRFS_BLOCK_GROUP_DATA)
3448                 fs_info->avail_data_alloc_bits |= extra_flags;
3449         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3450                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3451         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3452                 fs_info->avail_system_alloc_bits |= extra_flags;
3453         write_sequnlock(&fs_info->profiles_lock);
3454 }
3455
3456 /*
3457  * returns target flags in extended format or 0 if restripe for this
3458  * chunk_type is not in progress
3459  *
3460  * should be called with either volume_mutex or balance_lock held
3461  */
3462 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3463 {
3464         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3465         u64 target = 0;
3466
3467         if (!bctl)
3468                 return 0;
3469
3470         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3471             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3472                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3473         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3474                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3475                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3476         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3477                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3478                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3479         }
3480
3481         return target;
3482 }
3483
3484 /*
3485  * @flags: available profiles in extended format (see ctree.h)
3486  *
3487  * Returns reduced profile in chunk format.  If profile changing is in
3488  * progress (either running or paused) picks the target profile (if it's
3489  * already available), otherwise falls back to plain reducing.
3490  */
3491 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3492 {
3493         /*
3494          * we add in the count of missing devices because we want
3495          * to make sure that any RAID levels on a degraded FS
3496          * continue to be honored.
3497          */
3498         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3499                 root->fs_info->fs_devices->missing_devices;
3500         u64 target;
3501         u64 tmp;
3502
3503         /*
3504          * see if restripe for this chunk_type is in progress, if so
3505          * try to reduce to the target profile
3506          */
3507         spin_lock(&root->fs_info->balance_lock);
3508         target = get_restripe_target(root->fs_info, flags);
3509         if (target) {
3510                 /* pick target profile only if it's already available */
3511                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3512                         spin_unlock(&root->fs_info->balance_lock);
3513                         return extended_to_chunk(target);
3514                 }
3515         }
3516         spin_unlock(&root->fs_info->balance_lock);
3517
3518         /* First, mask out the RAID levels which aren't possible */
3519         if (num_devices == 1)
3520                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3521                            BTRFS_BLOCK_GROUP_RAID5);
3522         if (num_devices < 3)
3523                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3524         if (num_devices < 4)
3525                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3526
3527         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3528                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3529                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3530         flags &= ~tmp;
3531
3532         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3533                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3534         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3535                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3536         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3537                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3538         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3539                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3540         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3541                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3542
3543         return extended_to_chunk(flags | tmp);
3544 }
3545
3546 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3547 {
3548         unsigned seq;
3549         u64 flags;
3550
3551         do {
3552                 flags = orig_flags;
3553                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3554
3555                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3556                         flags |= root->fs_info->avail_data_alloc_bits;
3557                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3558                         flags |= root->fs_info->avail_system_alloc_bits;
3559                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3560                         flags |= root->fs_info->avail_metadata_alloc_bits;
3561         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3562
3563         return btrfs_reduce_alloc_profile(root, flags);
3564 }
3565
3566 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3567 {
3568         u64 flags;
3569         u64 ret;
3570
3571         if (data)
3572                 flags = BTRFS_BLOCK_GROUP_DATA;
3573         else if (root == root->fs_info->chunk_root)
3574                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3575         else
3576                 flags = BTRFS_BLOCK_GROUP_METADATA;
3577
3578         ret = get_alloc_profile(root, flags);
3579         return ret;
3580 }
3581
3582 /*
3583  * This will check the space that the inode allocates from to make sure we have
3584  * enough space for bytes.
3585  */
3586 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3587 {
3588         struct btrfs_space_info *data_sinfo;
3589         struct btrfs_root *root = BTRFS_I(inode)->root;
3590         struct btrfs_fs_info *fs_info = root->fs_info;
3591         u64 used;
3592         int ret = 0, committed = 0, alloc_chunk = 1;
3593
3594         /* make sure bytes are sectorsize aligned */
3595         bytes = ALIGN(bytes, root->sectorsize);
3596
3597         if (btrfs_is_free_space_inode(inode)) {
3598                 committed = 1;
3599                 ASSERT(current->journal_info);
3600         }
3601
3602         data_sinfo = fs_info->data_sinfo;
3603         if (!data_sinfo)
3604                 goto alloc;
3605
3606 again:
3607         /* make sure we have enough space to handle the data first */
3608         spin_lock(&data_sinfo->lock);
3609         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3610                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3611                 data_sinfo->bytes_may_use;
3612
3613         if (used + bytes > data_sinfo->total_bytes) {
3614                 struct btrfs_trans_handle *trans;
3615
3616                 /*
3617                  * if we don't have enough free bytes in this space then we need
3618                  * to alloc a new chunk.
3619                  */
3620                 if (!data_sinfo->full && alloc_chunk) {
3621                         u64 alloc_target;
3622
3623                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3624                         spin_unlock(&data_sinfo->lock);
3625 alloc:
3626                         alloc_target = btrfs_get_alloc_profile(root, 1);
3627                         /*
3628                          * It is ugly that we don't call nolock join
3629                          * transaction for the free space inode case here.
3630                          * But it is safe because we only do the data space
3631                          * reservation for the free space cache in the
3632                          * transaction context, the common join transaction
3633                          * just increase the counter of the current transaction
3634                          * handler, doesn't try to acquire the trans_lock of
3635                          * the fs.
3636                          */
3637                         trans = btrfs_join_transaction(root);
3638                         if (IS_ERR(trans))
3639                                 return PTR_ERR(trans);
3640
3641                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3642                                              alloc_target,
3643                                              CHUNK_ALLOC_NO_FORCE);
3644                         btrfs_end_transaction(trans, root);
3645                         if (ret < 0) {
3646                                 if (ret != -ENOSPC)
3647                                         return ret;
3648                                 else
3649                                         goto commit_trans;
3650                         }
3651
3652                         if (!data_sinfo)
3653                                 data_sinfo = fs_info->data_sinfo;
3654
3655                         goto again;
3656                 }
3657
3658                 /*
3659                  * If we don't have enough pinned space to deal with this
3660                  * allocation don't bother committing the transaction.
3661                  */
3662                 if (percpu_counter_compare(&data_sinfo->total_bytes_pinned,
3663                                            bytes) < 0)
3664                         committed = 1;
3665                 spin_unlock(&data_sinfo->lock);
3666
3667                 /* commit the current transaction and try again */
3668 commit_trans:
3669                 if (!committed &&
3670                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3671                         committed = 1;
3672
3673                         trans = btrfs_join_transaction(root);
3674                         if (IS_ERR(trans))
3675                                 return PTR_ERR(trans);
3676                         ret = btrfs_commit_transaction(trans, root);
3677                         if (ret)
3678                                 return ret;
3679                         goto again;
3680                 }
3681
3682                 trace_btrfs_space_reservation(root->fs_info,
3683                                               "space_info:enospc",
3684                                               data_sinfo->flags, bytes, 1);
3685                 return -ENOSPC;
3686         }
3687         data_sinfo->bytes_may_use += bytes;
3688         trace_btrfs_space_reservation(root->fs_info, "space_info",
3689                                       data_sinfo->flags, bytes, 1);
3690         spin_unlock(&data_sinfo->lock);
3691
3692         return 0;
3693 }
3694
3695 /*
3696  * Called if we need to clear a data reservation for this inode.
3697  */
3698 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3699 {
3700         struct btrfs_root *root = BTRFS_I(inode)->root;
3701         struct btrfs_space_info *data_sinfo;
3702
3703         /* make sure bytes are sectorsize aligned */
3704         bytes = ALIGN(bytes, root->sectorsize);
3705
3706         data_sinfo = root->fs_info->data_sinfo;
3707         spin_lock(&data_sinfo->lock);
3708         WARN_ON(data_sinfo->bytes_may_use < bytes);
3709         data_sinfo->bytes_may_use -= bytes;
3710         trace_btrfs_space_reservation(root->fs_info, "space_info",
3711                                       data_sinfo->flags, bytes, 0);
3712         spin_unlock(&data_sinfo->lock);
3713 }
3714
3715 static void force_metadata_allocation(struct btrfs_fs_info *info)
3716 {
3717         struct list_head *head = &info->space_info;
3718         struct btrfs_space_info *found;
3719
3720         rcu_read_lock();
3721         list_for_each_entry_rcu(found, head, list) {
3722                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3723                         found->force_alloc = CHUNK_ALLOC_FORCE;
3724         }
3725         rcu_read_unlock();
3726 }
3727
3728 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3729 {
3730         return (global->size << 1);
3731 }
3732
3733 static int should_alloc_chunk(struct btrfs_root *root,
3734                               struct btrfs_space_info *sinfo, int force)
3735 {
3736         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3737         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3738         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3739         u64 thresh;
3740
3741         if (force == CHUNK_ALLOC_FORCE)
3742                 return 1;
3743
3744         /*
3745          * We need to take into account the global rsv because for all intents
3746          * and purposes it's used space.  Don't worry about locking the
3747          * global_rsv, it doesn't change except when the transaction commits.
3748          */
3749         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3750                 num_allocated += calc_global_rsv_need_space(global_rsv);
3751
3752         /*
3753          * in limited mode, we want to have some free space up to
3754          * about 1% of the FS size.
3755          */
3756         if (force == CHUNK_ALLOC_LIMITED) {
3757                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3758                 thresh = max_t(u64, 64 * 1024 * 1024,
3759                                div_factor_fine(thresh, 1));
3760
3761                 if (num_bytes - num_allocated < thresh)
3762                         return 1;
3763         }
3764
3765         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3766                 return 0;
3767         return 1;
3768 }
3769
3770 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3771 {
3772         u64 num_dev;
3773
3774         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3775                     BTRFS_BLOCK_GROUP_RAID0 |
3776                     BTRFS_BLOCK_GROUP_RAID5 |
3777                     BTRFS_BLOCK_GROUP_RAID6))
3778                 num_dev = root->fs_info->fs_devices->rw_devices;
3779         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3780                 num_dev = 2;
3781         else
3782                 num_dev = 1;    /* DUP or single */
3783
3784         /* metadata for updaing devices and chunk tree */
3785         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3786 }
3787
3788 static void check_system_chunk(struct btrfs_trans_handle *trans,
3789                                struct btrfs_root *root, u64 type)
3790 {
3791         struct btrfs_space_info *info;
3792         u64 left;
3793         u64 thresh;
3794
3795         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3796         spin_lock(&info->lock);
3797         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3798                 info->bytes_reserved - info->bytes_readonly;
3799         spin_unlock(&info->lock);
3800
3801         thresh = get_system_chunk_thresh(root, type);
3802         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3803                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3804                         left, thresh, type);
3805                 dump_space_info(info, 0, 0);
3806         }
3807
3808         if (left < thresh) {
3809                 u64 flags;
3810
3811                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3812                 btrfs_alloc_chunk(trans, root, flags);
3813         }
3814 }
3815
3816 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3817                           struct btrfs_root *extent_root, u64 flags, int force)
3818 {
3819         struct btrfs_space_info *space_info;
3820         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3821         int wait_for_alloc = 0;
3822         int ret = 0;
3823
3824         /* Don't re-enter if we're already allocating a chunk */
3825         if (trans->allocating_chunk)
3826                 return -ENOSPC;
3827
3828         space_info = __find_space_info(extent_root->fs_info, flags);
3829         if (!space_info) {
3830                 ret = update_space_info(extent_root->fs_info, flags,
3831                                         0, 0, &space_info);
3832                 BUG_ON(ret); /* -ENOMEM */
3833         }
3834         BUG_ON(!space_info); /* Logic error */
3835
3836 again:
3837         spin_lock(&space_info->lock);
3838         if (force < space_info->force_alloc)
3839                 force = space_info->force_alloc;
3840         if (space_info->full) {
3841                 if (should_alloc_chunk(extent_root, space_info, force))
3842                         ret = -ENOSPC;
3843                 else
3844                         ret = 0;
3845                 spin_unlock(&space_info->lock);
3846                 return ret;
3847         }
3848
3849         if (!should_alloc_chunk(extent_root, space_info, force)) {
3850                 spin_unlock(&space_info->lock);
3851                 return 0;
3852         } else if (space_info->chunk_alloc) {
3853                 wait_for_alloc = 1;
3854         } else {
3855                 space_info->chunk_alloc = 1;
3856         }
3857
3858         spin_unlock(&space_info->lock);
3859
3860         mutex_lock(&fs_info->chunk_mutex);
3861
3862         /*
3863          * The chunk_mutex is held throughout the entirety of a chunk
3864          * allocation, so once we've acquired the chunk_mutex we know that the
3865          * other guy is done and we need to recheck and see if we should
3866          * allocate.
3867          */
3868         if (wait_for_alloc) {
3869                 mutex_unlock(&fs_info->chunk_mutex);
3870                 wait_for_alloc = 0;
3871                 goto again;
3872         }
3873
3874         trans->allocating_chunk = true;
3875
3876         /*
3877          * If we have mixed data/metadata chunks we want to make sure we keep
3878          * allocating mixed chunks instead of individual chunks.
3879          */
3880         if (btrfs_mixed_space_info(space_info))
3881                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3882
3883         /*
3884          * if we're doing a data chunk, go ahead and make sure that
3885          * we keep a reasonable number of metadata chunks allocated in the
3886          * FS as well.
3887          */
3888         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3889                 fs_info->data_chunk_allocations++;
3890                 if (!(fs_info->data_chunk_allocations %
3891                       fs_info->metadata_ratio))
3892                         force_metadata_allocation(fs_info);
3893         }
3894
3895         /*
3896          * Check if we have enough space in SYSTEM chunk because we may need
3897          * to update devices.
3898          */
3899         check_system_chunk(trans, extent_root, flags);
3900
3901         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3902         trans->allocating_chunk = false;
3903
3904         spin_lock(&space_info->lock);
3905         if (ret < 0 && ret != -ENOSPC)
3906                 goto out;
3907         if (ret)
3908                 space_info->full = 1;
3909         else
3910                 ret = 1;
3911
3912         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3913 out:
3914         space_info->chunk_alloc = 0;
3915         spin_unlock(&space_info->lock);
3916         mutex_unlock(&fs_info->chunk_mutex);
3917         return ret;
3918 }
3919
3920 static int can_overcommit(struct btrfs_root *root,
3921                           struct btrfs_space_info *space_info, u64 bytes,
3922                           enum btrfs_reserve_flush_enum flush)
3923 {
3924         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3925         u64 profile = btrfs_get_alloc_profile(root, 0);
3926         u64 space_size;
3927         u64 avail;
3928         u64 used;
3929
3930         used = space_info->bytes_used + space_info->bytes_reserved +
3931                 space_info->bytes_pinned + space_info->bytes_readonly;
3932
3933         /*
3934          * We only want to allow over committing if we have lots of actual space
3935          * free, but if we don't have enough space to handle the global reserve
3936          * space then we could end up having a real enospc problem when trying
3937          * to allocate a chunk or some other such important allocation.
3938          */
3939         spin_lock(&global_rsv->lock);
3940         space_size = calc_global_rsv_need_space(global_rsv);
3941         spin_unlock(&global_rsv->lock);
3942         if (used + space_size >= space_info->total_bytes)
3943                 return 0;
3944
3945         used += space_info->bytes_may_use;
3946
3947         spin_lock(&root->fs_info->free_chunk_lock);
3948         avail = root->fs_info->free_chunk_space;
3949         spin_unlock(&root->fs_info->free_chunk_lock);
3950
3951         /*
3952          * If we have dup, raid1 or raid10 then only half of the free
3953          * space is actually useable.  For raid56, the space info used
3954          * doesn't include the parity drive, so we don't have to
3955          * change the math
3956          */
3957         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3958                        BTRFS_BLOCK_GROUP_RAID1 |
3959                        BTRFS_BLOCK_GROUP_RAID10))
3960                 avail >>= 1;
3961
3962         /*
3963          * If we aren't flushing all things, let us overcommit up to
3964          * 1/2th of the space. If we can flush, don't let us overcommit
3965          * too much, let it overcommit up to 1/8 of the space.
3966          */
3967         if (flush == BTRFS_RESERVE_FLUSH_ALL)
3968                 avail >>= 3;
3969         else
3970                 avail >>= 1;
3971
3972         if (used + bytes < space_info->total_bytes + avail)
3973                 return 1;
3974         return 0;
3975 }
3976
3977 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3978                                          unsigned long nr_pages, int nr_items)
3979 {
3980         struct super_block *sb = root->fs_info->sb;
3981
3982         if (down_read_trylock(&sb->s_umount)) {
3983                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
3984                 up_read(&sb->s_umount);
3985         } else {
3986                 /*
3987                  * We needn't worry the filesystem going from r/w to r/o though
3988                  * we don't acquire ->s_umount mutex, because the filesystem
3989                  * should guarantee the delalloc inodes list be empty after
3990                  * the filesystem is readonly(all dirty pages are written to
3991                  * the disk).
3992                  */
3993                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
3994                 if (!current->journal_info)
3995                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
3996         }
3997 }
3998
3999 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4000 {
4001         u64 bytes;
4002         int nr;
4003
4004         bytes = btrfs_calc_trans_metadata_size(root, 1);
4005         nr = (int)div64_u64(to_reclaim, bytes);
4006         if (!nr)
4007                 nr = 1;
4008         return nr;
4009 }
4010
4011 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4012
4013 /*
4014  * shrink metadata reservation for delalloc
4015  */
4016 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4017                             bool wait_ordered)
4018 {
4019         struct btrfs_block_rsv *block_rsv;
4020         struct btrfs_space_info *space_info;
4021         struct btrfs_trans_handle *trans;
4022         u64 delalloc_bytes;
4023         u64 max_reclaim;
4024         long time_left;
4025         unsigned long nr_pages;
4026         int loops;
4027         int items;
4028         enum btrfs_reserve_flush_enum flush;
4029
4030         /* Calc the number of the pages we need flush for space reservation */
4031         items = calc_reclaim_items_nr(root, to_reclaim);
4032         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4033
4034         trans = (struct btrfs_trans_handle *)current->journal_info;
4035         block_rsv = &root->fs_info->delalloc_block_rsv;
4036         space_info = block_rsv->space_info;
4037
4038         delalloc_bytes = percpu_counter_sum_positive(
4039                                                 &root->fs_info->delalloc_bytes);
4040         if (delalloc_bytes == 0) {
4041                 if (trans)
4042                         return;
4043                 if (wait_ordered)
4044                         btrfs_wait_ordered_roots(root->fs_info, items);
4045                 return;
4046         }
4047
4048         loops = 0;
4049         while (delalloc_bytes && loops < 3) {
4050                 max_reclaim = min(delalloc_bytes, to_reclaim);
4051                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4052                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4053                 /*
4054                  * We need to wait for the async pages to actually start before
4055                  * we do anything.
4056                  */
4057                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4058                 if (!max_reclaim)
4059                         goto skip_async;
4060
4061                 if (max_reclaim <= nr_pages)
4062                         max_reclaim = 0;
4063                 else
4064                         max_reclaim -= nr_pages;
4065
4066                 wait_event(root->fs_info->async_submit_wait,
4067                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4068                            (int)max_reclaim);
4069 skip_async:
4070                 if (!trans)
4071                         flush = BTRFS_RESERVE_FLUSH_ALL;
4072                 else
4073                         flush = BTRFS_RESERVE_NO_FLUSH;
4074                 spin_lock(&space_info->lock);
4075                 if (can_overcommit(root, space_info, orig, flush)) {
4076                         spin_unlock(&space_info->lock);
4077                         break;
4078                 }
4079                 spin_unlock(&space_info->lock);
4080
4081                 loops++;
4082                 if (wait_ordered && !trans) {
4083                         btrfs_wait_ordered_roots(root->fs_info, items);
4084                 } else {
4085                         time_left = schedule_timeout_killable(1);
4086                         if (time_left)
4087                                 break;
4088                 }
4089                 delalloc_bytes = percpu_counter_sum_positive(
4090                                                 &root->fs_info->delalloc_bytes);
4091         }
4092 }
4093
4094 /**
4095  * maybe_commit_transaction - possibly commit the transaction if its ok to
4096  * @root - the root we're allocating for
4097  * @bytes - the number of bytes we want to reserve
4098  * @force - force the commit
4099  *
4100  * This will check to make sure that committing the transaction will actually
4101  * get us somewhere and then commit the transaction if it does.  Otherwise it
4102  * will return -ENOSPC.
4103  */
4104 static int may_commit_transaction(struct btrfs_root *root,
4105                                   struct btrfs_space_info *space_info,
4106                                   u64 bytes, int force)
4107 {
4108         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4109         struct btrfs_trans_handle *trans;
4110
4111         trans = (struct btrfs_trans_handle *)current->journal_info;
4112         if (trans)
4113                 return -EAGAIN;
4114
4115         if (force)
4116                 goto commit;
4117
4118         /* See if there is enough pinned space to make this reservation */
4119         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4120                                    bytes) >= 0)
4121                 goto commit;
4122
4123         /*
4124          * See if there is some space in the delayed insertion reservation for
4125          * this reservation.
4126          */
4127         if (space_info != delayed_rsv->space_info)
4128                 return -ENOSPC;
4129
4130         spin_lock(&delayed_rsv->lock);
4131         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4132                                    bytes - delayed_rsv->size) >= 0) {
4133                 spin_unlock(&delayed_rsv->lock);
4134                 return -ENOSPC;
4135         }
4136         spin_unlock(&delayed_rsv->lock);
4137
4138 commit:
4139         trans = btrfs_join_transaction(root);
4140         if (IS_ERR(trans))
4141                 return -ENOSPC;
4142
4143         return btrfs_commit_transaction(trans, root);
4144 }
4145
4146 enum flush_state {
4147         FLUSH_DELAYED_ITEMS_NR  =       1,
4148         FLUSH_DELAYED_ITEMS     =       2,
4149         FLUSH_DELALLOC          =       3,
4150         FLUSH_DELALLOC_WAIT     =       4,
4151         ALLOC_CHUNK             =       5,
4152         COMMIT_TRANS            =       6,
4153 };
4154
4155 static int flush_space(struct btrfs_root *root,
4156                        struct btrfs_space_info *space_info, u64 num_bytes,
4157                        u64 orig_bytes, int state)
4158 {
4159         struct btrfs_trans_handle *trans;
4160         int nr;
4161         int ret = 0;
4162
4163         switch (state) {
4164         case FLUSH_DELAYED_ITEMS_NR:
4165         case FLUSH_DELAYED_ITEMS:
4166                 if (state == FLUSH_DELAYED_ITEMS_NR)
4167                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4168                 else
4169                         nr = -1;
4170
4171                 trans = btrfs_join_transaction(root);
4172                 if (IS_ERR(trans)) {
4173                         ret = PTR_ERR(trans);
4174                         break;
4175                 }
4176                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4177                 btrfs_end_transaction(trans, root);
4178                 break;
4179         case FLUSH_DELALLOC:
4180         case FLUSH_DELALLOC_WAIT:
4181                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4182                                 state == FLUSH_DELALLOC_WAIT);
4183                 break;
4184         case ALLOC_CHUNK:
4185                 trans = btrfs_join_transaction(root);
4186                 if (IS_ERR(trans)) {
4187                         ret = PTR_ERR(trans);
4188                         break;
4189                 }
4190                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4191                                      btrfs_get_alloc_profile(root, 0),
4192                                      CHUNK_ALLOC_NO_FORCE);
4193                 btrfs_end_transaction(trans, root);
4194                 if (ret == -ENOSPC)
4195                         ret = 0;
4196                 break;
4197         case COMMIT_TRANS:
4198                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4199                 break;
4200         default:
4201                 ret = -ENOSPC;
4202                 break;
4203         }
4204
4205         return ret;
4206 }
4207
4208 static inline u64
4209 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4210                                  struct btrfs_space_info *space_info)
4211 {
4212         u64 used;
4213         u64 expected;
4214         u64 to_reclaim;
4215
4216         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4217                                 16 * 1024 * 1024);
4218         spin_lock(&space_info->lock);
4219         if (can_overcommit(root, space_info, to_reclaim,
4220                            BTRFS_RESERVE_FLUSH_ALL)) {
4221                 to_reclaim = 0;
4222                 goto out;
4223         }
4224
4225         used = space_info->bytes_used + space_info->bytes_reserved +
4226                space_info->bytes_pinned + space_info->bytes_readonly +
4227                space_info->bytes_may_use;
4228         if (can_overcommit(root, space_info, 1024 * 1024,
4229                            BTRFS_RESERVE_FLUSH_ALL))
4230                 expected = div_factor_fine(space_info->total_bytes, 95);
4231         else
4232                 expected = div_factor_fine(space_info->total_bytes, 90);
4233
4234         if (used > expected)
4235                 to_reclaim = used - expected;
4236         else
4237                 to_reclaim = 0;
4238         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4239                                      space_info->bytes_reserved);
4240 out:
4241         spin_unlock(&space_info->lock);
4242
4243         return to_reclaim;
4244 }
4245
4246 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4247                                         struct btrfs_fs_info *fs_info, u64 used)
4248 {
4249         return (used >= div_factor_fine(space_info->total_bytes, 98) &&
4250                 !btrfs_fs_closing(fs_info) &&
4251                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4252 }
4253
4254 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4255                                        struct btrfs_fs_info *fs_info)
4256 {
4257         u64 used;
4258
4259         spin_lock(&space_info->lock);
4260         used = space_info->bytes_used + space_info->bytes_reserved +
4261                space_info->bytes_pinned + space_info->bytes_readonly +
4262                space_info->bytes_may_use;
4263         if (need_do_async_reclaim(space_info, fs_info, used)) {
4264                 spin_unlock(&space_info->lock);
4265                 return 1;
4266         }
4267         spin_unlock(&space_info->lock);
4268
4269         return 0;
4270 }
4271
4272 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4273 {
4274         struct btrfs_fs_info *fs_info;
4275         struct btrfs_space_info *space_info;
4276         u64 to_reclaim;
4277         int flush_state;
4278
4279         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4280         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4281
4282         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4283                                                       space_info);
4284         if (!to_reclaim)
4285                 return;
4286
4287         flush_state = FLUSH_DELAYED_ITEMS_NR;
4288         do {
4289                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4290                             to_reclaim, flush_state);
4291                 flush_state++;
4292                 if (!btrfs_need_do_async_reclaim(space_info, fs_info))
4293                         return;
4294         } while (flush_state <= COMMIT_TRANS);
4295
4296         if (btrfs_need_do_async_reclaim(space_info, fs_info))
4297                 queue_work(system_unbound_wq, work);
4298 }
4299
4300 void btrfs_init_async_reclaim_work(struct work_struct *work)
4301 {
4302         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4303 }
4304
4305 /**
4306  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4307  * @root - the root we're allocating for
4308  * @block_rsv - the block_rsv we're allocating for
4309  * @orig_bytes - the number of bytes we want
4310  * @flush - whether or not we can flush to make our reservation
4311  *
4312  * This will reserve orgi_bytes number of bytes from the space info associated
4313  * with the block_rsv.  If there is not enough space it will make an attempt to
4314  * flush out space to make room.  It will do this by flushing delalloc if
4315  * possible or committing the transaction.  If flush is 0 then no attempts to
4316  * regain reservations will be made and this will fail if there is not enough
4317  * space already.
4318  */
4319 static int reserve_metadata_bytes(struct btrfs_root *root,
4320                                   struct btrfs_block_rsv *block_rsv,
4321                                   u64 orig_bytes,
4322                                   enum btrfs_reserve_flush_enum flush)
4323 {
4324         struct btrfs_space_info *space_info = block_rsv->space_info;
4325         u64 used;
4326         u64 num_bytes = orig_bytes;
4327         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4328         int ret = 0;
4329         bool flushing = false;
4330
4331 again:
4332         ret = 0;
4333         spin_lock(&space_info->lock);
4334         /*
4335          * We only want to wait if somebody other than us is flushing and we
4336          * are actually allowed to flush all things.
4337          */
4338         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4339                space_info->flush) {
4340                 spin_unlock(&space_info->lock);
4341                 /*
4342                  * If we have a trans handle we can't wait because the flusher
4343                  * may have to commit the transaction, which would mean we would
4344                  * deadlock since we are waiting for the flusher to finish, but
4345                  * hold the current transaction open.
4346                  */
4347                 if (current->journal_info)
4348                         return -EAGAIN;
4349                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4350                 /* Must have been killed, return */
4351                 if (ret)
4352                         return -EINTR;
4353
4354                 spin_lock(&space_info->lock);
4355         }
4356
4357         ret = -ENOSPC;
4358         used = space_info->bytes_used + space_info->bytes_reserved +
4359                 space_info->bytes_pinned + space_info->bytes_readonly +
4360                 space_info->bytes_may_use;
4361
4362         /*
4363          * The idea here is that we've not already over-reserved the block group
4364          * then we can go ahead and save our reservation first and then start
4365          * flushing if we need to.  Otherwise if we've already overcommitted
4366          * lets start flushing stuff first and then come back and try to make
4367          * our reservation.
4368          */
4369         if (used <= space_info->total_bytes) {
4370                 if (used + orig_bytes <= space_info->total_bytes) {
4371                         space_info->bytes_may_use += orig_bytes;
4372                         trace_btrfs_space_reservation(root->fs_info,
4373                                 "space_info", space_info->flags, orig_bytes, 1);
4374                         ret = 0;
4375                 } else {
4376                         /*
4377                          * Ok set num_bytes to orig_bytes since we aren't
4378                          * overocmmitted, this way we only try and reclaim what
4379                          * we need.
4380                          */
4381                         num_bytes = orig_bytes;
4382                 }
4383         } else {
4384                 /*
4385                  * Ok we're over committed, set num_bytes to the overcommitted
4386                  * amount plus the amount of bytes that we need for this
4387                  * reservation.
4388                  */
4389                 num_bytes = used - space_info->total_bytes +
4390                         (orig_bytes * 2);
4391         }
4392
4393         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4394                 space_info->bytes_may_use += orig_bytes;
4395                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4396                                               space_info->flags, orig_bytes,
4397                                               1);
4398                 ret = 0;
4399         }
4400
4401         /*
4402          * Couldn't make our reservation, save our place so while we're trying
4403          * to reclaim space we can actually use it instead of somebody else
4404          * stealing it from us.
4405          *
4406          * We make the other tasks wait for the flush only when we can flush
4407          * all things.
4408          */
4409         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4410                 flushing = true;
4411                 space_info->flush = 1;
4412         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4413                 used += orig_bytes;
4414                 if (need_do_async_reclaim(space_info, root->fs_info, used) &&
4415                     !work_busy(&root->fs_info->async_reclaim_work))
4416                         queue_work(system_unbound_wq,
4417                                    &root->fs_info->async_reclaim_work);
4418         }
4419         spin_unlock(&space_info->lock);
4420
4421         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4422                 goto out;
4423
4424         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4425                           flush_state);
4426         flush_state++;
4427
4428         /*
4429          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4430          * would happen. So skip delalloc flush.
4431          */
4432         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4433             (flush_state == FLUSH_DELALLOC ||
4434              flush_state == FLUSH_DELALLOC_WAIT))
4435                 flush_state = ALLOC_CHUNK;
4436
4437         if (!ret)
4438                 goto again;
4439         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4440                  flush_state < COMMIT_TRANS)
4441                 goto again;
4442         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4443                  flush_state <= COMMIT_TRANS)
4444                 goto again;
4445
4446 out:
4447         if (ret == -ENOSPC &&
4448             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4449                 struct btrfs_block_rsv *global_rsv =
4450                         &root->fs_info->global_block_rsv;
4451
4452                 if (block_rsv != global_rsv &&
4453                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4454                         ret = 0;
4455         }
4456         if (ret == -ENOSPC)
4457                 trace_btrfs_space_reservation(root->fs_info,
4458                                               "space_info:enospc",
4459                                               space_info->flags, orig_bytes, 1);
4460         if (flushing) {
4461                 spin_lock(&space_info->lock);
4462                 space_info->flush = 0;
4463                 wake_up_all(&space_info->wait);
4464                 spin_unlock(&space_info->lock);
4465         }
4466         return ret;
4467 }
4468
4469 static struct btrfs_block_rsv *get_block_rsv(
4470                                         const struct btrfs_trans_handle *trans,
4471                                         const struct btrfs_root *root)
4472 {
4473         struct btrfs_block_rsv *block_rsv = NULL;
4474
4475         if (root->ref_cows)
4476                 block_rsv = trans->block_rsv;
4477
4478         if (root == root->fs_info->csum_root && trans->adding_csums)
4479                 block_rsv = trans->block_rsv;
4480
4481         if (root == root->fs_info->uuid_root)
4482                 block_rsv = trans->block_rsv;
4483
4484         if (!block_rsv)
4485                 block_rsv = root->block_rsv;
4486
4487         if (!block_rsv)
4488                 block_rsv = &root->fs_info->empty_block_rsv;
4489
4490         return block_rsv;
4491 }
4492
4493 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4494                                u64 num_bytes)
4495 {
4496         int ret = -ENOSPC;
4497         spin_lock(&block_rsv->lock);
4498         if (block_rsv->reserved >= num_bytes) {
4499                 block_rsv->reserved -= num_bytes;
4500                 if (block_rsv->reserved < block_rsv->size)
4501                         block_rsv->full = 0;
4502                 ret = 0;
4503         }
4504         spin_unlock(&block_rsv->lock);
4505         return ret;
4506 }
4507
4508 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4509                                 u64 num_bytes, int update_size)
4510 {
4511         spin_lock(&block_rsv->lock);
4512         block_rsv->reserved += num_bytes;
4513         if (update_size)
4514                 block_rsv->size += num_bytes;
4515         else if (block_rsv->reserved >= block_rsv->size)
4516                 block_rsv->full = 1;
4517         spin_unlock(&block_rsv->lock);
4518 }
4519
4520 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4521                              struct btrfs_block_rsv *dest, u64 num_bytes,
4522                              int min_factor)
4523 {
4524         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4525         u64 min_bytes;
4526
4527         if (global_rsv->space_info != dest->space_info)
4528                 return -ENOSPC;
4529
4530         spin_lock(&global_rsv->lock);
4531         min_bytes = div_factor(global_rsv->size, min_factor);
4532         if (global_rsv->reserved < min_bytes + num_bytes) {
4533                 spin_unlock(&global_rsv->lock);
4534                 return -ENOSPC;
4535         }
4536         global_rsv->reserved -= num_bytes;
4537         if (global_rsv->reserved < global_rsv->size)
4538                 global_rsv->full = 0;
4539         spin_unlock(&global_rsv->lock);
4540
4541         block_rsv_add_bytes(dest, num_bytes, 1);
4542         return 0;
4543 }
4544
4545 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4546                                     struct btrfs_block_rsv *block_rsv,
4547                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4548 {
4549         struct btrfs_space_info *space_info = block_rsv->space_info;
4550
4551         spin_lock(&block_rsv->lock);
4552         if (num_bytes == (u64)-1)
4553                 num_bytes = block_rsv->size;
4554         block_rsv->size -= num_bytes;
4555         if (block_rsv->reserved >= block_rsv->size) {
4556                 num_bytes = block_rsv->reserved - block_rsv->size;
4557                 block_rsv->reserved = block_rsv->size;
4558                 block_rsv->full = 1;
4559         } else {
4560                 num_bytes = 0;
4561         }
4562         spin_unlock(&block_rsv->lock);
4563
4564         if (num_bytes > 0) {
4565                 if (dest) {
4566                         spin_lock(&dest->lock);
4567                         if (!dest->full) {
4568                                 u64 bytes_to_add;
4569
4570                                 bytes_to_add = dest->size - dest->reserved;
4571                                 bytes_to_add = min(num_bytes, bytes_to_add);
4572                                 dest->reserved += bytes_to_add;
4573                                 if (dest->reserved >= dest->size)
4574                                         dest->full = 1;
4575                                 num_bytes -= bytes_to_add;
4576                         }
4577                         spin_unlock(&dest->lock);
4578                 }
4579                 if (num_bytes) {
4580                         spin_lock(&space_info->lock);
4581                         space_info->bytes_may_use -= num_bytes;
4582                         trace_btrfs_space_reservation(fs_info, "space_info",
4583                                         space_info->flags, num_bytes, 0);
4584                         spin_unlock(&space_info->lock);
4585                 }
4586         }
4587 }
4588
4589 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4590                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4591 {
4592         int ret;
4593
4594         ret = block_rsv_use_bytes(src, num_bytes);
4595         if (ret)
4596                 return ret;
4597
4598         block_rsv_add_bytes(dst, num_bytes, 1);
4599         return 0;
4600 }
4601
4602 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4603 {
4604         memset(rsv, 0, sizeof(*rsv));
4605         spin_lock_init(&rsv->lock);
4606         rsv->type = type;
4607 }
4608
4609 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4610                                               unsigned short type)
4611 {
4612         struct btrfs_block_rsv *block_rsv;
4613         struct btrfs_fs_info *fs_info = root->fs_info;
4614
4615         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4616         if (!block_rsv)
4617                 return NULL;
4618
4619         btrfs_init_block_rsv(block_rsv, type);
4620         block_rsv->space_info = __find_space_info(fs_info,
4621                                                   BTRFS_BLOCK_GROUP_METADATA);
4622         return block_rsv;
4623 }
4624
4625 void btrfs_free_block_rsv(struct btrfs_root *root,
4626                           struct btrfs_block_rsv *rsv)
4627 {
4628         if (!rsv)
4629                 return;
4630         btrfs_block_rsv_release(root, rsv, (u64)-1);
4631         kfree(rsv);
4632 }
4633
4634 int btrfs_block_rsv_add(struct btrfs_root *root,
4635                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4636                         enum btrfs_reserve_flush_enum flush)
4637 {
4638         int ret;
4639
4640         if (num_bytes == 0)
4641                 return 0;
4642
4643         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4644         if (!ret) {
4645                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4646                 return 0;
4647         }
4648
4649         return ret;
4650 }
4651
4652 int btrfs_block_rsv_check(struct btrfs_root *root,
4653                           struct btrfs_block_rsv *block_rsv, int min_factor)
4654 {
4655         u64 num_bytes = 0;
4656         int ret = -ENOSPC;
4657
4658         if (!block_rsv)
4659                 return 0;
4660
4661         spin_lock(&block_rsv->lock);
4662         num_bytes = div_factor(block_rsv->size, min_factor);
4663         if (block_rsv->reserved >= num_bytes)
4664                 ret = 0;
4665         spin_unlock(&block_rsv->lock);
4666
4667         return ret;
4668 }
4669
4670 int btrfs_block_rsv_refill(struct btrfs_root *root,
4671                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4672                            enum btrfs_reserve_flush_enum flush)
4673 {
4674         u64 num_bytes = 0;
4675         int ret = -ENOSPC;
4676
4677         if (!block_rsv)
4678                 return 0;
4679
4680         spin_lock(&block_rsv->lock);
4681         num_bytes = min_reserved;
4682         if (block_rsv->reserved >= num_bytes)
4683                 ret = 0;
4684         else
4685                 num_bytes -= block_rsv->reserved;
4686         spin_unlock(&block_rsv->lock);
4687
4688         if (!ret)
4689                 return 0;
4690
4691         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4692         if (!ret) {
4693                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4694                 return 0;
4695         }
4696
4697         return ret;
4698 }
4699
4700 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4701                             struct btrfs_block_rsv *dst_rsv,
4702                             u64 num_bytes)
4703 {
4704         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4705 }
4706
4707 void btrfs_block_rsv_release(struct btrfs_root *root,
4708                              struct btrfs_block_rsv *block_rsv,
4709                              u64 num_bytes)
4710 {
4711         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4712         if (global_rsv == block_rsv ||
4713             block_rsv->space_info != global_rsv->space_info)
4714                 global_rsv = NULL;
4715         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4716                                 num_bytes);
4717 }
4718
4719 /*
4720  * helper to calculate size of global block reservation.
4721  * the desired value is sum of space used by extent tree,
4722  * checksum tree and root tree
4723  */
4724 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4725 {
4726         struct btrfs_space_info *sinfo;
4727         u64 num_bytes;
4728         u64 meta_used;
4729         u64 data_used;
4730         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4731
4732         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4733         spin_lock(&sinfo->lock);
4734         data_used = sinfo->bytes_used;
4735         spin_unlock(&sinfo->lock);
4736
4737         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4738         spin_lock(&sinfo->lock);
4739         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4740                 data_used = 0;
4741         meta_used = sinfo->bytes_used;
4742         spin_unlock(&sinfo->lock);
4743
4744         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4745                     csum_size * 2;
4746         num_bytes += div64_u64(data_used + meta_used, 50);
4747
4748         if (num_bytes * 3 > meta_used)
4749                 num_bytes = div64_u64(meta_used, 3);
4750
4751         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4752 }
4753
4754 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4755 {
4756         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4757         struct btrfs_space_info *sinfo = block_rsv->space_info;
4758         u64 num_bytes;
4759
4760         num_bytes = calc_global_metadata_size(fs_info);
4761
4762         spin_lock(&sinfo->lock);
4763         spin_lock(&block_rsv->lock);
4764
4765         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4766
4767         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4768                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4769                     sinfo->bytes_may_use;
4770
4771         if (sinfo->total_bytes > num_bytes) {
4772                 num_bytes = sinfo->total_bytes - num_bytes;
4773                 block_rsv->reserved += num_bytes;
4774                 sinfo->bytes_may_use += num_bytes;
4775                 trace_btrfs_space_reservation(fs_info, "space_info",
4776                                       sinfo->flags, num_bytes, 1);
4777         }
4778
4779         if (block_rsv->reserved >= block_rsv->size) {
4780                 num_bytes = block_rsv->reserved - block_rsv->size;
4781                 sinfo->bytes_may_use -= num_bytes;
4782                 trace_btrfs_space_reservation(fs_info, "space_info",
4783                                       sinfo->flags, num_bytes, 0);
4784                 block_rsv->reserved = block_rsv->size;
4785                 block_rsv->full = 1;
4786         }
4787
4788         spin_unlock(&block_rsv->lock);
4789         spin_unlock(&sinfo->lock);
4790 }
4791
4792 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4793 {
4794         struct btrfs_space_info *space_info;
4795
4796         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4797         fs_info->chunk_block_rsv.space_info = space_info;
4798
4799         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4800         fs_info->global_block_rsv.space_info = space_info;
4801         fs_info->delalloc_block_rsv.space_info = space_info;
4802         fs_info->trans_block_rsv.space_info = space_info;
4803         fs_info->empty_block_rsv.space_info = space_info;
4804         fs_info->delayed_block_rsv.space_info = space_info;
4805
4806         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4807         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4808         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4809         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4810         if (fs_info->quota_root)
4811                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
4812         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4813
4814         update_global_block_rsv(fs_info);
4815 }
4816
4817 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4818 {
4819         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4820                                 (u64)-1);
4821         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4822         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4823         WARN_ON(fs_info->trans_block_rsv.size > 0);
4824         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4825         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4826         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4827         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4828         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4829 }
4830
4831 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4832                                   struct btrfs_root *root)
4833 {
4834         if (!trans->block_rsv)
4835                 return;
4836
4837         if (!trans->bytes_reserved)
4838                 return;
4839
4840         trace_btrfs_space_reservation(root->fs_info, "transaction",
4841                                       trans->transid, trans->bytes_reserved, 0);
4842         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4843         trans->bytes_reserved = 0;
4844 }
4845
4846 /* Can only return 0 or -ENOSPC */
4847 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4848                                   struct inode *inode)
4849 {
4850         struct btrfs_root *root = BTRFS_I(inode)->root;
4851         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4852         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4853
4854         /*
4855          * We need to hold space in order to delete our orphan item once we've
4856          * added it, so this takes the reservation so we can release it later
4857          * when we are truly done with the orphan item.
4858          */
4859         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4860         trace_btrfs_space_reservation(root->fs_info, "orphan",
4861                                       btrfs_ino(inode), num_bytes, 1);
4862         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4863 }
4864
4865 void btrfs_orphan_release_metadata(struct inode *inode)
4866 {
4867         struct btrfs_root *root = BTRFS_I(inode)->root;
4868         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4869         trace_btrfs_space_reservation(root->fs_info, "orphan",
4870                                       btrfs_ino(inode), num_bytes, 0);
4871         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4872 }
4873
4874 /*
4875  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4876  * root: the root of the parent directory
4877  * rsv: block reservation
4878  * items: the number of items that we need do reservation
4879  * qgroup_reserved: used to return the reserved size in qgroup
4880  *
4881  * This function is used to reserve the space for snapshot/subvolume
4882  * creation and deletion. Those operations are different with the
4883  * common file/directory operations, they change two fs/file trees
4884  * and root tree, the number of items that the qgroup reserves is
4885  * different with the free space reservation. So we can not use
4886  * the space reseravtion mechanism in start_transaction().
4887  */
4888 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4889                                      struct btrfs_block_rsv *rsv,
4890                                      int items,
4891                                      u64 *qgroup_reserved,
4892                                      bool use_global_rsv)
4893 {
4894         u64 num_bytes;
4895         int ret;
4896         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4897
4898         if (root->fs_info->quota_enabled) {
4899                 /* One for parent inode, two for dir entries */
4900                 num_bytes = 3 * root->leafsize;
4901                 ret = btrfs_qgroup_reserve(root, num_bytes);
4902                 if (ret)
4903                         return ret;
4904         } else {
4905                 num_bytes = 0;
4906         }
4907
4908         *qgroup_reserved = num_bytes;
4909
4910         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4911         rsv->space_info = __find_space_info(root->fs_info,
4912                                             BTRFS_BLOCK_GROUP_METADATA);
4913         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4914                                   BTRFS_RESERVE_FLUSH_ALL);
4915
4916         if (ret == -ENOSPC && use_global_rsv)
4917                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
4918
4919         if (ret) {
4920                 if (*qgroup_reserved)
4921                         btrfs_qgroup_free(root, *qgroup_reserved);
4922         }
4923
4924         return ret;
4925 }
4926
4927 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4928                                       struct btrfs_block_rsv *rsv,
4929                                       u64 qgroup_reserved)
4930 {
4931         btrfs_block_rsv_release(root, rsv, (u64)-1);
4932         if (qgroup_reserved)
4933                 btrfs_qgroup_free(root, qgroup_reserved);
4934 }
4935
4936 /**
4937  * drop_outstanding_extent - drop an outstanding extent
4938  * @inode: the inode we're dropping the extent for
4939  *
4940  * This is called when we are freeing up an outstanding extent, either called
4941  * after an error or after an extent is written.  This will return the number of
4942  * reserved extents that need to be freed.  This must be called with
4943  * BTRFS_I(inode)->lock held.
4944  */
4945 static unsigned drop_outstanding_extent(struct inode *inode)
4946 {
4947         unsigned drop_inode_space = 0;
4948         unsigned dropped_extents = 0;
4949
4950         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4951         BTRFS_I(inode)->outstanding_extents--;
4952
4953         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4954             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4955                                &BTRFS_I(inode)->runtime_flags))
4956                 drop_inode_space = 1;
4957
4958         /*
4959          * If we have more or the same amount of outsanding extents than we have
4960          * reserved then we need to leave the reserved extents count alone.
4961          */
4962         if (BTRFS_I(inode)->outstanding_extents >=
4963             BTRFS_I(inode)->reserved_extents)
4964                 return drop_inode_space;
4965
4966         dropped_extents = BTRFS_I(inode)->reserved_extents -
4967                 BTRFS_I(inode)->outstanding_extents;
4968         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4969         return dropped_extents + drop_inode_space;
4970 }
4971
4972 /**
4973  * calc_csum_metadata_size - return the amount of metada space that must be
4974  *      reserved/free'd for the given bytes.
4975  * @inode: the inode we're manipulating
4976  * @num_bytes: the number of bytes in question
4977  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4978  *
4979  * This adjusts the number of csum_bytes in the inode and then returns the
4980  * correct amount of metadata that must either be reserved or freed.  We
4981  * calculate how many checksums we can fit into one leaf and then divide the
4982  * number of bytes that will need to be checksumed by this value to figure out
4983  * how many checksums will be required.  If we are adding bytes then the number
4984  * may go up and we will return the number of additional bytes that must be
4985  * reserved.  If it is going down we will return the number of bytes that must
4986  * be freed.
4987  *
4988  * This must be called with BTRFS_I(inode)->lock held.
4989  */
4990 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4991                                    int reserve)
4992 {
4993         struct btrfs_root *root = BTRFS_I(inode)->root;
4994         u64 csum_size;
4995         int num_csums_per_leaf;
4996         int num_csums;
4997         int old_csums;
4998
4999         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5000             BTRFS_I(inode)->csum_bytes == 0)
5001                 return 0;
5002
5003         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
5004         if (reserve)
5005                 BTRFS_I(inode)->csum_bytes += num_bytes;
5006         else
5007                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5008         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
5009         num_csums_per_leaf = (int)div64_u64(csum_size,
5010                                             sizeof(struct btrfs_csum_item) +
5011                                             sizeof(struct btrfs_disk_key));
5012         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
5013         num_csums = num_csums + num_csums_per_leaf - 1;
5014         num_csums = num_csums / num_csums_per_leaf;
5015
5016         old_csums = old_csums + num_csums_per_leaf - 1;
5017         old_csums = old_csums / num_csums_per_leaf;
5018
5019         /* No change, no need to reserve more */
5020         if (old_csums == num_csums)
5021                 return 0;
5022
5023         if (reserve)
5024                 return btrfs_calc_trans_metadata_size(root,
5025                                                       num_csums - old_csums);
5026
5027         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5028 }
5029
5030 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5031 {
5032         struct btrfs_root *root = BTRFS_I(inode)->root;
5033         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5034         u64 to_reserve = 0;
5035         u64 csum_bytes;
5036         unsigned nr_extents = 0;
5037         int extra_reserve = 0;
5038         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5039         int ret = 0;
5040         bool delalloc_lock = true;
5041         u64 to_free = 0;
5042         unsigned dropped;
5043
5044         /* If we are a free space inode we need to not flush since we will be in
5045          * the middle of a transaction commit.  We also don't need the delalloc
5046          * mutex since we won't race with anybody.  We need this mostly to make
5047          * lockdep shut its filthy mouth.
5048          */
5049         if (btrfs_is_free_space_inode(inode)) {
5050                 flush = BTRFS_RESERVE_NO_FLUSH;
5051                 delalloc_lock = false;
5052         }
5053
5054         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5055             btrfs_transaction_in_commit(root->fs_info))
5056                 schedule_timeout(1);
5057
5058         if (delalloc_lock)
5059                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5060
5061         num_bytes = ALIGN(num_bytes, root->sectorsize);
5062
5063         spin_lock(&BTRFS_I(inode)->lock);
5064         BTRFS_I(inode)->outstanding_extents++;
5065
5066         if (BTRFS_I(inode)->outstanding_extents >
5067             BTRFS_I(inode)->reserved_extents)
5068                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5069                         BTRFS_I(inode)->reserved_extents;
5070
5071         /*
5072          * Add an item to reserve for updating the inode when we complete the
5073          * delalloc io.
5074          */
5075         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5076                       &BTRFS_I(inode)->runtime_flags)) {
5077                 nr_extents++;
5078                 extra_reserve = 1;
5079         }
5080
5081         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5082         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5083         csum_bytes = BTRFS_I(inode)->csum_bytes;
5084         spin_unlock(&BTRFS_I(inode)->lock);
5085
5086         if (root->fs_info->quota_enabled) {
5087                 ret = btrfs_qgroup_reserve(root, num_bytes +
5088                                            nr_extents * root->leafsize);
5089                 if (ret)
5090                         goto out_fail;
5091         }
5092
5093         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5094         if (unlikely(ret)) {
5095                 if (root->fs_info->quota_enabled)
5096                         btrfs_qgroup_free(root, num_bytes +
5097                                                 nr_extents * root->leafsize);
5098                 goto out_fail;
5099         }
5100
5101         spin_lock(&BTRFS_I(inode)->lock);
5102         if (extra_reserve) {
5103                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5104                         &BTRFS_I(inode)->runtime_flags);
5105                 nr_extents--;
5106         }
5107         BTRFS_I(inode)->reserved_extents += nr_extents;
5108         spin_unlock(&BTRFS_I(inode)->lock);
5109
5110         if (delalloc_lock)
5111                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5112
5113         if (to_reserve)
5114                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5115                                               btrfs_ino(inode), to_reserve, 1);
5116         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5117
5118         return 0;
5119
5120 out_fail:
5121         spin_lock(&BTRFS_I(inode)->lock);
5122         dropped = drop_outstanding_extent(inode);
5123         /*
5124          * If the inodes csum_bytes is the same as the original
5125          * csum_bytes then we know we haven't raced with any free()ers
5126          * so we can just reduce our inodes csum bytes and carry on.
5127          */
5128         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5129                 calc_csum_metadata_size(inode, num_bytes, 0);
5130         } else {
5131                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5132                 u64 bytes;
5133
5134                 /*
5135                  * This is tricky, but first we need to figure out how much we
5136                  * free'd from any free-ers that occured during this
5137                  * reservation, so we reset ->csum_bytes to the csum_bytes
5138                  * before we dropped our lock, and then call the free for the
5139                  * number of bytes that were freed while we were trying our
5140                  * reservation.
5141                  */
5142                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5143                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5144                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5145
5146
5147                 /*
5148                  * Now we need to see how much we would have freed had we not
5149                  * been making this reservation and our ->csum_bytes were not
5150                  * artificially inflated.
5151                  */
5152                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5153                 bytes = csum_bytes - orig_csum_bytes;
5154                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5155
5156                 /*
5157                  * Now reset ->csum_bytes to what it should be.  If bytes is
5158                  * more than to_free then we would have free'd more space had we
5159                  * not had an artificially high ->csum_bytes, so we need to free
5160                  * the remainder.  If bytes is the same or less then we don't
5161                  * need to do anything, the other free-ers did the correct
5162                  * thing.
5163                  */
5164                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5165                 if (bytes > to_free)
5166                         to_free = bytes - to_free;
5167                 else
5168                         to_free = 0;
5169         }
5170         spin_unlock(&BTRFS_I(inode)->lock);
5171         if (dropped)
5172                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5173
5174         if (to_free) {
5175                 btrfs_block_rsv_release(root, block_rsv, to_free);
5176                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5177                                               btrfs_ino(inode), to_free, 0);
5178         }
5179         if (delalloc_lock)
5180                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5181         return ret;
5182 }
5183
5184 /**
5185  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5186  * @inode: the inode to release the reservation for
5187  * @num_bytes: the number of bytes we're releasing
5188  *
5189  * This will release the metadata reservation for an inode.  This can be called
5190  * once we complete IO for a given set of bytes to release their metadata
5191  * reservations.
5192  */
5193 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5194 {
5195         struct btrfs_root *root = BTRFS_I(inode)->root;
5196         u64 to_free = 0;
5197         unsigned dropped;
5198
5199         num_bytes = ALIGN(num_bytes, root->sectorsize);
5200         spin_lock(&BTRFS_I(inode)->lock);
5201         dropped = drop_outstanding_extent(inode);
5202
5203         if (num_bytes)
5204                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5205         spin_unlock(&BTRFS_I(inode)->lock);
5206         if (dropped > 0)
5207                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5208
5209         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5210                                       btrfs_ino(inode), to_free, 0);
5211         if (root->fs_info->quota_enabled) {
5212                 btrfs_qgroup_free(root, num_bytes +
5213                                         dropped * root->leafsize);
5214         }
5215
5216         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5217                                 to_free);
5218 }
5219
5220 /**
5221  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5222  * @inode: inode we're writing to
5223  * @num_bytes: the number of bytes we want to allocate
5224  *
5225  * This will do the following things
5226  *
5227  * o reserve space in the data space info for num_bytes
5228  * o reserve space in the metadata space info based on number of outstanding
5229  *   extents and how much csums will be needed
5230  * o add to the inodes ->delalloc_bytes
5231  * o add it to the fs_info's delalloc inodes list.
5232  *
5233  * This will return 0 for success and -ENOSPC if there is no space left.
5234  */
5235 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5236 {
5237         int ret;
5238
5239         ret = btrfs_check_data_free_space(inode, num_bytes);
5240         if (ret)
5241                 return ret;
5242
5243         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5244         if (ret) {
5245                 btrfs_free_reserved_data_space(inode, num_bytes);
5246                 return ret;
5247         }
5248
5249         return 0;
5250 }
5251
5252 /**
5253  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5254  * @inode: inode we're releasing space for
5255  * @num_bytes: the number of bytes we want to free up
5256  *
5257  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5258  * called in the case that we don't need the metadata AND data reservations
5259  * anymore.  So if there is an error or we insert an inline extent.
5260  *
5261  * This function will release the metadata space that was not used and will
5262  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5263  * list if there are no delalloc bytes left.
5264  */
5265 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5266 {
5267         btrfs_delalloc_release_metadata(inode, num_bytes);
5268         btrfs_free_reserved_data_space(inode, num_bytes);
5269 }
5270
5271 static int update_block_group(struct btrfs_root *root,
5272                               u64 bytenr, u64 num_bytes, int alloc)
5273 {
5274         struct btrfs_block_group_cache *cache = NULL;
5275         struct btrfs_fs_info *info = root->fs_info;
5276         u64 total = num_bytes;
5277         u64 old_val;
5278         u64 byte_in_group;
5279         int factor;
5280
5281         /* block accounting for super block */
5282         spin_lock(&info->delalloc_root_lock);
5283         old_val = btrfs_super_bytes_used(info->super_copy);
5284         if (alloc)
5285                 old_val += num_bytes;
5286         else
5287                 old_val -= num_bytes;
5288         btrfs_set_super_bytes_used(info->super_copy, old_val);
5289         spin_unlock(&info->delalloc_root_lock);
5290
5291         while (total) {
5292                 cache = btrfs_lookup_block_group(info, bytenr);
5293                 if (!cache)
5294                         return -ENOENT;
5295                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5296                                     BTRFS_BLOCK_GROUP_RAID1 |
5297                                     BTRFS_BLOCK_GROUP_RAID10))
5298                         factor = 2;
5299                 else
5300                         factor = 1;
5301                 /*
5302                  * If this block group has free space cache written out, we
5303                  * need to make sure to load it if we are removing space.  This
5304                  * is because we need the unpinning stage to actually add the
5305                  * space back to the block group, otherwise we will leak space.
5306                  */
5307                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5308                         cache_block_group(cache, 1);
5309
5310                 byte_in_group = bytenr - cache->key.objectid;
5311                 WARN_ON(byte_in_group > cache->key.offset);
5312
5313                 spin_lock(&cache->space_info->lock);
5314                 spin_lock(&cache->lock);
5315
5316                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5317                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5318                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5319
5320                 cache->dirty = 1;
5321                 old_val = btrfs_block_group_used(&cache->item);
5322                 num_bytes = min(total, cache->key.offset - byte_in_group);
5323                 if (alloc) {
5324                         old_val += num_bytes;
5325                         btrfs_set_block_group_used(&cache->item, old_val);
5326                         cache->reserved -= num_bytes;
5327                         cache->space_info->bytes_reserved -= num_bytes;
5328                         cache->space_info->bytes_used += num_bytes;
5329                         cache->space_info->disk_used += num_bytes * factor;
5330                         spin_unlock(&cache->lock);
5331                         spin_unlock(&cache->space_info->lock);
5332                 } else {
5333                         old_val -= num_bytes;
5334                         btrfs_set_block_group_used(&cache->item, old_val);
5335                         cache->pinned += num_bytes;
5336                         cache->space_info->bytes_pinned += num_bytes;
5337                         cache->space_info->bytes_used -= num_bytes;
5338                         cache->space_info->disk_used -= num_bytes * factor;
5339                         spin_unlock(&cache->lock);
5340                         spin_unlock(&cache->space_info->lock);
5341
5342                         set_extent_dirty(info->pinned_extents,
5343                                          bytenr, bytenr + num_bytes - 1,
5344                                          GFP_NOFS | __GFP_NOFAIL);
5345                 }
5346                 btrfs_put_block_group(cache);
5347                 total -= num_bytes;
5348                 bytenr += num_bytes;
5349         }
5350         return 0;
5351 }
5352
5353 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5354 {
5355         struct btrfs_block_group_cache *cache;
5356         u64 bytenr;
5357
5358         spin_lock(&root->fs_info->block_group_cache_lock);
5359         bytenr = root->fs_info->first_logical_byte;
5360         spin_unlock(&root->fs_info->block_group_cache_lock);
5361
5362         if (bytenr < (u64)-1)
5363                 return bytenr;
5364
5365         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5366         if (!cache)
5367                 return 0;
5368
5369         bytenr = cache->key.objectid;
5370         btrfs_put_block_group(cache);
5371
5372         return bytenr;
5373 }
5374
5375 static int pin_down_extent(struct btrfs_root *root,
5376                            struct btrfs_block_group_cache *cache,
5377                            u64 bytenr, u64 num_bytes, int reserved)
5378 {
5379         spin_lock(&cache->space_info->lock);
5380         spin_lock(&cache->lock);
5381         cache->pinned += num_bytes;
5382         cache->space_info->bytes_pinned += num_bytes;
5383         if (reserved) {
5384                 cache->reserved -= num_bytes;
5385                 cache->space_info->bytes_reserved -= num_bytes;
5386         }
5387         spin_unlock(&cache->lock);
5388         spin_unlock(&cache->space_info->lock);
5389
5390         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5391                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5392         if (reserved)
5393                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5394         return 0;
5395 }
5396
5397 /*
5398  * this function must be called within transaction
5399  */
5400 int btrfs_pin_extent(struct btrfs_root *root,
5401                      u64 bytenr, u64 num_bytes, int reserved)
5402 {
5403         struct btrfs_block_group_cache *cache;
5404
5405         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5406         BUG_ON(!cache); /* Logic error */
5407
5408         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5409
5410         btrfs_put_block_group(cache);
5411         return 0;
5412 }
5413
5414 /*
5415  * this function must be called within transaction
5416  */
5417 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5418                                     u64 bytenr, u64 num_bytes)
5419 {
5420         struct btrfs_block_group_cache *cache;
5421         int ret;
5422
5423         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5424         if (!cache)
5425                 return -EINVAL;
5426
5427         /*
5428          * pull in the free space cache (if any) so that our pin
5429          * removes the free space from the cache.  We have load_only set
5430          * to one because the slow code to read in the free extents does check
5431          * the pinned extents.
5432          */
5433         cache_block_group(cache, 1);
5434
5435         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5436
5437         /* remove us from the free space cache (if we're there at all) */
5438         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5439         btrfs_put_block_group(cache);
5440         return ret;
5441 }
5442
5443 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5444 {
5445         int ret;
5446         struct btrfs_block_group_cache *block_group;
5447         struct btrfs_caching_control *caching_ctl;
5448
5449         block_group = btrfs_lookup_block_group(root->fs_info, start);
5450         if (!block_group)
5451                 return -EINVAL;
5452
5453         cache_block_group(block_group, 0);
5454         caching_ctl = get_caching_control(block_group);
5455
5456         if (!caching_ctl) {
5457                 /* Logic error */
5458                 BUG_ON(!block_group_cache_done(block_group));
5459                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5460         } else {
5461                 mutex_lock(&caching_ctl->mutex);
5462
5463                 if (start >= caching_ctl->progress) {
5464                         ret = add_excluded_extent(root, start, num_bytes);
5465                 } else if (start + num_bytes <= caching_ctl->progress) {
5466                         ret = btrfs_remove_free_space(block_group,
5467                                                       start, num_bytes);
5468                 } else {
5469                         num_bytes = caching_ctl->progress - start;
5470                         ret = btrfs_remove_free_space(block_group,
5471                                                       start, num_bytes);
5472                         if (ret)
5473                                 goto out_lock;
5474
5475                         num_bytes = (start + num_bytes) -
5476                                 caching_ctl->progress;
5477                         start = caching_ctl->progress;
5478                         ret = add_excluded_extent(root, start, num_bytes);
5479                 }
5480 out_lock:
5481                 mutex_unlock(&caching_ctl->mutex);
5482                 put_caching_control(caching_ctl);
5483         }
5484         btrfs_put_block_group(block_group);
5485         return ret;
5486 }
5487
5488 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5489                                  struct extent_buffer *eb)
5490 {
5491         struct btrfs_file_extent_item *item;
5492         struct btrfs_key key;
5493         int found_type;
5494         int i;
5495
5496         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5497                 return 0;
5498
5499         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5500                 btrfs_item_key_to_cpu(eb, &key, i);
5501                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5502                         continue;
5503                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5504                 found_type = btrfs_file_extent_type(eb, item);
5505                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5506                         continue;
5507                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5508                         continue;
5509                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5510                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5511                 __exclude_logged_extent(log, key.objectid, key.offset);
5512         }
5513
5514         return 0;
5515 }
5516
5517 /**
5518  * btrfs_update_reserved_bytes - update the block_group and space info counters
5519  * @cache:      The cache we are manipulating
5520  * @num_bytes:  The number of bytes in question
5521  * @reserve:    One of the reservation enums
5522  *
5523  * This is called by the allocator when it reserves space, or by somebody who is
5524  * freeing space that was never actually used on disk.  For example if you
5525  * reserve some space for a new leaf in transaction A and before transaction A
5526  * commits you free that leaf, you call this with reserve set to 0 in order to
5527  * clear the reservation.
5528  *
5529  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5530  * ENOSPC accounting.  For data we handle the reservation through clearing the
5531  * delalloc bits in the io_tree.  We have to do this since we could end up
5532  * allocating less disk space for the amount of data we have reserved in the
5533  * case of compression.
5534  *
5535  * If this is a reservation and the block group has become read only we cannot
5536  * make the reservation and return -EAGAIN, otherwise this function always
5537  * succeeds.
5538  */
5539 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5540                                        u64 num_bytes, int reserve)
5541 {
5542         struct btrfs_space_info *space_info = cache->space_info;
5543         int ret = 0;
5544
5545         spin_lock(&space_info->lock);
5546         spin_lock(&cache->lock);
5547         if (reserve != RESERVE_FREE) {
5548                 if (cache->ro) {
5549                         ret = -EAGAIN;
5550                 } else {
5551                         cache->reserved += num_bytes;
5552                         space_info->bytes_reserved += num_bytes;
5553                         if (reserve == RESERVE_ALLOC) {
5554                                 trace_btrfs_space_reservation(cache->fs_info,
5555                                                 "space_info", space_info->flags,
5556                                                 num_bytes, 0);
5557                                 space_info->bytes_may_use -= num_bytes;
5558                         }
5559                 }
5560         } else {
5561                 if (cache->ro)
5562                         space_info->bytes_readonly += num_bytes;
5563                 cache->reserved -= num_bytes;
5564                 space_info->bytes_reserved -= num_bytes;
5565         }
5566         spin_unlock(&cache->lock);
5567         spin_unlock(&space_info->lock);
5568         return ret;
5569 }
5570
5571 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5572                                 struct btrfs_root *root)
5573 {
5574         struct btrfs_fs_info *fs_info = root->fs_info;
5575         struct btrfs_caching_control *next;
5576         struct btrfs_caching_control *caching_ctl;
5577         struct btrfs_block_group_cache *cache;
5578         struct btrfs_space_info *space_info;
5579
5580         down_write(&fs_info->commit_root_sem);
5581
5582         list_for_each_entry_safe(caching_ctl, next,
5583                                  &fs_info->caching_block_groups, list) {
5584                 cache = caching_ctl->block_group;
5585                 if (block_group_cache_done(cache)) {
5586                         cache->last_byte_to_unpin = (u64)-1;
5587                         list_del_init(&caching_ctl->list);
5588                         put_caching_control(caching_ctl);
5589                 } else {
5590                         cache->last_byte_to_unpin = caching_ctl->progress;
5591                 }
5592         }
5593
5594         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5595                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5596         else
5597                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5598
5599         up_write(&fs_info->commit_root_sem);
5600
5601         list_for_each_entry_rcu(space_info, &fs_info->space_info, list)
5602                 percpu_counter_set(&space_info->total_bytes_pinned, 0);
5603
5604         update_global_block_rsv(fs_info);
5605 }
5606
5607 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5608 {
5609         struct btrfs_fs_info *fs_info = root->fs_info;
5610         struct btrfs_block_group_cache *cache = NULL;
5611         struct btrfs_space_info *space_info;
5612         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5613         u64 len;
5614         bool readonly;
5615
5616         while (start <= end) {
5617                 readonly = false;
5618                 if (!cache ||
5619                     start >= cache->key.objectid + cache->key.offset) {
5620                         if (cache)
5621                                 btrfs_put_block_group(cache);
5622                         cache = btrfs_lookup_block_group(fs_info, start);
5623                         BUG_ON(!cache); /* Logic error */
5624                 }
5625
5626                 len = cache->key.objectid + cache->key.offset - start;
5627                 len = min(len, end + 1 - start);
5628
5629                 if (start < cache->last_byte_to_unpin) {
5630                         len = min(len, cache->last_byte_to_unpin - start);
5631                         btrfs_add_free_space(cache, start, len);
5632                 }
5633
5634                 start += len;
5635                 space_info = cache->space_info;
5636
5637                 spin_lock(&space_info->lock);
5638                 spin_lock(&cache->lock);
5639                 cache->pinned -= len;
5640                 space_info->bytes_pinned -= len;
5641                 if (cache->ro) {
5642                         space_info->bytes_readonly += len;
5643                         readonly = true;
5644                 }
5645                 spin_unlock(&cache->lock);
5646                 if (!readonly && global_rsv->space_info == space_info) {
5647                         spin_lock(&global_rsv->lock);
5648                         if (!global_rsv->full) {
5649                                 len = min(len, global_rsv->size -
5650                                           global_rsv->reserved);
5651                                 global_rsv->reserved += len;
5652                                 space_info->bytes_may_use += len;
5653                                 if (global_rsv->reserved >= global_rsv->size)
5654                                         global_rsv->full = 1;
5655                         }
5656                         spin_unlock(&global_rsv->lock);
5657                 }
5658                 spin_unlock(&space_info->lock);
5659         }
5660
5661         if (cache)
5662                 btrfs_put_block_group(cache);
5663         return 0;
5664 }
5665
5666 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5667                                struct btrfs_root *root)
5668 {
5669         struct btrfs_fs_info *fs_info = root->fs_info;
5670         struct extent_io_tree *unpin;
5671         u64 start;
5672         u64 end;
5673         int ret;
5674
5675         if (trans->aborted)
5676                 return 0;
5677
5678         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5679                 unpin = &fs_info->freed_extents[1];
5680         else
5681                 unpin = &fs_info->freed_extents[0];
5682
5683         while (1) {
5684                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5685                                             EXTENT_DIRTY, NULL);
5686                 if (ret)
5687                         break;
5688
5689                 if (btrfs_test_opt(root, DISCARD))
5690                         ret = btrfs_discard_extent(root, start,
5691                                                    end + 1 - start, NULL);
5692
5693                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5694                 unpin_extent_range(root, start, end);
5695                 cond_resched();
5696         }
5697
5698         return 0;
5699 }
5700
5701 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
5702                              u64 owner, u64 root_objectid)
5703 {
5704         struct btrfs_space_info *space_info;
5705         u64 flags;
5706
5707         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5708                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
5709                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
5710                 else
5711                         flags = BTRFS_BLOCK_GROUP_METADATA;
5712         } else {
5713                 flags = BTRFS_BLOCK_GROUP_DATA;
5714         }
5715
5716         space_info = __find_space_info(fs_info, flags);
5717         BUG_ON(!space_info); /* Logic bug */
5718         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
5719 }
5720
5721
5722 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5723                                 struct btrfs_root *root,
5724                                 u64 bytenr, u64 num_bytes, u64 parent,
5725                                 u64 root_objectid, u64 owner_objectid,
5726                                 u64 owner_offset, int refs_to_drop,
5727                                 struct btrfs_delayed_extent_op *extent_op)
5728 {
5729         struct btrfs_key key;
5730         struct btrfs_path *path;
5731         struct btrfs_fs_info *info = root->fs_info;
5732         struct btrfs_root *extent_root = info->extent_root;
5733         struct extent_buffer *leaf;
5734         struct btrfs_extent_item *ei;
5735         struct btrfs_extent_inline_ref *iref;
5736         int ret;
5737         int is_data;
5738         int extent_slot = 0;
5739         int found_extent = 0;
5740         int num_to_del = 1;
5741         u32 item_size;
5742         u64 refs;
5743         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5744                                                  SKINNY_METADATA);
5745
5746         path = btrfs_alloc_path();
5747         if (!path)
5748                 return -ENOMEM;
5749
5750         path->reada = 1;
5751         path->leave_spinning = 1;
5752
5753         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5754         BUG_ON(!is_data && refs_to_drop != 1);
5755
5756         if (is_data)
5757                 skinny_metadata = 0;
5758
5759         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5760                                     bytenr, num_bytes, parent,
5761                                     root_objectid, owner_objectid,
5762                                     owner_offset);
5763         if (ret == 0) {
5764                 extent_slot = path->slots[0];
5765                 while (extent_slot >= 0) {
5766                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5767                                               extent_slot);
5768                         if (key.objectid != bytenr)
5769                                 break;
5770                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5771                             key.offset == num_bytes) {
5772                                 found_extent = 1;
5773                                 break;
5774                         }
5775                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5776                             key.offset == owner_objectid) {
5777                                 found_extent = 1;
5778                                 break;
5779                         }
5780                         if (path->slots[0] - extent_slot > 5)
5781                                 break;
5782                         extent_slot--;
5783                 }
5784 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5785                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5786                 if (found_extent && item_size < sizeof(*ei))
5787                         found_extent = 0;
5788 #endif
5789                 if (!found_extent) {
5790                         BUG_ON(iref);
5791                         ret = remove_extent_backref(trans, extent_root, path,
5792                                                     NULL, refs_to_drop,
5793                                                     is_data);
5794                         if (ret) {
5795                                 btrfs_abort_transaction(trans, extent_root, ret);
5796                                 goto out;
5797                         }
5798                         btrfs_release_path(path);
5799                         path->leave_spinning = 1;
5800
5801                         key.objectid = bytenr;
5802                         key.type = BTRFS_EXTENT_ITEM_KEY;
5803                         key.offset = num_bytes;
5804
5805                         if (!is_data && skinny_metadata) {
5806                                 key.type = BTRFS_METADATA_ITEM_KEY;
5807                                 key.offset = owner_objectid;
5808                         }
5809
5810                         ret = btrfs_search_slot(trans, extent_root,
5811                                                 &key, path, -1, 1);
5812                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5813                                 /*
5814                                  * Couldn't find our skinny metadata item,
5815                                  * see if we have ye olde extent item.
5816                                  */
5817                                 path->slots[0]--;
5818                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5819                                                       path->slots[0]);
5820                                 if (key.objectid == bytenr &&
5821                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5822                                     key.offset == num_bytes)
5823                                         ret = 0;
5824                         }
5825
5826                         if (ret > 0 && skinny_metadata) {
5827                                 skinny_metadata = false;
5828                                 key.objectid = bytenr;
5829                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5830                                 key.offset = num_bytes;
5831                                 btrfs_release_path(path);
5832                                 ret = btrfs_search_slot(trans, extent_root,
5833                                                         &key, path, -1, 1);
5834                         }
5835
5836                         if (ret) {
5837                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5838                                         ret, bytenr);
5839                                 if (ret > 0)
5840                                         btrfs_print_leaf(extent_root,
5841                                                          path->nodes[0]);
5842                         }
5843                         if (ret < 0) {
5844                                 btrfs_abort_transaction(trans, extent_root, ret);
5845                                 goto out;
5846                         }
5847                         extent_slot = path->slots[0];
5848                 }
5849         } else if (WARN_ON(ret == -ENOENT)) {
5850                 btrfs_print_leaf(extent_root, path->nodes[0]);
5851                 btrfs_err(info,
5852                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5853                         bytenr, parent, root_objectid, owner_objectid,
5854                         owner_offset);
5855                 btrfs_abort_transaction(trans, extent_root, ret);
5856                 goto out;
5857         } else {
5858                 btrfs_abort_transaction(trans, extent_root, ret);
5859                 goto out;
5860         }
5861
5862         leaf = path->nodes[0];
5863         item_size = btrfs_item_size_nr(leaf, extent_slot);
5864 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5865         if (item_size < sizeof(*ei)) {
5866                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5867                 ret = convert_extent_item_v0(trans, extent_root, path,
5868                                              owner_objectid, 0);
5869                 if (ret < 0) {
5870                         btrfs_abort_transaction(trans, extent_root, ret);
5871                         goto out;
5872                 }
5873
5874                 btrfs_release_path(path);
5875                 path->leave_spinning = 1;
5876
5877                 key.objectid = bytenr;
5878                 key.type = BTRFS_EXTENT_ITEM_KEY;
5879                 key.offset = num_bytes;
5880
5881                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5882                                         -1, 1);
5883                 if (ret) {
5884                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5885                                 ret, bytenr);
5886                         btrfs_print_leaf(extent_root, path->nodes[0]);
5887                 }
5888                 if (ret < 0) {
5889                         btrfs_abort_transaction(trans, extent_root, ret);
5890                         goto out;
5891                 }
5892
5893                 extent_slot = path->slots[0];
5894                 leaf = path->nodes[0];
5895                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5896         }
5897 #endif
5898         BUG_ON(item_size < sizeof(*ei));
5899         ei = btrfs_item_ptr(leaf, extent_slot,
5900                             struct btrfs_extent_item);
5901         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
5902             key.type == BTRFS_EXTENT_ITEM_KEY) {
5903                 struct btrfs_tree_block_info *bi;
5904                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5905                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5906                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5907         }
5908
5909         refs = btrfs_extent_refs(leaf, ei);
5910         if (refs < refs_to_drop) {
5911                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
5912                           "for bytenr %Lu\n", refs_to_drop, refs, bytenr);
5913                 ret = -EINVAL;
5914                 btrfs_abort_transaction(trans, extent_root, ret);
5915                 goto out;
5916         }
5917         refs -= refs_to_drop;
5918
5919         if (refs > 0) {
5920                 if (extent_op)
5921                         __run_delayed_extent_op(extent_op, leaf, ei);
5922                 /*
5923                  * In the case of inline back ref, reference count will
5924                  * be updated by remove_extent_backref
5925                  */
5926                 if (iref) {
5927                         BUG_ON(!found_extent);
5928                 } else {
5929                         btrfs_set_extent_refs(leaf, ei, refs);
5930                         btrfs_mark_buffer_dirty(leaf);
5931                 }
5932                 if (found_extent) {
5933                         ret = remove_extent_backref(trans, extent_root, path,
5934                                                     iref, refs_to_drop,
5935                                                     is_data);
5936                         if (ret) {
5937                                 btrfs_abort_transaction(trans, extent_root, ret);
5938                                 goto out;
5939                         }
5940                 }
5941                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
5942                                  root_objectid);
5943         } else {
5944                 if (found_extent) {
5945                         BUG_ON(is_data && refs_to_drop !=
5946                                extent_data_ref_count(root, path, iref));
5947                         if (iref) {
5948                                 BUG_ON(path->slots[0] != extent_slot);
5949                         } else {
5950                                 BUG_ON(path->slots[0] != extent_slot + 1);
5951                                 path->slots[0] = extent_slot;
5952                                 num_to_del = 2;
5953                         }
5954                 }
5955
5956                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5957                                       num_to_del);
5958                 if (ret) {
5959                         btrfs_abort_transaction(trans, extent_root, ret);
5960                         goto out;
5961                 }
5962                 btrfs_release_path(path);
5963
5964                 if (is_data) {
5965                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5966                         if (ret) {
5967                                 btrfs_abort_transaction(trans, extent_root, ret);
5968                                 goto out;
5969                         }
5970                 }
5971
5972                 ret = update_block_group(root, bytenr, num_bytes, 0);
5973                 if (ret) {
5974                         btrfs_abort_transaction(trans, extent_root, ret);
5975                         goto out;
5976                 }
5977         }
5978 out:
5979         btrfs_free_path(path);
5980         return ret;
5981 }
5982
5983 /*
5984  * when we free an block, it is possible (and likely) that we free the last
5985  * delayed ref for that extent as well.  This searches the delayed ref tree for
5986  * a given extent, and if there are no other delayed refs to be processed, it
5987  * removes it from the tree.
5988  */
5989 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5990                                       struct btrfs_root *root, u64 bytenr)
5991 {
5992         struct btrfs_delayed_ref_head *head;
5993         struct btrfs_delayed_ref_root *delayed_refs;
5994         int ret = 0;
5995
5996         delayed_refs = &trans->transaction->delayed_refs;
5997         spin_lock(&delayed_refs->lock);
5998         head = btrfs_find_delayed_ref_head(trans, bytenr);
5999         if (!head)
6000                 goto out_delayed_unlock;
6001
6002         spin_lock(&head->lock);
6003         if (rb_first(&head->ref_root))
6004                 goto out;
6005
6006         if (head->extent_op) {
6007                 if (!head->must_insert_reserved)
6008                         goto out;
6009                 btrfs_free_delayed_extent_op(head->extent_op);
6010                 head->extent_op = NULL;
6011         }
6012
6013         /*
6014          * waiting for the lock here would deadlock.  If someone else has it
6015          * locked they are already in the process of dropping it anyway
6016          */
6017         if (!mutex_trylock(&head->mutex))
6018                 goto out;
6019
6020         /*
6021          * at this point we have a head with no other entries.  Go
6022          * ahead and process it.
6023          */
6024         head->node.in_tree = 0;
6025         rb_erase(&head->href_node, &delayed_refs->href_root);
6026
6027         atomic_dec(&delayed_refs->num_entries);
6028
6029         /*
6030          * we don't take a ref on the node because we're removing it from the
6031          * tree, so we just steal the ref the tree was holding.
6032          */
6033         delayed_refs->num_heads--;
6034         if (head->processing == 0)
6035                 delayed_refs->num_heads_ready--;
6036         head->processing = 0;
6037         spin_unlock(&head->lock);
6038         spin_unlock(&delayed_refs->lock);
6039
6040         BUG_ON(head->extent_op);
6041         if (head->must_insert_reserved)
6042                 ret = 1;
6043
6044         mutex_unlock(&head->mutex);
6045         btrfs_put_delayed_ref(&head->node);
6046         return ret;
6047 out:
6048         spin_unlock(&head->lock);
6049
6050 out_delayed_unlock:
6051         spin_unlock(&delayed_refs->lock);
6052         return 0;
6053 }
6054
6055 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6056                            struct btrfs_root *root,
6057                            struct extent_buffer *buf,
6058                            u64 parent, int last_ref)
6059 {
6060         struct btrfs_block_group_cache *cache = NULL;
6061         int pin = 1;
6062         int ret;
6063
6064         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6065                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6066                                         buf->start, buf->len,
6067                                         parent, root->root_key.objectid,
6068                                         btrfs_header_level(buf),
6069                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6070                 BUG_ON(ret); /* -ENOMEM */
6071         }
6072
6073         if (!last_ref)
6074                 return;
6075
6076         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6077
6078         if (btrfs_header_generation(buf) == trans->transid) {
6079                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6080                         ret = check_ref_cleanup(trans, root, buf->start);
6081                         if (!ret)
6082                                 goto out;
6083                 }
6084
6085                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6086                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6087                         goto out;
6088                 }
6089
6090                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6091
6092                 btrfs_add_free_space(cache, buf->start, buf->len);
6093                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
6094                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6095                 pin = 0;
6096         }
6097 out:
6098         if (pin)
6099                 add_pinned_bytes(root->fs_info, buf->len,
6100                                  btrfs_header_level(buf),
6101                                  root->root_key.objectid);
6102
6103         /*
6104          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6105          * anymore.
6106          */
6107         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6108         btrfs_put_block_group(cache);
6109 }
6110
6111 /* Can return -ENOMEM */
6112 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6113                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6114                       u64 owner, u64 offset, int for_cow)
6115 {
6116         int ret;
6117         struct btrfs_fs_info *fs_info = root->fs_info;
6118
6119         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6120
6121         /*
6122          * tree log blocks never actually go into the extent allocation
6123          * tree, just update pinning info and exit early.
6124          */
6125         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6126                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6127                 /* unlocks the pinned mutex */
6128                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6129                 ret = 0;
6130         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6131                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6132                                         num_bytes,
6133                                         parent, root_objectid, (int)owner,
6134                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
6135         } else {
6136                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6137                                                 num_bytes,
6138                                                 parent, root_objectid, owner,
6139                                                 offset, BTRFS_DROP_DELAYED_REF,
6140                                                 NULL, for_cow);
6141         }
6142         return ret;
6143 }
6144
6145 static u64 stripe_align(struct btrfs_root *root,
6146                         struct btrfs_block_group_cache *cache,
6147                         u64 val, u64 num_bytes)
6148 {
6149         u64 ret = ALIGN(val, root->stripesize);
6150         return ret;
6151 }
6152
6153 /*
6154  * when we wait for progress in the block group caching, its because
6155  * our allocation attempt failed at least once.  So, we must sleep
6156  * and let some progress happen before we try again.
6157  *
6158  * This function will sleep at least once waiting for new free space to
6159  * show up, and then it will check the block group free space numbers
6160  * for our min num_bytes.  Another option is to have it go ahead
6161  * and look in the rbtree for a free extent of a given size, but this
6162  * is a good start.
6163  *
6164  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6165  * any of the information in this block group.
6166  */
6167 static noinline void
6168 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6169                                 u64 num_bytes)
6170 {
6171         struct btrfs_caching_control *caching_ctl;
6172
6173         caching_ctl = get_caching_control(cache);
6174         if (!caching_ctl)
6175                 return;
6176
6177         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6178                    (cache->free_space_ctl->free_space >= num_bytes));
6179
6180         put_caching_control(caching_ctl);
6181 }
6182
6183 static noinline int
6184 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6185 {
6186         struct btrfs_caching_control *caching_ctl;
6187         int ret = 0;
6188
6189         caching_ctl = get_caching_control(cache);
6190         if (!caching_ctl)
6191                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6192
6193         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6194         if (cache->cached == BTRFS_CACHE_ERROR)
6195                 ret = -EIO;
6196         put_caching_control(caching_ctl);
6197         return ret;
6198 }
6199
6200 int __get_raid_index(u64 flags)
6201 {
6202         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6203                 return BTRFS_RAID_RAID10;
6204         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6205                 return BTRFS_RAID_RAID1;
6206         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6207                 return BTRFS_RAID_DUP;
6208         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6209                 return BTRFS_RAID_RAID0;
6210         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6211                 return BTRFS_RAID_RAID5;
6212         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6213                 return BTRFS_RAID_RAID6;
6214
6215         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6216 }
6217
6218 int get_block_group_index(struct btrfs_block_group_cache *cache)
6219 {
6220         return __get_raid_index(cache->flags);
6221 }
6222
6223 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6224         [BTRFS_RAID_RAID10]     = "raid10",
6225         [BTRFS_RAID_RAID1]      = "raid1",
6226         [BTRFS_RAID_DUP]        = "dup",
6227         [BTRFS_RAID_RAID0]      = "raid0",
6228         [BTRFS_RAID_SINGLE]     = "single",
6229         [BTRFS_RAID_RAID5]      = "raid5",
6230         [BTRFS_RAID_RAID6]      = "raid6",
6231 };
6232
6233 static const char *get_raid_name(enum btrfs_raid_types type)
6234 {
6235         if (type >= BTRFS_NR_RAID_TYPES)
6236                 return NULL;
6237
6238         return btrfs_raid_type_names[type];
6239 }
6240
6241 enum btrfs_loop_type {
6242         LOOP_CACHING_NOWAIT = 0,
6243         LOOP_CACHING_WAIT = 1,
6244         LOOP_ALLOC_CHUNK = 2,
6245         LOOP_NO_EMPTY_SIZE = 3,
6246 };
6247
6248 /*
6249  * walks the btree of allocated extents and find a hole of a given size.
6250  * The key ins is changed to record the hole:
6251  * ins->objectid == start position
6252  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6253  * ins->offset == the size of the hole.
6254  * Any available blocks before search_start are skipped.
6255  *
6256  * If there is no suitable free space, we will record the max size of
6257  * the free space extent currently.
6258  */
6259 static noinline int find_free_extent(struct btrfs_root *orig_root,
6260                                      u64 num_bytes, u64 empty_size,
6261                                      u64 hint_byte, struct btrfs_key *ins,
6262                                      u64 flags)
6263 {
6264         int ret = 0;
6265         struct btrfs_root *root = orig_root->fs_info->extent_root;
6266         struct btrfs_free_cluster *last_ptr = NULL;
6267         struct btrfs_block_group_cache *block_group = NULL;
6268         u64 search_start = 0;
6269         u64 max_extent_size = 0;
6270         int empty_cluster = 2 * 1024 * 1024;
6271         struct btrfs_space_info *space_info;
6272         int loop = 0;
6273         int index = __get_raid_index(flags);
6274         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6275                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6276         bool failed_cluster_refill = false;
6277         bool failed_alloc = false;
6278         bool use_cluster = true;
6279         bool have_caching_bg = false;
6280
6281         WARN_ON(num_bytes < root->sectorsize);
6282         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
6283         ins->objectid = 0;
6284         ins->offset = 0;
6285
6286         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6287
6288         space_info = __find_space_info(root->fs_info, flags);
6289         if (!space_info) {
6290                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6291                 return -ENOSPC;
6292         }
6293
6294         /*
6295          * If the space info is for both data and metadata it means we have a
6296          * small filesystem and we can't use the clustering stuff.
6297          */
6298         if (btrfs_mixed_space_info(space_info))
6299                 use_cluster = false;
6300
6301         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6302                 last_ptr = &root->fs_info->meta_alloc_cluster;
6303                 if (!btrfs_test_opt(root, SSD))
6304                         empty_cluster = 64 * 1024;
6305         }
6306
6307         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6308             btrfs_test_opt(root, SSD)) {
6309                 last_ptr = &root->fs_info->data_alloc_cluster;
6310         }
6311
6312         if (last_ptr) {
6313                 spin_lock(&last_ptr->lock);
6314                 if (last_ptr->block_group)
6315                         hint_byte = last_ptr->window_start;
6316                 spin_unlock(&last_ptr->lock);
6317         }
6318
6319         search_start = max(search_start, first_logical_byte(root, 0));
6320         search_start = max(search_start, hint_byte);
6321
6322         if (!last_ptr)
6323                 empty_cluster = 0;
6324
6325         if (search_start == hint_byte) {
6326                 block_group = btrfs_lookup_block_group(root->fs_info,
6327                                                        search_start);
6328                 /*
6329                  * we don't want to use the block group if it doesn't match our
6330                  * allocation bits, or if its not cached.
6331                  *
6332                  * However if we are re-searching with an ideal block group
6333                  * picked out then we don't care that the block group is cached.
6334                  */
6335                 if (block_group && block_group_bits(block_group, flags) &&
6336                     block_group->cached != BTRFS_CACHE_NO) {
6337                         down_read(&space_info->groups_sem);
6338                         if (list_empty(&block_group->list) ||
6339                             block_group->ro) {
6340                                 /*
6341                                  * someone is removing this block group,
6342                                  * we can't jump into the have_block_group
6343                                  * target because our list pointers are not
6344                                  * valid
6345                                  */
6346                                 btrfs_put_block_group(block_group);
6347                                 up_read(&space_info->groups_sem);
6348                         } else {
6349                                 index = get_block_group_index(block_group);
6350                                 goto have_block_group;
6351                         }
6352                 } else if (block_group) {
6353                         btrfs_put_block_group(block_group);
6354                 }
6355         }
6356 search:
6357         have_caching_bg = false;
6358         down_read(&space_info->groups_sem);
6359         list_for_each_entry(block_group, &space_info->block_groups[index],
6360                             list) {
6361                 u64 offset;
6362                 int cached;
6363
6364                 btrfs_get_block_group(block_group);
6365                 search_start = block_group->key.objectid;
6366
6367                 /*
6368                  * this can happen if we end up cycling through all the
6369                  * raid types, but we want to make sure we only allocate
6370                  * for the proper type.
6371                  */
6372                 if (!block_group_bits(block_group, flags)) {
6373                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6374                                 BTRFS_BLOCK_GROUP_RAID1 |
6375                                 BTRFS_BLOCK_GROUP_RAID5 |
6376                                 BTRFS_BLOCK_GROUP_RAID6 |
6377                                 BTRFS_BLOCK_GROUP_RAID10;
6378
6379                         /*
6380                          * if they asked for extra copies and this block group
6381                          * doesn't provide them, bail.  This does allow us to
6382                          * fill raid0 from raid1.
6383                          */
6384                         if ((flags & extra) && !(block_group->flags & extra))
6385                                 goto loop;
6386                 }
6387
6388 have_block_group:
6389                 cached = block_group_cache_done(block_group);
6390                 if (unlikely(!cached)) {
6391                         ret = cache_block_group(block_group, 0);
6392                         BUG_ON(ret < 0);
6393                         ret = 0;
6394                 }
6395
6396                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6397                         goto loop;
6398                 if (unlikely(block_group->ro))
6399                         goto loop;
6400
6401                 /*
6402                  * Ok we want to try and use the cluster allocator, so
6403                  * lets look there
6404                  */
6405                 if (last_ptr) {
6406                         struct btrfs_block_group_cache *used_block_group;
6407                         unsigned long aligned_cluster;
6408                         /*
6409                          * the refill lock keeps out other
6410                          * people trying to start a new cluster
6411                          */
6412                         spin_lock(&last_ptr->refill_lock);
6413                         used_block_group = last_ptr->block_group;
6414                         if (used_block_group != block_group &&
6415                             (!used_block_group ||
6416                              used_block_group->ro ||
6417                              !block_group_bits(used_block_group, flags)))
6418                                 goto refill_cluster;
6419
6420                         if (used_block_group != block_group)
6421                                 btrfs_get_block_group(used_block_group);
6422
6423                         offset = btrfs_alloc_from_cluster(used_block_group,
6424                                                 last_ptr,
6425                                                 num_bytes,
6426                                                 used_block_group->key.objectid,
6427                                                 &max_extent_size);
6428                         if (offset) {
6429                                 /* we have a block, we're done */
6430                                 spin_unlock(&last_ptr->refill_lock);
6431                                 trace_btrfs_reserve_extent_cluster(root,
6432                                                 used_block_group,
6433                                                 search_start, num_bytes);
6434                                 if (used_block_group != block_group) {
6435                                         btrfs_put_block_group(block_group);
6436                                         block_group = used_block_group;
6437                                 }
6438                                 goto checks;
6439                         }
6440
6441                         WARN_ON(last_ptr->block_group != used_block_group);
6442                         if (used_block_group != block_group)
6443                                 btrfs_put_block_group(used_block_group);
6444 refill_cluster:
6445                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6446                          * set up a new clusters, so lets just skip it
6447                          * and let the allocator find whatever block
6448                          * it can find.  If we reach this point, we
6449                          * will have tried the cluster allocator
6450                          * plenty of times and not have found
6451                          * anything, so we are likely way too
6452                          * fragmented for the clustering stuff to find
6453                          * anything.
6454                          *
6455                          * However, if the cluster is taken from the
6456                          * current block group, release the cluster
6457                          * first, so that we stand a better chance of
6458                          * succeeding in the unclustered
6459                          * allocation.  */
6460                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6461                             last_ptr->block_group != block_group) {
6462                                 spin_unlock(&last_ptr->refill_lock);
6463                                 goto unclustered_alloc;
6464                         }
6465
6466                         /*
6467                          * this cluster didn't work out, free it and
6468                          * start over
6469                          */
6470                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6471
6472                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6473                                 spin_unlock(&last_ptr->refill_lock);
6474                                 goto unclustered_alloc;
6475                         }
6476
6477                         aligned_cluster = max_t(unsigned long,
6478                                                 empty_cluster + empty_size,
6479                                               block_group->full_stripe_len);
6480
6481                         /* allocate a cluster in this block group */
6482                         ret = btrfs_find_space_cluster(root, block_group,
6483                                                        last_ptr, search_start,
6484                                                        num_bytes,
6485                                                        aligned_cluster);
6486                         if (ret == 0) {
6487                                 /*
6488                                  * now pull our allocation out of this
6489                                  * cluster
6490                                  */
6491                                 offset = btrfs_alloc_from_cluster(block_group,
6492                                                         last_ptr,
6493                                                         num_bytes,
6494                                                         search_start,
6495                                                         &max_extent_size);
6496                                 if (offset) {
6497                                         /* we found one, proceed */
6498                                         spin_unlock(&last_ptr->refill_lock);
6499                                         trace_btrfs_reserve_extent_cluster(root,
6500                                                 block_group, search_start,
6501                                                 num_bytes);
6502                                         goto checks;
6503                                 }
6504                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6505                                    && !failed_cluster_refill) {
6506                                 spin_unlock(&last_ptr->refill_lock);
6507
6508                                 failed_cluster_refill = true;
6509                                 wait_block_group_cache_progress(block_group,
6510                                        num_bytes + empty_cluster + empty_size);
6511                                 goto have_block_group;
6512                         }
6513
6514                         /*
6515                          * at this point we either didn't find a cluster
6516                          * or we weren't able to allocate a block from our
6517                          * cluster.  Free the cluster we've been trying
6518                          * to use, and go to the next block group
6519                          */
6520                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6521                         spin_unlock(&last_ptr->refill_lock);
6522                         goto loop;
6523                 }
6524
6525 unclustered_alloc:
6526                 spin_lock(&block_group->free_space_ctl->tree_lock);
6527                 if (cached &&
6528                     block_group->free_space_ctl->free_space <
6529                     num_bytes + empty_cluster + empty_size) {
6530                         if (block_group->free_space_ctl->free_space >
6531                             max_extent_size)
6532                                 max_extent_size =
6533                                         block_group->free_space_ctl->free_space;
6534                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6535                         goto loop;
6536                 }
6537                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6538
6539                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6540                                                     num_bytes, empty_size,
6541                                                     &max_extent_size);
6542                 /*
6543                  * If we didn't find a chunk, and we haven't failed on this
6544                  * block group before, and this block group is in the middle of
6545                  * caching and we are ok with waiting, then go ahead and wait
6546                  * for progress to be made, and set failed_alloc to true.
6547                  *
6548                  * If failed_alloc is true then we've already waited on this
6549                  * block group once and should move on to the next block group.
6550                  */
6551                 if (!offset && !failed_alloc && !cached &&
6552                     loop > LOOP_CACHING_NOWAIT) {
6553                         wait_block_group_cache_progress(block_group,
6554                                                 num_bytes + empty_size);
6555                         failed_alloc = true;
6556                         goto have_block_group;
6557                 } else if (!offset) {
6558                         if (!cached)
6559                                 have_caching_bg = true;
6560                         goto loop;
6561                 }
6562 checks:
6563                 search_start = stripe_align(root, block_group,
6564                                             offset, num_bytes);
6565
6566                 /* move on to the next group */
6567                 if (search_start + num_bytes >
6568                     block_group->key.objectid + block_group->key.offset) {
6569                         btrfs_add_free_space(block_group, offset, num_bytes);
6570                         goto loop;
6571                 }
6572
6573                 if (offset < search_start)
6574                         btrfs_add_free_space(block_group, offset,
6575                                              search_start - offset);
6576                 BUG_ON(offset > search_start);
6577
6578                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
6579                                                   alloc_type);
6580                 if (ret == -EAGAIN) {
6581                         btrfs_add_free_space(block_group, offset, num_bytes);
6582                         goto loop;
6583                 }
6584
6585                 /* we are all good, lets return */
6586                 ins->objectid = search_start;
6587                 ins->offset = num_bytes;
6588
6589                 trace_btrfs_reserve_extent(orig_root, block_group,
6590                                            search_start, num_bytes);
6591                 btrfs_put_block_group(block_group);
6592                 break;
6593 loop:
6594                 failed_cluster_refill = false;
6595                 failed_alloc = false;
6596                 BUG_ON(index != get_block_group_index(block_group));
6597                 btrfs_put_block_group(block_group);
6598         }
6599         up_read(&space_info->groups_sem);
6600
6601         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6602                 goto search;
6603
6604         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6605                 goto search;
6606
6607         /*
6608          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6609          *                      caching kthreads as we move along
6610          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6611          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6612          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6613          *                      again
6614          */
6615         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6616                 index = 0;
6617                 loop++;
6618                 if (loop == LOOP_ALLOC_CHUNK) {
6619                         struct btrfs_trans_handle *trans;
6620
6621                         trans = btrfs_join_transaction(root);
6622                         if (IS_ERR(trans)) {
6623                                 ret = PTR_ERR(trans);
6624                                 goto out;
6625                         }
6626
6627                         ret = do_chunk_alloc(trans, root, flags,
6628                                              CHUNK_ALLOC_FORCE);
6629                         /*
6630                          * Do not bail out on ENOSPC since we
6631                          * can do more things.
6632                          */
6633                         if (ret < 0 && ret != -ENOSPC)
6634                                 btrfs_abort_transaction(trans,
6635                                                         root, ret);
6636                         else
6637                                 ret = 0;
6638                         btrfs_end_transaction(trans, root);
6639                         if (ret)
6640                                 goto out;
6641                 }
6642
6643                 if (loop == LOOP_NO_EMPTY_SIZE) {
6644                         empty_size = 0;
6645                         empty_cluster = 0;
6646                 }
6647
6648                 goto search;
6649         } else if (!ins->objectid) {
6650                 ret = -ENOSPC;
6651         } else if (ins->objectid) {
6652                 ret = 0;
6653         }
6654 out:
6655         if (ret == -ENOSPC)
6656                 ins->offset = max_extent_size;
6657         return ret;
6658 }
6659
6660 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6661                             int dump_block_groups)
6662 {
6663         struct btrfs_block_group_cache *cache;
6664         int index = 0;
6665
6666         spin_lock(&info->lock);
6667         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
6668                info->flags,
6669                info->total_bytes - info->bytes_used - info->bytes_pinned -
6670                info->bytes_reserved - info->bytes_readonly,
6671                (info->full) ? "" : "not ");
6672         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
6673                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6674                info->total_bytes, info->bytes_used, info->bytes_pinned,
6675                info->bytes_reserved, info->bytes_may_use,
6676                info->bytes_readonly);
6677         spin_unlock(&info->lock);
6678
6679         if (!dump_block_groups)
6680                 return;
6681
6682         down_read(&info->groups_sem);
6683 again:
6684         list_for_each_entry(cache, &info->block_groups[index], list) {
6685                 spin_lock(&cache->lock);
6686                 printk(KERN_INFO "BTRFS: "
6687                            "block group %llu has %llu bytes, "
6688                            "%llu used %llu pinned %llu reserved %s\n",
6689                        cache->key.objectid, cache->key.offset,
6690                        btrfs_block_group_used(&cache->item), cache->pinned,
6691                        cache->reserved, cache->ro ? "[readonly]" : "");
6692                 btrfs_dump_free_space(cache, bytes);
6693                 spin_unlock(&cache->lock);
6694         }
6695         if (++index < BTRFS_NR_RAID_TYPES)
6696                 goto again;
6697         up_read(&info->groups_sem);
6698 }
6699
6700 int btrfs_reserve_extent(struct btrfs_root *root,
6701                          u64 num_bytes, u64 min_alloc_size,
6702                          u64 empty_size, u64 hint_byte,
6703                          struct btrfs_key *ins, int is_data)
6704 {
6705         bool final_tried = false;
6706         u64 flags;
6707         int ret;
6708
6709         flags = btrfs_get_alloc_profile(root, is_data);
6710 again:
6711         WARN_ON(num_bytes < root->sectorsize);
6712         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
6713                                flags);
6714
6715         if (ret == -ENOSPC) {
6716                 if (!final_tried && ins->offset) {
6717                         num_bytes = min(num_bytes >> 1, ins->offset);
6718                         num_bytes = round_down(num_bytes, root->sectorsize);
6719                         num_bytes = max(num_bytes, min_alloc_size);
6720                         if (num_bytes == min_alloc_size)
6721                                 final_tried = true;
6722                         goto again;
6723                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6724                         struct btrfs_space_info *sinfo;
6725
6726                         sinfo = __find_space_info(root->fs_info, flags);
6727                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6728                                 flags, num_bytes);
6729                         if (sinfo)
6730                                 dump_space_info(sinfo, num_bytes, 1);
6731                 }
6732         }
6733
6734         return ret;
6735 }
6736
6737 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6738                                         u64 start, u64 len, int pin)
6739 {
6740         struct btrfs_block_group_cache *cache;
6741         int ret = 0;
6742
6743         cache = btrfs_lookup_block_group(root->fs_info, start);
6744         if (!cache) {
6745                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6746                         start);
6747                 return -ENOSPC;
6748         }
6749
6750         if (btrfs_test_opt(root, DISCARD))
6751                 ret = btrfs_discard_extent(root, start, len, NULL);
6752
6753         if (pin)
6754                 pin_down_extent(root, cache, start, len, 1);
6755         else {
6756                 btrfs_add_free_space(cache, start, len);
6757                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6758         }
6759         btrfs_put_block_group(cache);
6760
6761         trace_btrfs_reserved_extent_free(root, start, len);
6762
6763         return ret;
6764 }
6765
6766 int btrfs_free_reserved_extent(struct btrfs_root *root,
6767                                         u64 start, u64 len)
6768 {
6769         return __btrfs_free_reserved_extent(root, start, len, 0);
6770 }
6771
6772 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6773                                        u64 start, u64 len)
6774 {
6775         return __btrfs_free_reserved_extent(root, start, len, 1);
6776 }
6777
6778 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6779                                       struct btrfs_root *root,
6780                                       u64 parent, u64 root_objectid,
6781                                       u64 flags, u64 owner, u64 offset,
6782                                       struct btrfs_key *ins, int ref_mod)
6783 {
6784         int ret;
6785         struct btrfs_fs_info *fs_info = root->fs_info;
6786         struct btrfs_extent_item *extent_item;
6787         struct btrfs_extent_inline_ref *iref;
6788         struct btrfs_path *path;
6789         struct extent_buffer *leaf;
6790         int type;
6791         u32 size;
6792
6793         if (parent > 0)
6794                 type = BTRFS_SHARED_DATA_REF_KEY;
6795         else
6796                 type = BTRFS_EXTENT_DATA_REF_KEY;
6797
6798         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6799
6800         path = btrfs_alloc_path();
6801         if (!path)
6802                 return -ENOMEM;
6803
6804         path->leave_spinning = 1;
6805         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6806                                       ins, size);
6807         if (ret) {
6808                 btrfs_free_path(path);
6809                 return ret;
6810         }
6811
6812         leaf = path->nodes[0];
6813         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6814                                      struct btrfs_extent_item);
6815         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6816         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6817         btrfs_set_extent_flags(leaf, extent_item,
6818                                flags | BTRFS_EXTENT_FLAG_DATA);
6819
6820         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6821         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6822         if (parent > 0) {
6823                 struct btrfs_shared_data_ref *ref;
6824                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6825                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6826                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6827         } else {
6828                 struct btrfs_extent_data_ref *ref;
6829                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6830                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6831                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6832                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6833                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6834         }
6835
6836         btrfs_mark_buffer_dirty(path->nodes[0]);
6837         btrfs_free_path(path);
6838
6839         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6840         if (ret) { /* -ENOENT, logic error */
6841                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6842                         ins->objectid, ins->offset);
6843                 BUG();
6844         }
6845         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6846         return ret;
6847 }
6848
6849 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6850                                      struct btrfs_root *root,
6851                                      u64 parent, u64 root_objectid,
6852                                      u64 flags, struct btrfs_disk_key *key,
6853                                      int level, struct btrfs_key *ins)
6854 {
6855         int ret;
6856         struct btrfs_fs_info *fs_info = root->fs_info;
6857         struct btrfs_extent_item *extent_item;
6858         struct btrfs_tree_block_info *block_info;
6859         struct btrfs_extent_inline_ref *iref;
6860         struct btrfs_path *path;
6861         struct extent_buffer *leaf;
6862         u32 size = sizeof(*extent_item) + sizeof(*iref);
6863         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6864                                                  SKINNY_METADATA);
6865
6866         if (!skinny_metadata)
6867                 size += sizeof(*block_info);
6868
6869         path = btrfs_alloc_path();
6870         if (!path) {
6871                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
6872                                                    root->leafsize);
6873                 return -ENOMEM;
6874         }
6875
6876         path->leave_spinning = 1;
6877         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6878                                       ins, size);
6879         if (ret) {
6880                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
6881                                                    root->leafsize);
6882                 btrfs_free_path(path);
6883                 return ret;
6884         }
6885
6886         leaf = path->nodes[0];
6887         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6888                                      struct btrfs_extent_item);
6889         btrfs_set_extent_refs(leaf, extent_item, 1);
6890         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6891         btrfs_set_extent_flags(leaf, extent_item,
6892                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6893
6894         if (skinny_metadata) {
6895                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6896         } else {
6897                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6898                 btrfs_set_tree_block_key(leaf, block_info, key);
6899                 btrfs_set_tree_block_level(leaf, block_info, level);
6900                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6901         }
6902
6903         if (parent > 0) {
6904                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6905                 btrfs_set_extent_inline_ref_type(leaf, iref,
6906                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6907                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6908         } else {
6909                 btrfs_set_extent_inline_ref_type(leaf, iref,
6910                                                  BTRFS_TREE_BLOCK_REF_KEY);
6911                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6912         }
6913
6914         btrfs_mark_buffer_dirty(leaf);
6915         btrfs_free_path(path);
6916
6917         ret = update_block_group(root, ins->objectid, root->leafsize, 1);
6918         if (ret) { /* -ENOENT, logic error */
6919                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6920                         ins->objectid, ins->offset);
6921                 BUG();
6922         }
6923
6924         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->leafsize);
6925         return ret;
6926 }
6927
6928 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6929                                      struct btrfs_root *root,
6930                                      u64 root_objectid, u64 owner,
6931                                      u64 offset, struct btrfs_key *ins)
6932 {
6933         int ret;
6934
6935         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6936
6937         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6938                                          ins->offset, 0,
6939                                          root_objectid, owner, offset,
6940                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6941         return ret;
6942 }
6943
6944 /*
6945  * this is used by the tree logging recovery code.  It records that
6946  * an extent has been allocated and makes sure to clear the free
6947  * space cache bits as well
6948  */
6949 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6950                                    struct btrfs_root *root,
6951                                    u64 root_objectid, u64 owner, u64 offset,
6952                                    struct btrfs_key *ins)
6953 {
6954         int ret;
6955         struct btrfs_block_group_cache *block_group;
6956
6957         /*
6958          * Mixed block groups will exclude before processing the log so we only
6959          * need to do the exlude dance if this fs isn't mixed.
6960          */
6961         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
6962                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
6963                 if (ret)
6964                         return ret;
6965         }
6966
6967         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6968         if (!block_group)
6969                 return -EINVAL;
6970
6971         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6972                                           RESERVE_ALLOC_NO_ACCOUNT);
6973         BUG_ON(ret); /* logic error */
6974         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6975                                          0, owner, offset, ins, 1);
6976         btrfs_put_block_group(block_group);
6977         return ret;
6978 }
6979
6980 static struct extent_buffer *
6981 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6982                       u64 bytenr, u32 blocksize, int level)
6983 {
6984         struct extent_buffer *buf;
6985
6986         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6987         if (!buf)
6988                 return ERR_PTR(-ENOMEM);
6989         btrfs_set_header_generation(buf, trans->transid);
6990         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6991         btrfs_tree_lock(buf);
6992         clean_tree_block(trans, root, buf);
6993         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6994
6995         btrfs_set_lock_blocking(buf);
6996         btrfs_set_buffer_uptodate(buf);
6997
6998         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6999                 /*
7000                  * we allow two log transactions at a time, use different
7001                  * EXENT bit to differentiate dirty pages.
7002                  */
7003                 if (root->log_transid % 2 == 0)
7004                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7005                                         buf->start + buf->len - 1, GFP_NOFS);
7006                 else
7007                         set_extent_new(&root->dirty_log_pages, buf->start,
7008                                         buf->start + buf->len - 1, GFP_NOFS);
7009         } else {
7010                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7011                          buf->start + buf->len - 1, GFP_NOFS);
7012         }
7013         trans->blocks_used++;
7014         /* this returns a buffer locked for blocking */
7015         return buf;
7016 }
7017
7018 static struct btrfs_block_rsv *
7019 use_block_rsv(struct btrfs_trans_handle *trans,
7020               struct btrfs_root *root, u32 blocksize)
7021 {
7022         struct btrfs_block_rsv *block_rsv;
7023         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7024         int ret;
7025         bool global_updated = false;
7026
7027         block_rsv = get_block_rsv(trans, root);
7028
7029         if (unlikely(block_rsv->size == 0))
7030                 goto try_reserve;
7031 again:
7032         ret = block_rsv_use_bytes(block_rsv, blocksize);
7033         if (!ret)
7034                 return block_rsv;
7035
7036         if (block_rsv->failfast)
7037                 return ERR_PTR(ret);
7038
7039         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7040                 global_updated = true;
7041                 update_global_block_rsv(root->fs_info);
7042                 goto again;
7043         }
7044
7045         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7046                 static DEFINE_RATELIMIT_STATE(_rs,
7047                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7048                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7049                 if (__ratelimit(&_rs))
7050                         WARN(1, KERN_DEBUG
7051                                 "BTRFS: block rsv returned %d\n", ret);
7052         }
7053 try_reserve:
7054         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7055                                      BTRFS_RESERVE_NO_FLUSH);
7056         if (!ret)
7057                 return block_rsv;
7058         /*
7059          * If we couldn't reserve metadata bytes try and use some from
7060          * the global reserve if its space type is the same as the global
7061          * reservation.
7062          */
7063         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7064             block_rsv->space_info == global_rsv->space_info) {
7065                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7066                 if (!ret)
7067                         return global_rsv;
7068         }
7069         return ERR_PTR(ret);
7070 }
7071
7072 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7073                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7074 {
7075         block_rsv_add_bytes(block_rsv, blocksize, 0);
7076         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7077 }
7078
7079 /*
7080  * finds a free extent and does all the dirty work required for allocation
7081  * returns the key for the extent through ins, and a tree buffer for
7082  * the first block of the extent through buf.
7083  *
7084  * returns the tree buffer or NULL.
7085  */
7086 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
7087                                         struct btrfs_root *root, u32 blocksize,
7088                                         u64 parent, u64 root_objectid,
7089                                         struct btrfs_disk_key *key, int level,
7090                                         u64 hint, u64 empty_size)
7091 {
7092         struct btrfs_key ins;
7093         struct btrfs_block_rsv *block_rsv;
7094         struct extent_buffer *buf;
7095         u64 flags = 0;
7096         int ret;
7097         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7098                                                  SKINNY_METADATA);
7099
7100         block_rsv = use_block_rsv(trans, root, blocksize);
7101         if (IS_ERR(block_rsv))
7102                 return ERR_CAST(block_rsv);
7103
7104         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7105                                    empty_size, hint, &ins, 0);
7106         if (ret) {
7107                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7108                 return ERR_PTR(ret);
7109         }
7110
7111         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
7112                                     blocksize, level);
7113         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
7114
7115         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7116                 if (parent == 0)
7117                         parent = ins.objectid;
7118                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7119         } else
7120                 BUG_ON(parent > 0);
7121
7122         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7123                 struct btrfs_delayed_extent_op *extent_op;
7124                 extent_op = btrfs_alloc_delayed_extent_op();
7125                 BUG_ON(!extent_op); /* -ENOMEM */
7126                 if (key)
7127                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7128                 else
7129                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7130                 extent_op->flags_to_set = flags;
7131                 if (skinny_metadata)
7132                         extent_op->update_key = 0;
7133                 else
7134                         extent_op->update_key = 1;
7135                 extent_op->update_flags = 1;
7136                 extent_op->is_data = 0;
7137                 extent_op->level = level;
7138
7139                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7140                                         ins.objectid,
7141                                         ins.offset, parent, root_objectid,
7142                                         level, BTRFS_ADD_DELAYED_EXTENT,
7143                                         extent_op, 0);
7144                 BUG_ON(ret); /* -ENOMEM */
7145         }
7146         return buf;
7147 }
7148
7149 struct walk_control {
7150         u64 refs[BTRFS_MAX_LEVEL];
7151         u64 flags[BTRFS_MAX_LEVEL];
7152         struct btrfs_key update_progress;
7153         int stage;
7154         int level;
7155         int shared_level;
7156         int update_ref;
7157         int keep_locks;
7158         int reada_slot;
7159         int reada_count;
7160         int for_reloc;
7161 };
7162
7163 #define DROP_REFERENCE  1
7164 #define UPDATE_BACKREF  2
7165
7166 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7167                                      struct btrfs_root *root,
7168                                      struct walk_control *wc,
7169                                      struct btrfs_path *path)
7170 {
7171         u64 bytenr;
7172         u64 generation;
7173         u64 refs;
7174         u64 flags;
7175         u32 nritems;
7176         u32 blocksize;
7177         struct btrfs_key key;
7178         struct extent_buffer *eb;
7179         int ret;
7180         int slot;
7181         int nread = 0;
7182
7183         if (path->slots[wc->level] < wc->reada_slot) {
7184                 wc->reada_count = wc->reada_count * 2 / 3;
7185                 wc->reada_count = max(wc->reada_count, 2);
7186         } else {
7187                 wc->reada_count = wc->reada_count * 3 / 2;
7188                 wc->reada_count = min_t(int, wc->reada_count,
7189                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7190         }
7191
7192         eb = path->nodes[wc->level];
7193         nritems = btrfs_header_nritems(eb);
7194         blocksize = btrfs_level_size(root, wc->level - 1);
7195
7196         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7197                 if (nread >= wc->reada_count)
7198                         break;
7199
7200                 cond_resched();
7201                 bytenr = btrfs_node_blockptr(eb, slot);
7202                 generation = btrfs_node_ptr_generation(eb, slot);
7203
7204                 if (slot == path->slots[wc->level])
7205                         goto reada;
7206
7207                 if (wc->stage == UPDATE_BACKREF &&
7208                     generation <= root->root_key.offset)
7209                         continue;
7210
7211                 /* We don't lock the tree block, it's OK to be racy here */
7212                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7213                                                wc->level - 1, 1, &refs,
7214                                                &flags);
7215                 /* We don't care about errors in readahead. */
7216                 if (ret < 0)
7217                         continue;
7218                 BUG_ON(refs == 0);
7219
7220                 if (wc->stage == DROP_REFERENCE) {
7221                         if (refs == 1)
7222                                 goto reada;
7223
7224                         if (wc->level == 1 &&
7225                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7226                                 continue;
7227                         if (!wc->update_ref ||
7228                             generation <= root->root_key.offset)
7229                                 continue;
7230                         btrfs_node_key_to_cpu(eb, &key, slot);
7231                         ret = btrfs_comp_cpu_keys(&key,
7232                                                   &wc->update_progress);
7233                         if (ret < 0)
7234                                 continue;
7235                 } else {
7236                         if (wc->level == 1 &&
7237                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7238                                 continue;
7239                 }
7240 reada:
7241                 ret = readahead_tree_block(root, bytenr, blocksize,
7242                                            generation);
7243                 if (ret)
7244                         break;
7245                 nread++;
7246         }
7247         wc->reada_slot = slot;
7248 }
7249
7250 /*
7251  * helper to process tree block while walking down the tree.
7252  *
7253  * when wc->stage == UPDATE_BACKREF, this function updates
7254  * back refs for pointers in the block.
7255  *
7256  * NOTE: return value 1 means we should stop walking down.
7257  */
7258 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7259                                    struct btrfs_root *root,
7260                                    struct btrfs_path *path,
7261                                    struct walk_control *wc, int lookup_info)
7262 {
7263         int level = wc->level;
7264         struct extent_buffer *eb = path->nodes[level];
7265         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7266         int ret;
7267
7268         if (wc->stage == UPDATE_BACKREF &&
7269             btrfs_header_owner(eb) != root->root_key.objectid)
7270                 return 1;
7271
7272         /*
7273          * when reference count of tree block is 1, it won't increase
7274          * again. once full backref flag is set, we never clear it.
7275          */
7276         if (lookup_info &&
7277             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7278              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7279                 BUG_ON(!path->locks[level]);
7280                 ret = btrfs_lookup_extent_info(trans, root,
7281                                                eb->start, level, 1,
7282                                                &wc->refs[level],
7283                                                &wc->flags[level]);
7284                 BUG_ON(ret == -ENOMEM);
7285                 if (ret)
7286                         return ret;
7287                 BUG_ON(wc->refs[level] == 0);
7288         }
7289
7290         if (wc->stage == DROP_REFERENCE) {
7291                 if (wc->refs[level] > 1)
7292                         return 1;
7293
7294                 if (path->locks[level] && !wc->keep_locks) {
7295                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7296                         path->locks[level] = 0;
7297                 }
7298                 return 0;
7299         }
7300
7301         /* wc->stage == UPDATE_BACKREF */
7302         if (!(wc->flags[level] & flag)) {
7303                 BUG_ON(!path->locks[level]);
7304                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
7305                 BUG_ON(ret); /* -ENOMEM */
7306                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
7307                 BUG_ON(ret); /* -ENOMEM */
7308                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7309                                                   eb->len, flag,
7310                                                   btrfs_header_level(eb), 0);
7311                 BUG_ON(ret); /* -ENOMEM */
7312                 wc->flags[level] |= flag;
7313         }
7314
7315         /*
7316          * the block is shared by multiple trees, so it's not good to
7317          * keep the tree lock
7318          */
7319         if (path->locks[level] && level > 0) {
7320                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7321                 path->locks[level] = 0;
7322         }
7323         return 0;
7324 }
7325
7326 /*
7327  * helper to process tree block pointer.
7328  *
7329  * when wc->stage == DROP_REFERENCE, this function checks
7330  * reference count of the block pointed to. if the block
7331  * is shared and we need update back refs for the subtree
7332  * rooted at the block, this function changes wc->stage to
7333  * UPDATE_BACKREF. if the block is shared and there is no
7334  * need to update back, this function drops the reference
7335  * to the block.
7336  *
7337  * NOTE: return value 1 means we should stop walking down.
7338  */
7339 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7340                                  struct btrfs_root *root,
7341                                  struct btrfs_path *path,
7342                                  struct walk_control *wc, int *lookup_info)
7343 {
7344         u64 bytenr;
7345         u64 generation;
7346         u64 parent;
7347         u32 blocksize;
7348         struct btrfs_key key;
7349         struct extent_buffer *next;
7350         int level = wc->level;
7351         int reada = 0;
7352         int ret = 0;
7353
7354         generation = btrfs_node_ptr_generation(path->nodes[level],
7355                                                path->slots[level]);
7356         /*
7357          * if the lower level block was created before the snapshot
7358          * was created, we know there is no need to update back refs
7359          * for the subtree
7360          */
7361         if (wc->stage == UPDATE_BACKREF &&
7362             generation <= root->root_key.offset) {
7363                 *lookup_info = 1;
7364                 return 1;
7365         }
7366
7367         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7368         blocksize = btrfs_level_size(root, level - 1);
7369
7370         next = btrfs_find_tree_block(root, bytenr, blocksize);
7371         if (!next) {
7372                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7373                 if (!next)
7374                         return -ENOMEM;
7375                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
7376                                                level - 1);
7377                 reada = 1;
7378         }
7379         btrfs_tree_lock(next);
7380         btrfs_set_lock_blocking(next);
7381
7382         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7383                                        &wc->refs[level - 1],
7384                                        &wc->flags[level - 1]);
7385         if (ret < 0) {
7386                 btrfs_tree_unlock(next);
7387                 return ret;
7388         }
7389
7390         if (unlikely(wc->refs[level - 1] == 0)) {
7391                 btrfs_err(root->fs_info, "Missing references.");
7392                 BUG();
7393         }
7394         *lookup_info = 0;
7395
7396         if (wc->stage == DROP_REFERENCE) {
7397                 if (wc->refs[level - 1] > 1) {
7398                         if (level == 1 &&
7399                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7400                                 goto skip;
7401
7402                         if (!wc->update_ref ||
7403                             generation <= root->root_key.offset)
7404                                 goto skip;
7405
7406                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7407                                               path->slots[level]);
7408                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7409                         if (ret < 0)
7410                                 goto skip;
7411
7412                         wc->stage = UPDATE_BACKREF;
7413                         wc->shared_level = level - 1;
7414                 }
7415         } else {
7416                 if (level == 1 &&
7417                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7418                         goto skip;
7419         }
7420
7421         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7422                 btrfs_tree_unlock(next);
7423                 free_extent_buffer(next);
7424                 next = NULL;
7425                 *lookup_info = 1;
7426         }
7427
7428         if (!next) {
7429                 if (reada && level == 1)
7430                         reada_walk_down(trans, root, wc, path);
7431                 next = read_tree_block(root, bytenr, blocksize, generation);
7432                 if (!next || !extent_buffer_uptodate(next)) {
7433                         free_extent_buffer(next);
7434                         return -EIO;
7435                 }
7436                 btrfs_tree_lock(next);
7437                 btrfs_set_lock_blocking(next);
7438         }
7439
7440         level--;
7441         BUG_ON(level != btrfs_header_level(next));
7442         path->nodes[level] = next;
7443         path->slots[level] = 0;
7444         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7445         wc->level = level;
7446         if (wc->level == 1)
7447                 wc->reada_slot = 0;
7448         return 0;
7449 skip:
7450         wc->refs[level - 1] = 0;
7451         wc->flags[level - 1] = 0;
7452         if (wc->stage == DROP_REFERENCE) {
7453                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7454                         parent = path->nodes[level]->start;
7455                 } else {
7456                         BUG_ON(root->root_key.objectid !=
7457                                btrfs_header_owner(path->nodes[level]));
7458                         parent = 0;
7459                 }
7460
7461                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7462                                 root->root_key.objectid, level - 1, 0, 0);
7463                 BUG_ON(ret); /* -ENOMEM */
7464         }
7465         btrfs_tree_unlock(next);
7466         free_extent_buffer(next);
7467         *lookup_info = 1;
7468         return 1;
7469 }
7470
7471 /*
7472  * helper to process tree block while walking up the tree.
7473  *
7474  * when wc->stage == DROP_REFERENCE, this function drops
7475  * reference count on the block.
7476  *
7477  * when wc->stage == UPDATE_BACKREF, this function changes
7478  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7479  * to UPDATE_BACKREF previously while processing the block.
7480  *
7481  * NOTE: return value 1 means we should stop walking up.
7482  */
7483 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7484                                  struct btrfs_root *root,
7485                                  struct btrfs_path *path,
7486                                  struct walk_control *wc)
7487 {
7488         int ret;
7489         int level = wc->level;
7490         struct extent_buffer *eb = path->nodes[level];
7491         u64 parent = 0;
7492
7493         if (wc->stage == UPDATE_BACKREF) {
7494                 BUG_ON(wc->shared_level < level);
7495                 if (level < wc->shared_level)
7496                         goto out;
7497
7498                 ret = find_next_key(path, level + 1, &wc->update_progress);
7499                 if (ret > 0)
7500                         wc->update_ref = 0;
7501
7502                 wc->stage = DROP_REFERENCE;
7503                 wc->shared_level = -1;
7504                 path->slots[level] = 0;
7505
7506                 /*
7507                  * check reference count again if the block isn't locked.
7508                  * we should start walking down the tree again if reference
7509                  * count is one.
7510                  */
7511                 if (!path->locks[level]) {
7512                         BUG_ON(level == 0);
7513                         btrfs_tree_lock(eb);
7514                         btrfs_set_lock_blocking(eb);
7515                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7516
7517                         ret = btrfs_lookup_extent_info(trans, root,
7518                                                        eb->start, level, 1,
7519                                                        &wc->refs[level],
7520                                                        &wc->flags[level]);
7521                         if (ret < 0) {
7522                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7523                                 path->locks[level] = 0;
7524                                 return ret;
7525                         }
7526                         BUG_ON(wc->refs[level] == 0);
7527                         if (wc->refs[level] == 1) {
7528                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7529                                 path->locks[level] = 0;
7530                                 return 1;
7531                         }
7532                 }
7533         }
7534
7535         /* wc->stage == DROP_REFERENCE */
7536         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7537
7538         if (wc->refs[level] == 1) {
7539                 if (level == 0) {
7540                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7541                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7542                                                     wc->for_reloc);
7543                         else
7544                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7545                                                     wc->for_reloc);
7546                         BUG_ON(ret); /* -ENOMEM */
7547                 }
7548                 /* make block locked assertion in clean_tree_block happy */
7549                 if (!path->locks[level] &&
7550                     btrfs_header_generation(eb) == trans->transid) {
7551                         btrfs_tree_lock(eb);
7552                         btrfs_set_lock_blocking(eb);
7553                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7554                 }
7555                 clean_tree_block(trans, root, eb);
7556         }
7557
7558         if (eb == root->node) {
7559                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7560                         parent = eb->start;
7561                 else
7562                         BUG_ON(root->root_key.objectid !=
7563                                btrfs_header_owner(eb));
7564         } else {
7565                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7566                         parent = path->nodes[level + 1]->start;
7567                 else
7568                         BUG_ON(root->root_key.objectid !=
7569                                btrfs_header_owner(path->nodes[level + 1]));
7570         }
7571
7572         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7573 out:
7574         wc->refs[level] = 0;
7575         wc->flags[level] = 0;
7576         return 0;
7577 }
7578
7579 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7580                                    struct btrfs_root *root,
7581                                    struct btrfs_path *path,
7582                                    struct walk_control *wc)
7583 {
7584         int level = wc->level;
7585         int lookup_info = 1;
7586         int ret;
7587
7588         while (level >= 0) {
7589                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7590                 if (ret > 0)
7591                         break;
7592
7593                 if (level == 0)
7594                         break;
7595
7596                 if (path->slots[level] >=
7597                     btrfs_header_nritems(path->nodes[level]))
7598                         break;
7599
7600                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7601                 if (ret > 0) {
7602                         path->slots[level]++;
7603                         continue;
7604                 } else if (ret < 0)
7605                         return ret;
7606                 level = wc->level;
7607         }
7608         return 0;
7609 }
7610
7611 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7612                                  struct btrfs_root *root,
7613                                  struct btrfs_path *path,
7614                                  struct walk_control *wc, int max_level)
7615 {
7616         int level = wc->level;
7617         int ret;
7618
7619         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7620         while (level < max_level && path->nodes[level]) {
7621                 wc->level = level;
7622                 if (path->slots[level] + 1 <
7623                     btrfs_header_nritems(path->nodes[level])) {
7624                         path->slots[level]++;
7625                         return 0;
7626                 } else {
7627                         ret = walk_up_proc(trans, root, path, wc);
7628                         if (ret > 0)
7629                                 return 0;
7630
7631                         if (path->locks[level]) {
7632                                 btrfs_tree_unlock_rw(path->nodes[level],
7633                                                      path->locks[level]);
7634                                 path->locks[level] = 0;
7635                         }
7636                         free_extent_buffer(path->nodes[level]);
7637                         path->nodes[level] = NULL;
7638                         level++;
7639                 }
7640         }
7641         return 1;
7642 }
7643
7644 /*
7645  * drop a subvolume tree.
7646  *
7647  * this function traverses the tree freeing any blocks that only
7648  * referenced by the tree.
7649  *
7650  * when a shared tree block is found. this function decreases its
7651  * reference count by one. if update_ref is true, this function
7652  * also make sure backrefs for the shared block and all lower level
7653  * blocks are properly updated.
7654  *
7655  * If called with for_reloc == 0, may exit early with -EAGAIN
7656  */
7657 int btrfs_drop_snapshot(struct btrfs_root *root,
7658                          struct btrfs_block_rsv *block_rsv, int update_ref,
7659                          int for_reloc)
7660 {
7661         struct btrfs_path *path;
7662         struct btrfs_trans_handle *trans;
7663         struct btrfs_root *tree_root = root->fs_info->tree_root;
7664         struct btrfs_root_item *root_item = &root->root_item;
7665         struct walk_control *wc;
7666         struct btrfs_key key;
7667         int err = 0;
7668         int ret;
7669         int level;
7670         bool root_dropped = false;
7671
7672         path = btrfs_alloc_path();
7673         if (!path) {
7674                 err = -ENOMEM;
7675                 goto out;
7676         }
7677
7678         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7679         if (!wc) {
7680                 btrfs_free_path(path);
7681                 err = -ENOMEM;
7682                 goto out;
7683         }
7684
7685         trans = btrfs_start_transaction(tree_root, 0);
7686         if (IS_ERR(trans)) {
7687                 err = PTR_ERR(trans);
7688                 goto out_free;
7689         }
7690
7691         if (block_rsv)
7692                 trans->block_rsv = block_rsv;
7693
7694         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7695                 level = btrfs_header_level(root->node);
7696                 path->nodes[level] = btrfs_lock_root_node(root);
7697                 btrfs_set_lock_blocking(path->nodes[level]);
7698                 path->slots[level] = 0;
7699                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7700                 memset(&wc->update_progress, 0,
7701                        sizeof(wc->update_progress));
7702         } else {
7703                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7704                 memcpy(&wc->update_progress, &key,
7705                        sizeof(wc->update_progress));
7706
7707                 level = root_item->drop_level;
7708                 BUG_ON(level == 0);
7709                 path->lowest_level = level;
7710                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7711                 path->lowest_level = 0;
7712                 if (ret < 0) {
7713                         err = ret;
7714                         goto out_end_trans;
7715                 }
7716                 WARN_ON(ret > 0);
7717
7718                 /*
7719                  * unlock our path, this is safe because only this
7720                  * function is allowed to delete this snapshot
7721                  */
7722                 btrfs_unlock_up_safe(path, 0);
7723
7724                 level = btrfs_header_level(root->node);
7725                 while (1) {
7726                         btrfs_tree_lock(path->nodes[level]);
7727                         btrfs_set_lock_blocking(path->nodes[level]);
7728                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7729
7730                         ret = btrfs_lookup_extent_info(trans, root,
7731                                                 path->nodes[level]->start,
7732                                                 level, 1, &wc->refs[level],
7733                                                 &wc->flags[level]);
7734                         if (ret < 0) {
7735                                 err = ret;
7736                                 goto out_end_trans;
7737                         }
7738                         BUG_ON(wc->refs[level] == 0);
7739
7740                         if (level == root_item->drop_level)
7741                                 break;
7742
7743                         btrfs_tree_unlock(path->nodes[level]);
7744                         path->locks[level] = 0;
7745                         WARN_ON(wc->refs[level] != 1);
7746                         level--;
7747                 }
7748         }
7749
7750         wc->level = level;
7751         wc->shared_level = -1;
7752         wc->stage = DROP_REFERENCE;
7753         wc->update_ref = update_ref;
7754         wc->keep_locks = 0;
7755         wc->for_reloc = for_reloc;
7756         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7757
7758         while (1) {
7759
7760                 ret = walk_down_tree(trans, root, path, wc);
7761                 if (ret < 0) {
7762                         err = ret;
7763                         break;
7764                 }
7765
7766                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7767                 if (ret < 0) {
7768                         err = ret;
7769                         break;
7770                 }
7771
7772                 if (ret > 0) {
7773                         BUG_ON(wc->stage != DROP_REFERENCE);
7774                         break;
7775                 }
7776
7777                 if (wc->stage == DROP_REFERENCE) {
7778                         level = wc->level;
7779                         btrfs_node_key(path->nodes[level],
7780                                        &root_item->drop_progress,
7781                                        path->slots[level]);
7782                         root_item->drop_level = level;
7783                 }
7784
7785                 BUG_ON(wc->level == 0);
7786                 if (btrfs_should_end_transaction(trans, tree_root) ||
7787                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
7788                         ret = btrfs_update_root(trans, tree_root,
7789                                                 &root->root_key,
7790                                                 root_item);
7791                         if (ret) {
7792                                 btrfs_abort_transaction(trans, tree_root, ret);
7793                                 err = ret;
7794                                 goto out_end_trans;
7795                         }
7796
7797                         btrfs_end_transaction_throttle(trans, tree_root);
7798                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
7799                                 pr_debug("BTRFS: drop snapshot early exit\n");
7800                                 err = -EAGAIN;
7801                                 goto out_free;
7802                         }
7803
7804                         trans = btrfs_start_transaction(tree_root, 0);
7805                         if (IS_ERR(trans)) {
7806                                 err = PTR_ERR(trans);
7807                                 goto out_free;
7808                         }
7809                         if (block_rsv)
7810                                 trans->block_rsv = block_rsv;
7811                 }
7812         }
7813         btrfs_release_path(path);
7814         if (err)
7815                 goto out_end_trans;
7816
7817         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7818         if (ret) {
7819                 btrfs_abort_transaction(trans, tree_root, ret);
7820                 goto out_end_trans;
7821         }
7822
7823         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7824                 ret = btrfs_find_root(tree_root, &root->root_key, path,
7825                                       NULL, NULL);
7826                 if (ret < 0) {
7827                         btrfs_abort_transaction(trans, tree_root, ret);
7828                         err = ret;
7829                         goto out_end_trans;
7830                 } else if (ret > 0) {
7831                         /* if we fail to delete the orphan item this time
7832                          * around, it'll get picked up the next time.
7833                          *
7834                          * The most common failure here is just -ENOENT.
7835                          */
7836                         btrfs_del_orphan_item(trans, tree_root,
7837                                               root->root_key.objectid);
7838                 }
7839         }
7840
7841         if (root->in_radix) {
7842                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
7843         } else {
7844                 free_extent_buffer(root->node);
7845                 free_extent_buffer(root->commit_root);
7846                 btrfs_put_fs_root(root);
7847         }
7848         root_dropped = true;
7849 out_end_trans:
7850         btrfs_end_transaction_throttle(trans, tree_root);
7851 out_free:
7852         kfree(wc);
7853         btrfs_free_path(path);
7854 out:
7855         /*
7856          * So if we need to stop dropping the snapshot for whatever reason we
7857          * need to make sure to add it back to the dead root list so that we
7858          * keep trying to do the work later.  This also cleans up roots if we
7859          * don't have it in the radix (like when we recover after a power fail
7860          * or unmount) so we don't leak memory.
7861          */
7862         if (!for_reloc && root_dropped == false)
7863                 btrfs_add_dead_root(root);
7864         if (err && err != -EAGAIN)
7865                 btrfs_std_error(root->fs_info, err);
7866         return err;
7867 }
7868
7869 /*
7870  * drop subtree rooted at tree block 'node'.
7871  *
7872  * NOTE: this function will unlock and release tree block 'node'
7873  * only used by relocation code
7874  */
7875 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7876                         struct btrfs_root *root,
7877                         struct extent_buffer *node,
7878                         struct extent_buffer *parent)
7879 {
7880         struct btrfs_path *path;
7881         struct walk_control *wc;
7882         int level;
7883         int parent_level;
7884         int ret = 0;
7885         int wret;
7886
7887         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7888
7889         path = btrfs_alloc_path();
7890         if (!path)
7891                 return -ENOMEM;
7892
7893         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7894         if (!wc) {
7895                 btrfs_free_path(path);
7896                 return -ENOMEM;
7897         }
7898
7899         btrfs_assert_tree_locked(parent);
7900         parent_level = btrfs_header_level(parent);
7901         extent_buffer_get(parent);
7902         path->nodes[parent_level] = parent;
7903         path->slots[parent_level] = btrfs_header_nritems(parent);
7904
7905         btrfs_assert_tree_locked(node);
7906         level = btrfs_header_level(node);
7907         path->nodes[level] = node;
7908         path->slots[level] = 0;
7909         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7910
7911         wc->refs[parent_level] = 1;
7912         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7913         wc->level = level;
7914         wc->shared_level = -1;
7915         wc->stage = DROP_REFERENCE;
7916         wc->update_ref = 0;
7917         wc->keep_locks = 1;
7918         wc->for_reloc = 1;
7919         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7920
7921         while (1) {
7922                 wret = walk_down_tree(trans, root, path, wc);
7923                 if (wret < 0) {
7924                         ret = wret;
7925                         break;
7926                 }
7927
7928                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7929                 if (wret < 0)
7930                         ret = wret;
7931                 if (wret != 0)
7932                         break;
7933         }
7934
7935         kfree(wc);
7936         btrfs_free_path(path);
7937         return ret;
7938 }
7939
7940 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7941 {
7942         u64 num_devices;
7943         u64 stripped;
7944
7945         /*
7946          * if restripe for this chunk_type is on pick target profile and
7947          * return, otherwise do the usual balance
7948          */
7949         stripped = get_restripe_target(root->fs_info, flags);
7950         if (stripped)
7951                 return extended_to_chunk(stripped);
7952
7953         /*
7954          * we add in the count of missing devices because we want
7955          * to make sure that any RAID levels on a degraded FS
7956          * continue to be honored.
7957          */
7958         num_devices = root->fs_info->fs_devices->rw_devices +
7959                 root->fs_info->fs_devices->missing_devices;
7960
7961         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7962                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7963                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7964
7965         if (num_devices == 1) {
7966                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7967                 stripped = flags & ~stripped;
7968
7969                 /* turn raid0 into single device chunks */
7970                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7971                         return stripped;
7972
7973                 /* turn mirroring into duplication */
7974                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7975                              BTRFS_BLOCK_GROUP_RAID10))
7976                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7977         } else {
7978                 /* they already had raid on here, just return */
7979                 if (flags & stripped)
7980                         return flags;
7981
7982                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7983                 stripped = flags & ~stripped;
7984
7985                 /* switch duplicated blocks with raid1 */
7986                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7987                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7988
7989                 /* this is drive concat, leave it alone */
7990         }
7991
7992         return flags;
7993 }
7994
7995 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7996 {
7997         struct btrfs_space_info *sinfo = cache->space_info;
7998         u64 num_bytes;
7999         u64 min_allocable_bytes;
8000         int ret = -ENOSPC;
8001
8002
8003         /*
8004          * We need some metadata space and system metadata space for
8005          * allocating chunks in some corner cases until we force to set
8006          * it to be readonly.
8007          */
8008         if ((sinfo->flags &
8009              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8010             !force)
8011                 min_allocable_bytes = 1 * 1024 * 1024;
8012         else
8013                 min_allocable_bytes = 0;
8014
8015         spin_lock(&sinfo->lock);
8016         spin_lock(&cache->lock);
8017
8018         if (cache->ro) {
8019                 ret = 0;
8020                 goto out;
8021         }
8022
8023         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8024                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8025
8026         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8027             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8028             min_allocable_bytes <= sinfo->total_bytes) {
8029                 sinfo->bytes_readonly += num_bytes;
8030                 cache->ro = 1;
8031                 ret = 0;
8032         }
8033 out:
8034         spin_unlock(&cache->lock);
8035         spin_unlock(&sinfo->lock);
8036         return ret;
8037 }
8038
8039 int btrfs_set_block_group_ro(struct btrfs_root *root,
8040                              struct btrfs_block_group_cache *cache)
8041
8042 {
8043         struct btrfs_trans_handle *trans;
8044         u64 alloc_flags;
8045         int ret;
8046
8047         BUG_ON(cache->ro);
8048
8049         trans = btrfs_join_transaction(root);
8050         if (IS_ERR(trans))
8051                 return PTR_ERR(trans);
8052
8053         alloc_flags = update_block_group_flags(root, cache->flags);
8054         if (alloc_flags != cache->flags) {
8055                 ret = do_chunk_alloc(trans, root, alloc_flags,
8056                                      CHUNK_ALLOC_FORCE);
8057                 if (ret < 0)
8058                         goto out;
8059         }
8060
8061         ret = set_block_group_ro(cache, 0);
8062         if (!ret)
8063                 goto out;
8064         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8065         ret = do_chunk_alloc(trans, root, alloc_flags,
8066                              CHUNK_ALLOC_FORCE);
8067         if (ret < 0)
8068                 goto out;
8069         ret = set_block_group_ro(cache, 0);
8070 out:
8071         btrfs_end_transaction(trans, root);
8072         return ret;
8073 }
8074
8075 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8076                             struct btrfs_root *root, u64 type)
8077 {
8078         u64 alloc_flags = get_alloc_profile(root, type);
8079         return do_chunk_alloc(trans, root, alloc_flags,
8080                               CHUNK_ALLOC_FORCE);
8081 }
8082
8083 /*
8084  * helper to account the unused space of all the readonly block group in the
8085  * list. takes mirrors into account.
8086  */
8087 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
8088 {
8089         struct btrfs_block_group_cache *block_group;
8090         u64 free_bytes = 0;
8091         int factor;
8092
8093         list_for_each_entry(block_group, groups_list, list) {
8094                 spin_lock(&block_group->lock);
8095
8096                 if (!block_group->ro) {
8097                         spin_unlock(&block_group->lock);
8098                         continue;
8099                 }
8100
8101                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8102                                           BTRFS_BLOCK_GROUP_RAID10 |
8103                                           BTRFS_BLOCK_GROUP_DUP))
8104                         factor = 2;
8105                 else
8106                         factor = 1;
8107
8108                 free_bytes += (block_group->key.offset -
8109                                btrfs_block_group_used(&block_group->item)) *
8110                                factor;
8111
8112                 spin_unlock(&block_group->lock);
8113         }
8114
8115         return free_bytes;
8116 }
8117
8118 /*
8119  * helper to account the unused space of all the readonly block group in the
8120  * space_info. takes mirrors into account.
8121  */
8122 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8123 {
8124         int i;
8125         u64 free_bytes = 0;
8126
8127         spin_lock(&sinfo->lock);
8128
8129         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
8130                 if (!list_empty(&sinfo->block_groups[i]))
8131                         free_bytes += __btrfs_get_ro_block_group_free_space(
8132                                                 &sinfo->block_groups[i]);
8133
8134         spin_unlock(&sinfo->lock);
8135
8136         return free_bytes;
8137 }
8138
8139 void btrfs_set_block_group_rw(struct btrfs_root *root,
8140                               struct btrfs_block_group_cache *cache)
8141 {
8142         struct btrfs_space_info *sinfo = cache->space_info;
8143         u64 num_bytes;
8144
8145         BUG_ON(!cache->ro);
8146
8147         spin_lock(&sinfo->lock);
8148         spin_lock(&cache->lock);
8149         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8150                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8151         sinfo->bytes_readonly -= num_bytes;
8152         cache->ro = 0;
8153         spin_unlock(&cache->lock);
8154         spin_unlock(&sinfo->lock);
8155 }
8156
8157 /*
8158  * checks to see if its even possible to relocate this block group.
8159  *
8160  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8161  * ok to go ahead and try.
8162  */
8163 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8164 {
8165         struct btrfs_block_group_cache *block_group;
8166         struct btrfs_space_info *space_info;
8167         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8168         struct btrfs_device *device;
8169         struct btrfs_trans_handle *trans;
8170         u64 min_free;
8171         u64 dev_min = 1;
8172         u64 dev_nr = 0;
8173         u64 target;
8174         int index;
8175         int full = 0;
8176         int ret = 0;
8177
8178         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8179
8180         /* odd, couldn't find the block group, leave it alone */
8181         if (!block_group)
8182                 return -1;
8183
8184         min_free = btrfs_block_group_used(&block_group->item);
8185
8186         /* no bytes used, we're good */
8187         if (!min_free)
8188                 goto out;
8189
8190         space_info = block_group->space_info;
8191         spin_lock(&space_info->lock);
8192
8193         full = space_info->full;
8194
8195         /*
8196          * if this is the last block group we have in this space, we can't
8197          * relocate it unless we're able to allocate a new chunk below.
8198          *
8199          * Otherwise, we need to make sure we have room in the space to handle
8200          * all of the extents from this block group.  If we can, we're good
8201          */
8202         if ((space_info->total_bytes != block_group->key.offset) &&
8203             (space_info->bytes_used + space_info->bytes_reserved +
8204              space_info->bytes_pinned + space_info->bytes_readonly +
8205              min_free < space_info->total_bytes)) {
8206                 spin_unlock(&space_info->lock);
8207                 goto out;
8208         }
8209         spin_unlock(&space_info->lock);
8210
8211         /*
8212          * ok we don't have enough space, but maybe we have free space on our
8213          * devices to allocate new chunks for relocation, so loop through our
8214          * alloc devices and guess if we have enough space.  if this block
8215          * group is going to be restriped, run checks against the target
8216          * profile instead of the current one.
8217          */
8218         ret = -1;
8219
8220         /*
8221          * index:
8222          *      0: raid10
8223          *      1: raid1
8224          *      2: dup
8225          *      3: raid0
8226          *      4: single
8227          */
8228         target = get_restripe_target(root->fs_info, block_group->flags);
8229         if (target) {
8230                 index = __get_raid_index(extended_to_chunk(target));
8231         } else {
8232                 /*
8233                  * this is just a balance, so if we were marked as full
8234                  * we know there is no space for a new chunk
8235                  */
8236                 if (full)
8237                         goto out;
8238
8239                 index = get_block_group_index(block_group);
8240         }
8241
8242         if (index == BTRFS_RAID_RAID10) {
8243                 dev_min = 4;
8244                 /* Divide by 2 */
8245                 min_free >>= 1;
8246         } else if (index == BTRFS_RAID_RAID1) {
8247                 dev_min = 2;
8248         } else if (index == BTRFS_RAID_DUP) {
8249                 /* Multiply by 2 */
8250                 min_free <<= 1;
8251         } else if (index == BTRFS_RAID_RAID0) {
8252                 dev_min = fs_devices->rw_devices;
8253                 do_div(min_free, dev_min);
8254         }
8255
8256         /* We need to do this so that we can look at pending chunks */
8257         trans = btrfs_join_transaction(root);
8258         if (IS_ERR(trans)) {
8259                 ret = PTR_ERR(trans);
8260                 goto out;
8261         }
8262
8263         mutex_lock(&root->fs_info->chunk_mutex);
8264         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8265                 u64 dev_offset;
8266
8267                 /*
8268                  * check to make sure we can actually find a chunk with enough
8269                  * space to fit our block group in.
8270                  */
8271                 if (device->total_bytes > device->bytes_used + min_free &&
8272                     !device->is_tgtdev_for_dev_replace) {
8273                         ret = find_free_dev_extent(trans, device, min_free,
8274                                                    &dev_offset, NULL);
8275                         if (!ret)
8276                                 dev_nr++;
8277
8278                         if (dev_nr >= dev_min)
8279                                 break;
8280
8281                         ret = -1;
8282                 }
8283         }
8284         mutex_unlock(&root->fs_info->chunk_mutex);
8285         btrfs_end_transaction(trans, root);
8286 out:
8287         btrfs_put_block_group(block_group);
8288         return ret;
8289 }
8290
8291 static int find_first_block_group(struct btrfs_root *root,
8292                 struct btrfs_path *path, struct btrfs_key *key)
8293 {
8294         int ret = 0;
8295         struct btrfs_key found_key;
8296         struct extent_buffer *leaf;
8297         int slot;
8298
8299         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8300         if (ret < 0)
8301                 goto out;
8302
8303         while (1) {
8304                 slot = path->slots[0];
8305                 leaf = path->nodes[0];
8306                 if (slot >= btrfs_header_nritems(leaf)) {
8307                         ret = btrfs_next_leaf(root, path);
8308                         if (ret == 0)
8309                                 continue;
8310                         if (ret < 0)
8311                                 goto out;
8312                         break;
8313                 }
8314                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8315
8316                 if (found_key.objectid >= key->objectid &&
8317                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8318                         ret = 0;
8319                         goto out;
8320                 }
8321                 path->slots[0]++;
8322         }
8323 out:
8324         return ret;
8325 }
8326
8327 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8328 {
8329         struct btrfs_block_group_cache *block_group;
8330         u64 last = 0;
8331
8332         while (1) {
8333                 struct inode *inode;
8334
8335                 block_group = btrfs_lookup_first_block_group(info, last);
8336                 while (block_group) {
8337                         spin_lock(&block_group->lock);
8338                         if (block_group->iref)
8339                                 break;
8340                         spin_unlock(&block_group->lock);
8341                         block_group = next_block_group(info->tree_root,
8342                                                        block_group);
8343                 }
8344                 if (!block_group) {
8345                         if (last == 0)
8346                                 break;
8347                         last = 0;
8348                         continue;
8349                 }
8350
8351                 inode = block_group->inode;
8352                 block_group->iref = 0;
8353                 block_group->inode = NULL;
8354                 spin_unlock(&block_group->lock);
8355                 iput(inode);
8356                 last = block_group->key.objectid + block_group->key.offset;
8357                 btrfs_put_block_group(block_group);
8358         }
8359 }
8360
8361 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8362 {
8363         struct btrfs_block_group_cache *block_group;
8364         struct btrfs_space_info *space_info;
8365         struct btrfs_caching_control *caching_ctl;
8366         struct rb_node *n;
8367
8368         down_write(&info->commit_root_sem);
8369         while (!list_empty(&info->caching_block_groups)) {
8370                 caching_ctl = list_entry(info->caching_block_groups.next,
8371                                          struct btrfs_caching_control, list);
8372                 list_del(&caching_ctl->list);
8373                 put_caching_control(caching_ctl);
8374         }
8375         up_write(&info->commit_root_sem);
8376
8377         spin_lock(&info->block_group_cache_lock);
8378         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8379                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8380                                        cache_node);
8381                 rb_erase(&block_group->cache_node,
8382                          &info->block_group_cache_tree);
8383                 spin_unlock(&info->block_group_cache_lock);
8384
8385                 down_write(&block_group->space_info->groups_sem);
8386                 list_del(&block_group->list);
8387                 up_write(&block_group->space_info->groups_sem);
8388
8389                 if (block_group->cached == BTRFS_CACHE_STARTED)
8390                         wait_block_group_cache_done(block_group);
8391
8392                 /*
8393                  * We haven't cached this block group, which means we could
8394                  * possibly have excluded extents on this block group.
8395                  */
8396                 if (block_group->cached == BTRFS_CACHE_NO ||
8397                     block_group->cached == BTRFS_CACHE_ERROR)
8398                         free_excluded_extents(info->extent_root, block_group);
8399
8400                 btrfs_remove_free_space_cache(block_group);
8401                 btrfs_put_block_group(block_group);
8402
8403                 spin_lock(&info->block_group_cache_lock);
8404         }
8405         spin_unlock(&info->block_group_cache_lock);
8406
8407         /* now that all the block groups are freed, go through and
8408          * free all the space_info structs.  This is only called during
8409          * the final stages of unmount, and so we know nobody is
8410          * using them.  We call synchronize_rcu() once before we start,
8411          * just to be on the safe side.
8412          */
8413         synchronize_rcu();
8414
8415         release_global_block_rsv(info);
8416
8417         while (!list_empty(&info->space_info)) {
8418                 int i;
8419
8420                 space_info = list_entry(info->space_info.next,
8421                                         struct btrfs_space_info,
8422                                         list);
8423                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8424                         if (WARN_ON(space_info->bytes_pinned > 0 ||
8425                             space_info->bytes_reserved > 0 ||
8426                             space_info->bytes_may_use > 0)) {
8427                                 dump_space_info(space_info, 0, 0);
8428                         }
8429                 }
8430                 list_del(&space_info->list);
8431                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
8432                         struct kobject *kobj;
8433                         kobj = &space_info->block_group_kobjs[i];
8434                         if (kobj->parent) {
8435                                 kobject_del(kobj);
8436                                 kobject_put(kobj);
8437                         }
8438                 }
8439                 kobject_del(&space_info->kobj);
8440                 kobject_put(&space_info->kobj);
8441         }
8442         return 0;
8443 }
8444
8445 static void __link_block_group(struct btrfs_space_info *space_info,
8446                                struct btrfs_block_group_cache *cache)
8447 {
8448         int index = get_block_group_index(cache);
8449         bool first = false;
8450
8451         down_write(&space_info->groups_sem);
8452         if (list_empty(&space_info->block_groups[index]))
8453                 first = true;
8454         list_add_tail(&cache->list, &space_info->block_groups[index]);
8455         up_write(&space_info->groups_sem);
8456
8457         if (first) {
8458                 struct kobject *kobj = &space_info->block_group_kobjs[index];
8459                 int ret;
8460
8461                 kobject_get(&space_info->kobj); /* put in release */
8462                 ret = kobject_add(kobj, &space_info->kobj, "%s",
8463                                   get_raid_name(index));
8464                 if (ret) {
8465                         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
8466                         kobject_put(&space_info->kobj);
8467                 }
8468         }
8469 }
8470
8471 static struct btrfs_block_group_cache *
8472 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
8473 {
8474         struct btrfs_block_group_cache *cache;
8475
8476         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8477         if (!cache)
8478                 return NULL;
8479
8480         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8481                                         GFP_NOFS);
8482         if (!cache->free_space_ctl) {
8483                 kfree(cache);
8484                 return NULL;
8485         }
8486
8487         cache->key.objectid = start;
8488         cache->key.offset = size;
8489         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8490
8491         cache->sectorsize = root->sectorsize;
8492         cache->fs_info = root->fs_info;
8493         cache->full_stripe_len = btrfs_full_stripe_len(root,
8494                                                &root->fs_info->mapping_tree,
8495                                                start);
8496         atomic_set(&cache->count, 1);
8497         spin_lock_init(&cache->lock);
8498         INIT_LIST_HEAD(&cache->list);
8499         INIT_LIST_HEAD(&cache->cluster_list);
8500         INIT_LIST_HEAD(&cache->new_bg_list);
8501         btrfs_init_free_space_ctl(cache);
8502
8503         return cache;
8504 }
8505
8506 int btrfs_read_block_groups(struct btrfs_root *root)
8507 {
8508         struct btrfs_path *path;
8509         int ret;
8510         struct btrfs_block_group_cache *cache;
8511         struct btrfs_fs_info *info = root->fs_info;
8512         struct btrfs_space_info *space_info;
8513         struct btrfs_key key;
8514         struct btrfs_key found_key;
8515         struct extent_buffer *leaf;
8516         int need_clear = 0;
8517         u64 cache_gen;
8518
8519         root = info->extent_root;
8520         key.objectid = 0;
8521         key.offset = 0;
8522         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8523         path = btrfs_alloc_path();
8524         if (!path)
8525                 return -ENOMEM;
8526         path->reada = 1;
8527
8528         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
8529         if (btrfs_test_opt(root, SPACE_CACHE) &&
8530             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
8531                 need_clear = 1;
8532         if (btrfs_test_opt(root, CLEAR_CACHE))
8533                 need_clear = 1;
8534
8535         while (1) {
8536                 ret = find_first_block_group(root, path, &key);
8537                 if (ret > 0)
8538                         break;
8539                 if (ret != 0)
8540                         goto error;
8541
8542                 leaf = path->nodes[0];
8543                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8544
8545                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
8546                                                        found_key.offset);
8547                 if (!cache) {
8548                         ret = -ENOMEM;
8549                         goto error;
8550                 }
8551
8552                 if (need_clear) {
8553                         /*
8554                          * When we mount with old space cache, we need to
8555                          * set BTRFS_DC_CLEAR and set dirty flag.
8556                          *
8557                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8558                          *    truncate the old free space cache inode and
8559                          *    setup a new one.
8560                          * b) Setting 'dirty flag' makes sure that we flush
8561                          *    the new space cache info onto disk.
8562                          */
8563                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8564                         if (btrfs_test_opt(root, SPACE_CACHE))
8565                                 cache->dirty = 1;
8566                 }
8567
8568                 read_extent_buffer(leaf, &cache->item,
8569                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8570                                    sizeof(cache->item));
8571                 cache->flags = btrfs_block_group_flags(&cache->item);
8572
8573                 key.objectid = found_key.objectid + found_key.offset;
8574                 btrfs_release_path(path);
8575
8576                 /*
8577                  * We need to exclude the super stripes now so that the space
8578                  * info has super bytes accounted for, otherwise we'll think
8579                  * we have more space than we actually do.
8580                  */
8581                 ret = exclude_super_stripes(root, cache);
8582                 if (ret) {
8583                         /*
8584                          * We may have excluded something, so call this just in
8585                          * case.
8586                          */
8587                         free_excluded_extents(root, cache);
8588                         btrfs_put_block_group(cache);
8589                         goto error;
8590                 }
8591
8592                 /*
8593                  * check for two cases, either we are full, and therefore
8594                  * don't need to bother with the caching work since we won't
8595                  * find any space, or we are empty, and we can just add all
8596                  * the space in and be done with it.  This saves us _alot_ of
8597                  * time, particularly in the full case.
8598                  */
8599                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8600                         cache->last_byte_to_unpin = (u64)-1;
8601                         cache->cached = BTRFS_CACHE_FINISHED;
8602                         free_excluded_extents(root, cache);
8603                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8604                         cache->last_byte_to_unpin = (u64)-1;
8605                         cache->cached = BTRFS_CACHE_FINISHED;
8606                         add_new_free_space(cache, root->fs_info,
8607                                            found_key.objectid,
8608                                            found_key.objectid +
8609                                            found_key.offset);
8610                         free_excluded_extents(root, cache);
8611                 }
8612
8613                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8614                 if (ret) {
8615                         btrfs_remove_free_space_cache(cache);
8616                         btrfs_put_block_group(cache);
8617                         goto error;
8618                 }
8619
8620                 ret = update_space_info(info, cache->flags, found_key.offset,
8621                                         btrfs_block_group_used(&cache->item),
8622                                         &space_info);
8623                 if (ret) {
8624                         btrfs_remove_free_space_cache(cache);
8625                         spin_lock(&info->block_group_cache_lock);
8626                         rb_erase(&cache->cache_node,
8627                                  &info->block_group_cache_tree);
8628                         spin_unlock(&info->block_group_cache_lock);
8629                         btrfs_put_block_group(cache);
8630                         goto error;
8631                 }
8632
8633                 cache->space_info = space_info;
8634                 spin_lock(&cache->space_info->lock);
8635                 cache->space_info->bytes_readonly += cache->bytes_super;
8636                 spin_unlock(&cache->space_info->lock);
8637
8638                 __link_block_group(space_info, cache);
8639
8640                 set_avail_alloc_bits(root->fs_info, cache->flags);
8641                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8642                         set_block_group_ro(cache, 1);
8643         }
8644
8645         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8646                 if (!(get_alloc_profile(root, space_info->flags) &
8647                       (BTRFS_BLOCK_GROUP_RAID10 |
8648                        BTRFS_BLOCK_GROUP_RAID1 |
8649                        BTRFS_BLOCK_GROUP_RAID5 |
8650                        BTRFS_BLOCK_GROUP_RAID6 |
8651                        BTRFS_BLOCK_GROUP_DUP)))
8652                         continue;
8653                 /*
8654                  * avoid allocating from un-mirrored block group if there are
8655                  * mirrored block groups.
8656                  */
8657                 list_for_each_entry(cache,
8658                                 &space_info->block_groups[BTRFS_RAID_RAID0],
8659                                 list)
8660                         set_block_group_ro(cache, 1);
8661                 list_for_each_entry(cache,
8662                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
8663                                 list)
8664                         set_block_group_ro(cache, 1);
8665         }
8666
8667         init_global_block_rsv(info);
8668         ret = 0;
8669 error:
8670         btrfs_free_path(path);
8671         return ret;
8672 }
8673
8674 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8675                                        struct btrfs_root *root)
8676 {
8677         struct btrfs_block_group_cache *block_group, *tmp;
8678         struct btrfs_root *extent_root = root->fs_info->extent_root;
8679         struct btrfs_block_group_item item;
8680         struct btrfs_key key;
8681         int ret = 0;
8682
8683         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8684                                  new_bg_list) {
8685                 list_del_init(&block_group->new_bg_list);
8686
8687                 if (ret)
8688                         continue;
8689
8690                 spin_lock(&block_group->lock);
8691                 memcpy(&item, &block_group->item, sizeof(item));
8692                 memcpy(&key, &block_group->key, sizeof(key));
8693                 spin_unlock(&block_group->lock);
8694
8695                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8696                                         sizeof(item));
8697                 if (ret)
8698                         btrfs_abort_transaction(trans, extent_root, ret);
8699                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
8700                                                key.objectid, key.offset);
8701                 if (ret)
8702                         btrfs_abort_transaction(trans, extent_root, ret);
8703         }
8704 }
8705
8706 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8707                            struct btrfs_root *root, u64 bytes_used,
8708                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8709                            u64 size)
8710 {
8711         int ret;
8712         struct btrfs_root *extent_root;
8713         struct btrfs_block_group_cache *cache;
8714
8715         extent_root = root->fs_info->extent_root;
8716
8717         root->fs_info->last_trans_log_full_commit = trans->transid;
8718
8719         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
8720         if (!cache)
8721                 return -ENOMEM;
8722
8723         btrfs_set_block_group_used(&cache->item, bytes_used);
8724         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8725         btrfs_set_block_group_flags(&cache->item, type);
8726
8727         cache->flags = type;
8728         cache->last_byte_to_unpin = (u64)-1;
8729         cache->cached = BTRFS_CACHE_FINISHED;
8730         ret = exclude_super_stripes(root, cache);
8731         if (ret) {
8732                 /*
8733                  * We may have excluded something, so call this just in
8734                  * case.
8735                  */
8736                 free_excluded_extents(root, cache);
8737                 btrfs_put_block_group(cache);
8738                 return ret;
8739         }
8740
8741         add_new_free_space(cache, root->fs_info, chunk_offset,
8742                            chunk_offset + size);
8743
8744         free_excluded_extents(root, cache);
8745
8746         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8747         if (ret) {
8748                 btrfs_remove_free_space_cache(cache);
8749                 btrfs_put_block_group(cache);
8750                 return ret;
8751         }
8752
8753         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8754                                 &cache->space_info);
8755         if (ret) {
8756                 btrfs_remove_free_space_cache(cache);
8757                 spin_lock(&root->fs_info->block_group_cache_lock);
8758                 rb_erase(&cache->cache_node,
8759                          &root->fs_info->block_group_cache_tree);
8760                 spin_unlock(&root->fs_info->block_group_cache_lock);
8761                 btrfs_put_block_group(cache);
8762                 return ret;
8763         }
8764         update_global_block_rsv(root->fs_info);
8765
8766         spin_lock(&cache->space_info->lock);
8767         cache->space_info->bytes_readonly += cache->bytes_super;
8768         spin_unlock(&cache->space_info->lock);
8769
8770         __link_block_group(cache->space_info, cache);
8771
8772         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8773
8774         set_avail_alloc_bits(extent_root->fs_info, type);
8775
8776         return 0;
8777 }
8778
8779 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8780 {
8781         u64 extra_flags = chunk_to_extended(flags) &
8782                                 BTRFS_EXTENDED_PROFILE_MASK;
8783
8784         write_seqlock(&fs_info->profiles_lock);
8785         if (flags & BTRFS_BLOCK_GROUP_DATA)
8786                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8787         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8788                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8789         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8790                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8791         write_sequnlock(&fs_info->profiles_lock);
8792 }
8793
8794 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8795                              struct btrfs_root *root, u64 group_start)
8796 {
8797         struct btrfs_path *path;
8798         struct btrfs_block_group_cache *block_group;
8799         struct btrfs_free_cluster *cluster;
8800         struct btrfs_root *tree_root = root->fs_info->tree_root;
8801         struct btrfs_key key;
8802         struct inode *inode;
8803         int ret;
8804         int index;
8805         int factor;
8806
8807         root = root->fs_info->extent_root;
8808
8809         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8810         BUG_ON(!block_group);
8811         BUG_ON(!block_group->ro);
8812
8813         /*
8814          * Free the reserved super bytes from this block group before
8815          * remove it.
8816          */
8817         free_excluded_extents(root, block_group);
8818
8819         memcpy(&key, &block_group->key, sizeof(key));
8820         index = get_block_group_index(block_group);
8821         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8822                                   BTRFS_BLOCK_GROUP_RAID1 |
8823                                   BTRFS_BLOCK_GROUP_RAID10))
8824                 factor = 2;
8825         else
8826                 factor = 1;
8827
8828         /* make sure this block group isn't part of an allocation cluster */
8829         cluster = &root->fs_info->data_alloc_cluster;
8830         spin_lock(&cluster->refill_lock);
8831         btrfs_return_cluster_to_free_space(block_group, cluster);
8832         spin_unlock(&cluster->refill_lock);
8833
8834         /*
8835          * make sure this block group isn't part of a metadata
8836          * allocation cluster
8837          */
8838         cluster = &root->fs_info->meta_alloc_cluster;
8839         spin_lock(&cluster->refill_lock);
8840         btrfs_return_cluster_to_free_space(block_group, cluster);
8841         spin_unlock(&cluster->refill_lock);
8842
8843         path = btrfs_alloc_path();
8844         if (!path) {
8845                 ret = -ENOMEM;
8846                 goto out;
8847         }
8848
8849         inode = lookup_free_space_inode(tree_root, block_group, path);
8850         if (!IS_ERR(inode)) {
8851                 ret = btrfs_orphan_add(trans, inode);
8852                 if (ret) {
8853                         btrfs_add_delayed_iput(inode);
8854                         goto out;
8855                 }
8856                 clear_nlink(inode);
8857                 /* One for the block groups ref */
8858                 spin_lock(&block_group->lock);
8859                 if (block_group->iref) {
8860                         block_group->iref = 0;
8861                         block_group->inode = NULL;
8862                         spin_unlock(&block_group->lock);
8863                         iput(inode);
8864                 } else {
8865                         spin_unlock(&block_group->lock);
8866                 }
8867                 /* One for our lookup ref */
8868                 btrfs_add_delayed_iput(inode);
8869         }
8870
8871         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8872         key.offset = block_group->key.objectid;
8873         key.type = 0;
8874
8875         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8876         if (ret < 0)
8877                 goto out;
8878         if (ret > 0)
8879                 btrfs_release_path(path);
8880         if (ret == 0) {
8881                 ret = btrfs_del_item(trans, tree_root, path);
8882                 if (ret)
8883                         goto out;
8884                 btrfs_release_path(path);
8885         }
8886
8887         spin_lock(&root->fs_info->block_group_cache_lock);
8888         rb_erase(&block_group->cache_node,
8889                  &root->fs_info->block_group_cache_tree);
8890
8891         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8892                 root->fs_info->first_logical_byte = (u64)-1;
8893         spin_unlock(&root->fs_info->block_group_cache_lock);
8894
8895         down_write(&block_group->space_info->groups_sem);
8896         /*
8897          * we must use list_del_init so people can check to see if they
8898          * are still on the list after taking the semaphore
8899          */
8900         list_del_init(&block_group->list);
8901         if (list_empty(&block_group->space_info->block_groups[index])) {
8902                 kobject_del(&block_group->space_info->block_group_kobjs[index]);
8903                 kobject_put(&block_group->space_info->block_group_kobjs[index]);
8904                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8905         }
8906         up_write(&block_group->space_info->groups_sem);
8907
8908         if (block_group->cached == BTRFS_CACHE_STARTED)
8909                 wait_block_group_cache_done(block_group);
8910
8911         btrfs_remove_free_space_cache(block_group);
8912
8913         spin_lock(&block_group->space_info->lock);
8914         block_group->space_info->total_bytes -= block_group->key.offset;
8915         block_group->space_info->bytes_readonly -= block_group->key.offset;
8916         block_group->space_info->disk_total -= block_group->key.offset * factor;
8917         spin_unlock(&block_group->space_info->lock);
8918
8919         memcpy(&key, &block_group->key, sizeof(key));
8920
8921         btrfs_clear_space_info_full(root->fs_info);
8922
8923         btrfs_put_block_group(block_group);
8924         btrfs_put_block_group(block_group);
8925
8926         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8927         if (ret > 0)
8928                 ret = -EIO;
8929         if (ret < 0)
8930                 goto out;
8931
8932         ret = btrfs_del_item(trans, root, path);
8933 out:
8934         btrfs_free_path(path);
8935         return ret;
8936 }
8937
8938 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8939 {
8940         struct btrfs_space_info *space_info;
8941         struct btrfs_super_block *disk_super;
8942         u64 features;
8943         u64 flags;
8944         int mixed = 0;
8945         int ret;
8946
8947         disk_super = fs_info->super_copy;
8948         if (!btrfs_super_root(disk_super))
8949                 return 1;
8950
8951         features = btrfs_super_incompat_flags(disk_super);
8952         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8953                 mixed = 1;
8954
8955         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8956         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8957         if (ret)
8958                 goto out;
8959
8960         if (mixed) {
8961                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8962                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8963         } else {
8964                 flags = BTRFS_BLOCK_GROUP_METADATA;
8965                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8966                 if (ret)
8967                         goto out;
8968
8969                 flags = BTRFS_BLOCK_GROUP_DATA;
8970                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8971         }
8972 out:
8973         return ret;
8974 }
8975
8976 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8977 {
8978         return unpin_extent_range(root, start, end);
8979 }
8980
8981 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8982                                u64 num_bytes, u64 *actual_bytes)
8983 {
8984         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8985 }
8986
8987 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8988 {
8989         struct btrfs_fs_info *fs_info = root->fs_info;
8990         struct btrfs_block_group_cache *cache = NULL;
8991         u64 group_trimmed;
8992         u64 start;
8993         u64 end;
8994         u64 trimmed = 0;
8995         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8996         int ret = 0;
8997
8998         /*
8999          * try to trim all FS space, our block group may start from non-zero.
9000          */
9001         if (range->len == total_bytes)
9002                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
9003         else
9004                 cache = btrfs_lookup_block_group(fs_info, range->start);
9005
9006         while (cache) {
9007                 if (cache->key.objectid >= (range->start + range->len)) {
9008                         btrfs_put_block_group(cache);
9009                         break;
9010                 }
9011
9012                 start = max(range->start, cache->key.objectid);
9013                 end = min(range->start + range->len,
9014                                 cache->key.objectid + cache->key.offset);
9015
9016                 if (end - start >= range->minlen) {
9017                         if (!block_group_cache_done(cache)) {
9018                                 ret = cache_block_group(cache, 0);
9019                                 if (ret) {
9020                                         btrfs_put_block_group(cache);
9021                                         break;
9022                                 }
9023                                 ret = wait_block_group_cache_done(cache);
9024                                 if (ret) {
9025                                         btrfs_put_block_group(cache);
9026                                         break;
9027                                 }
9028                         }
9029                         ret = btrfs_trim_block_group(cache,
9030                                                      &group_trimmed,
9031                                                      start,
9032                                                      end,
9033                                                      range->minlen);
9034
9035                         trimmed += group_trimmed;
9036                         if (ret) {
9037                                 btrfs_put_block_group(cache);
9038                                 break;
9039                         }
9040                 }
9041
9042                 cache = next_block_group(fs_info->tree_root, cache);
9043         }
9044
9045         range->len = trimmed;
9046         return ret;
9047 }
9048
9049 /*
9050  * btrfs_{start,end}_write() is similar to mnt_{want, drop}_write(),
9051  * they are used to prevent the some tasks writing data into the page cache
9052  * by nocow before the subvolume is snapshoted, but flush the data into
9053  * the disk after the snapshot creation.
9054  */
9055 void btrfs_end_nocow_write(struct btrfs_root *root)
9056 {
9057         percpu_counter_dec(&root->subv_writers->counter);
9058         /*
9059          * Make sure counter is updated before we wake up
9060          * waiters.
9061          */
9062         smp_mb();
9063         if (waitqueue_active(&root->subv_writers->wait))
9064                 wake_up(&root->subv_writers->wait);
9065 }
9066
9067 int btrfs_start_nocow_write(struct btrfs_root *root)
9068 {
9069         if (unlikely(atomic_read(&root->will_be_snapshoted)))
9070                 return 0;
9071
9072         percpu_counter_inc(&root->subv_writers->counter);
9073         /*
9074          * Make sure counter is updated before we check for snapshot creation.
9075          */
9076         smp_mb();
9077         if (unlikely(atomic_read(&root->will_be_snapshoted))) {
9078                 btrfs_end_nocow_write(root);
9079                 return 0;
9080         }
9081         return 1;
9082 }