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 "compat.h"
  29 #include "hash.h"
  30 #include "ctree.h"
  31 #include "disk-io.h"
  32 #include "print-tree.h"
  33 #include "transaction.h"
  34 #include "volumes.h"
  35 #include "raid56.h"
  36 #include "locking.h"
  37 #include "free-space-cache.h"
  38 #include "math.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->extent_commit_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                                 caching_ctl->progress = last;
 447                                 btrfs_release_path(path);
 448                                 up_read(&fs_info->extent_commit_sem);
 449                                 mutex_unlock(&caching_ctl->mutex);
 450                                 cond_resched();
 451                                 goto again;
 452                         }
 453
 454                         ret = btrfs_next_leaf(extent_root, path);
 455                         if (ret < 0)
 456                                 goto err;
 457                         if (ret)
 458                                 break;
 459                         leaf = path->nodes[0];
 460                         nritems = btrfs_header_nritems(leaf);
 461                         continue;
 462                 }
 463
 464                 if (key.objectid < last) {
 465                         key.objectid = last;
 466                         key.offset = 0;
 467                         key.type = BTRFS_EXTENT_ITEM_KEY;
 468
 469                         caching_ctl->progress = last;
 470                         btrfs_release_path(path);
 471                         goto next;
 472                 }
 473
 474                 if (key.objectid < block_group->key.objectid) {
 475                         path->slots[0]++;
 476                         continue;
 477                 }
 478
 479                 if (key.objectid >= block_group->key.objectid +
 480                     block_group->key.offset)
 481                         break;
 482
 483                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 484                     key.type == BTRFS_METADATA_ITEM_KEY) {
 485                         total_found += add_new_free_space(block_group,
 486                                                           fs_info, last,
 487                                                           key.objectid);
 488                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 489                                 last = key.objectid +
 490                                         fs_info->tree_root->leafsize;
 491                         else
 492                                 last = key.objectid + key.offset;
 493
 494                         if (total_found > (1024 * 1024 * 2)) {
 495                                 total_found = 0;
 496                                 wake_up(&caching_ctl->wait);
 497                         }
 498                 }
 499                 path->slots[0]++;
 500         }
 501         ret = 0;
 502
 503         total_found += add_new_free_space(block_group, fs_info, last,
 504                                           block_group->key.objectid +
 505                                           block_group->key.offset);
 506         caching_ctl->progress = (u64)-1;
 507
 508         spin_lock(&block_group->lock);
 509         block_group->caching_ctl = NULL;
 510         block_group->cached = BTRFS_CACHE_FINISHED;
 511         spin_unlock(&block_group->lock);
 512
 513 err:
 514         btrfs_free_path(path);
 515         up_read(&fs_info->extent_commit_sem);
 516
 517         free_excluded_extents(extent_root, block_group);
 518
 519         mutex_unlock(&caching_ctl->mutex);
 520 out:
 521         if (ret) {
 522                 spin_lock(&block_group->lock);
 523                 block_group->caching_ctl = NULL;
 524                 block_group->cached = BTRFS_CACHE_ERROR;
 525                 spin_unlock(&block_group->lock);
 526         }
 527         wake_up(&caching_ctl->wait);
 528
 529         put_caching_control(caching_ctl);
 530         btrfs_put_block_group(block_group);
 531 }
 532
 533 static int cache_block_group(struct btrfs_block_group_cache *cache,
 534                              int load_cache_only)
 535 {
 536         DEFINE_WAIT(wait);
 537         struct btrfs_fs_info *fs_info = cache->fs_info;
 538         struct btrfs_caching_control *caching_ctl;
 539         int ret = 0;
 540
 541         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 542         if (!caching_ctl)
 543                 return -ENOMEM;
 544
 545         INIT_LIST_HEAD(&caching_ctl->list);
 546         mutex_init(&caching_ctl->mutex);
 547         init_waitqueue_head(&caching_ctl->wait);
 548         caching_ctl->block_group = cache;
 549         caching_ctl->progress = cache->key.objectid;
 550         atomic_set(&caching_ctl->count, 1);
 551         caching_ctl->work.func = caching_thread;
 552
 553         spin_lock(&cache->lock);
 554         /*
 555          * This should be a rare occasion, but this could happen I think in the
 556          * case where one thread starts to load the space cache info, and then
 557          * some other thread starts a transaction commit which tries to do an
 558          * allocation while the other thread is still loading the space cache
 559          * info.  The previous loop should have kept us from choosing this block
 560          * group, but if we've moved to the state where we will wait on caching
 561          * block groups we need to first check if we're doing a fast load here,
 562          * so we can wait for it to finish, otherwise we could end up allocating
 563          * from a block group who's cache gets evicted for one reason or
 564          * another.
 565          */
 566         while (cache->cached == BTRFS_CACHE_FAST) {
 567                 struct btrfs_caching_control *ctl;
 568
 569                 ctl = cache->caching_ctl;
 570                 atomic_inc(&ctl->count);
 571                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 572                 spin_unlock(&cache->lock);
 573
 574                 schedule();
 575
 576                 finish_wait(&ctl->wait, &wait);
 577                 put_caching_control(ctl);
 578                 spin_lock(&cache->lock);
 579         }
 580
 581         if (cache->cached != BTRFS_CACHE_NO) {
 582                 spin_unlock(&cache->lock);
 583                 kfree(caching_ctl);
 584                 return 0;
 585         }
 586         WARN_ON(cache->caching_ctl);
 587         cache->caching_ctl = caching_ctl;
 588         cache->cached = BTRFS_CACHE_FAST;
 589         spin_unlock(&cache->lock);
 590
 591         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 592                 ret = load_free_space_cache(fs_info, cache);
 593
 594                 spin_lock(&cache->lock);
 595                 if (ret == 1) {
 596                         cache->caching_ctl = NULL;
 597                         cache->cached = BTRFS_CACHE_FINISHED;
 598                         cache->last_byte_to_unpin = (u64)-1;
 599                 } else {
 600                         if (load_cache_only) {
 601                                 cache->caching_ctl = NULL;
 602                                 cache->cached = BTRFS_CACHE_NO;
 603                         } else {
 604                                 cache->cached = BTRFS_CACHE_STARTED;
 605                         }
 606                 }
 607                 spin_unlock(&cache->lock);
 608                 wake_up(&caching_ctl->wait);
 609                 if (ret == 1) {
 610                         put_caching_control(caching_ctl);
 611                         free_excluded_extents(fs_info->extent_root, cache);
 612                         return 0;
 613                 }
 614         } else {
 615                 /*
 616                  * We are not going to do the fast caching, set cached to the
 617                  * appropriate value and wakeup any waiters.
 618                  */
 619                 spin_lock(&cache->lock);
 620                 if (load_cache_only) {
 621                         cache->caching_ctl = NULL;
 622                         cache->cached = BTRFS_CACHE_NO;
 623                 } else {
 624                         cache->cached = BTRFS_CACHE_STARTED;
 625                 }
 626                 spin_unlock(&cache->lock);
 627                 wake_up(&caching_ctl->wait);
 628         }
 629
 630         if (load_cache_only) {
 631                 put_caching_control(caching_ctl);
 632                 return 0;
 633         }
 634
 635         down_write(&fs_info->extent_commit_sem);
 636         atomic_inc(&caching_ctl->count);
 637         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 638         up_write(&fs_info->extent_commit_sem);
 639
 640         btrfs_get_block_group(cache);
 641
 642         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 643
 644         return ret;
 645 }
 646
 647 /*
 648  * return the block group that starts at or after bytenr
 649  */
 650 static struct btrfs_block_group_cache *
 651 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 652 {
 653         struct btrfs_block_group_cache *cache;
 654
 655         cache = block_group_cache_tree_search(info, bytenr, 0);
 656
 657         return cache;
 658 }
 659
 660 /*
 661  * return the block group that contains the given bytenr
 662  */
 663 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 664                                                  struct btrfs_fs_info *info,
 665                                                  u64 bytenr)
 666 {
 667         struct btrfs_block_group_cache *cache;
 668
 669         cache = block_group_cache_tree_search(info, bytenr, 1);
 670
 671         return cache;
 672 }
 673
 674 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 675                                                   u64 flags)
 676 {
 677         struct list_head *head = &info->space_info;
 678         struct btrfs_space_info *found;
 679
 680         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 681
 682         rcu_read_lock();
 683         list_for_each_entry_rcu(found, head, list) {
 684                 if (found->flags & flags) {
 685                         rcu_read_unlock();
 686                         return found;
 687                 }
 688         }
 689         rcu_read_unlock();
 690         return NULL;
 691 }
 692
 693 /*
 694  * after adding space to the filesystem, we need to clear the full flags
 695  * on all the space infos.
 696  */
 697 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 698 {
 699         struct list_head *head = &info->space_info;
 700         struct btrfs_space_info *found;
 701
 702         rcu_read_lock();
 703         list_for_each_entry_rcu(found, head, list)
 704                 found->full = 0;
 705         rcu_read_unlock();
 706 }
 707
 708 /* simple helper to search for an existing extent at a given offset */
 709 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 710 {
 711         int ret;
 712         struct btrfs_key key;
 713         struct btrfs_path *path;
 714
 715         path = btrfs_alloc_path();
 716         if (!path)
 717                 return -ENOMEM;
 718
 719         key.objectid = start;
 720         key.offset = len;
 721         key.type = BTRFS_EXTENT_ITEM_KEY;
 722         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 723                                 0, 0);
 724         if (ret > 0) {
 725                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 726                 if (key.objectid == start &&
 727                     key.type == BTRFS_METADATA_ITEM_KEY)
 728                         ret = 0;
 729         }
 730         btrfs_free_path(path);
 731         return ret;
 732 }
 733
 734 /*
 735  * helper function to lookup reference count and flags of a tree block.
 736  *
 737  * the head node for delayed ref is used to store the sum of all the
 738  * reference count modifications queued up in the rbtree. the head
 739  * node may also store the extent flags to set. This way you can check
 740  * to see what the reference count and extent flags would be if all of
 741  * the delayed refs are not processed.
 742  */
 743 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 744                              struct btrfs_root *root, u64 bytenr,
 745                              u64 offset, int metadata, u64 *refs, u64 *flags)
 746 {
 747         struct btrfs_delayed_ref_head *head;
 748         struct btrfs_delayed_ref_root *delayed_refs;
 749         struct btrfs_path *path;
 750         struct btrfs_extent_item *ei;
 751         struct extent_buffer *leaf;
 752         struct btrfs_key key;
 753         u32 item_size;
 754         u64 num_refs;
 755         u64 extent_flags;
 756         int ret;
 757
 758         /*
 759          * If we don't have skinny metadata, don't bother doing anything
 760          * different
 761          */
 762         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 763                 offset = root->leafsize;
 764                 metadata = 0;
 765         }
 766
 767         path = btrfs_alloc_path();
 768         if (!path)
 769                 return -ENOMEM;
 770
 771         if (metadata) {
 772                 key.objectid = bytenr;
 773                 key.type = BTRFS_METADATA_ITEM_KEY;
 774                 key.offset = offset;
 775         } else {
 776                 key.objectid = bytenr;
 777                 key.type = BTRFS_EXTENT_ITEM_KEY;
 778                 key.offset = offset;
 779         }
 780
 781         if (!trans) {
 782                 path->skip_locking = 1;
 783                 path->search_commit_root = 1;
 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                 metadata = 0;
 793                 if (path->slots[0]) {
 794                         path->slots[0]--;
 795                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 796                                               path->slots[0]);
 797                         if (key.objectid == bytenr &&
 798                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 799                             key.offset == root->leafsize)
 800                                 ret = 0;
 801                 }
 802                 if (ret) {
 803                         key.objectid = bytenr;
 804                         key.type = BTRFS_EXTENT_ITEM_KEY;
 805                         key.offset = root->leafsize;
 806                         btrfs_release_path(path);
 807                         goto again;
 808                 }
 809         }
 810
 811         if (ret == 0) {
 812                 leaf = path->nodes[0];
 813                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 814                 if (item_size >= sizeof(*ei)) {
 815                         ei = btrfs_item_ptr(leaf, path->slots[0],
 816                                             struct btrfs_extent_item);
 817                         num_refs = btrfs_extent_refs(leaf, ei);
 818                         extent_flags = btrfs_extent_flags(leaf, ei);
 819                 } else {
 820 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 821                         struct btrfs_extent_item_v0 *ei0;
 822                         BUG_ON(item_size != sizeof(*ei0));
 823                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 824                                              struct btrfs_extent_item_v0);
 825                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 826                         /* FIXME: this isn't correct for data */
 827                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 828 #else
 829                         BUG();
 830 #endif
 831                 }
 832                 BUG_ON(num_refs == 0);
 833         } else {
 834                 num_refs = 0;
 835                 extent_flags = 0;
 836                 ret = 0;
 837         }
 838
 839         if (!trans)
 840                 goto out;
 841
 842         delayed_refs = &trans->transaction->delayed_refs;
 843         spin_lock(&delayed_refs->lock);
 844         head = btrfs_find_delayed_ref_head(trans, bytenr);
 845         if (head) {
 846                 if (!mutex_trylock(&head->mutex)) {
 847                         atomic_inc(&head->node.refs);
 848                         spin_unlock(&delayed_refs->lock);
 849
 850                         btrfs_release_path(path);
 851
 852                         /*
 853                          * Mutex was contended, block until it's released and try
 854                          * again
 855                          */
 856                         mutex_lock(&head->mutex);
 857                         mutex_unlock(&head->mutex);
 858                         btrfs_put_delayed_ref(&head->node);
 859                         goto again;
 860                 }
 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                 mutex_unlock(&head->mutex);
 868         }
 869         spin_unlock(&delayed_refs->lock);
 870 out:
 871         WARN_ON(num_refs == 0);
 872         if (refs)
 873                 *refs = num_refs;
 874         if (flags)
 875                 *flags = extent_flags;
 876 out_free:
 877         btrfs_free_path(path);
 878         return ret;
 879 }
 880
 881 /*
 882  * Back reference rules.  Back refs have three main goals:
 883  *
 884  * 1) differentiate between all holders of references to an extent so that
 885  *    when a reference is dropped we can make sure it was a valid reference
 886  *    before freeing the extent.
 887  *
 888  * 2) Provide enough information to quickly find the holders of an extent
 889  *    if we notice a given block is corrupted or bad.
 890  *
 891  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 892  *    maintenance.  This is actually the same as #2, but with a slightly
 893  *    different use case.
 894  *
 895  * There are two kinds of back refs. The implicit back refs is optimized
 896  * for pointers in non-shared tree blocks. For a given pointer in a block,
 897  * back refs of this kind provide information about the block's owner tree
 898  * and the pointer's key. These information allow us to find the block by
 899  * b-tree searching. The full back refs is for pointers in tree blocks not
 900  * referenced by their owner trees. The location of tree block is recorded
 901  * in the back refs. Actually the full back refs is generic, and can be
 902  * used in all cases the implicit back refs is used. The major shortcoming
 903  * of the full back refs is its overhead. Every time a tree block gets
 904  * COWed, we have to update back refs entry for all pointers in it.
 905  *
 906  * For a newly allocated tree block, we use implicit back refs for
 907  * pointers in it. This means most tree related operations only involve
 908  * implicit back refs. For a tree block created in old transaction, the
 909  * only way to drop a reference to it is COW it. So we can detect the
 910  * event that tree block loses its owner tree's reference and do the
 911  * back refs conversion.
 912  *
 913  * When a tree block is COW'd through a tree, there are four cases:
 914  *
 915  * The reference count of the block is one and the tree is the block's
 916  * owner tree. Nothing to do in this case.
 917  *
 918  * The reference count of the block is one and the tree is not the
 919  * block's owner tree. In this case, full back refs is used for pointers
 920  * in the block. Remove these full back refs, add implicit back refs for
 921  * every pointers in the new block.
 922  *
 923  * The reference count of the block is greater than one and the tree is
 924  * the block's owner tree. In this case, implicit back refs is used for
 925  * pointers in the block. Add full back refs for every pointers in the
 926  * block, increase lower level extents' reference counts. The original
 927  * implicit back refs are entailed to the new block.
 928  *
 929  * The reference count of the block is greater than one and the tree is
 930  * not the block's owner tree. Add implicit back refs for every pointer in
 931  * the new block, increase lower level extents' reference count.
 932  *
 933  * Back Reference Key composing:
 934  *
 935  * The key objectid corresponds to the first byte in the extent,
 936  * The key type is used to differentiate between types of back refs.
 937  * There are different meanings of the key offset for different types
 938  * of back refs.
 939  *
 940  * File extents can be referenced by:
 941  *
 942  * - multiple snapshots, subvolumes, or different generations in one subvol
 943  * - different files inside a single subvolume
 944  * - different offsets inside a file (bookend extents in file.c)
 945  *
 946  * The extent ref structure for the implicit back refs has fields for:
 947  *
 948  * - Objectid of the subvolume root
 949  * - objectid of the file holding the reference
 950  * - original offset in the file
 951  * - how many bookend extents
 952  *
 953  * The key offset for the implicit back refs is hash of the first
 954  * three fields.
 955  *
 956  * The extent ref structure for the full back refs has field for:
 957  *
 958  * - number of pointers in the tree leaf
 959  *
 960  * The key offset for the implicit back refs is the first byte of
 961  * the tree leaf
 962  *
 963  * When a file extent is allocated, The implicit back refs is used.
 964  * the fields are filled in:
 965  *
 966  *     (root_key.objectid, inode objectid, offset in file, 1)
 967  *
 968  * When a file extent is removed file truncation, we find the
 969  * corresponding implicit back refs and check the following fields:
 970  *
 971  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 972  *
 973  * Btree extents can be referenced by:
 974  *
 975  * - Different subvolumes
 976  *
 977  * Both the implicit back refs and the full back refs for tree blocks
 978  * only consist of key. The key offset for the implicit back refs is
 979  * objectid of block's owner tree. The key offset for the full back refs
 980  * is the first byte of parent block.
 981  *
 982  * When implicit back refs is used, information about the lowest key and
 983  * level of the tree block are required. These information are stored in
 984  * tree block info structure.
 985  */
 986
 987 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 988 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 989                                   struct btrfs_root *root,
 990                                   struct btrfs_path *path,
 991                                   u64 owner, u32 extra_size)
 992 {
 993         struct btrfs_extent_item *item;
 994         struct btrfs_extent_item_v0 *ei0;
 995         struct btrfs_extent_ref_v0 *ref0;
 996         struct btrfs_tree_block_info *bi;
 997         struct extent_buffer *leaf;
 998         struct btrfs_key key;
 999         struct btrfs_key found_key;
1000         u32 new_size = sizeof(*item);
1001         u64 refs;
1002         int ret;
1003
1004         leaf = path->nodes[0];
1005         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1006
1007         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1008         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1009                              struct btrfs_extent_item_v0);
1010         refs = btrfs_extent_refs_v0(leaf, ei0);
1011
1012         if (owner == (u64)-1) {
1013                 while (1) {
1014                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1015                                 ret = btrfs_next_leaf(root, path);
1016                                 if (ret < 0)
1017                                         return ret;
1018                                 BUG_ON(ret > 0); /* Corruption */
1019                                 leaf = path->nodes[0];
1020                         }
1021                         btrfs_item_key_to_cpu(leaf, &found_key,
1022                                               path->slots[0]);
1023                         BUG_ON(key.objectid != found_key.objectid);
1024                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1025                                 path->slots[0]++;
1026                                 continue;
1027                         }
1028                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1029                                               struct btrfs_extent_ref_v0);
1030                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1031                         break;
1032                 }
1033         }
1034         btrfs_release_path(path);
1035
1036         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1037                 new_size += sizeof(*bi);
1038
1039         new_size -= sizeof(*ei0);
1040         ret = btrfs_search_slot(trans, root, &key, path,
1041                                 new_size + extra_size, 1);
1042         if (ret < 0)
1043                 return ret;
1044         BUG_ON(ret); /* Corruption */
1045
1046         btrfs_extend_item(root, path, new_size);
1047
1048         leaf = path->nodes[0];
1049         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1050         btrfs_set_extent_refs(leaf, item, refs);
1051         /* FIXME: get real generation */
1052         btrfs_set_extent_generation(leaf, item, 0);
1053         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1054                 btrfs_set_extent_flags(leaf, item,
1055                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1056                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1057                 bi = (struct btrfs_tree_block_info *)(item + 1);
1058                 /* FIXME: get first key of the block */
1059                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1060                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1061         } else {
1062                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1063         }
1064         btrfs_mark_buffer_dirty(leaf);
1065         return 0;
1066 }
1067 #endif
1068
1069 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1070 {
1071         u32 high_crc = ~(u32)0;
1072         u32 low_crc = ~(u32)0;
1073         __le64 lenum;
1074
1075         lenum = cpu_to_le64(root_objectid);
1076         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1077         lenum = cpu_to_le64(owner);
1078         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1079         lenum = cpu_to_le64(offset);
1080         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1081
1082         return ((u64)high_crc << 31) ^ (u64)low_crc;
1083 }
1084
1085 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1086                                      struct btrfs_extent_data_ref *ref)
1087 {
1088         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1089                                     btrfs_extent_data_ref_objectid(leaf, ref),
1090                                     btrfs_extent_data_ref_offset(leaf, ref));
1091 }
1092
1093 static int match_extent_data_ref(struct extent_buffer *leaf,
1094                                  struct btrfs_extent_data_ref *ref,
1095                                  u64 root_objectid, u64 owner, u64 offset)
1096 {
1097         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1098             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1099             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1100                 return 0;
1101         return 1;
1102 }
1103
1104 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1105                                            struct btrfs_root *root,
1106                                            struct btrfs_path *path,
1107                                            u64 bytenr, u64 parent,
1108                                            u64 root_objectid,
1109                                            u64 owner, u64 offset)
1110 {
1111         struct btrfs_key key;
1112         struct btrfs_extent_data_ref *ref;
1113         struct extent_buffer *leaf;
1114         u32 nritems;
1115         int ret;
1116         int recow;
1117         int err = -ENOENT;
1118
1119         key.objectid = bytenr;
1120         if (parent) {
1121                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1122                 key.offset = parent;
1123         } else {
1124                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1125                 key.offset = hash_extent_data_ref(root_objectid,
1126                                                   owner, offset);
1127         }
1128 again:
1129         recow = 0;
1130         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1131         if (ret < 0) {
1132                 err = ret;
1133                 goto fail;
1134         }
1135
1136         if (parent) {
1137                 if (!ret)
1138                         return 0;
1139 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1140                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1141                 btrfs_release_path(path);
1142                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1143                 if (ret < 0) {
1144                         err = ret;
1145                         goto fail;
1146                 }
1147                 if (!ret)
1148                         return 0;
1149 #endif
1150                 goto fail;
1151         }
1152
1153         leaf = path->nodes[0];
1154         nritems = btrfs_header_nritems(leaf);
1155         while (1) {
1156                 if (path->slots[0] >= nritems) {
1157                         ret = btrfs_next_leaf(root, path);
1158                         if (ret < 0)
1159                                 err = ret;
1160                         if (ret)
1161                                 goto fail;
1162
1163                         leaf = path->nodes[0];
1164                         nritems = btrfs_header_nritems(leaf);
1165                         recow = 1;
1166                 }
1167
1168                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1169                 if (key.objectid != bytenr ||
1170                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1171                         goto fail;
1172
1173                 ref = btrfs_item_ptr(leaf, path->slots[0],
1174                                      struct btrfs_extent_data_ref);
1175
1176                 if (match_extent_data_ref(leaf, ref, root_objectid,
1177                                           owner, offset)) {
1178                         if (recow) {
1179                                 btrfs_release_path(path);
1180                                 goto again;
1181                         }
1182                         err = 0;
1183                         break;
1184                 }
1185                 path->slots[0]++;
1186         }
1187 fail:
1188         return err;
1189 }
1190
1191 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1192                                            struct btrfs_root *root,
1193                                            struct btrfs_path *path,
1194                                            u64 bytenr, u64 parent,
1195                                            u64 root_objectid, u64 owner,
1196                                            u64 offset, int refs_to_add)
1197 {
1198         struct btrfs_key key;
1199         struct extent_buffer *leaf;
1200         u32 size;
1201         u32 num_refs;
1202         int ret;
1203
1204         key.objectid = bytenr;
1205         if (parent) {
1206                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1207                 key.offset = parent;
1208                 size = sizeof(struct btrfs_shared_data_ref);
1209         } else {
1210                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1211                 key.offset = hash_extent_data_ref(root_objectid,
1212                                                   owner, offset);
1213                 size = sizeof(struct btrfs_extent_data_ref);
1214         }
1215
1216         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1217         if (ret && ret != -EEXIST)
1218                 goto fail;
1219
1220         leaf = path->nodes[0];
1221         if (parent) {
1222                 struct btrfs_shared_data_ref *ref;
1223                 ref = btrfs_item_ptr(leaf, path->slots[0],
1224                                      struct btrfs_shared_data_ref);
1225                 if (ret == 0) {
1226                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1227                 } else {
1228                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1229                         num_refs += refs_to_add;
1230                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1231                 }
1232         } else {
1233                 struct btrfs_extent_data_ref *ref;
1234                 while (ret == -EEXIST) {
1235                         ref = btrfs_item_ptr(leaf, path->slots[0],
1236                                              struct btrfs_extent_data_ref);
1237                         if (match_extent_data_ref(leaf, ref, root_objectid,
1238                                                   owner, offset))
1239                                 break;
1240                         btrfs_release_path(path);
1241                         key.offset++;
1242                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1243                                                       size);
1244                         if (ret && ret != -EEXIST)
1245                                 goto fail;
1246
1247                         leaf = path->nodes[0];
1248                 }
1249                 ref = btrfs_item_ptr(leaf, path->slots[0],
1250                                      struct btrfs_extent_data_ref);
1251                 if (ret == 0) {
1252                         btrfs_set_extent_data_ref_root(leaf, ref,
1253                                                        root_objectid);
1254                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1255                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1256                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1257                 } else {
1258                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1259                         num_refs += refs_to_add;
1260                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1261                 }
1262         }
1263         btrfs_mark_buffer_dirty(leaf);
1264         ret = 0;
1265 fail:
1266         btrfs_release_path(path);
1267         return ret;
1268 }
1269
1270 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1271                                            struct btrfs_root *root,
1272                                            struct btrfs_path *path,
1273                                            int refs_to_drop)
1274 {
1275         struct btrfs_key key;
1276         struct btrfs_extent_data_ref *ref1 = NULL;
1277         struct btrfs_shared_data_ref *ref2 = NULL;
1278         struct extent_buffer *leaf;
1279         u32 num_refs = 0;
1280         int ret = 0;
1281
1282         leaf = path->nodes[0];
1283         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1284
1285         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1286                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1287                                       struct btrfs_extent_data_ref);
1288                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1289         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1290                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1291                                       struct btrfs_shared_data_ref);
1292                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1293 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1294         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1295                 struct btrfs_extent_ref_v0 *ref0;
1296                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1297                                       struct btrfs_extent_ref_v0);
1298                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1299 #endif
1300         } else {
1301                 BUG();
1302         }
1303
1304         BUG_ON(num_refs < refs_to_drop);
1305         num_refs -= refs_to_drop;
1306
1307         if (num_refs == 0) {
1308                 ret = btrfs_del_item(trans, root, path);
1309         } else {
1310                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1311                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1312                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1313                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1314 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1315                 else {
1316                         struct btrfs_extent_ref_v0 *ref0;
1317                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1318                                         struct btrfs_extent_ref_v0);
1319                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1320                 }
1321 #endif
1322                 btrfs_mark_buffer_dirty(leaf);
1323         }
1324         return ret;
1325 }
1326
1327 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1328                                           struct btrfs_path *path,
1329                                           struct btrfs_extent_inline_ref *iref)
1330 {
1331         struct btrfs_key key;
1332         struct extent_buffer *leaf;
1333         struct btrfs_extent_data_ref *ref1;
1334         struct btrfs_shared_data_ref *ref2;
1335         u32 num_refs = 0;
1336
1337         leaf = path->nodes[0];
1338         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1339         if (iref) {
1340                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1341                     BTRFS_EXTENT_DATA_REF_KEY) {
1342                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1343                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1344                 } else {
1345                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1346                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1347                 }
1348         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1349                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1350                                       struct btrfs_extent_data_ref);
1351                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1352         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1353                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1354                                       struct btrfs_shared_data_ref);
1355                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1356 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1357         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1358                 struct btrfs_extent_ref_v0 *ref0;
1359                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1360                                       struct btrfs_extent_ref_v0);
1361                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1362 #endif
1363         } else {
1364                 WARN_ON(1);
1365         }
1366         return num_refs;
1367 }
1368
1369 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1370                                           struct btrfs_root *root,
1371                                           struct btrfs_path *path,
1372                                           u64 bytenr, u64 parent,
1373                                           u64 root_objectid)
1374 {
1375         struct btrfs_key key;
1376         int ret;
1377
1378         key.objectid = bytenr;
1379         if (parent) {
1380                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1381                 key.offset = parent;
1382         } else {
1383                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1384                 key.offset = root_objectid;
1385         }
1386
1387         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1388         if (ret > 0)
1389                 ret = -ENOENT;
1390 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1391         if (ret == -ENOENT && parent) {
1392                 btrfs_release_path(path);
1393                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1394                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1395                 if (ret > 0)
1396                         ret = -ENOENT;
1397         }
1398 #endif
1399         return ret;
1400 }
1401
1402 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1403                                           struct btrfs_root *root,
1404                                           struct btrfs_path *path,
1405                                           u64 bytenr, u64 parent,
1406                                           u64 root_objectid)
1407 {
1408         struct btrfs_key key;
1409         int ret;
1410
1411         key.objectid = bytenr;
1412         if (parent) {
1413                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1414                 key.offset = parent;
1415         } else {
1416                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1417                 key.offset = root_objectid;
1418         }
1419
1420         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1421         btrfs_release_path(path);
1422         return ret;
1423 }
1424
1425 static inline int extent_ref_type(u64 parent, u64 owner)
1426 {
1427         int type;
1428         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1429                 if (parent > 0)
1430                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1431                 else
1432                         type = BTRFS_TREE_BLOCK_REF_KEY;
1433         } else {
1434                 if (parent > 0)
1435                         type = BTRFS_SHARED_DATA_REF_KEY;
1436                 else
1437                         type = BTRFS_EXTENT_DATA_REF_KEY;
1438         }
1439         return type;
1440 }
1441
1442 static int find_next_key(struct btrfs_path *path, int level,
1443                          struct btrfs_key *key)
1444
1445 {
1446         for (; level < BTRFS_MAX_LEVEL; level++) {
1447                 if (!path->nodes[level])
1448                         break;
1449                 if (path->slots[level] + 1 >=
1450                     btrfs_header_nritems(path->nodes[level]))
1451                         continue;
1452                 if (level == 0)
1453                         btrfs_item_key_to_cpu(path->nodes[level], key,
1454                                               path->slots[level] + 1);
1455                 else
1456                         btrfs_node_key_to_cpu(path->nodes[level], key,
1457                                               path->slots[level] + 1);
1458                 return 0;
1459         }
1460         return 1;
1461 }
1462
1463 /*
1464  * look for inline back ref. if back ref is found, *ref_ret is set
1465  * to the address of inline back ref, and 0 is returned.
1466  *
1467  * if back ref isn't found, *ref_ret is set to the address where it
1468  * should be inserted, and -ENOENT is returned.
1469  *
1470  * if insert is true and there are too many inline back refs, the path
1471  * points to the extent item, and -EAGAIN is returned.
1472  *
1473  * NOTE: inline back refs are ordered in the same way that back ref
1474  *       items in the tree are ordered.
1475  */
1476 static noinline_for_stack
1477 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1478                                  struct btrfs_root *root,
1479                                  struct btrfs_path *path,
1480                                  struct btrfs_extent_inline_ref **ref_ret,
1481                                  u64 bytenr, u64 num_bytes,
1482                                  u64 parent, u64 root_objectid,
1483                                  u64 owner, u64 offset, int insert)
1484 {
1485         struct btrfs_key key;
1486         struct extent_buffer *leaf;
1487         struct btrfs_extent_item *ei;
1488         struct btrfs_extent_inline_ref *iref;
1489         u64 flags;
1490         u64 item_size;
1491         unsigned long ptr;
1492         unsigned long end;
1493         int extra_size;
1494         int type;
1495         int want;
1496         int ret;
1497         int err = 0;
1498         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1499                                                  SKINNY_METADATA);
1500
1501         key.objectid = bytenr;
1502         key.type = BTRFS_EXTENT_ITEM_KEY;
1503         key.offset = num_bytes;
1504
1505         want = extent_ref_type(parent, owner);
1506         if (insert) {
1507                 extra_size = btrfs_extent_inline_ref_size(want);
1508                 path->keep_locks = 1;
1509         } else
1510                 extra_size = -1;
1511
1512         /*
1513          * Owner is our parent level, so we can just add one to get the level
1514          * for the block we are interested in.
1515          */
1516         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1517                 key.type = BTRFS_METADATA_ITEM_KEY;
1518                 key.offset = owner;
1519         }
1520
1521 again:
1522         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1523         if (ret < 0) {
1524                 err = ret;
1525                 goto out;
1526         }
1527
1528         /*
1529          * We may be a newly converted file system which still has the old fat
1530          * extent entries for metadata, so try and see if we have one of those.
1531          */
1532         if (ret > 0 && skinny_metadata) {
1533                 skinny_metadata = false;
1534                 if (path->slots[0]) {
1535                         path->slots[0]--;
1536                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1537                                               path->slots[0]);
1538                         if (key.objectid == bytenr &&
1539                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1540                             key.offset == num_bytes)
1541                                 ret = 0;
1542                 }
1543                 if (ret) {
1544                         key.type = BTRFS_EXTENT_ITEM_KEY;
1545                         key.offset = num_bytes;
1546                         btrfs_release_path(path);
1547                         goto again;
1548                 }
1549         }
1550
1551         if (ret && !insert) {
1552                 err = -ENOENT;
1553                 goto out;
1554         } else if (ret) {
1555                 err = -EIO;
1556                 WARN_ON(1);
1557                 goto out;
1558         }
1559
1560         leaf = path->nodes[0];
1561         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1562 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1563         if (item_size < sizeof(*ei)) {
1564                 if (!insert) {
1565                         err = -ENOENT;
1566                         goto out;
1567                 }
1568                 ret = convert_extent_item_v0(trans, root, path, owner,
1569                                              extra_size);
1570                 if (ret < 0) {
1571                         err = ret;
1572                         goto out;
1573                 }
1574                 leaf = path->nodes[0];
1575                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1576         }
1577 #endif
1578         BUG_ON(item_size < sizeof(*ei));
1579
1580         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1581         flags = btrfs_extent_flags(leaf, ei);
1582
1583         ptr = (unsigned long)(ei + 1);
1584         end = (unsigned long)ei + item_size;
1585
1586         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1587                 ptr += sizeof(struct btrfs_tree_block_info);
1588                 BUG_ON(ptr > end);
1589         }
1590
1591         err = -ENOENT;
1592         while (1) {
1593                 if (ptr >= end) {
1594                         WARN_ON(ptr > end);
1595                         break;
1596                 }
1597                 iref = (struct btrfs_extent_inline_ref *)ptr;
1598                 type = btrfs_extent_inline_ref_type(leaf, iref);
1599                 if (want < type)
1600                         break;
1601                 if (want > type) {
1602                         ptr += btrfs_extent_inline_ref_size(type);
1603                         continue;
1604                 }
1605
1606                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1607                         struct btrfs_extent_data_ref *dref;
1608                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1609                         if (match_extent_data_ref(leaf, dref, root_objectid,
1610                                                   owner, offset)) {
1611                                 err = 0;
1612                                 break;
1613                         }
1614                         if (hash_extent_data_ref_item(leaf, dref) <
1615                             hash_extent_data_ref(root_objectid, owner, offset))
1616                                 break;
1617                 } else {
1618                         u64 ref_offset;
1619                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1620                         if (parent > 0) {
1621                                 if (parent == ref_offset) {
1622                                         err = 0;
1623                                         break;
1624                                 }
1625                                 if (ref_offset < parent)
1626                                         break;
1627                         } else {
1628                                 if (root_objectid == ref_offset) {
1629                                         err = 0;
1630                                         break;
1631                                 }
1632                                 if (ref_offset < root_objectid)
1633                                         break;
1634                         }
1635                 }
1636                 ptr += btrfs_extent_inline_ref_size(type);
1637         }
1638         if (err == -ENOENT && insert) {
1639                 if (item_size + extra_size >=
1640                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1641                         err = -EAGAIN;
1642                         goto out;
1643                 }
1644                 /*
1645                  * To add new inline back ref, we have to make sure
1646                  * there is no corresponding back ref item.
1647                  * For simplicity, we just do not add new inline back
1648                  * ref if there is any kind of item for this block
1649                  */
1650                 if (find_next_key(path, 0, &key) == 0 &&
1651                     key.objectid == bytenr &&
1652                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1653                         err = -EAGAIN;
1654                         goto out;
1655                 }
1656         }
1657         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1658 out:
1659         if (insert) {
1660                 path->keep_locks = 0;
1661                 btrfs_unlock_up_safe(path, 1);
1662         }
1663         return err;
1664 }
1665
1666 /*
1667  * helper to add new inline back ref
1668  */
1669 static noinline_for_stack
1670 void setup_inline_extent_backref(struct btrfs_root *root,
1671                                  struct btrfs_path *path,
1672                                  struct btrfs_extent_inline_ref *iref,
1673                                  u64 parent, u64 root_objectid,
1674                                  u64 owner, u64 offset, int refs_to_add,
1675                                  struct btrfs_delayed_extent_op *extent_op)
1676 {
1677         struct extent_buffer *leaf;
1678         struct btrfs_extent_item *ei;
1679         unsigned long ptr;
1680         unsigned long end;
1681         unsigned long item_offset;
1682         u64 refs;
1683         int size;
1684         int type;
1685
1686         leaf = path->nodes[0];
1687         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1688         item_offset = (unsigned long)iref - (unsigned long)ei;
1689
1690         type = extent_ref_type(parent, owner);
1691         size = btrfs_extent_inline_ref_size(type);
1692
1693         btrfs_extend_item(root, path, size);
1694
1695         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1696         refs = btrfs_extent_refs(leaf, ei);
1697         refs += refs_to_add;
1698         btrfs_set_extent_refs(leaf, ei, refs);
1699         if (extent_op)
1700                 __run_delayed_extent_op(extent_op, leaf, ei);
1701
1702         ptr = (unsigned long)ei + item_offset;
1703         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1704         if (ptr < end - size)
1705                 memmove_extent_buffer(leaf, ptr + size, ptr,
1706                                       end - size - ptr);
1707
1708         iref = (struct btrfs_extent_inline_ref *)ptr;
1709         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1710         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1711                 struct btrfs_extent_data_ref *dref;
1712                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1713                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1714                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1715                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1716                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1717         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1718                 struct btrfs_shared_data_ref *sref;
1719                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1720                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1721                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1722         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1723                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1724         } else {
1725                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1726         }
1727         btrfs_mark_buffer_dirty(leaf);
1728 }
1729
1730 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1731                                  struct btrfs_root *root,
1732                                  struct btrfs_path *path,
1733                                  struct btrfs_extent_inline_ref **ref_ret,
1734                                  u64 bytenr, u64 num_bytes, u64 parent,
1735                                  u64 root_objectid, u64 owner, u64 offset)
1736 {
1737         int ret;
1738
1739         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1740                                            bytenr, num_bytes, parent,
1741                                            root_objectid, owner, offset, 0);
1742         if (ret != -ENOENT)
1743                 return ret;
1744
1745         btrfs_release_path(path);
1746         *ref_ret = NULL;
1747
1748         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1749                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1750                                             root_objectid);
1751         } else {
1752                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1753                                              root_objectid, owner, offset);
1754         }
1755         return ret;
1756 }
1757
1758 /*
1759  * helper to update/remove inline back ref
1760  */
1761 static noinline_for_stack
1762 void update_inline_extent_backref(struct btrfs_root *root,
1763                                   struct btrfs_path *path,
1764                                   struct btrfs_extent_inline_ref *iref,
1765                                   int refs_to_mod,
1766                                   struct btrfs_delayed_extent_op *extent_op)
1767 {
1768         struct extent_buffer *leaf;
1769         struct btrfs_extent_item *ei;
1770         struct btrfs_extent_data_ref *dref = NULL;
1771         struct btrfs_shared_data_ref *sref = NULL;
1772         unsigned long ptr;
1773         unsigned long end;
1774         u32 item_size;
1775         int size;
1776         int type;
1777         u64 refs;
1778
1779         leaf = path->nodes[0];
1780         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1781         refs = btrfs_extent_refs(leaf, ei);
1782         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1783         refs += refs_to_mod;
1784         btrfs_set_extent_refs(leaf, ei, refs);
1785         if (extent_op)
1786                 __run_delayed_extent_op(extent_op, leaf, ei);
1787
1788         type = btrfs_extent_inline_ref_type(leaf, iref);
1789
1790         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1791                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1792                 refs = btrfs_extent_data_ref_count(leaf, dref);
1793         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1794                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1795                 refs = btrfs_shared_data_ref_count(leaf, sref);
1796         } else {
1797                 refs = 1;
1798                 BUG_ON(refs_to_mod != -1);
1799         }
1800
1801         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1802         refs += refs_to_mod;
1803
1804         if (refs > 0) {
1805                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1806                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1807                 else
1808                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1809         } else {
1810                 size =  btrfs_extent_inline_ref_size(type);
1811                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1812                 ptr = (unsigned long)iref;
1813                 end = (unsigned long)ei + item_size;
1814                 if (ptr + size < end)
1815                         memmove_extent_buffer(leaf, ptr, ptr + size,
1816                                               end - ptr - size);
1817                 item_size -= size;
1818                 btrfs_truncate_item(root, path, item_size, 1);
1819         }
1820         btrfs_mark_buffer_dirty(leaf);
1821 }
1822
1823 static noinline_for_stack
1824 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1825                                  struct btrfs_root *root,
1826                                  struct btrfs_path *path,
1827                                  u64 bytenr, u64 num_bytes, u64 parent,
1828                                  u64 root_objectid, u64 owner,
1829                                  u64 offset, int refs_to_add,
1830                                  struct btrfs_delayed_extent_op *extent_op)
1831 {
1832         struct btrfs_extent_inline_ref *iref;
1833         int ret;
1834
1835         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1836                                            bytenr, num_bytes, parent,
1837                                            root_objectid, owner, offset, 1);
1838         if (ret == 0) {
1839                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1840                 update_inline_extent_backref(root, path, iref,
1841                                              refs_to_add, extent_op);
1842         } else if (ret == -ENOENT) {
1843                 setup_inline_extent_backref(root, path, iref, parent,
1844                                             root_objectid, owner, offset,
1845                                             refs_to_add, extent_op);
1846                 ret = 0;
1847         }
1848         return ret;
1849 }
1850
1851 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1852                                  struct btrfs_root *root,
1853                                  struct btrfs_path *path,
1854                                  u64 bytenr, u64 parent, u64 root_objectid,
1855                                  u64 owner, u64 offset, int refs_to_add)
1856 {
1857         int ret;
1858         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1859                 BUG_ON(refs_to_add != 1);
1860                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1861                                             parent, root_objectid);
1862         } else {
1863                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1864                                              parent, root_objectid,
1865                                              owner, offset, refs_to_add);
1866         }
1867         return ret;
1868 }
1869
1870 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1871                                  struct btrfs_root *root,
1872                                  struct btrfs_path *path,
1873                                  struct btrfs_extent_inline_ref *iref,
1874                                  int refs_to_drop, int is_data)
1875 {
1876         int ret = 0;
1877
1878         BUG_ON(!is_data && refs_to_drop != 1);
1879         if (iref) {
1880                 update_inline_extent_backref(root, path, iref,
1881                                              -refs_to_drop, NULL);
1882         } else if (is_data) {
1883                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1884         } else {
1885                 ret = btrfs_del_item(trans, root, path);
1886         }
1887         return ret;
1888 }
1889
1890 static int btrfs_issue_discard(struct block_device *bdev,
1891                                 u64 start, u64 len)
1892 {
1893         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1894 }
1895
1896 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1897                                 u64 num_bytes, u64 *actual_bytes)
1898 {
1899         int ret;
1900         u64 discarded_bytes = 0;
1901         struct btrfs_bio *bbio = NULL;
1902
1903
1904         /* Tell the block device(s) that the sectors can be discarded */
1905         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1906                               bytenr, &num_bytes, &bbio, 0);
1907         /* Error condition is -ENOMEM */
1908         if (!ret) {
1909                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1910                 int i;
1911
1912
1913                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1914                         if (!stripe->dev->can_discard)
1915                                 continue;
1916
1917                         ret = btrfs_issue_discard(stripe->dev->bdev,
1918                                                   stripe->physical,
1919                                                   stripe->length);
1920                         if (!ret)
1921                                 discarded_bytes += stripe->length;
1922                         else if (ret != -EOPNOTSUPP)
1923                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1924
1925                         /*
1926                          * Just in case we get back EOPNOTSUPP for some reason,
1927                          * just ignore the return value so we don't screw up
1928                          * people calling discard_extent.
1929                          */
1930                         ret = 0;
1931                 }
1932                 kfree(bbio);
1933         }
1934
1935         if (actual_bytes)
1936                 *actual_bytes = discarded_bytes;
1937
1938
1939         if (ret == -EOPNOTSUPP)
1940                 ret = 0;
1941         return ret;
1942 }
1943
1944 /* Can return -ENOMEM */
1945 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1946                          struct btrfs_root *root,
1947                          u64 bytenr, u64 num_bytes, u64 parent,
1948                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1949 {
1950         int ret;
1951         struct btrfs_fs_info *fs_info = root->fs_info;
1952
1953         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1954                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1955
1956         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1957                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1958                                         num_bytes,
1959                                         parent, root_objectid, (int)owner,
1960                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1961         } else {
1962                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1963                                         num_bytes,
1964                                         parent, root_objectid, owner, offset,
1965                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1966         }
1967         return ret;
1968 }
1969
1970 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1971                                   struct btrfs_root *root,
1972                                   u64 bytenr, u64 num_bytes,
1973                                   u64 parent, u64 root_objectid,
1974                                   u64 owner, u64 offset, int refs_to_add,
1975                                   struct btrfs_delayed_extent_op *extent_op)
1976 {
1977         struct btrfs_path *path;
1978         struct extent_buffer *leaf;
1979         struct btrfs_extent_item *item;
1980         u64 refs;
1981         int ret;
1982
1983         path = btrfs_alloc_path();
1984         if (!path)
1985                 return -ENOMEM;
1986
1987         path->reada = 1;
1988         path->leave_spinning = 1;
1989         /* this will setup the path even if it fails to insert the back ref */
1990         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1991                                            path, bytenr, num_bytes, parent,
1992                                            root_objectid, owner, offset,
1993                                            refs_to_add, extent_op);
1994         if (ret != -EAGAIN)
1995                 goto out;
1996
1997         leaf = path->nodes[0];
1998         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1999         refs = btrfs_extent_refs(leaf, item);
2000         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2001         if (extent_op)
2002                 __run_delayed_extent_op(extent_op, leaf, item);
2003
2004         btrfs_mark_buffer_dirty(leaf);
2005         btrfs_release_path(path);
2006
2007         path->reada = 1;
2008         path->leave_spinning = 1;
2009
2010         /* now insert the actual backref */
2011         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2012                                     path, bytenr, parent, root_objectid,
2013                                     owner, offset, refs_to_add);
2014         if (ret)
2015                 btrfs_abort_transaction(trans, root, ret);
2016 out:
2017         btrfs_free_path(path);
2018         return ret;
2019 }
2020
2021 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2022                                 struct btrfs_root *root,
2023                                 struct btrfs_delayed_ref_node *node,
2024                                 struct btrfs_delayed_extent_op *extent_op,
2025                                 int insert_reserved)
2026 {
2027         int ret = 0;
2028         struct btrfs_delayed_data_ref *ref;
2029         struct btrfs_key ins;
2030         u64 parent = 0;
2031         u64 ref_root = 0;
2032         u64 flags = 0;
2033
2034         ins.objectid = node->bytenr;
2035         ins.offset = node->num_bytes;
2036         ins.type = BTRFS_EXTENT_ITEM_KEY;
2037
2038         ref = btrfs_delayed_node_to_data_ref(node);
2039         trace_run_delayed_data_ref(node, ref, node->action);
2040
2041         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2042                 parent = ref->parent;
2043         else
2044                 ref_root = ref->root;
2045
2046         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2047                 if (extent_op)
2048                         flags |= extent_op->flags_to_set;
2049                 ret = alloc_reserved_file_extent(trans, root,
2050                                                  parent, ref_root, flags,
2051                                                  ref->objectid, ref->offset,
2052                                                  &ins, node->ref_mod);
2053         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2054                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2055                                              node->num_bytes, parent,
2056                                              ref_root, ref->objectid,
2057                                              ref->offset, node->ref_mod,
2058                                              extent_op);
2059         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2060                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2061                                           node->num_bytes, parent,
2062                                           ref_root, ref->objectid,
2063                                           ref->offset, node->ref_mod,
2064                                           extent_op);
2065         } else {
2066                 BUG();
2067         }
2068         return ret;
2069 }
2070
2071 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2072                                     struct extent_buffer *leaf,
2073                                     struct btrfs_extent_item *ei)
2074 {
2075         u64 flags = btrfs_extent_flags(leaf, ei);
2076         if (extent_op->update_flags) {
2077                 flags |= extent_op->flags_to_set;
2078                 btrfs_set_extent_flags(leaf, ei, flags);
2079         }
2080
2081         if (extent_op->update_key) {
2082                 struct btrfs_tree_block_info *bi;
2083                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2084                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2085                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2086         }
2087 }
2088
2089 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2090                                  struct btrfs_root *root,
2091                                  struct btrfs_delayed_ref_node *node,
2092                                  struct btrfs_delayed_extent_op *extent_op)
2093 {
2094         struct btrfs_key key;
2095         struct btrfs_path *path;
2096         struct btrfs_extent_item *ei;
2097         struct extent_buffer *leaf;
2098         u32 item_size;
2099         int ret;
2100         int err = 0;
2101         int metadata = !extent_op->is_data;
2102
2103         if (trans->aborted)
2104                 return 0;
2105
2106         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2107                 metadata = 0;
2108
2109         path = btrfs_alloc_path();
2110         if (!path)
2111                 return -ENOMEM;
2112
2113         key.objectid = node->bytenr;
2114
2115         if (metadata) {
2116                 key.type = BTRFS_METADATA_ITEM_KEY;
2117                 key.offset = extent_op->level;
2118         } else {
2119                 key.type = BTRFS_EXTENT_ITEM_KEY;
2120                 key.offset = node->num_bytes;
2121         }
2122
2123 again:
2124         path->reada = 1;
2125         path->leave_spinning = 1;
2126         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2127                                 path, 0, 1);
2128         if (ret < 0) {
2129                 err = ret;
2130                 goto out;
2131         }
2132         if (ret > 0) {
2133                 if (metadata) {
2134                         btrfs_release_path(path);
2135                         metadata = 0;
2136
2137                         key.offset = node->num_bytes;
2138                         key.type = BTRFS_EXTENT_ITEM_KEY;
2139                         goto again;
2140                 }
2141                 err = -EIO;
2142                 goto out;
2143         }
2144
2145         leaf = path->nodes[0];
2146         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2147 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2148         if (item_size < sizeof(*ei)) {
2149                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2150                                              path, (u64)-1, 0);
2151                 if (ret < 0) {
2152                         err = ret;
2153                         goto out;
2154                 }
2155                 leaf = path->nodes[0];
2156                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2157         }
2158 #endif
2159         BUG_ON(item_size < sizeof(*ei));
2160         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2161         __run_delayed_extent_op(extent_op, leaf, ei);
2162
2163         btrfs_mark_buffer_dirty(leaf);
2164 out:
2165         btrfs_free_path(path);
2166         return err;
2167 }
2168
2169 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2170                                 struct btrfs_root *root,
2171                                 struct btrfs_delayed_ref_node *node,
2172                                 struct btrfs_delayed_extent_op *extent_op,
2173                                 int insert_reserved)
2174 {
2175         int ret = 0;
2176         struct btrfs_delayed_tree_ref *ref;
2177         struct btrfs_key ins;
2178         u64 parent = 0;
2179         u64 ref_root = 0;
2180         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2181                                                  SKINNY_METADATA);
2182
2183         ref = btrfs_delayed_node_to_tree_ref(node);
2184         trace_run_delayed_tree_ref(node, ref, node->action);
2185
2186         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2187                 parent = ref->parent;
2188         else
2189                 ref_root = ref->root;
2190
2191         ins.objectid = node->bytenr;
2192         if (skinny_metadata) {
2193                 ins.offset = ref->level;
2194                 ins.type = BTRFS_METADATA_ITEM_KEY;
2195         } else {
2196                 ins.offset = node->num_bytes;
2197                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2198         }
2199
2200         BUG_ON(node->ref_mod != 1);
2201         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2202                 BUG_ON(!extent_op || !extent_op->update_flags);
2203                 ret = alloc_reserved_tree_block(trans, root,
2204                                                 parent, ref_root,
2205                                                 extent_op->flags_to_set,
2206                                                 &extent_op->key,
2207                                                 ref->level, &ins);
2208         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2209                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2210                                              node->num_bytes, parent, ref_root,
2211                                              ref->level, 0, 1, extent_op);
2212         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2213                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2214                                           node->num_bytes, parent, ref_root,
2215                                           ref->level, 0, 1, extent_op);
2216         } else {
2217                 BUG();
2218         }
2219         return ret;
2220 }
2221
2222 /* helper function to actually process a single delayed ref entry */
2223 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2224                                struct btrfs_root *root,
2225                                struct btrfs_delayed_ref_node *node,
2226                                struct btrfs_delayed_extent_op *extent_op,
2227                                int insert_reserved)
2228 {
2229         int ret = 0;
2230
2231         if (trans->aborted) {
2232                 if (insert_reserved)
2233                         btrfs_pin_extent(root, node->bytenr,
2234                                          node->num_bytes, 1);
2235                 return 0;
2236         }
2237
2238         if (btrfs_delayed_ref_is_head(node)) {
2239                 struct btrfs_delayed_ref_head *head;
2240                 /*
2241                  * we've hit the end of the chain and we were supposed
2242                  * to insert this extent into the tree.  But, it got
2243                  * deleted before we ever needed to insert it, so all
2244                  * we have to do is clean up the accounting
2245                  */
2246                 BUG_ON(extent_op);
2247                 head = btrfs_delayed_node_to_head(node);
2248                 trace_run_delayed_ref_head(node, head, node->action);
2249
2250                 if (insert_reserved) {
2251                         btrfs_pin_extent(root, node->bytenr,
2252                                          node->num_bytes, 1);
2253                         if (head->is_data) {
2254                                 ret = btrfs_del_csums(trans, root,
2255                                                       node->bytenr,
2256                                                       node->num_bytes);
2257                         }
2258                 }
2259                 return ret;
2260         }
2261
2262         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2263             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2264                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2265                                            insert_reserved);
2266         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2267                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2268                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2269                                            insert_reserved);
2270         else
2271                 BUG();
2272         return ret;
2273 }
2274
2275 static noinline struct btrfs_delayed_ref_node *
2276 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2277 {
2278         struct rb_node *node;
2279         struct btrfs_delayed_ref_node *ref;
2280         int action = BTRFS_ADD_DELAYED_REF;
2281 again:
2282         /*
2283          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2284          * this prevents ref count from going down to zero when
2285          * there still are pending delayed ref.
2286          */
2287         node = rb_prev(&head->node.rb_node);
2288         while (1) {
2289                 if (!node)
2290                         break;
2291                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2292                                 rb_node);
2293                 if (ref->bytenr != head->node.bytenr)
2294                         break;
2295                 if (ref->action == action)
2296                         return ref;
2297                 node = rb_prev(node);
2298         }
2299         if (action == BTRFS_ADD_DELAYED_REF) {
2300                 action = BTRFS_DROP_DELAYED_REF;
2301                 goto again;
2302         }
2303         return NULL;
2304 }
2305
2306 /*
2307  * Returns 0 on success or if called with an already aborted transaction.
2308  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2309  */
2310 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2311                                        struct btrfs_root *root,
2312                                        struct list_head *cluster)
2313 {
2314         struct btrfs_delayed_ref_root *delayed_refs;
2315         struct btrfs_delayed_ref_node *ref;
2316         struct btrfs_delayed_ref_head *locked_ref = NULL;
2317         struct btrfs_delayed_extent_op *extent_op;
2318         struct btrfs_fs_info *fs_info = root->fs_info;
2319         int ret;
2320         int count = 0;
2321         int must_insert_reserved = 0;
2322
2323         delayed_refs = &trans->transaction->delayed_refs;
2324         while (1) {
2325                 if (!locked_ref) {
2326                         /* pick a new head ref from the cluster list */
2327                         if (list_empty(cluster))
2328                                 break;
2329
2330                         locked_ref = list_entry(cluster->next,
2331                                      struct btrfs_delayed_ref_head, cluster);
2332
2333                         /* grab the lock that says we are going to process
2334                          * all the refs for this head */
2335                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2336
2337                         /*
2338                          * we may have dropped the spin lock to get the head
2339                          * mutex lock, and that might have given someone else
2340                          * time to free the head.  If that's true, it has been
2341                          * removed from our list and we can move on.
2342                          */
2343                         if (ret == -EAGAIN) {
2344                                 locked_ref = NULL;
2345                                 count++;
2346                                 continue;
2347                         }
2348                 }
2349
2350                 /*
2351                  * We need to try and merge add/drops of the same ref since we
2352                  * can run into issues with relocate dropping the implicit ref
2353                  * and then it being added back again before the drop can
2354                  * finish.  If we merged anything we need to re-loop so we can
2355                  * get a good ref.
2356                  */
2357                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2358                                          locked_ref);
2359
2360                 /*
2361                  * locked_ref is the head node, so we have to go one
2362                  * node back for any delayed ref updates
2363                  */
2364                 ref = select_delayed_ref(locked_ref);
2365
2366                 if (ref && ref->seq &&
2367                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2368                         /*
2369                          * there are still refs with lower seq numbers in the
2370                          * process of being added. Don't run this ref yet.
2371                          */
2372                         list_del_init(&locked_ref->cluster);
2373                         btrfs_delayed_ref_unlock(locked_ref);
2374                         locked_ref = NULL;
2375                         delayed_refs->num_heads_ready++;
2376                         spin_unlock(&delayed_refs->lock);
2377                         cond_resched();
2378                         spin_lock(&delayed_refs->lock);
2379                         continue;
2380                 }
2381
2382                 /*
2383                  * record the must insert reserved flag before we
2384                  * drop the spin lock.
2385                  */
2386                 must_insert_reserved = locked_ref->must_insert_reserved;
2387                 locked_ref->must_insert_reserved = 0;
2388
2389                 extent_op = locked_ref->extent_op;
2390                 locked_ref->extent_op = NULL;
2391
2392                 if (!ref) {
2393                         /* All delayed refs have been processed, Go ahead
2394                          * and send the head node to run_one_delayed_ref,
2395                          * so that any accounting fixes can happen
2396                          */
2397                         ref = &locked_ref->node;
2398
2399                         if (extent_op && must_insert_reserved) {
2400                                 btrfs_free_delayed_extent_op(extent_op);
2401                                 extent_op = NULL;
2402                         }
2403
2404                         if (extent_op) {
2405                                 spin_unlock(&delayed_refs->lock);
2406
2407                                 ret = run_delayed_extent_op(trans, root,
2408                                                             ref, extent_op);
2409                                 btrfs_free_delayed_extent_op(extent_op);
2410
2411                                 if (ret) {
2412                                         /*
2413                                          * Need to reset must_insert_reserved if
2414                                          * there was an error so the abort stuff
2415                                          * can cleanup the reserved space
2416                                          * properly.
2417                                          */
2418                                         if (must_insert_reserved)
2419                                                 locked_ref->must_insert_reserved = 1;
2420                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2421                                         spin_lock(&delayed_refs->lock);
2422                                         btrfs_delayed_ref_unlock(locked_ref);
2423                                         return ret;
2424                                 }
2425
2426                                 goto next;
2427                         }
2428                 }
2429
2430                 ref->in_tree = 0;
2431                 rb_erase(&ref->rb_node, &delayed_refs->root);
2432                 delayed_refs->num_entries--;
2433                 if (!btrfs_delayed_ref_is_head(ref)) {
2434                         /*
2435                          * when we play the delayed ref, also correct the
2436                          * ref_mod on head
2437                          */
2438                         switch (ref->action) {
2439                         case BTRFS_ADD_DELAYED_REF:
2440                         case BTRFS_ADD_DELAYED_EXTENT:
2441                                 locked_ref->node.ref_mod -= ref->ref_mod;
2442                                 break;
2443                         case BTRFS_DROP_DELAYED_REF:
2444                                 locked_ref->node.ref_mod += ref->ref_mod;
2445                                 break;
2446                         default:
2447                                 WARN_ON(1);
2448                         }
2449                 } else {
2450                         list_del_init(&locked_ref->cluster);
2451                 }
2452                 spin_unlock(&delayed_refs->lock);
2453
2454                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2455                                           must_insert_reserved);
2456
2457                 btrfs_free_delayed_extent_op(extent_op);
2458                 if (ret) {
2459                         btrfs_delayed_ref_unlock(locked_ref);
2460                         btrfs_put_delayed_ref(ref);
2461                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2462                         spin_lock(&delayed_refs->lock);
2463                         return ret;
2464                 }
2465
2466                 /*
2467                  * If this node is a head, that means all the refs in this head
2468                  * have been dealt with, and we will pick the next head to deal
2469                  * with, so we must unlock the head and drop it from the cluster
2470                  * list before we release it.
2471                  */
2472                 if (btrfs_delayed_ref_is_head(ref)) {
2473                         btrfs_delayed_ref_unlock(locked_ref);
2474                         locked_ref = NULL;
2475                 }
2476                 btrfs_put_delayed_ref(ref);
2477                 count++;
2478 next:
2479                 cond_resched();
2480                 spin_lock(&delayed_refs->lock);
2481         }
2482         return count;
2483 }
2484
2485 #ifdef SCRAMBLE_DELAYED_REFS
2486 /*
2487  * Normally delayed refs get processed in ascending bytenr order. This
2488  * correlates in most cases to the order added. To expose dependencies on this
2489  * order, we start to process the tree in the middle instead of the beginning
2490  */
2491 static u64 find_middle(struct rb_root *root)
2492 {
2493         struct rb_node *n = root->rb_node;
2494         struct btrfs_delayed_ref_node *entry;
2495         int alt = 1;
2496         u64 middle;
2497         u64 first = 0, last = 0;
2498
2499         n = rb_first(root);
2500         if (n) {
2501                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2502                 first = entry->bytenr;
2503         }
2504         n = rb_last(root);
2505         if (n) {
2506                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2507                 last = entry->bytenr;
2508         }
2509         n = root->rb_node;
2510
2511         while (n) {
2512                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2513                 WARN_ON(!entry->in_tree);
2514
2515                 middle = entry->bytenr;
2516
2517                 if (alt)
2518                         n = n->rb_left;
2519                 else
2520                         n = n->rb_right;
2521
2522                 alt = 1 - alt;
2523         }
2524         return middle;
2525 }
2526 #endif
2527
2528 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2529                                          struct btrfs_fs_info *fs_info)
2530 {
2531         struct qgroup_update *qgroup_update;
2532         int ret = 0;
2533
2534         if (list_empty(&trans->qgroup_ref_list) !=
2535             !trans->delayed_ref_elem.seq) {
2536                 /* list without seq or seq without list */
2537                 btrfs_err(fs_info,
2538                         "qgroup accounting update error, list is%s empty, seq is %#x.%x",
2539                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2540                         (u32)(trans->delayed_ref_elem.seq >> 32),
2541                         (u32)trans->delayed_ref_elem.seq);
2542                 BUG();
2543         }
2544
2545         if (!trans->delayed_ref_elem.seq)
2546                 return 0;
2547
2548         while (!list_empty(&trans->qgroup_ref_list)) {
2549                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2550                                                  struct qgroup_update, list);
2551                 list_del(&qgroup_update->list);
2552                 if (!ret)
2553                         ret = btrfs_qgroup_account_ref(
2554                                         trans, fs_info, qgroup_update->node,
2555                                         qgroup_update->extent_op);
2556                 kfree(qgroup_update);
2557         }
2558
2559         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2560
2561         return ret;
2562 }
2563
2564 static int refs_newer(struct btrfs_delayed_ref_root *delayed_refs, int seq,
2565                       int count)
2566 {
2567         int val = atomic_read(&delayed_refs->ref_seq);
2568
2569         if (val < seq || val >= seq + count)
2570                 return 1;
2571         return 0;
2572 }
2573
2574 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2575 {
2576         u64 num_bytes;
2577
2578         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2579                              sizeof(struct btrfs_extent_inline_ref));
2580         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2581                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2582
2583         /*
2584          * We don't ever fill up leaves all the way so multiply by 2 just to be
2585          * closer to what we're really going to want to ouse.
2586          */
2587         return div64_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2588 }
2589
2590 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2591                                        struct btrfs_root *root)
2592 {
2593         struct btrfs_block_rsv *global_rsv;
2594         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2595         u64 num_bytes;
2596         int ret = 0;
2597
2598         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2599         num_heads = heads_to_leaves(root, num_heads);
2600         if (num_heads > 1)
2601                 num_bytes += (num_heads - 1) * root->leafsize;
2602         num_bytes <<= 1;
2603         global_rsv = &root->fs_info->global_block_rsv;
2604
2605         /*
2606          * If we can't allocate any more chunks lets make sure we have _lots_ of
2607          * wiggle room since running delayed refs can create more delayed refs.
2608          */
2609         if (global_rsv->space_info->full)
2610                 num_bytes <<= 1;
2611
2612         spin_lock(&global_rsv->lock);
2613         if (global_rsv->reserved <= num_bytes)
2614                 ret = 1;
2615         spin_unlock(&global_rsv->lock);
2616         return ret;
2617 }
2618
2619 /*
2620  * this starts processing the delayed reference count updates and
2621  * extent insertions we have queued up so far.  count can be
2622  * 0, which means to process everything in the tree at the start
2623  * of the run (but not newly added entries), or it can be some target
2624  * number you'd like to process.
2625  *
2626  * Returns 0 on success or if called with an aborted transaction
2627  * Returns <0 on error and aborts the transaction
2628  */
2629 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2630                            struct btrfs_root *root, unsigned long count)
2631 {
2632         struct rb_node *node;
2633         struct btrfs_delayed_ref_root *delayed_refs;
2634         struct btrfs_delayed_ref_node *ref;
2635         struct list_head cluster;
2636         int ret;
2637         u64 delayed_start;
2638         int run_all = count == (unsigned long)-1;
2639         int run_most = 0;
2640         int loops;
2641
2642         /* We'll clean this up in btrfs_cleanup_transaction */
2643         if (trans->aborted)
2644                 return 0;
2645
2646         if (root == root->fs_info->extent_root)
2647                 root = root->fs_info->tree_root;
2648
2649         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2650
2651         delayed_refs = &trans->transaction->delayed_refs;
2652         INIT_LIST_HEAD(&cluster);
2653         if (count == 0) {
2654                 count = delayed_refs->num_entries * 2;
2655                 run_most = 1;
2656         }
2657
2658         if (!run_all && !run_most) {
2659                 int old;
2660                 int seq = atomic_read(&delayed_refs->ref_seq);
2661
2662 progress:
2663                 old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2664                 if (old) {
2665                         DEFINE_WAIT(__wait);
2666                         if (delayed_refs->flushing ||
2667                             !btrfs_should_throttle_delayed_refs(trans, root))
2668                                 return 0;
2669
2670                         prepare_to_wait(&delayed_refs->wait, &__wait,
2671                                         TASK_UNINTERRUPTIBLE);
2672
2673                         old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2674                         if (old) {
2675                                 schedule();
2676                                 finish_wait(&delayed_refs->wait, &__wait);
2677
2678                                 if (!refs_newer(delayed_refs, seq, 256))
2679                                         goto progress;
2680                                 else
2681                                         return 0;
2682                         } else {
2683                                 finish_wait(&delayed_refs->wait, &__wait);
2684                                 goto again;
2685                         }
2686                 }
2687
2688         } else {
2689                 atomic_inc(&delayed_refs->procs_running_refs);
2690         }
2691
2692 again:
2693         loops = 0;
2694         spin_lock(&delayed_refs->lock);
2695
2696 #ifdef SCRAMBLE_DELAYED_REFS
2697         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2698 #endif
2699
2700         while (1) {
2701                 if (!(run_all || run_most) &&
2702                     !btrfs_should_throttle_delayed_refs(trans, root))
2703                         break;
2704
2705                 /*
2706                  * go find something we can process in the rbtree.  We start at
2707                  * the beginning of the tree, and then build a cluster
2708                  * of refs to process starting at the first one we are able to
2709                  * lock
2710                  */
2711                 delayed_start = delayed_refs->run_delayed_start;
2712                 ret = btrfs_find_ref_cluster(trans, &cluster,
2713                                              delayed_refs->run_delayed_start);
2714                 if (ret)
2715                         break;
2716
2717                 ret = run_clustered_refs(trans, root, &cluster);
2718                 if (ret < 0) {
2719                         btrfs_release_ref_cluster(&cluster);
2720                         spin_unlock(&delayed_refs->lock);
2721                         btrfs_abort_transaction(trans, root, ret);
2722                         atomic_dec(&delayed_refs->procs_running_refs);
2723                         wake_up(&delayed_refs->wait);
2724                         return ret;
2725                 }
2726
2727                 atomic_add(ret, &delayed_refs->ref_seq);
2728
2729                 count -= min_t(unsigned long, ret, count);
2730
2731                 if (count == 0)
2732                         break;
2733
2734                 if (delayed_start >= delayed_refs->run_delayed_start) {
2735                         if (loops == 0) {
2736                                 /*
2737                                  * btrfs_find_ref_cluster looped. let's do one
2738                                  * more cycle. if we don't run any delayed ref
2739                                  * during that cycle (because we can't because
2740                                  * all of them are blocked), bail out.
2741                                  */
2742                                 loops = 1;
2743                         } else {
2744                                 /*
2745                                  * no runnable refs left, stop trying
2746                                  */
2747                                 BUG_ON(run_all);
2748                                 break;
2749                         }
2750                 }
2751                 if (ret) {
2752                         /* refs were run, let's reset staleness detection */
2753                         loops = 0;
2754                 }
2755         }
2756
2757         if (run_all) {
2758                 if (!list_empty(&trans->new_bgs)) {
2759                         spin_unlock(&delayed_refs->lock);
2760                         btrfs_create_pending_block_groups(trans, root);
2761                         spin_lock(&delayed_refs->lock);
2762                 }
2763
2764                 node = rb_first(&delayed_refs->root);
2765                 if (!node)
2766                         goto out;
2767                 count = (unsigned long)-1;
2768
2769                 while (node) {
2770                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2771                                        rb_node);
2772                         if (btrfs_delayed_ref_is_head(ref)) {
2773                                 struct btrfs_delayed_ref_head *head;
2774
2775                                 head = btrfs_delayed_node_to_head(ref);
2776                                 atomic_inc(&ref->refs);
2777
2778                                 spin_unlock(&delayed_refs->lock);
2779                                 /*
2780                                  * Mutex was contended, block until it's
2781                                  * released and try again
2782                                  */
2783                                 mutex_lock(&head->mutex);
2784                                 mutex_unlock(&head->mutex);
2785
2786                                 btrfs_put_delayed_ref(ref);
2787                                 cond_resched();
2788                                 goto again;
2789                         }
2790                         node = rb_next(node);
2791                 }
2792                 spin_unlock(&delayed_refs->lock);
2793                 schedule_timeout(1);
2794                 goto again;
2795         }
2796 out:
2797         atomic_dec(&delayed_refs->procs_running_refs);
2798         smp_mb();
2799         if (waitqueue_active(&delayed_refs->wait))
2800                 wake_up(&delayed_refs->wait);
2801
2802         spin_unlock(&delayed_refs->lock);
2803         assert_qgroups_uptodate(trans);
2804         return 0;
2805 }
2806
2807 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2808                                 struct btrfs_root *root,
2809                                 u64 bytenr, u64 num_bytes, u64 flags,
2810                                 int level, int is_data)
2811 {
2812         struct btrfs_delayed_extent_op *extent_op;
2813         int ret;
2814
2815         extent_op = btrfs_alloc_delayed_extent_op();
2816         if (!extent_op)
2817                 return -ENOMEM;
2818
2819         extent_op->flags_to_set = flags;
2820         extent_op->update_flags = 1;
2821         extent_op->update_key = 0;
2822         extent_op->is_data = is_data ? 1 : 0;
2823         extent_op->level = level;
2824
2825         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2826                                           num_bytes, extent_op);
2827         if (ret)
2828                 btrfs_free_delayed_extent_op(extent_op);
2829         return ret;
2830 }
2831
2832 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2833                                       struct btrfs_root *root,
2834                                       struct btrfs_path *path,
2835                                       u64 objectid, u64 offset, u64 bytenr)
2836 {
2837         struct btrfs_delayed_ref_head *head;
2838         struct btrfs_delayed_ref_node *ref;
2839         struct btrfs_delayed_data_ref *data_ref;
2840         struct btrfs_delayed_ref_root *delayed_refs;
2841         struct rb_node *node;
2842         int ret = 0;
2843
2844         ret = -ENOENT;
2845         delayed_refs = &trans->transaction->delayed_refs;
2846         spin_lock(&delayed_refs->lock);
2847         head = btrfs_find_delayed_ref_head(trans, bytenr);
2848         if (!head)
2849                 goto out;
2850
2851         if (!mutex_trylock(&head->mutex)) {
2852                 atomic_inc(&head->node.refs);
2853                 spin_unlock(&delayed_refs->lock);
2854
2855                 btrfs_release_path(path);
2856
2857                 /*
2858                  * Mutex was contended, block until it's released and let
2859                  * caller try again
2860                  */
2861                 mutex_lock(&head->mutex);
2862                 mutex_unlock(&head->mutex);
2863                 btrfs_put_delayed_ref(&head->node);
2864                 return -EAGAIN;
2865         }
2866
2867         node = rb_prev(&head->node.rb_node);
2868         if (!node)
2869                 goto out_unlock;
2870
2871         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2872
2873         if (ref->bytenr != bytenr)
2874                 goto out_unlock;
2875
2876         ret = 1;
2877         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2878                 goto out_unlock;
2879
2880         data_ref = btrfs_delayed_node_to_data_ref(ref);
2881
2882         node = rb_prev(node);
2883         if (node) {
2884                 int seq = ref->seq;
2885
2886                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2887                 if (ref->bytenr == bytenr && ref->seq == seq)
2888                         goto out_unlock;
2889         }
2890
2891         if (data_ref->root != root->root_key.objectid ||
2892             data_ref->objectid != objectid || data_ref->offset != offset)
2893                 goto out_unlock;
2894
2895         ret = 0;
2896 out_unlock:
2897         mutex_unlock(&head->mutex);
2898 out:
2899         spin_unlock(&delayed_refs->lock);
2900         return ret;
2901 }
2902
2903 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2904                                         struct btrfs_root *root,
2905                                         struct btrfs_path *path,
2906                                         u64 objectid, u64 offset, u64 bytenr)
2907 {
2908         struct btrfs_root *extent_root = root->fs_info->extent_root;
2909         struct extent_buffer *leaf;
2910         struct btrfs_extent_data_ref *ref;
2911         struct btrfs_extent_inline_ref *iref;
2912         struct btrfs_extent_item *ei;
2913         struct btrfs_key key;
2914         u32 item_size;
2915         int ret;
2916
2917         key.objectid = bytenr;
2918         key.offset = (u64)-1;
2919         key.type = BTRFS_EXTENT_ITEM_KEY;
2920
2921         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2922         if (ret < 0)
2923                 goto out;
2924         BUG_ON(ret == 0); /* Corruption */
2925
2926         ret = -ENOENT;
2927         if (path->slots[0] == 0)
2928                 goto out;
2929
2930         path->slots[0]--;
2931         leaf = path->nodes[0];
2932         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2933
2934         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2935                 goto out;
2936
2937         ret = 1;
2938         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2939 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2940         if (item_size < sizeof(*ei)) {
2941                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2942                 goto out;
2943         }
2944 #endif
2945         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2946
2947         if (item_size != sizeof(*ei) +
2948             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2949                 goto out;
2950
2951         if (btrfs_extent_generation(leaf, ei) <=
2952             btrfs_root_last_snapshot(&root->root_item))
2953                 goto out;
2954
2955         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2956         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2957             BTRFS_EXTENT_DATA_REF_KEY)
2958                 goto out;
2959
2960         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2961         if (btrfs_extent_refs(leaf, ei) !=
2962             btrfs_extent_data_ref_count(leaf, ref) ||
2963             btrfs_extent_data_ref_root(leaf, ref) !=
2964             root->root_key.objectid ||
2965             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2966             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2967                 goto out;
2968
2969         ret = 0;
2970 out:
2971         return ret;
2972 }
2973
2974 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2975                           struct btrfs_root *root,
2976                           u64 objectid, u64 offset, u64 bytenr)
2977 {
2978         struct btrfs_path *path;
2979         int ret;
2980         int ret2;
2981
2982         path = btrfs_alloc_path();
2983         if (!path)
2984                 return -ENOENT;
2985
2986         do {
2987                 ret = check_committed_ref(trans, root, path, objectid,
2988                                           offset, bytenr);
2989                 if (ret && ret != -ENOENT)
2990                         goto out;
2991
2992                 ret2 = check_delayed_ref(trans, root, path, objectid,
2993                                          offset, bytenr);
2994         } while (ret2 == -EAGAIN);
2995
2996         if (ret2 && ret2 != -ENOENT) {
2997                 ret = ret2;
2998                 goto out;
2999         }
3000
3001         if (ret != -ENOENT || ret2 != -ENOENT)
3002                 ret = 0;
3003 out:
3004         btrfs_free_path(path);
3005         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3006                 WARN_ON(ret > 0);
3007         return ret;
3008 }
3009
3010 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3011                            struct btrfs_root *root,
3012                            struct extent_buffer *buf,
3013                            int full_backref, int inc, int for_cow)
3014 {
3015         u64 bytenr;
3016         u64 num_bytes;
3017         u64 parent;
3018         u64 ref_root;
3019         u32 nritems;
3020         struct btrfs_key key;
3021         struct btrfs_file_extent_item *fi;
3022         int i;
3023         int level;
3024         int ret = 0;
3025         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3026                             u64, u64, u64, u64, u64, u64, int);
3027
3028         ref_root = btrfs_header_owner(buf);
3029         nritems = btrfs_header_nritems(buf);
3030         level = btrfs_header_level(buf);
3031
3032         if (!root->ref_cows && level == 0)
3033                 return 0;
3034
3035         if (inc)
3036                 process_func = btrfs_inc_extent_ref;
3037         else
3038                 process_func = btrfs_free_extent;
3039
3040         if (full_backref)
3041                 parent = buf->start;
3042         else
3043                 parent = 0;
3044
3045         for (i = 0; i < nritems; i++) {
3046                 if (level == 0) {
3047                         btrfs_item_key_to_cpu(buf, &key, i);
3048                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
3049                                 continue;
3050                         fi = btrfs_item_ptr(buf, i,
3051                                             struct btrfs_file_extent_item);
3052                         if (btrfs_file_extent_type(buf, fi) ==
3053                             BTRFS_FILE_EXTENT_INLINE)
3054                                 continue;
3055                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3056                         if (bytenr == 0)
3057                                 continue;
3058
3059                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3060                         key.offset -= btrfs_file_extent_offset(buf, fi);
3061                         ret = process_func(trans, root, bytenr, num_bytes,
3062                                            parent, ref_root, key.objectid,
3063                                            key.offset, for_cow);
3064                         if (ret)
3065                                 goto fail;
3066                 } else {
3067                         bytenr = btrfs_node_blockptr(buf, i);
3068                         num_bytes = btrfs_level_size(root, level - 1);
3069                         ret = process_func(trans, root, bytenr, num_bytes,
3070                                            parent, ref_root, level - 1, 0,
3071                                            for_cow);
3072                         if (ret)
3073                                 goto fail;
3074                 }
3075         }
3076         return 0;
3077 fail:
3078         return ret;
3079 }
3080
3081 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3082                   struct extent_buffer *buf, int full_backref, int for_cow)
3083 {
3084         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
3085 }
3086
3087 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3088                   struct extent_buffer *buf, int full_backref, int for_cow)
3089 {
3090         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
3091 }
3092
3093 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3094                                  struct btrfs_root *root,
3095                                  struct btrfs_path *path,
3096                                  struct btrfs_block_group_cache *cache)
3097 {
3098         int ret;
3099         struct btrfs_root *extent_root = root->fs_info->extent_root;
3100         unsigned long bi;
3101         struct extent_buffer *leaf;
3102
3103         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3104         if (ret < 0)
3105                 goto fail;
3106         BUG_ON(ret); /* Corruption */
3107
3108         leaf = path->nodes[0];
3109         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3110         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3111         btrfs_mark_buffer_dirty(leaf);
3112         btrfs_release_path(path);
3113 fail:
3114         if (ret) {
3115                 btrfs_abort_transaction(trans, root, ret);
3116                 return ret;
3117         }
3118         return 0;
3119
3120 }
3121
3122 static struct btrfs_block_group_cache *
3123 next_block_group(struct btrfs_root *root,
3124                  struct btrfs_block_group_cache *cache)
3125 {
3126         struct rb_node *node;
3127         spin_lock(&root->fs_info->block_group_cache_lock);
3128         node = rb_next(&cache->cache_node);
3129         btrfs_put_block_group(cache);
3130         if (node) {
3131                 cache = rb_entry(node, struct btrfs_block_group_cache,
3132                                  cache_node);
3133                 btrfs_get_block_group(cache);
3134         } else
3135                 cache = NULL;
3136         spin_unlock(&root->fs_info->block_group_cache_lock);
3137         return cache;
3138 }
3139
3140 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3141                             struct btrfs_trans_handle *trans,
3142                             struct btrfs_path *path)
3143 {
3144         struct btrfs_root *root = block_group->fs_info->tree_root;
3145         struct inode *inode = NULL;
3146         u64 alloc_hint = 0;
3147         int dcs = BTRFS_DC_ERROR;
3148         int num_pages = 0;
3149         int retries = 0;
3150         int ret = 0;
3151
3152         /*
3153          * If this block group is smaller than 100 megs don't bother caching the
3154          * block group.
3155          */
3156         if (block_group->key.offset < (100 * 1024 * 1024)) {
3157                 spin_lock(&block_group->lock);
3158                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3159                 spin_unlock(&block_group->lock);
3160                 return 0;
3161         }
3162
3163 again:
3164         inode = lookup_free_space_inode(root, block_group, path);
3165         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3166                 ret = PTR_ERR(inode);
3167                 btrfs_release_path(path);
3168                 goto out;
3169         }
3170
3171         if (IS_ERR(inode)) {
3172                 BUG_ON(retries);
3173                 retries++;
3174
3175                 if (block_group->ro)
3176                         goto out_free;
3177
3178                 ret = create_free_space_inode(root, trans, block_group, path);
3179                 if (ret)
3180                         goto out_free;
3181                 goto again;
3182         }
3183
3184         /* We've already setup this transaction, go ahead and exit */
3185         if (block_group->cache_generation == trans->transid &&
3186             i_size_read(inode)) {
3187                 dcs = BTRFS_DC_SETUP;
3188                 goto out_put;
3189         }
3190
3191         /*
3192          * We want to set the generation to 0, that way if anything goes wrong
3193          * from here on out we know not to trust this cache when we load up next
3194          * time.
3195          */
3196         BTRFS_I(inode)->generation = 0;
3197         ret = btrfs_update_inode(trans, root, inode);
3198         WARN_ON(ret);
3199
3200         if (i_size_read(inode) > 0) {
3201                 ret = btrfs_check_trunc_cache_free_space(root,
3202                                         &root->fs_info->global_block_rsv);
3203                 if (ret)
3204                         goto out_put;
3205
3206                 ret = btrfs_truncate_free_space_cache(root, trans, inode);
3207                 if (ret)
3208                         goto out_put;
3209         }
3210
3211         spin_lock(&block_group->lock);
3212         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3213             !btrfs_test_opt(root, SPACE_CACHE)) {
3214                 /*
3215                  * don't bother trying to write stuff out _if_
3216                  * a) we're not cached,
3217                  * b) we're with nospace_cache mount option.
3218                  */
3219                 dcs = BTRFS_DC_WRITTEN;
3220                 spin_unlock(&block_group->lock);
3221                 goto out_put;
3222         }
3223         spin_unlock(&block_group->lock);
3224
3225         /*
3226          * Try to preallocate enough space based on how big the block group is.
3227          * Keep in mind this has to include any pinned space which could end up
3228          * taking up quite a bit since it's not folded into the other space
3229          * cache.
3230          */
3231         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3232         if (!num_pages)
3233                 num_pages = 1;
3234
3235         num_pages *= 16;
3236         num_pages *= PAGE_CACHE_SIZE;
3237
3238         ret = btrfs_check_data_free_space(inode, num_pages);
3239         if (ret)
3240                 goto out_put;
3241
3242         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3243                                               num_pages, num_pages,
3244                                               &alloc_hint);
3245         if (!ret)
3246                 dcs = BTRFS_DC_SETUP;
3247         btrfs_free_reserved_data_space(inode, num_pages);
3248
3249 out_put:
3250         iput(inode);
3251 out_free:
3252         btrfs_release_path(path);
3253 out:
3254         spin_lock(&block_group->lock);
3255         if (!ret && dcs == BTRFS_DC_SETUP)
3256                 block_group->cache_generation = trans->transid;
3257         block_group->disk_cache_state = dcs;
3258         spin_unlock(&block_group->lock);
3259
3260         return ret;
3261 }
3262
3263 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3264                                    struct btrfs_root *root)
3265 {
3266         struct btrfs_block_group_cache *cache;
3267         int err = 0;
3268         struct btrfs_path *path;
3269         u64 last = 0;
3270
3271         path = btrfs_alloc_path();
3272         if (!path)
3273                 return -ENOMEM;
3274
3275 again:
3276         while (1) {
3277                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3278                 while (cache) {
3279                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3280                                 break;
3281                         cache = next_block_group(root, cache);
3282                 }
3283                 if (!cache) {
3284                         if (last == 0)
3285                                 break;
3286                         last = 0;
3287                         continue;
3288                 }
3289                 err = cache_save_setup(cache, trans, path);
3290                 last = cache->key.objectid + cache->key.offset;
3291                 btrfs_put_block_group(cache);
3292         }
3293
3294         while (1) {
3295                 if (last == 0) {
3296                         err = btrfs_run_delayed_refs(trans, root,
3297                                                      (unsigned long)-1);
3298                         if (err) /* File system offline */
3299                                 goto out;
3300                 }
3301
3302                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3303                 while (cache) {
3304                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3305                                 btrfs_put_block_group(cache);
3306                                 goto again;
3307                         }
3308
3309                         if (cache->dirty)
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                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3321                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3322                 cache->dirty = 0;
3323                 last = cache->key.objectid + cache->key.offset;
3324
3325                 err = write_one_cache_group(trans, root, path, cache);
3326                 btrfs_put_block_group(cache);
3327                 if (err) /* File system offline */
3328                         goto out;
3329         }
3330
3331         while (1) {
3332                 /*
3333                  * I don't think this is needed since we're just marking our
3334                  * preallocated extent as written, but just in case it can't
3335                  * hurt.
3336                  */
3337                 if (last == 0) {
3338                         err = btrfs_run_delayed_refs(trans, root,
3339                                                      (unsigned long)-1);
3340                         if (err) /* File system offline */
3341                                 goto out;
3342                 }
3343
3344                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3345                 while (cache) {
3346                         /*
3347                          * Really this shouldn't happen, but it could if we
3348                          * couldn't write the entire preallocated extent and
3349                          * splitting the extent resulted in a new block.
3350                          */
3351                         if (cache->dirty) {
3352                                 btrfs_put_block_group(cache);
3353                                 goto again;
3354                         }
3355                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3356                                 break;
3357                         cache = next_block_group(root, cache);
3358                 }
3359                 if (!cache) {
3360                         if (last == 0)
3361                                 break;
3362                         last = 0;
3363                         continue;
3364                 }
3365
3366                 err = btrfs_write_out_cache(root, trans, cache, path);
3367
3368                 /*
3369                  * If we didn't have an error then the cache state is still
3370                  * NEED_WRITE, so we can set it to WRITTEN.
3371                  */
3372                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3373                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3374                 last = cache->key.objectid + cache->key.offset;
3375                 btrfs_put_block_group(cache);
3376         }
3377 out:
3378
3379         btrfs_free_path(path);
3380         return err;
3381 }
3382
3383 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3384 {
3385         struct btrfs_block_group_cache *block_group;
3386         int readonly = 0;
3387
3388         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3389         if (!block_group || block_group->ro)
3390                 readonly = 1;
3391         if (block_group)
3392                 btrfs_put_block_group(block_group);
3393         return readonly;
3394 }
3395
3396 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3397                              u64 total_bytes, u64 bytes_used,
3398                              struct btrfs_space_info **space_info)
3399 {
3400         struct btrfs_space_info *found;
3401         int i;
3402         int factor;
3403         int ret;
3404
3405         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3406                      BTRFS_BLOCK_GROUP_RAID10))
3407                 factor = 2;
3408         else
3409                 factor = 1;
3410
3411         found = __find_space_info(info, flags);
3412         if (found) {
3413                 spin_lock(&found->lock);
3414                 found->total_bytes += total_bytes;
3415                 found->disk_total += total_bytes * factor;
3416                 found->bytes_used += bytes_used;
3417                 found->disk_used += bytes_used * factor;
3418                 found->full = 0;
3419                 spin_unlock(&found->lock);
3420                 *space_info = found;
3421                 return 0;
3422         }
3423         found = kzalloc(sizeof(*found), GFP_NOFS);
3424         if (!found)
3425                 return -ENOMEM;
3426
3427         ret = percpu_counter_init(&found->total_bytes_pinned, 0);
3428         if (ret) {
3429                 kfree(found);
3430                 return ret;
3431         }
3432
3433         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3434                 INIT_LIST_HEAD(&found->block_groups[i]);
3435         init_rwsem(&found->groups_sem);
3436         spin_lock_init(&found->lock);
3437         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3438         found->total_bytes = total_bytes;
3439         found->disk_total = total_bytes * factor;
3440         found->bytes_used = bytes_used;
3441         found->disk_used = bytes_used * factor;
3442         found->bytes_pinned = 0;
3443         found->bytes_reserved = 0;
3444         found->bytes_readonly = 0;
3445         found->bytes_may_use = 0;
3446         found->full = 0;
3447         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3448         found->chunk_alloc = 0;
3449         found->flush = 0;
3450         init_waitqueue_head(&found->wait);
3451         *space_info = found;
3452         list_add_rcu(&found->list, &info->space_info);
3453         if (flags & BTRFS_BLOCK_GROUP_DATA)
3454                 info->data_sinfo = found;
3455         return 0;
3456 }
3457
3458 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3459 {
3460         u64 extra_flags = chunk_to_extended(flags) &
3461                                 BTRFS_EXTENDED_PROFILE_MASK;
3462
3463         write_seqlock(&fs_info->profiles_lock);
3464         if (flags & BTRFS_BLOCK_GROUP_DATA)
3465                 fs_info->avail_data_alloc_bits |= extra_flags;
3466         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3467                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3468         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3469                 fs_info->avail_system_alloc_bits |= extra_flags;
3470         write_sequnlock(&fs_info->profiles_lock);
3471 }
3472
3473 /*
3474  * returns target flags in extended format or 0 if restripe for this
3475  * chunk_type is not in progress
3476  *
3477  * should be called with either volume_mutex or balance_lock held
3478  */
3479 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3480 {
3481         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3482         u64 target = 0;
3483
3484         if (!bctl)
3485                 return 0;
3486
3487         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3488             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3489                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3490         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3491                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3492                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3493         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3494                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3495                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3496         }
3497
3498         return target;
3499 }
3500
3501 /*
3502  * @flags: available profiles in extended format (see ctree.h)
3503  *
3504  * Returns reduced profile in chunk format.  If profile changing is in
3505  * progress (either running or paused) picks the target profile (if it's
3506  * already available), otherwise falls back to plain reducing.
3507  */
3508 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3509 {
3510         /*
3511          * we add in the count of missing devices because we want
3512          * to make sure that any RAID levels on a degraded FS
3513          * continue to be honored.
3514          */
3515         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3516                 root->fs_info->fs_devices->missing_devices;
3517         u64 target;
3518         u64 tmp;
3519
3520         /*
3521          * see if restripe for this chunk_type is in progress, if so
3522          * try to reduce to the target profile
3523          */
3524         spin_lock(&root->fs_info->balance_lock);
3525         target = get_restripe_target(root->fs_info, flags);
3526         if (target) {
3527                 /* pick target profile only if it's already available */
3528                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3529                         spin_unlock(&root->fs_info->balance_lock);
3530                         return extended_to_chunk(target);
3531                 }
3532         }
3533         spin_unlock(&root->fs_info->balance_lock);
3534
3535         /* First, mask out the RAID levels which aren't possible */
3536         if (num_devices == 1)
3537                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3538                            BTRFS_BLOCK_GROUP_RAID5);
3539         if (num_devices < 3)
3540                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3541         if (num_devices < 4)
3542                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3543
3544         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3545                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3546                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3547         flags &= ~tmp;
3548
3549         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3550                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3551         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3552                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3553         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3554                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3555         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3556                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3557         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3558                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3559
3560         return extended_to_chunk(flags | tmp);
3561 }
3562
3563 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3564 {
3565         unsigned seq;
3566
3567         do {
3568                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3569
3570                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3571                         flags |= root->fs_info->avail_data_alloc_bits;
3572                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3573                         flags |= root->fs_info->avail_system_alloc_bits;
3574                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3575                         flags |= root->fs_info->avail_metadata_alloc_bits;
3576         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3577
3578         return btrfs_reduce_alloc_profile(root, flags);
3579 }
3580
3581 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3582 {
3583         u64 flags;
3584         u64 ret;
3585
3586         if (data)
3587                 flags = BTRFS_BLOCK_GROUP_DATA;
3588         else if (root == root->fs_info->chunk_root)
3589                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3590         else
3591                 flags = BTRFS_BLOCK_GROUP_METADATA;
3592
3593         ret = get_alloc_profile(root, flags);
3594         return ret;
3595 }
3596
3597 /*
3598  * This will check the space that the inode allocates from to make sure we have
3599  * enough space for bytes.
3600  */
3601 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3602 {
3603         struct btrfs_space_info *data_sinfo;
3604         struct btrfs_root *root = BTRFS_I(inode)->root;
3605         struct btrfs_fs_info *fs_info = root->fs_info;
3606         u64 used;
3607         int ret = 0, committed = 0, alloc_chunk = 1;
3608
3609         /* make sure bytes are sectorsize aligned */
3610         bytes = ALIGN(bytes, root->sectorsize);
3611
3612         if (root == root->fs_info->tree_root ||
3613             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3614                 alloc_chunk = 0;
3615                 committed = 1;
3616         }
3617
3618         data_sinfo = fs_info->data_sinfo;
3619         if (!data_sinfo)
3620                 goto alloc;
3621
3622 again:
3623         /* make sure we have enough space to handle the data first */
3624         spin_lock(&data_sinfo->lock);
3625         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3626                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3627                 data_sinfo->bytes_may_use;
3628
3629         if (used + bytes > data_sinfo->total_bytes) {
3630                 struct btrfs_trans_handle *trans;
3631
3632                 /*
3633                  * if we don't have enough free bytes in this space then we need
3634                  * to alloc a new chunk.
3635                  */
3636                 if (!data_sinfo->full && alloc_chunk) {
3637                         u64 alloc_target;
3638
3639                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3640                         spin_unlock(&data_sinfo->lock);
3641 alloc:
3642                         alloc_target = btrfs_get_alloc_profile(root, 1);
3643                         trans = btrfs_join_transaction(root);
3644                         if (IS_ERR(trans))
3645                                 return PTR_ERR(trans);
3646
3647                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3648                                              alloc_target,
3649                                              CHUNK_ALLOC_NO_FORCE);
3650                         btrfs_end_transaction(trans, root);
3651                         if (ret < 0) {
3652                                 if (ret != -ENOSPC)
3653                                         return ret;
3654                                 else
3655                                         goto commit_trans;
3656                         }
3657
3658                         if (!data_sinfo)
3659                                 data_sinfo = fs_info->data_sinfo;
3660
3661                         goto again;
3662                 }
3663
3664                 /*
3665                  * If we don't have enough pinned space to deal with this
3666                  * allocation don't bother committing the transaction.
3667                  */
3668                 if (percpu_counter_compare(&data_sinfo->total_bytes_pinned,
3669                                            bytes) < 0)
3670                         committed = 1;
3671                 spin_unlock(&data_sinfo->lock);
3672
3673                 /* commit the current transaction and try again */
3674 commit_trans:
3675                 if (!committed &&
3676                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3677                         committed = 1;
3678
3679                         trans = btrfs_join_transaction(root);
3680                         if (IS_ERR(trans))
3681                                 return PTR_ERR(trans);
3682                         ret = btrfs_commit_transaction(trans, root);
3683                         if (ret)
3684                                 return ret;
3685                         goto again;
3686                 }
3687
3688                 return -ENOSPC;
3689         }
3690         data_sinfo->bytes_may_use += bytes;
3691         trace_btrfs_space_reservation(root->fs_info, "space_info",
3692                                       data_sinfo->flags, bytes, 1);
3693         spin_unlock(&data_sinfo->lock);
3694
3695         return 0;
3696 }
3697
3698 /*
3699  * Called if we need to clear a data reservation for this inode.
3700  */
3701 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3702 {
3703         struct btrfs_root *root = BTRFS_I(inode)->root;
3704         struct btrfs_space_info *data_sinfo;
3705
3706         /* make sure bytes are sectorsize aligned */
3707         bytes = ALIGN(bytes, root->sectorsize);
3708
3709         data_sinfo = root->fs_info->data_sinfo;
3710         spin_lock(&data_sinfo->lock);
3711         WARN_ON(data_sinfo->bytes_may_use < bytes);
3712         data_sinfo->bytes_may_use -= bytes;
3713         trace_btrfs_space_reservation(root->fs_info, "space_info",
3714                                       data_sinfo->flags, bytes, 0);
3715         spin_unlock(&data_sinfo->lock);
3716 }
3717
3718 static void force_metadata_allocation(struct btrfs_fs_info *info)
3719 {
3720         struct list_head *head = &info->space_info;
3721         struct btrfs_space_info *found;
3722
3723         rcu_read_lock();
3724         list_for_each_entry_rcu(found, head, list) {
3725                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3726                         found->force_alloc = CHUNK_ALLOC_FORCE;
3727         }
3728         rcu_read_unlock();
3729 }
3730
3731 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3732 {
3733         return (global->size << 1);
3734 }
3735
3736 static int should_alloc_chunk(struct btrfs_root *root,
3737                               struct btrfs_space_info *sinfo, int force)
3738 {
3739         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3740         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3741         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3742         u64 thresh;
3743
3744         if (force == CHUNK_ALLOC_FORCE)
3745                 return 1;
3746
3747         /*
3748          * We need to take into account the global rsv because for all intents
3749          * and purposes it's used space.  Don't worry about locking the
3750          * global_rsv, it doesn't change except when the transaction commits.
3751          */
3752         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3753                 num_allocated += calc_global_rsv_need_space(global_rsv);
3754
3755         /*
3756          * in limited mode, we want to have some free space up to
3757          * about 1% of the FS size.
3758          */
3759         if (force == CHUNK_ALLOC_LIMITED) {
3760                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3761                 thresh = max_t(u64, 64 * 1024 * 1024,
3762                                div_factor_fine(thresh, 1));
3763
3764                 if (num_bytes - num_allocated < thresh)
3765                         return 1;
3766         }
3767
3768         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3769                 return 0;
3770         return 1;
3771 }
3772
3773 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3774 {
3775         u64 num_dev;
3776
3777         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3778                     BTRFS_BLOCK_GROUP_RAID0 |
3779                     BTRFS_BLOCK_GROUP_RAID5 |
3780                     BTRFS_BLOCK_GROUP_RAID6))
3781                 num_dev = root->fs_info->fs_devices->rw_devices;
3782         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3783                 num_dev = 2;
3784         else
3785                 num_dev = 1;    /* DUP or single */
3786
3787         /* metadata for updaing devices and chunk tree */
3788         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3789 }
3790
3791 static void check_system_chunk(struct btrfs_trans_handle *trans,
3792                                struct btrfs_root *root, u64 type)
3793 {
3794         struct btrfs_space_info *info;
3795         u64 left;
3796         u64 thresh;
3797
3798         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3799         spin_lock(&info->lock);
3800         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3801                 info->bytes_reserved - info->bytes_readonly;
3802         spin_unlock(&info->lock);
3803
3804         thresh = get_system_chunk_thresh(root, type);
3805         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3806                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3807                         left, thresh, type);
3808                 dump_space_info(info, 0, 0);
3809         }
3810
3811         if (left < thresh) {
3812                 u64 flags;
3813
3814                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3815                 btrfs_alloc_chunk(trans, root, flags);
3816         }
3817 }
3818
3819 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3820                           struct btrfs_root *extent_root, u64 flags, int force)
3821 {
3822         struct btrfs_space_info *space_info;
3823         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3824         int wait_for_alloc = 0;
3825         int ret = 0;
3826
3827         /* Don't re-enter if we're already allocating a chunk */
3828         if (trans->allocating_chunk)
3829                 return -ENOSPC;
3830
3831         space_info = __find_space_info(extent_root->fs_info, flags);
3832         if (!space_info) {
3833                 ret = update_space_info(extent_root->fs_info, flags,
3834                                         0, 0, &space_info);
3835                 BUG_ON(ret); /* -ENOMEM */
3836         }
3837         BUG_ON(!space_info); /* Logic error */
3838
3839 again:
3840         spin_lock(&space_info->lock);
3841         if (force < space_info->force_alloc)
3842                 force = space_info->force_alloc;
3843         if (space_info->full) {
3844                 if (should_alloc_chunk(extent_root, space_info, force))
3845                         ret = -ENOSPC;
3846                 else
3847                         ret = 0;
3848                 spin_unlock(&space_info->lock);
3849                 return ret;
3850         }
3851
3852         if (!should_alloc_chunk(extent_root, space_info, force)) {
3853                 spin_unlock(&space_info->lock);
3854                 return 0;
3855         } else if (space_info->chunk_alloc) {
3856                 wait_for_alloc = 1;
3857         } else {
3858                 space_info->chunk_alloc = 1;
3859         }
3860
3861         spin_unlock(&space_info->lock);
3862
3863         mutex_lock(&fs_info->chunk_mutex);
3864
3865         /*
3866          * The chunk_mutex is held throughout the entirety of a chunk
3867          * allocation, so once we've acquired the chunk_mutex we know that the
3868          * other guy is done and we need to recheck and see if we should
3869          * allocate.
3870          */
3871         if (wait_for_alloc) {
3872                 mutex_unlock(&fs_info->chunk_mutex);
3873                 wait_for_alloc = 0;
3874                 goto again;
3875         }
3876
3877         trans->allocating_chunk = true;
3878
3879         /*
3880          * If we have mixed data/metadata chunks we want to make sure we keep
3881          * allocating mixed chunks instead of individual chunks.
3882          */
3883         if (btrfs_mixed_space_info(space_info))
3884                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3885
3886         /*
3887          * if we're doing a data chunk, go ahead and make sure that
3888          * we keep a reasonable number of metadata chunks allocated in the
3889          * FS as well.
3890          */
3891         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3892                 fs_info->data_chunk_allocations++;
3893                 if (!(fs_info->data_chunk_allocations %
3894                       fs_info->metadata_ratio))
3895                         force_metadata_allocation(fs_info);
3896         }
3897
3898         /*
3899          * Check if we have enough space in SYSTEM chunk because we may need
3900          * to update devices.
3901          */
3902         check_system_chunk(trans, extent_root, flags);
3903
3904         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3905         trans->allocating_chunk = false;
3906
3907         spin_lock(&space_info->lock);
3908         if (ret < 0 && ret != -ENOSPC)
3909                 goto out;
3910         if (ret)
3911                 space_info->full = 1;
3912         else
3913                 ret = 1;
3914
3915         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3916 out:
3917         space_info->chunk_alloc = 0;
3918         spin_unlock(&space_info->lock);
3919         mutex_unlock(&fs_info->chunk_mutex);
3920         return ret;
3921 }
3922
3923 static int can_overcommit(struct btrfs_root *root,
3924                           struct btrfs_space_info *space_info, u64 bytes,
3925                           enum btrfs_reserve_flush_enum flush)
3926 {
3927         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3928         u64 profile = btrfs_get_alloc_profile(root, 0);
3929         u64 space_size;
3930         u64 avail;
3931         u64 used;
3932
3933         used = space_info->bytes_used + space_info->bytes_reserved +
3934                 space_info->bytes_pinned + space_info->bytes_readonly;
3935
3936         /*
3937          * We only want to allow over committing if we have lots of actual space
3938          * free, but if we don't have enough space to handle the global reserve
3939          * space then we could end up having a real enospc problem when trying
3940          * to allocate a chunk or some other such important allocation.
3941          */
3942         spin_lock(&global_rsv->lock);
3943         space_size = calc_global_rsv_need_space(global_rsv);
3944         spin_unlock(&global_rsv->lock);
3945         if (used + space_size >= space_info->total_bytes)
3946                 return 0;
3947
3948         used += space_info->bytes_may_use;
3949
3950         spin_lock(&root->fs_info->free_chunk_lock);
3951         avail = root->fs_info->free_chunk_space;
3952         spin_unlock(&root->fs_info->free_chunk_lock);
3953
3954         /*
3955          * If we have dup, raid1 or raid10 then only half of the free
3956          * space is actually useable.  For raid56, the space info used
3957          * doesn't include the parity drive, so we don't have to
3958          * change the math
3959          */
3960         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3961                        BTRFS_BLOCK_GROUP_RAID1 |
3962                        BTRFS_BLOCK_GROUP_RAID10))
3963                 avail >>= 1;
3964
3965         /*
3966          * If we aren't flushing all things, let us overcommit up to
3967          * 1/2th of the space. If we can flush, don't let us overcommit
3968          * too much, let it overcommit up to 1/8 of the space.
3969          */
3970         if (flush == BTRFS_RESERVE_FLUSH_ALL)
3971                 avail >>= 3;
3972         else
3973                 avail >>= 1;
3974
3975         if (used + bytes < space_info->total_bytes + avail)
3976                 return 1;
3977         return 0;
3978 }
3979
3980 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3981                                          unsigned long nr_pages)
3982 {
3983         struct super_block *sb = root->fs_info->sb;
3984
3985         if (down_read_trylock(&sb->s_umount)) {
3986                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
3987                 up_read(&sb->s_umount);
3988         } else {
3989                 /*
3990                  * We needn't worry the filesystem going from r/w to r/o though
3991                  * we don't acquire ->s_umount mutex, because the filesystem
3992                  * should guarantee the delalloc inodes list be empty after
3993                  * the filesystem is readonly(all dirty pages are written to
3994                  * the disk).
3995                  */
3996                 btrfs_start_all_delalloc_inodes(root->fs_info, 0);
3997                 if (!current->journal_info)
3998                         btrfs_wait_all_ordered_extents(root->fs_info);
3999         }
4000 }
4001
4002 /*
4003  * shrink metadata reservation for delalloc
4004  */
4005 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4006                             bool wait_ordered)
4007 {
4008         struct btrfs_block_rsv *block_rsv;
4009         struct btrfs_space_info *space_info;
4010         struct btrfs_trans_handle *trans;
4011         u64 delalloc_bytes;
4012         u64 max_reclaim;
4013         long time_left;
4014         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
4015         int loops = 0;
4016         enum btrfs_reserve_flush_enum flush;
4017
4018         trans = (struct btrfs_trans_handle *)current->journal_info;
4019         block_rsv = &root->fs_info->delalloc_block_rsv;
4020         space_info = block_rsv->space_info;
4021
4022         smp_mb();
4023         delalloc_bytes = percpu_counter_sum_positive(
4024                                                 &root->fs_info->delalloc_bytes);
4025         if (delalloc_bytes == 0) {
4026                 if (trans)
4027                         return;
4028                 btrfs_wait_all_ordered_extents(root->fs_info);
4029                 return;
4030         }
4031
4032         while (delalloc_bytes && loops < 3) {
4033                 max_reclaim = min(delalloc_bytes, to_reclaim);
4034                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4035                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
4036                 /*
4037                  * We need to wait for the async pages to actually start before
4038                  * we do anything.
4039                  */
4040                 wait_event(root->fs_info->async_submit_wait,
4041                            !atomic_read(&root->fs_info->async_delalloc_pages));
4042
4043                 if (!trans)
4044                         flush = BTRFS_RESERVE_FLUSH_ALL;
4045                 else
4046                         flush = BTRFS_RESERVE_NO_FLUSH;
4047                 spin_lock(&space_info->lock);
4048                 if (can_overcommit(root, space_info, orig, flush)) {
4049                         spin_unlock(&space_info->lock);
4050                         break;
4051                 }
4052                 spin_unlock(&space_info->lock);
4053
4054                 loops++;
4055                 if (wait_ordered && !trans) {
4056                         btrfs_wait_all_ordered_extents(root->fs_info);
4057                 } else {
4058                         time_left = schedule_timeout_killable(1);
4059                         if (time_left)
4060                                 break;
4061                 }
4062                 smp_mb();
4063                 delalloc_bytes = percpu_counter_sum_positive(
4064                                                 &root->fs_info->delalloc_bytes);
4065         }
4066 }
4067
4068 /**
4069  * maybe_commit_transaction - possibly commit the transaction if its ok to
4070  * @root - the root we're allocating for
4071  * @bytes - the number of bytes we want to reserve
4072  * @force - force the commit
4073  *
4074  * This will check to make sure that committing the transaction will actually
4075  * get us somewhere and then commit the transaction if it does.  Otherwise it
4076  * will return -ENOSPC.
4077  */
4078 static int may_commit_transaction(struct btrfs_root *root,
4079                                   struct btrfs_space_info *space_info,
4080                                   u64 bytes, int force)
4081 {
4082         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4083         struct btrfs_trans_handle *trans;
4084
4085         trans = (struct btrfs_trans_handle *)current->journal_info;
4086         if (trans)
4087                 return -EAGAIN;
4088
4089         if (force)
4090                 goto commit;
4091
4092         /* See if there is enough pinned space to make this reservation */
4093         spin_lock(&space_info->lock);
4094         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4095                                    bytes) >= 0) {
4096                 spin_unlock(&space_info->lock);
4097                 goto commit;
4098         }
4099         spin_unlock(&space_info->lock);
4100
4101         /*
4102          * See if there is some space in the delayed insertion reservation for
4103          * this reservation.
4104          */
4105         if (space_info != delayed_rsv->space_info)
4106                 return -ENOSPC;
4107
4108         spin_lock(&space_info->lock);
4109         spin_lock(&delayed_rsv->lock);
4110         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4111                                    bytes - delayed_rsv->size) >= 0) {
4112                 spin_unlock(&delayed_rsv->lock);
4113                 spin_unlock(&space_info->lock);
4114                 return -ENOSPC;
4115         }
4116         spin_unlock(&delayed_rsv->lock);
4117         spin_unlock(&space_info->lock);
4118
4119 commit:
4120         trans = btrfs_join_transaction(root);
4121         if (IS_ERR(trans))
4122                 return -ENOSPC;
4123
4124         return btrfs_commit_transaction(trans, root);
4125 }
4126
4127 enum flush_state {
4128         FLUSH_DELAYED_ITEMS_NR  =       1,
4129         FLUSH_DELAYED_ITEMS     =       2,
4130         FLUSH_DELALLOC          =       3,
4131         FLUSH_DELALLOC_WAIT     =       4,
4132         ALLOC_CHUNK             =       5,
4133         COMMIT_TRANS            =       6,
4134 };
4135
4136 static int flush_space(struct btrfs_root *root,
4137                        struct btrfs_space_info *space_info, u64 num_bytes,
4138                        u64 orig_bytes, int state)
4139 {
4140         struct btrfs_trans_handle *trans;
4141         int nr;
4142         int ret = 0;
4143
4144         switch (state) {
4145         case FLUSH_DELAYED_ITEMS_NR:
4146         case FLUSH_DELAYED_ITEMS:
4147                 if (state == FLUSH_DELAYED_ITEMS_NR) {
4148                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
4149
4150                         nr = (int)div64_u64(num_bytes, bytes);
4151                         if (!nr)
4152                                 nr = 1;
4153                         nr *= 2;
4154                 } else {
4155                         nr = -1;
4156                 }
4157                 trans = btrfs_join_transaction(root);
4158                 if (IS_ERR(trans)) {
4159                         ret = PTR_ERR(trans);
4160                         break;
4161                 }
4162                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4163                 btrfs_end_transaction(trans, root);
4164                 break;
4165         case FLUSH_DELALLOC:
4166         case FLUSH_DELALLOC_WAIT:
4167                 shrink_delalloc(root, num_bytes, orig_bytes,
4168                                 state == FLUSH_DELALLOC_WAIT);
4169                 break;
4170         case ALLOC_CHUNK:
4171                 trans = btrfs_join_transaction(root);
4172                 if (IS_ERR(trans)) {
4173                         ret = PTR_ERR(trans);
4174                         break;
4175                 }
4176                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4177                                      btrfs_get_alloc_profile(root, 0),
4178                                      CHUNK_ALLOC_NO_FORCE);
4179                 btrfs_end_transaction(trans, root);
4180                 if (ret == -ENOSPC)
4181                         ret = 0;
4182                 break;
4183         case COMMIT_TRANS:
4184                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4185                 break;
4186         default:
4187                 ret = -ENOSPC;
4188                 break;
4189         }
4190
4191         return ret;
4192 }
4193 /**
4194  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4195  * @root - the root we're allocating for
4196  * @block_rsv - the block_rsv we're allocating for
4197  * @orig_bytes - the number of bytes we want
4198  * @flush - whether or not we can flush to make our reservation
4199  *
4200  * This will reserve orgi_bytes number of bytes from the space info associated
4201  * with the block_rsv.  If there is not enough space it will make an attempt to
4202  * flush out space to make room.  It will do this by flushing delalloc if
4203  * possible or committing the transaction.  If flush is 0 then no attempts to
4204  * regain reservations will be made and this will fail if there is not enough
4205  * space already.
4206  */
4207 static int reserve_metadata_bytes(struct btrfs_root *root,
4208                                   struct btrfs_block_rsv *block_rsv,
4209                                   u64 orig_bytes,
4210                                   enum btrfs_reserve_flush_enum flush)
4211 {
4212         struct btrfs_space_info *space_info = block_rsv->space_info;
4213         u64 used;
4214         u64 num_bytes = orig_bytes;
4215         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4216         int ret = 0;
4217         bool flushing = false;
4218
4219 again:
4220         ret = 0;
4221         spin_lock(&space_info->lock);
4222         /*
4223          * We only want to wait if somebody other than us is flushing and we
4224          * are actually allowed to flush all things.
4225          */
4226         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4227                space_info->flush) {
4228                 spin_unlock(&space_info->lock);
4229                 /*
4230                  * If we have a trans handle we can't wait because the flusher
4231                  * may have to commit the transaction, which would mean we would
4232                  * deadlock since we are waiting for the flusher to finish, but
4233                  * hold the current transaction open.
4234                  */
4235                 if (current->journal_info)
4236                         return -EAGAIN;
4237                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4238                 /* Must have been killed, return */
4239                 if (ret)
4240                         return -EINTR;
4241
4242                 spin_lock(&space_info->lock);
4243         }
4244
4245         ret = -ENOSPC;
4246         used = space_info->bytes_used + space_info->bytes_reserved +
4247                 space_info->bytes_pinned + space_info->bytes_readonly +
4248                 space_info->bytes_may_use;
4249
4250         /*
4251          * The idea here is that we've not already over-reserved the block group
4252          * then we can go ahead and save our reservation first and then start
4253          * flushing if we need to.  Otherwise if we've already overcommitted
4254          * lets start flushing stuff first and then come back and try to make
4255          * our reservation.
4256          */
4257         if (used <= space_info->total_bytes) {
4258                 if (used + orig_bytes <= space_info->total_bytes) {
4259                         space_info->bytes_may_use += orig_bytes;
4260                         trace_btrfs_space_reservation(root->fs_info,
4261                                 "space_info", space_info->flags, orig_bytes, 1);
4262                         ret = 0;
4263                 } else {
4264                         /*
4265                          * Ok set num_bytes to orig_bytes since we aren't
4266                          * overocmmitted, this way we only try and reclaim what
4267                          * we need.
4268                          */
4269                         num_bytes = orig_bytes;
4270                 }
4271         } else {
4272                 /*
4273                  * Ok we're over committed, set num_bytes to the overcommitted
4274                  * amount plus the amount of bytes that we need for this
4275                  * reservation.
4276                  */
4277                 num_bytes = used - space_info->total_bytes +
4278                         (orig_bytes * 2);
4279         }
4280
4281         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4282                 space_info->bytes_may_use += orig_bytes;
4283                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4284                                               space_info->flags, orig_bytes,
4285                                               1);
4286                 ret = 0;
4287         }
4288
4289         /*
4290          * Couldn't make our reservation, save our place so while we're trying
4291          * to reclaim space we can actually use it instead of somebody else
4292          * stealing it from us.
4293          *
4294          * We make the other tasks wait for the flush only when we can flush
4295          * all things.
4296          */
4297         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4298                 flushing = true;
4299                 space_info->flush = 1;
4300         }
4301
4302         spin_unlock(&space_info->lock);
4303
4304         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4305                 goto out;
4306
4307         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4308                           flush_state);
4309         flush_state++;
4310
4311         /*
4312          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4313          * would happen. So skip delalloc flush.
4314          */
4315         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4316             (flush_state == FLUSH_DELALLOC ||
4317              flush_state == FLUSH_DELALLOC_WAIT))
4318                 flush_state = ALLOC_CHUNK;
4319
4320         if (!ret)
4321                 goto again;
4322         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4323                  flush_state < COMMIT_TRANS)
4324                 goto again;
4325         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4326                  flush_state <= COMMIT_TRANS)
4327                 goto again;
4328
4329 out:
4330         if (ret == -ENOSPC &&
4331             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4332                 struct btrfs_block_rsv *global_rsv =
4333                         &root->fs_info->global_block_rsv;
4334
4335                 if (block_rsv != global_rsv &&
4336                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4337                         ret = 0;
4338         }
4339         if (flushing) {
4340                 spin_lock(&space_info->lock);
4341                 space_info->flush = 0;
4342                 wake_up_all(&space_info->wait);
4343                 spin_unlock(&space_info->lock);
4344         }
4345         return ret;
4346 }
4347
4348 static struct btrfs_block_rsv *get_block_rsv(
4349                                         const struct btrfs_trans_handle *trans,
4350                                         const struct btrfs_root *root)
4351 {
4352         struct btrfs_block_rsv *block_rsv = NULL;
4353
4354         if (root->ref_cows)
4355                 block_rsv = trans->block_rsv;
4356
4357         if (root == root->fs_info->csum_root && trans->adding_csums)
4358                 block_rsv = trans->block_rsv;
4359
4360         if (root == root->fs_info->uuid_root)
4361                 block_rsv = trans->block_rsv;
4362
4363         if (!block_rsv)
4364                 block_rsv = root->block_rsv;
4365
4366         if (!block_rsv)
4367                 block_rsv = &root->fs_info->empty_block_rsv;
4368
4369         return block_rsv;
4370 }
4371
4372 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4373                                u64 num_bytes)
4374 {
4375         int ret = -ENOSPC;
4376         spin_lock(&block_rsv->lock);
4377         if (block_rsv->reserved >= num_bytes) {
4378                 block_rsv->reserved -= num_bytes;
4379                 if (block_rsv->reserved < block_rsv->size)
4380                         block_rsv->full = 0;
4381                 ret = 0;
4382         }
4383         spin_unlock(&block_rsv->lock);
4384         return ret;
4385 }
4386
4387 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4388                                 u64 num_bytes, int update_size)
4389 {
4390         spin_lock(&block_rsv->lock);
4391         block_rsv->reserved += num_bytes;
4392         if (update_size)
4393                 block_rsv->size += num_bytes;
4394         else if (block_rsv->reserved >= block_rsv->size)
4395                 block_rsv->full = 1;
4396         spin_unlock(&block_rsv->lock);
4397 }
4398
4399 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4400                              struct btrfs_block_rsv *dest, u64 num_bytes,
4401                              int min_factor)
4402 {
4403         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4404         u64 min_bytes;
4405
4406         if (global_rsv->space_info != dest->space_info)
4407                 return -ENOSPC;
4408
4409         spin_lock(&global_rsv->lock);
4410         min_bytes = div_factor(global_rsv->size, min_factor);
4411         if (global_rsv->reserved < min_bytes + num_bytes) {
4412                 spin_unlock(&global_rsv->lock);
4413                 return -ENOSPC;
4414         }
4415         global_rsv->reserved -= num_bytes;
4416         if (global_rsv->reserved < global_rsv->size)
4417                 global_rsv->full = 0;
4418         spin_unlock(&global_rsv->lock);
4419
4420         block_rsv_add_bytes(dest, num_bytes, 1);
4421         return 0;
4422 }
4423
4424 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4425                                     struct btrfs_block_rsv *block_rsv,
4426                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4427 {
4428         struct btrfs_space_info *space_info = block_rsv->space_info;
4429
4430         spin_lock(&block_rsv->lock);
4431         if (num_bytes == (u64)-1)
4432                 num_bytes = block_rsv->size;
4433         block_rsv->size -= num_bytes;
4434         if (block_rsv->reserved >= block_rsv->size) {
4435                 num_bytes = block_rsv->reserved - block_rsv->size;
4436                 block_rsv->reserved = block_rsv->size;
4437                 block_rsv->full = 1;
4438         } else {
4439                 num_bytes = 0;
4440         }
4441         spin_unlock(&block_rsv->lock);
4442
4443         if (num_bytes > 0) {
4444                 if (dest) {
4445                         spin_lock(&dest->lock);
4446                         if (!dest->full) {
4447                                 u64 bytes_to_add;
4448
4449                                 bytes_to_add = dest->size - dest->reserved;
4450                                 bytes_to_add = min(num_bytes, bytes_to_add);
4451                                 dest->reserved += bytes_to_add;
4452                                 if (dest->reserved >= dest->size)
4453                                         dest->full = 1;
4454                                 num_bytes -= bytes_to_add;
4455                         }
4456                         spin_unlock(&dest->lock);
4457                 }
4458                 if (num_bytes) {
4459                         spin_lock(&space_info->lock);
4460                         space_info->bytes_may_use -= num_bytes;
4461                         trace_btrfs_space_reservation(fs_info, "space_info",
4462                                         space_info->flags, num_bytes, 0);
4463                         spin_unlock(&space_info->lock);
4464                 }
4465         }
4466 }
4467
4468 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4469                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4470 {
4471         int ret;
4472
4473         ret = block_rsv_use_bytes(src, num_bytes);
4474         if (ret)
4475                 return ret;
4476
4477         block_rsv_add_bytes(dst, num_bytes, 1);
4478         return 0;
4479 }
4480
4481 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4482 {
4483         memset(rsv, 0, sizeof(*rsv));
4484         spin_lock_init(&rsv->lock);
4485         rsv->type = type;
4486 }
4487
4488 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4489                                               unsigned short type)
4490 {
4491         struct btrfs_block_rsv *block_rsv;
4492         struct btrfs_fs_info *fs_info = root->fs_info;
4493
4494         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4495         if (!block_rsv)
4496                 return NULL;
4497
4498         btrfs_init_block_rsv(block_rsv, type);
4499         block_rsv->space_info = __find_space_info(fs_info,
4500                                                   BTRFS_BLOCK_GROUP_METADATA);
4501         return block_rsv;
4502 }
4503
4504 void btrfs_free_block_rsv(struct btrfs_root *root,
4505                           struct btrfs_block_rsv *rsv)
4506 {
4507         if (!rsv)
4508                 return;
4509         btrfs_block_rsv_release(root, rsv, (u64)-1);
4510         kfree(rsv);
4511 }
4512
4513 int btrfs_block_rsv_add(struct btrfs_root *root,
4514                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4515                         enum btrfs_reserve_flush_enum flush)
4516 {
4517         int ret;
4518
4519         if (num_bytes == 0)
4520                 return 0;
4521
4522         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4523         if (!ret) {
4524                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4525                 return 0;
4526         }
4527
4528         return ret;
4529 }
4530
4531 int btrfs_block_rsv_check(struct btrfs_root *root,
4532                           struct btrfs_block_rsv *block_rsv, int min_factor)
4533 {
4534         u64 num_bytes = 0;
4535         int ret = -ENOSPC;
4536
4537         if (!block_rsv)
4538                 return 0;
4539
4540         spin_lock(&block_rsv->lock);
4541         num_bytes = div_factor(block_rsv->size, min_factor);
4542         if (block_rsv->reserved >= num_bytes)
4543                 ret = 0;
4544         spin_unlock(&block_rsv->lock);
4545
4546         return ret;
4547 }
4548
4549 int btrfs_block_rsv_refill(struct btrfs_root *root,
4550                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4551                            enum btrfs_reserve_flush_enum flush)
4552 {
4553         u64 num_bytes = 0;
4554         int ret = -ENOSPC;
4555
4556         if (!block_rsv)
4557                 return 0;
4558
4559         spin_lock(&block_rsv->lock);
4560         num_bytes = min_reserved;
4561         if (block_rsv->reserved >= num_bytes)
4562                 ret = 0;
4563         else
4564                 num_bytes -= block_rsv->reserved;
4565         spin_unlock(&block_rsv->lock);
4566
4567         if (!ret)
4568                 return 0;
4569
4570         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4571         if (!ret) {
4572                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4573                 return 0;
4574         }
4575
4576         return ret;
4577 }
4578
4579 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4580                             struct btrfs_block_rsv *dst_rsv,
4581                             u64 num_bytes)
4582 {
4583         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4584 }
4585
4586 void btrfs_block_rsv_release(struct btrfs_root *root,
4587                              struct btrfs_block_rsv *block_rsv,
4588                              u64 num_bytes)
4589 {
4590         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4591         if (global_rsv->full || global_rsv == block_rsv ||
4592             block_rsv->space_info != global_rsv->space_info)
4593                 global_rsv = NULL;
4594         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4595                                 num_bytes);
4596 }
4597
4598 /*
4599  * helper to calculate size of global block reservation.
4600  * the desired value is sum of space used by extent tree,
4601  * checksum tree and root tree
4602  */
4603 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4604 {
4605         struct btrfs_space_info *sinfo;
4606         u64 num_bytes;
4607         u64 meta_used;
4608         u64 data_used;
4609         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4610
4611         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4612         spin_lock(&sinfo->lock);
4613         data_used = sinfo->bytes_used;
4614         spin_unlock(&sinfo->lock);
4615
4616         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4617         spin_lock(&sinfo->lock);
4618         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4619                 data_used = 0;
4620         meta_used = sinfo->bytes_used;
4621         spin_unlock(&sinfo->lock);
4622
4623         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4624                     csum_size * 2;
4625         num_bytes += div64_u64(data_used + meta_used, 50);
4626
4627         if (num_bytes * 3 > meta_used)
4628                 num_bytes = div64_u64(meta_used, 3);
4629
4630         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4631 }
4632
4633 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4634 {
4635         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4636         struct btrfs_space_info *sinfo = block_rsv->space_info;
4637         u64 num_bytes;
4638
4639         num_bytes = calc_global_metadata_size(fs_info);
4640
4641         spin_lock(&sinfo->lock);
4642         spin_lock(&block_rsv->lock);
4643
4644         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4645
4646         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4647                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4648                     sinfo->bytes_may_use;
4649
4650         if (sinfo->total_bytes > num_bytes) {
4651                 num_bytes = sinfo->total_bytes - num_bytes;
4652                 block_rsv->reserved += num_bytes;
4653                 sinfo->bytes_may_use += num_bytes;
4654                 trace_btrfs_space_reservation(fs_info, "space_info",
4655                                       sinfo->flags, num_bytes, 1);
4656         }
4657
4658         if (block_rsv->reserved >= block_rsv->size) {
4659                 num_bytes = block_rsv->reserved - block_rsv->size;
4660                 sinfo->bytes_may_use -= num_bytes;
4661                 trace_btrfs_space_reservation(fs_info, "space_info",
4662                                       sinfo->flags, num_bytes, 0);
4663                 block_rsv->reserved = block_rsv->size;
4664                 block_rsv->full = 1;
4665         }
4666
4667         spin_unlock(&block_rsv->lock);
4668         spin_unlock(&sinfo->lock);
4669 }
4670
4671 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4672 {
4673         struct btrfs_space_info *space_info;
4674
4675         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4676         fs_info->chunk_block_rsv.space_info = space_info;
4677
4678         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4679         fs_info->global_block_rsv.space_info = space_info;
4680         fs_info->delalloc_block_rsv.space_info = space_info;
4681         fs_info->trans_block_rsv.space_info = space_info;
4682         fs_info->empty_block_rsv.space_info = space_info;
4683         fs_info->delayed_block_rsv.space_info = space_info;
4684
4685         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4686         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4687         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4688         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4689         if (fs_info->quota_root)
4690                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
4691         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4692
4693         update_global_block_rsv(fs_info);
4694 }
4695
4696 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4697 {
4698         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4699                                 (u64)-1);
4700         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4701         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4702         WARN_ON(fs_info->trans_block_rsv.size > 0);
4703         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4704         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4705         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4706         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4707         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4708 }
4709
4710 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4711                                   struct btrfs_root *root)
4712 {
4713         if (!trans->block_rsv)
4714                 return;
4715
4716         if (!trans->bytes_reserved)
4717                 return;
4718
4719         trace_btrfs_space_reservation(root->fs_info, "transaction",
4720                                       trans->transid, trans->bytes_reserved, 0);
4721         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4722         trans->bytes_reserved = 0;
4723 }
4724
4725 /* Can only return 0 or -ENOSPC */
4726 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4727                                   struct inode *inode)
4728 {
4729         struct btrfs_root *root = BTRFS_I(inode)->root;
4730         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4731         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4732
4733         /*
4734          * We need to hold space in order to delete our orphan item once we've
4735          * added it, so this takes the reservation so we can release it later
4736          * when we are truly done with the orphan item.
4737          */
4738         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4739         trace_btrfs_space_reservation(root->fs_info, "orphan",
4740                                       btrfs_ino(inode), num_bytes, 1);
4741         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4742 }
4743
4744 void btrfs_orphan_release_metadata(struct inode *inode)
4745 {
4746         struct btrfs_root *root = BTRFS_I(inode)->root;
4747         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4748         trace_btrfs_space_reservation(root->fs_info, "orphan",
4749                                       btrfs_ino(inode), num_bytes, 0);
4750         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4751 }
4752
4753 /*
4754  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4755  * root: the root of the parent directory
4756  * rsv: block reservation
4757  * items: the number of items that we need do reservation
4758  * qgroup_reserved: used to return the reserved size in qgroup
4759  *
4760  * This function is used to reserve the space for snapshot/subvolume
4761  * creation and deletion. Those operations are different with the
4762  * common file/directory operations, they change two fs/file trees
4763  * and root tree, the number of items that the qgroup reserves is
4764  * different with the free space reservation. So we can not use
4765  * the space reseravtion mechanism in start_transaction().
4766  */
4767 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4768                                      struct btrfs_block_rsv *rsv,
4769                                      int items,
4770                                      u64 *qgroup_reserved,
4771                                      bool use_global_rsv)
4772 {
4773         u64 num_bytes;
4774         int ret;
4775         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4776
4777         if (root->fs_info->quota_enabled) {
4778                 /* One for parent inode, two for dir entries */
4779                 num_bytes = 3 * root->leafsize;
4780                 ret = btrfs_qgroup_reserve(root, num_bytes);
4781                 if (ret)
4782                         return ret;
4783         } else {
4784                 num_bytes = 0;
4785         }
4786
4787         *qgroup_reserved = num_bytes;
4788
4789         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4790         rsv->space_info = __find_space_info(root->fs_info,
4791                                             BTRFS_BLOCK_GROUP_METADATA);
4792         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4793                                   BTRFS_RESERVE_FLUSH_ALL);
4794
4795         if (ret == -ENOSPC && use_global_rsv)
4796                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
4797
4798         if (ret) {
4799                 if (*qgroup_reserved)
4800                         btrfs_qgroup_free(root, *qgroup_reserved);
4801         }
4802
4803         return ret;
4804 }
4805
4806 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4807                                       struct btrfs_block_rsv *rsv,
4808                                       u64 qgroup_reserved)
4809 {
4810         btrfs_block_rsv_release(root, rsv, (u64)-1);
4811         if (qgroup_reserved)
4812                 btrfs_qgroup_free(root, qgroup_reserved);
4813 }
4814
4815 /**
4816  * drop_outstanding_extent - drop an outstanding extent
4817  * @inode: the inode we're dropping the extent for
4818  *
4819  * This is called when we are freeing up an outstanding extent, either called
4820  * after an error or after an extent is written.  This will return the number of
4821  * reserved extents that need to be freed.  This must be called with
4822  * BTRFS_I(inode)->lock held.
4823  */
4824 static unsigned drop_outstanding_extent(struct inode *inode)
4825 {
4826         unsigned drop_inode_space = 0;
4827         unsigned dropped_extents = 0;
4828
4829         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4830         BTRFS_I(inode)->outstanding_extents--;
4831
4832         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4833             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4834                                &BTRFS_I(inode)->runtime_flags))
4835                 drop_inode_space = 1;
4836
4837         /*
4838          * If we have more or the same amount of outsanding extents than we have
4839          * reserved then we need to leave the reserved extents count alone.
4840          */
4841         if (BTRFS_I(inode)->outstanding_extents >=
4842             BTRFS_I(inode)->reserved_extents)
4843                 return drop_inode_space;
4844
4845         dropped_extents = BTRFS_I(inode)->reserved_extents -
4846                 BTRFS_I(inode)->outstanding_extents;
4847         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4848         return dropped_extents + drop_inode_space;
4849 }
4850
4851 /**
4852  * calc_csum_metadata_size - return the amount of metada space that must be
4853  *      reserved/free'd for the given bytes.
4854  * @inode: the inode we're manipulating
4855  * @num_bytes: the number of bytes in question
4856  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4857  *
4858  * This adjusts the number of csum_bytes in the inode and then returns the
4859  * correct amount of metadata that must either be reserved or freed.  We
4860  * calculate how many checksums we can fit into one leaf and then divide the
4861  * number of bytes that will need to be checksumed by this value to figure out
4862  * how many checksums will be required.  If we are adding bytes then the number
4863  * may go up and we will return the number of additional bytes that must be
4864  * reserved.  If it is going down we will return the number of bytes that must
4865  * be freed.
4866  *
4867  * This must be called with BTRFS_I(inode)->lock held.
4868  */
4869 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4870                                    int reserve)
4871 {
4872         struct btrfs_root *root = BTRFS_I(inode)->root;
4873         u64 csum_size;
4874         int num_csums_per_leaf;
4875         int num_csums;
4876         int old_csums;
4877
4878         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4879             BTRFS_I(inode)->csum_bytes == 0)
4880                 return 0;
4881
4882         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4883         if (reserve)
4884                 BTRFS_I(inode)->csum_bytes += num_bytes;
4885         else
4886                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4887         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4888         num_csums_per_leaf = (int)div64_u64(csum_size,
4889                                             sizeof(struct btrfs_csum_item) +
4890                                             sizeof(struct btrfs_disk_key));
4891         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4892         num_csums = num_csums + num_csums_per_leaf - 1;
4893         num_csums = num_csums / num_csums_per_leaf;
4894
4895         old_csums = old_csums + num_csums_per_leaf - 1;
4896         old_csums = old_csums / num_csums_per_leaf;
4897
4898         /* No change, no need to reserve more */
4899         if (old_csums == num_csums)
4900                 return 0;
4901
4902         if (reserve)
4903                 return btrfs_calc_trans_metadata_size(root,
4904                                                       num_csums - old_csums);
4905
4906         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4907 }
4908
4909 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4910 {
4911         struct btrfs_root *root = BTRFS_I(inode)->root;
4912         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4913         u64 to_reserve = 0;
4914         u64 csum_bytes;
4915         unsigned nr_extents = 0;
4916         int extra_reserve = 0;
4917         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
4918         int ret = 0;
4919         bool delalloc_lock = true;
4920         u64 to_free = 0;
4921         unsigned dropped;
4922
4923         /* If we are a free space inode we need to not flush since we will be in
4924          * the middle of a transaction commit.  We also don't need the delalloc
4925          * mutex since we won't race with anybody.  We need this mostly to make
4926          * lockdep shut its filthy mouth.
4927          */
4928         if (btrfs_is_free_space_inode(inode)) {
4929                 flush = BTRFS_RESERVE_NO_FLUSH;
4930                 delalloc_lock = false;
4931         }
4932
4933         if (flush != BTRFS_RESERVE_NO_FLUSH &&
4934             btrfs_transaction_in_commit(root->fs_info))
4935                 schedule_timeout(1);
4936
4937         if (delalloc_lock)
4938                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4939
4940         num_bytes = ALIGN(num_bytes, root->sectorsize);
4941
4942         spin_lock(&BTRFS_I(inode)->lock);
4943         BTRFS_I(inode)->outstanding_extents++;
4944
4945         if (BTRFS_I(inode)->outstanding_extents >
4946             BTRFS_I(inode)->reserved_extents)
4947                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4948                         BTRFS_I(inode)->reserved_extents;
4949
4950         /*
4951          * Add an item to reserve for updating the inode when we complete the
4952          * delalloc io.
4953          */
4954         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4955                       &BTRFS_I(inode)->runtime_flags)) {
4956                 nr_extents++;
4957                 extra_reserve = 1;
4958         }
4959
4960         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4961         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4962         csum_bytes = BTRFS_I(inode)->csum_bytes;
4963         spin_unlock(&BTRFS_I(inode)->lock);
4964
4965         if (root->fs_info->quota_enabled) {
4966                 ret = btrfs_qgroup_reserve(root, num_bytes +
4967                                            nr_extents * root->leafsize);
4968                 if (ret)
4969                         goto out_fail;
4970         }
4971
4972         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4973         if (unlikely(ret)) {
4974                 if (root->fs_info->quota_enabled)
4975                         btrfs_qgroup_free(root, num_bytes +
4976                                                 nr_extents * root->leafsize);
4977                 goto out_fail;
4978         }
4979
4980         spin_lock(&BTRFS_I(inode)->lock);
4981         if (extra_reserve) {
4982                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4983                         &BTRFS_I(inode)->runtime_flags);
4984                 nr_extents--;
4985         }
4986         BTRFS_I(inode)->reserved_extents += nr_extents;
4987         spin_unlock(&BTRFS_I(inode)->lock);
4988
4989         if (delalloc_lock)
4990                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4991
4992         if (to_reserve)
4993                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4994                                               btrfs_ino(inode), to_reserve, 1);
4995         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4996
4997         return 0;
4998
4999 out_fail:
5000         spin_lock(&BTRFS_I(inode)->lock);
5001         dropped = drop_outstanding_extent(inode);
5002         /*
5003          * If the inodes csum_bytes is the same as the original
5004          * csum_bytes then we know we haven't raced with any free()ers
5005          * so we can just reduce our inodes csum bytes and carry on.
5006          */
5007         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5008                 calc_csum_metadata_size(inode, num_bytes, 0);
5009         } else {
5010                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5011                 u64 bytes;
5012
5013                 /*
5014                  * This is tricky, but first we need to figure out how much we
5015                  * free'd from any free-ers that occured during this
5016                  * reservation, so we reset ->csum_bytes to the csum_bytes
5017                  * before we dropped our lock, and then call the free for the
5018                  * number of bytes that were freed while we were trying our
5019                  * reservation.
5020                  */
5021                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5022                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5023                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5024
5025
5026                 /*
5027                  * Now we need to see how much we would have freed had we not
5028                  * been making this reservation and our ->csum_bytes were not
5029                  * artificially inflated.
5030                  */
5031                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5032                 bytes = csum_bytes - orig_csum_bytes;
5033                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5034
5035                 /*
5036                  * Now reset ->csum_bytes to what it should be.  If bytes is
5037                  * more than to_free then we would have free'd more space had we
5038                  * not had an artificially high ->csum_bytes, so we need to free
5039                  * the remainder.  If bytes is the same or less then we don't
5040                  * need to do anything, the other free-ers did the correct
5041                  * thing.
5042                  */
5043                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5044                 if (bytes > to_free)
5045                         to_free = bytes - to_free;
5046                 else
5047                         to_free = 0;
5048         }
5049         spin_unlock(&BTRFS_I(inode)->lock);
5050         if (dropped)
5051                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5052
5053         if (to_free) {
5054                 btrfs_block_rsv_release(root, block_rsv, to_free);
5055                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5056                                               btrfs_ino(inode), to_free, 0);
5057         }
5058         if (delalloc_lock)
5059                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5060         return ret;
5061 }
5062
5063 /**
5064  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5065  * @inode: the inode to release the reservation for
5066  * @num_bytes: the number of bytes we're releasing
5067  *
5068  * This will release the metadata reservation for an inode.  This can be called
5069  * once we complete IO for a given set of bytes to release their metadata
5070  * reservations.
5071  */
5072 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5073 {
5074         struct btrfs_root *root = BTRFS_I(inode)->root;
5075         u64 to_free = 0;
5076         unsigned dropped;
5077
5078         num_bytes = ALIGN(num_bytes, root->sectorsize);
5079         spin_lock(&BTRFS_I(inode)->lock);
5080         dropped = drop_outstanding_extent(inode);
5081
5082         if (num_bytes)
5083                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5084         spin_unlock(&BTRFS_I(inode)->lock);
5085         if (dropped > 0)
5086                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5087
5088         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5089                                       btrfs_ino(inode), to_free, 0);
5090         if (root->fs_info->quota_enabled) {
5091                 btrfs_qgroup_free(root, num_bytes +
5092                                         dropped * root->leafsize);
5093         }
5094
5095         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5096                                 to_free);
5097 }
5098
5099 /**
5100  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5101  * @inode: inode we're writing to
5102  * @num_bytes: the number of bytes we want to allocate
5103  *
5104  * This will do the following things
5105  *
5106  * o reserve space in the data space info for num_bytes
5107  * o reserve space in the metadata space info based on number of outstanding
5108  *   extents and how much csums will be needed
5109  * o add to the inodes ->delalloc_bytes
5110  * o add it to the fs_info's delalloc inodes list.
5111  *
5112  * This will return 0 for success and -ENOSPC if there is no space left.
5113  */
5114 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5115 {
5116         int ret;
5117
5118         ret = btrfs_check_data_free_space(inode, num_bytes);
5119         if (ret)
5120                 return ret;
5121
5122         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5123         if (ret) {
5124                 btrfs_free_reserved_data_space(inode, num_bytes);
5125                 return ret;
5126         }
5127
5128         return 0;
5129 }
5130
5131 /**
5132  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5133  * @inode: inode we're releasing space for
5134  * @num_bytes: the number of bytes we want to free up
5135  *
5136  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5137  * called in the case that we don't need the metadata AND data reservations
5138  * anymore.  So if there is an error or we insert an inline extent.
5139  *
5140  * This function will release the metadata space that was not used and will
5141  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5142  * list if there are no delalloc bytes left.
5143  */
5144 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5145 {
5146         btrfs_delalloc_release_metadata(inode, num_bytes);
5147         btrfs_free_reserved_data_space(inode, num_bytes);
5148 }
5149
5150 static int update_block_group(struct btrfs_root *root,
5151                               u64 bytenr, u64 num_bytes, int alloc)
5152 {
5153         struct btrfs_block_group_cache *cache = NULL;
5154         struct btrfs_fs_info *info = root->fs_info;
5155         u64 total = num_bytes;
5156         u64 old_val;
5157         u64 byte_in_group;
5158         int factor;
5159
5160         /* block accounting for super block */
5161         spin_lock(&info->delalloc_root_lock);
5162         old_val = btrfs_super_bytes_used(info->super_copy);
5163         if (alloc)
5164                 old_val += num_bytes;
5165         else
5166                 old_val -= num_bytes;
5167         btrfs_set_super_bytes_used(info->super_copy, old_val);
5168         spin_unlock(&info->delalloc_root_lock);
5169
5170         while (total) {
5171                 cache = btrfs_lookup_block_group(info, bytenr);
5172                 if (!cache)
5173                         return -ENOENT;
5174                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5175                                     BTRFS_BLOCK_GROUP_RAID1 |
5176                                     BTRFS_BLOCK_GROUP_RAID10))
5177                         factor = 2;
5178                 else
5179                         factor = 1;
5180                 /*
5181                  * If this block group has free space cache written out, we
5182                  * need to make sure to load it if we are removing space.  This
5183                  * is because we need the unpinning stage to actually add the
5184                  * space back to the block group, otherwise we will leak space.
5185                  */
5186                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5187                         cache_block_group(cache, 1);
5188
5189                 byte_in_group = bytenr - cache->key.objectid;
5190                 WARN_ON(byte_in_group > cache->key.offset);
5191
5192                 spin_lock(&cache->space_info->lock);
5193                 spin_lock(&cache->lock);
5194
5195                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5196                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5197                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5198
5199                 cache->dirty = 1;
5200                 old_val = btrfs_block_group_used(&cache->item);
5201                 num_bytes = min(total, cache->key.offset - byte_in_group);
5202                 if (alloc) {
5203                         old_val += num_bytes;
5204                         btrfs_set_block_group_used(&cache->item, old_val);
5205                         cache->reserved -= num_bytes;
5206                         cache->space_info->bytes_reserved -= num_bytes;
5207                         cache->space_info->bytes_used += num_bytes;
5208                         cache->space_info->disk_used += num_bytes * factor;
5209                         spin_unlock(&cache->lock);
5210                         spin_unlock(&cache->space_info->lock);
5211                 } else {
5212                         old_val -= num_bytes;
5213                         btrfs_set_block_group_used(&cache->item, old_val);
5214                         cache->pinned += num_bytes;
5215                         cache->space_info->bytes_pinned += num_bytes;
5216                         cache->space_info->bytes_used -= num_bytes;
5217                         cache->space_info->disk_used -= num_bytes * factor;
5218                         spin_unlock(&cache->lock);
5219                         spin_unlock(&cache->space_info->lock);
5220
5221                         set_extent_dirty(info->pinned_extents,
5222                                          bytenr, bytenr + num_bytes - 1,
5223                                          GFP_NOFS | __GFP_NOFAIL);
5224                 }
5225                 btrfs_put_block_group(cache);
5226                 total -= num_bytes;
5227                 bytenr += num_bytes;
5228         }
5229         return 0;
5230 }
5231
5232 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5233 {
5234         struct btrfs_block_group_cache *cache;
5235         u64 bytenr;
5236
5237         spin_lock(&root->fs_info->block_group_cache_lock);
5238         bytenr = root->fs_info->first_logical_byte;
5239         spin_unlock(&root->fs_info->block_group_cache_lock);
5240
5241         if (bytenr < (u64)-1)
5242                 return bytenr;
5243
5244         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5245         if (!cache)
5246                 return 0;
5247
5248         bytenr = cache->key.objectid;
5249         btrfs_put_block_group(cache);
5250
5251         return bytenr;
5252 }
5253
5254 static int pin_down_extent(struct btrfs_root *root,
5255                            struct btrfs_block_group_cache *cache,
5256                            u64 bytenr, u64 num_bytes, int reserved)
5257 {
5258         spin_lock(&cache->space_info->lock);
5259         spin_lock(&cache->lock);
5260         cache->pinned += num_bytes;
5261         cache->space_info->bytes_pinned += num_bytes;
5262         if (reserved) {
5263                 cache->reserved -= num_bytes;
5264                 cache->space_info->bytes_reserved -= num_bytes;
5265         }
5266         spin_unlock(&cache->lock);
5267         spin_unlock(&cache->space_info->lock);
5268
5269         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5270                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5271         if (reserved)
5272                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5273         return 0;
5274 }
5275
5276 /*
5277  * this function must be called within transaction
5278  */
5279 int btrfs_pin_extent(struct btrfs_root *root,
5280                      u64 bytenr, u64 num_bytes, int reserved)
5281 {
5282         struct btrfs_block_group_cache *cache;
5283
5284         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5285         BUG_ON(!cache); /* Logic error */
5286
5287         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5288
5289         btrfs_put_block_group(cache);
5290         return 0;
5291 }
5292
5293 /*
5294  * this function must be called within transaction
5295  */
5296 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5297                                     u64 bytenr, u64 num_bytes)
5298 {
5299         struct btrfs_block_group_cache *cache;
5300         int ret;
5301
5302         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5303         if (!cache)
5304                 return -EINVAL;
5305
5306         /*
5307          * pull in the free space cache (if any) so that our pin
5308          * removes the free space from the cache.  We have load_only set
5309          * to one because the slow code to read in the free extents does check
5310          * the pinned extents.
5311          */
5312         cache_block_group(cache, 1);
5313
5314         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5315
5316         /* remove us from the free space cache (if we're there at all) */
5317         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5318         btrfs_put_block_group(cache);
5319         return ret;
5320 }
5321
5322 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5323 {
5324         int ret;
5325         struct btrfs_block_group_cache *block_group;
5326         struct btrfs_caching_control *caching_ctl;
5327
5328         block_group = btrfs_lookup_block_group(root->fs_info, start);
5329         if (!block_group)
5330                 return -EINVAL;
5331
5332         cache_block_group(block_group, 0);
5333         caching_ctl = get_caching_control(block_group);
5334
5335         if (!caching_ctl) {
5336                 /* Logic error */
5337                 BUG_ON(!block_group_cache_done(block_group));
5338                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5339         } else {
5340                 mutex_lock(&caching_ctl->mutex);
5341
5342                 if (start >= caching_ctl->progress) {
5343                         ret = add_excluded_extent(root, start, num_bytes);
5344                 } else if (start + num_bytes <= caching_ctl->progress) {
5345                         ret = btrfs_remove_free_space(block_group,
5346                                                       start, num_bytes);
5347                 } else {
5348                         num_bytes = caching_ctl->progress - start;
5349                         ret = btrfs_remove_free_space(block_group,
5350                                                       start, num_bytes);
5351                         if (ret)
5352                                 goto out_lock;
5353
5354                         num_bytes = (start + num_bytes) -
5355                                 caching_ctl->progress;
5356                         start = caching_ctl->progress;
5357                         ret = add_excluded_extent(root, start, num_bytes);
5358                 }
5359 out_lock:
5360                 mutex_unlock(&caching_ctl->mutex);
5361                 put_caching_control(caching_ctl);
5362         }
5363         btrfs_put_block_group(block_group);
5364         return ret;
5365 }
5366
5367 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5368                                  struct extent_buffer *eb)
5369 {
5370         struct btrfs_file_extent_item *item;
5371         struct btrfs_key key;
5372         int found_type;
5373         int i;
5374
5375         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5376                 return 0;
5377
5378         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5379                 btrfs_item_key_to_cpu(eb, &key, i);
5380                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5381                         continue;
5382                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5383                 found_type = btrfs_file_extent_type(eb, item);
5384                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5385                         continue;
5386                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5387                         continue;
5388                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5389                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5390                 __exclude_logged_extent(log, key.objectid, key.offset);
5391         }
5392
5393         return 0;
5394 }
5395
5396 /**
5397  * btrfs_update_reserved_bytes - update the block_group and space info counters
5398  * @cache:      The cache we are manipulating
5399  * @num_bytes:  The number of bytes in question
5400  * @reserve:    One of the reservation enums
5401  *
5402  * This is called by the allocator when it reserves space, or by somebody who is
5403  * freeing space that was never actually used on disk.  For example if you
5404  * reserve some space for a new leaf in transaction A and before transaction A
5405  * commits you free that leaf, you call this with reserve set to 0 in order to
5406  * clear the reservation.
5407  *
5408  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5409  * ENOSPC accounting.  For data we handle the reservation through clearing the
5410  * delalloc bits in the io_tree.  We have to do this since we could end up
5411  * allocating less disk space for the amount of data we have reserved in the
5412  * case of compression.
5413  *
5414  * If this is a reservation and the block group has become read only we cannot
5415  * make the reservation and return -EAGAIN, otherwise this function always
5416  * succeeds.
5417  */
5418 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5419                                        u64 num_bytes, int reserve)
5420 {
5421         struct btrfs_space_info *space_info = cache->space_info;
5422         int ret = 0;
5423
5424         spin_lock(&space_info->lock);
5425         spin_lock(&cache->lock);
5426         if (reserve != RESERVE_FREE) {
5427                 if (cache->ro) {
5428                         ret = -EAGAIN;
5429                 } else {
5430                         cache->reserved += num_bytes;
5431                         space_info->bytes_reserved += num_bytes;
5432                         if (reserve == RESERVE_ALLOC) {
5433                                 trace_btrfs_space_reservation(cache->fs_info,
5434                                                 "space_info", space_info->flags,
5435                                                 num_bytes, 0);
5436                                 space_info->bytes_may_use -= num_bytes;
5437                         }
5438                 }
5439         } else {
5440                 if (cache->ro)
5441                         space_info->bytes_readonly += num_bytes;
5442                 cache->reserved -= num_bytes;
5443                 space_info->bytes_reserved -= num_bytes;
5444         }
5445         spin_unlock(&cache->lock);
5446         spin_unlock(&space_info->lock);
5447         return ret;
5448 }
5449
5450 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5451                                 struct btrfs_root *root)
5452 {
5453         struct btrfs_fs_info *fs_info = root->fs_info;
5454         struct btrfs_caching_control *next;
5455         struct btrfs_caching_control *caching_ctl;
5456         struct btrfs_block_group_cache *cache;
5457         struct btrfs_space_info *space_info;
5458
5459         down_write(&fs_info->extent_commit_sem);
5460
5461         list_for_each_entry_safe(caching_ctl, next,
5462                                  &fs_info->caching_block_groups, list) {
5463                 cache = caching_ctl->block_group;
5464                 if (block_group_cache_done(cache)) {
5465                         cache->last_byte_to_unpin = (u64)-1;
5466                         list_del_init(&caching_ctl->list);
5467                         put_caching_control(caching_ctl);
5468                 } else {
5469                         cache->last_byte_to_unpin = caching_ctl->progress;
5470                 }
5471         }
5472
5473         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5474                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5475         else
5476                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5477
5478         up_write(&fs_info->extent_commit_sem);
5479
5480         list_for_each_entry_rcu(space_info, &fs_info->space_info, list)
5481                 percpu_counter_set(&space_info->total_bytes_pinned, 0);
5482
5483         update_global_block_rsv(fs_info);
5484 }
5485
5486 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5487 {
5488         struct btrfs_fs_info *fs_info = root->fs_info;
5489         struct btrfs_block_group_cache *cache = NULL;
5490         struct btrfs_space_info *space_info;
5491         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5492         u64 len;
5493         bool readonly;
5494
5495         while (start <= end) {
5496                 readonly = false;
5497                 if (!cache ||
5498                     start >= cache->key.objectid + cache->key.offset) {
5499                         if (cache)
5500                                 btrfs_put_block_group(cache);
5501                         cache = btrfs_lookup_block_group(fs_info, start);
5502                         BUG_ON(!cache); /* Logic error */
5503                 }
5504
5505                 len = cache->key.objectid + cache->key.offset - start;
5506                 len = min(len, end + 1 - start);
5507
5508                 if (start < cache->last_byte_to_unpin) {
5509                         len = min(len, cache->last_byte_to_unpin - start);
5510                         btrfs_add_free_space(cache, start, len);
5511                 }
5512
5513                 start += len;
5514                 space_info = cache->space_info;
5515
5516                 spin_lock(&space_info->lock);
5517                 spin_lock(&cache->lock);
5518                 cache->pinned -= len;
5519                 space_info->bytes_pinned -= len;
5520                 if (cache->ro) {
5521                         space_info->bytes_readonly += len;
5522                         readonly = true;
5523                 }
5524                 spin_unlock(&cache->lock);
5525                 if (!readonly && global_rsv->space_info == space_info) {
5526                         spin_lock(&global_rsv->lock);
5527                         if (!global_rsv->full) {
5528                                 len = min(len, global_rsv->size -
5529                                           global_rsv->reserved);
5530                                 global_rsv->reserved += len;
5531                                 space_info->bytes_may_use += len;
5532                                 if (global_rsv->reserved >= global_rsv->size)
5533                                         global_rsv->full = 1;
5534                         }
5535                         spin_unlock(&global_rsv->lock);
5536                 }
5537                 spin_unlock(&space_info->lock);
5538         }
5539
5540         if (cache)
5541                 btrfs_put_block_group(cache);
5542         return 0;
5543 }
5544
5545 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5546                                struct btrfs_root *root)
5547 {
5548         struct btrfs_fs_info *fs_info = root->fs_info;
5549         struct extent_io_tree *unpin;
5550         u64 start;
5551         u64 end;
5552         int ret;
5553
5554         if (trans->aborted)
5555                 return 0;
5556
5557         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5558                 unpin = &fs_info->freed_extents[1];
5559         else
5560                 unpin = &fs_info->freed_extents[0];
5561
5562         while (1) {
5563                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5564                                             EXTENT_DIRTY, NULL);
5565                 if (ret)
5566                         break;
5567
5568                 if (btrfs_test_opt(root, DISCARD))
5569                         ret = btrfs_discard_extent(root, start,
5570                                                    end + 1 - start, NULL);
5571
5572                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5573                 unpin_extent_range(root, start, end);
5574                 cond_resched();
5575         }
5576
5577         return 0;
5578 }
5579
5580 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
5581                              u64 owner, u64 root_objectid)
5582 {
5583         struct btrfs_space_info *space_info;
5584         u64 flags;
5585
5586         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5587                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
5588                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
5589                 else
5590                         flags = BTRFS_BLOCK_GROUP_METADATA;
5591         } else {
5592                 flags = BTRFS_BLOCK_GROUP_DATA;
5593         }
5594
5595         space_info = __find_space_info(fs_info, flags);
5596         BUG_ON(!space_info); /* Logic bug */
5597         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
5598 }
5599
5600
5601 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5602                                 struct btrfs_root *root,
5603                                 u64 bytenr, u64 num_bytes, u64 parent,
5604                                 u64 root_objectid, u64 owner_objectid,
5605                                 u64 owner_offset, int refs_to_drop,
5606                                 struct btrfs_delayed_extent_op *extent_op)
5607 {
5608         struct btrfs_key key;
5609         struct btrfs_path *path;
5610         struct btrfs_fs_info *info = root->fs_info;
5611         struct btrfs_root *extent_root = info->extent_root;
5612         struct extent_buffer *leaf;
5613         struct btrfs_extent_item *ei;
5614         struct btrfs_extent_inline_ref *iref;
5615         int ret;
5616         int is_data;
5617         int extent_slot = 0;
5618         int found_extent = 0;
5619         int num_to_del = 1;
5620         u32 item_size;
5621         u64 refs;
5622         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5623                                                  SKINNY_METADATA);
5624
5625         path = btrfs_alloc_path();
5626         if (!path)
5627                 return -ENOMEM;
5628
5629         path->reada = 1;
5630         path->leave_spinning = 1;
5631
5632         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5633         BUG_ON(!is_data && refs_to_drop != 1);
5634
5635         if (is_data)
5636                 skinny_metadata = 0;
5637
5638         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5639                                     bytenr, num_bytes, parent,
5640                                     root_objectid, owner_objectid,
5641                                     owner_offset);
5642         if (ret == 0) {
5643                 extent_slot = path->slots[0];
5644                 while (extent_slot >= 0) {
5645                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5646                                               extent_slot);
5647                         if (key.objectid != bytenr)
5648                                 break;
5649                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5650                             key.offset == num_bytes) {
5651                                 found_extent = 1;
5652                                 break;
5653                         }
5654                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5655                             key.offset == owner_objectid) {
5656                                 found_extent = 1;
5657                                 break;
5658                         }
5659                         if (path->slots[0] - extent_slot > 5)
5660                                 break;
5661                         extent_slot--;
5662                 }
5663 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5664                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5665                 if (found_extent && item_size < sizeof(*ei))
5666                         found_extent = 0;
5667 #endif
5668                 if (!found_extent) {
5669                         BUG_ON(iref);
5670                         ret = remove_extent_backref(trans, extent_root, path,
5671                                                     NULL, refs_to_drop,
5672                                                     is_data);
5673                         if (ret) {
5674                                 btrfs_abort_transaction(trans, extent_root, ret);
5675                                 goto out;
5676                         }
5677                         btrfs_release_path(path);
5678                         path->leave_spinning = 1;
5679
5680                         key.objectid = bytenr;
5681                         key.type = BTRFS_EXTENT_ITEM_KEY;
5682                         key.offset = num_bytes;
5683
5684                         if (!is_data && skinny_metadata) {
5685                                 key.type = BTRFS_METADATA_ITEM_KEY;
5686                                 key.offset = owner_objectid;
5687                         }
5688
5689                         ret = btrfs_search_slot(trans, extent_root,
5690                                                 &key, path, -1, 1);
5691                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5692                                 /*
5693                                  * Couldn't find our skinny metadata item,
5694                                  * see if we have ye olde extent item.
5695                                  */
5696                                 path->slots[0]--;
5697                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5698                                                       path->slots[0]);
5699                                 if (key.objectid == bytenr &&
5700                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5701                                     key.offset == num_bytes)
5702                                         ret = 0;
5703                         }
5704
5705                         if (ret > 0 && skinny_metadata) {
5706                                 skinny_metadata = false;
5707                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5708                                 key.offset = num_bytes;
5709                                 btrfs_release_path(path);
5710                                 ret = btrfs_search_slot(trans, extent_root,
5711                                                         &key, path, -1, 1);
5712                         }
5713
5714                         if (ret) {
5715                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5716                                         ret, bytenr);
5717                                 if (ret > 0)
5718                                         btrfs_print_leaf(extent_root,
5719                                                          path->nodes[0]);
5720                         }
5721                         if (ret < 0) {
5722                                 btrfs_abort_transaction(trans, extent_root, ret);
5723                                 goto out;
5724                         }
5725                         extent_slot = path->slots[0];
5726                 }
5727         } else if (ret == -ENOENT) {
5728                 btrfs_print_leaf(extent_root, path->nodes[0]);
5729                 WARN_ON(1);
5730                 btrfs_err(info,
5731                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5732                         bytenr, parent, root_objectid, owner_objectid,
5733                         owner_offset);
5734         } else {
5735                 btrfs_abort_transaction(trans, extent_root, ret);
5736                 goto out;
5737         }
5738
5739         leaf = path->nodes[0];
5740         item_size = btrfs_item_size_nr(leaf, extent_slot);
5741 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5742         if (item_size < sizeof(*ei)) {
5743                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5744                 ret = convert_extent_item_v0(trans, extent_root, path,
5745                                              owner_objectid, 0);
5746                 if (ret < 0) {
5747                         btrfs_abort_transaction(trans, extent_root, ret);
5748                         goto out;
5749                 }
5750
5751                 btrfs_release_path(path);
5752                 path->leave_spinning = 1;
5753
5754                 key.objectid = bytenr;
5755                 key.type = BTRFS_EXTENT_ITEM_KEY;
5756                 key.offset = num_bytes;
5757
5758                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5759                                         -1, 1);
5760                 if (ret) {
5761                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5762                                 ret, bytenr);
5763                         btrfs_print_leaf(extent_root, path->nodes[0]);
5764                 }
5765                 if (ret < 0) {
5766                         btrfs_abort_transaction(trans, extent_root, ret);
5767                         goto out;
5768                 }
5769
5770                 extent_slot = path->slots[0];
5771                 leaf = path->nodes[0];
5772                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5773         }
5774 #endif
5775         BUG_ON(item_size < sizeof(*ei));
5776         ei = btrfs_item_ptr(leaf, extent_slot,
5777                             struct btrfs_extent_item);
5778         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
5779             key.type == BTRFS_EXTENT_ITEM_KEY) {
5780                 struct btrfs_tree_block_info *bi;
5781                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5782                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5783                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5784         }
5785
5786         refs = btrfs_extent_refs(leaf, ei);
5787         if (refs < refs_to_drop) {
5788                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
5789                           "for bytenr %Lu\n", refs_to_drop, refs, bytenr);
5790                 ret = -EINVAL;
5791                 btrfs_abort_transaction(trans, extent_root, ret);
5792                 goto out;
5793         }
5794         refs -= refs_to_drop;
5795
5796         if (refs > 0) {
5797                 if (extent_op)
5798                         __run_delayed_extent_op(extent_op, leaf, ei);
5799                 /*
5800                  * In the case of inline back ref, reference count will
5801                  * be updated by remove_extent_backref
5802                  */
5803                 if (iref) {
5804                         BUG_ON(!found_extent);
5805                 } else {
5806                         btrfs_set_extent_refs(leaf, ei, refs);
5807                         btrfs_mark_buffer_dirty(leaf);
5808                 }
5809                 if (found_extent) {
5810                         ret = remove_extent_backref(trans, extent_root, path,
5811                                                     iref, refs_to_drop,
5812                                                     is_data);
5813                         if (ret) {
5814                                 btrfs_abort_transaction(trans, extent_root, ret);
5815                                 goto out;
5816                         }
5817                 }
5818                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
5819                                  root_objectid);
5820         } else {
5821                 if (found_extent) {
5822                         BUG_ON(is_data && refs_to_drop !=
5823                                extent_data_ref_count(root, path, iref));
5824                         if (iref) {
5825                                 BUG_ON(path->slots[0] != extent_slot);
5826                         } else {
5827                                 BUG_ON(path->slots[0] != extent_slot + 1);
5828                                 path->slots[0] = extent_slot;
5829                                 num_to_del = 2;
5830                         }
5831                 }
5832
5833                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5834                                       num_to_del);
5835                 if (ret) {
5836                         btrfs_abort_transaction(trans, extent_root, ret);
5837                         goto out;
5838                 }
5839                 btrfs_release_path(path);
5840
5841                 if (is_data) {
5842                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5843                         if (ret) {
5844                                 btrfs_abort_transaction(trans, extent_root, ret);
5845                                 goto out;
5846                         }
5847                 }
5848
5849                 ret = update_block_group(root, bytenr, num_bytes, 0);
5850                 if (ret) {
5851                         btrfs_abort_transaction(trans, extent_root, ret);
5852                         goto out;
5853                 }
5854         }
5855 out:
5856         btrfs_free_path(path);
5857         return ret;
5858 }
5859
5860 /*
5861  * when we free an block, it is possible (and likely) that we free the last
5862  * delayed ref for that extent as well.  This searches the delayed ref tree for
5863  * a given extent, and if there are no other delayed refs to be processed, it
5864  * removes it from the tree.
5865  */
5866 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5867                                       struct btrfs_root *root, u64 bytenr)
5868 {
5869         struct btrfs_delayed_ref_head *head;
5870         struct btrfs_delayed_ref_root *delayed_refs;
5871         struct btrfs_delayed_ref_node *ref;
5872         struct rb_node *node;
5873         int ret = 0;
5874
5875         delayed_refs = &trans->transaction->delayed_refs;
5876         spin_lock(&delayed_refs->lock);
5877         head = btrfs_find_delayed_ref_head(trans, bytenr);
5878         if (!head)
5879                 goto out;
5880
5881         node = rb_prev(&head->node.rb_node);
5882         if (!node)
5883                 goto out;
5884
5885         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5886
5887         /* there are still entries for this ref, we can't drop it */
5888         if (ref->bytenr == bytenr)
5889                 goto out;
5890
5891         if (head->extent_op) {
5892                 if (!head->must_insert_reserved)
5893                         goto out;
5894                 btrfs_free_delayed_extent_op(head->extent_op);
5895                 head->extent_op = NULL;
5896         }
5897
5898         /*
5899          * waiting for the lock here would deadlock.  If someone else has it
5900          * locked they are already in the process of dropping it anyway
5901          */
5902         if (!mutex_trylock(&head->mutex))
5903                 goto out;
5904
5905         /*
5906          * at this point we have a head with no other entries.  Go
5907          * ahead and process it.
5908          */
5909         head->node.in_tree = 0;
5910         rb_erase(&head->node.rb_node, &delayed_refs->root);
5911
5912         delayed_refs->num_entries--;
5913
5914         /*
5915          * we don't take a ref on the node because we're removing it from the
5916          * tree, so we just steal the ref the tree was holding.
5917          */
5918         delayed_refs->num_heads--;
5919         if (list_empty(&head->cluster))
5920                 delayed_refs->num_heads_ready--;
5921
5922         list_del_init(&head->cluster);
5923         spin_unlock(&delayed_refs->lock);
5924
5925         BUG_ON(head->extent_op);
5926         if (head->must_insert_reserved)
5927                 ret = 1;
5928
5929         mutex_unlock(&head->mutex);
5930         btrfs_put_delayed_ref(&head->node);
5931         return ret;
5932 out:
5933         spin_unlock(&delayed_refs->lock);
5934         return 0;
5935 }
5936
5937 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5938                            struct btrfs_root *root,
5939                            struct extent_buffer *buf,
5940                            u64 parent, int last_ref)
5941 {
5942         struct btrfs_block_group_cache *cache = NULL;
5943         int pin = 1;
5944         int ret;
5945
5946         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5947                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5948                                         buf->start, buf->len,
5949                                         parent, root->root_key.objectid,
5950                                         btrfs_header_level(buf),
5951                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5952                 BUG_ON(ret); /* -ENOMEM */
5953         }
5954
5955         if (!last_ref)
5956                 return;
5957
5958         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5959
5960         if (btrfs_header_generation(buf) == trans->transid) {
5961                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5962                         ret = check_ref_cleanup(trans, root, buf->start);
5963                         if (!ret)
5964                                 goto out;
5965                 }
5966
5967                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5968                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5969                         goto out;
5970                 }
5971
5972                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5973
5974                 btrfs_add_free_space(cache, buf->start, buf->len);
5975                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5976                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
5977                 pin = 0;
5978         }
5979 out:
5980         if (pin)
5981                 add_pinned_bytes(root->fs_info, buf->len,
5982                                  btrfs_header_level(buf),
5983                                  root->root_key.objectid);
5984
5985         /*
5986          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5987          * anymore.
5988          */
5989         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5990         btrfs_put_block_group(cache);
5991 }
5992
5993 /* Can return -ENOMEM */
5994 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5995                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5996                       u64 owner, u64 offset, int for_cow)
5997 {
5998         int ret;
5999         struct btrfs_fs_info *fs_info = root->fs_info;
6000
6001         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6002
6003         /*
6004          * tree log blocks never actually go into the extent allocation
6005          * tree, just update pinning info and exit early.
6006          */
6007         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6008                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6009                 /* unlocks the pinned mutex */
6010                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6011                 ret = 0;
6012         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6013                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6014                                         num_bytes,
6015                                         parent, root_objectid, (int)owner,
6016                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
6017         } else {
6018                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6019                                                 num_bytes,
6020                                                 parent, root_objectid, owner,
6021                                                 offset, BTRFS_DROP_DELAYED_REF,
6022                                                 NULL, for_cow);
6023         }
6024         return ret;
6025 }
6026
6027 static u64 stripe_align(struct btrfs_root *root,
6028                         struct btrfs_block_group_cache *cache,
6029                         u64 val, u64 num_bytes)
6030 {
6031         u64 ret = ALIGN(val, root->stripesize);
6032         return ret;
6033 }
6034
6035 /*
6036  * when we wait for progress in the block group caching, its because
6037  * our allocation attempt failed at least once.  So, we must sleep
6038  * and let some progress happen before we try again.
6039  *
6040  * This function will sleep at least once waiting for new free space to
6041  * show up, and then it will check the block group free space numbers
6042  * for our min num_bytes.  Another option is to have it go ahead
6043  * and look in the rbtree for a free extent of a given size, but this
6044  * is a good start.
6045  *
6046  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6047  * any of the information in this block group.
6048  */
6049 static noinline void
6050 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6051                                 u64 num_bytes)
6052 {
6053         struct btrfs_caching_control *caching_ctl;
6054
6055         caching_ctl = get_caching_control(cache);
6056         if (!caching_ctl)
6057                 return;
6058
6059         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6060                    (cache->free_space_ctl->free_space >= num_bytes));
6061
6062         put_caching_control(caching_ctl);
6063 }
6064
6065 static noinline int
6066 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6067 {
6068         struct btrfs_caching_control *caching_ctl;
6069         int ret = 0;
6070
6071         caching_ctl = get_caching_control(cache);
6072         if (!caching_ctl)
6073                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6074
6075         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6076         if (cache->cached == BTRFS_CACHE_ERROR)
6077                 ret = -EIO;
6078         put_caching_control(caching_ctl);
6079         return ret;
6080 }
6081
6082 int __get_raid_index(u64 flags)
6083 {
6084         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6085                 return BTRFS_RAID_RAID10;
6086         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6087                 return BTRFS_RAID_RAID1;
6088         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6089                 return BTRFS_RAID_DUP;
6090         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6091                 return BTRFS_RAID_RAID0;
6092         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6093                 return BTRFS_RAID_RAID5;
6094         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6095                 return BTRFS_RAID_RAID6;
6096
6097         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6098 }
6099
6100 static int get_block_group_index(struct btrfs_block_group_cache *cache)
6101 {
6102         return __get_raid_index(cache->flags);
6103 }
6104
6105 enum btrfs_loop_type {
6106         LOOP_CACHING_NOWAIT = 0,
6107         LOOP_CACHING_WAIT = 1,
6108         LOOP_ALLOC_CHUNK = 2,
6109         LOOP_NO_EMPTY_SIZE = 3,
6110 };
6111
6112 /*
6113  * walks the btree of allocated extents and find a hole of a given size.
6114  * The key ins is changed to record the hole:
6115  * ins->objectid == start position
6116  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6117  * ins->offset == the size of the hole.
6118  * Any available blocks before search_start are skipped.
6119  *
6120  * If there is no suitable free space, we will record the max size of
6121  * the free space extent currently.
6122  */
6123 static noinline int find_free_extent(struct btrfs_root *orig_root,
6124                                      u64 num_bytes, u64 empty_size,
6125                                      u64 hint_byte, struct btrfs_key *ins,
6126                                      u64 flags)
6127 {
6128         int ret = 0;
6129         struct btrfs_root *root = orig_root->fs_info->extent_root;
6130         struct btrfs_free_cluster *last_ptr = NULL;
6131         struct btrfs_block_group_cache *block_group = NULL;
6132         struct btrfs_block_group_cache *used_block_group;
6133         u64 search_start = 0;
6134         u64 max_extent_size = 0;
6135         int empty_cluster = 2 * 1024 * 1024;
6136         struct btrfs_space_info *space_info;
6137         int loop = 0;
6138         int index = __get_raid_index(flags);
6139         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6140                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6141         bool found_uncached_bg = false;
6142         bool failed_cluster_refill = false;
6143         bool failed_alloc = false;
6144         bool use_cluster = true;
6145         bool have_caching_bg = false;
6146
6147         WARN_ON(num_bytes < root->sectorsize);
6148         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
6149         ins->objectid = 0;
6150         ins->offset = 0;
6151
6152         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6153
6154         space_info = __find_space_info(root->fs_info, flags);
6155         if (!space_info) {
6156                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6157                 return -ENOSPC;
6158         }
6159
6160         /*
6161          * If the space info is for both data and metadata it means we have a
6162          * small filesystem and we can't use the clustering stuff.
6163          */
6164         if (btrfs_mixed_space_info(space_info))
6165                 use_cluster = false;
6166
6167         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6168                 last_ptr = &root->fs_info->meta_alloc_cluster;
6169                 if (!btrfs_test_opt(root, SSD))
6170                         empty_cluster = 64 * 1024;
6171         }
6172
6173         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6174             btrfs_test_opt(root, SSD)) {
6175                 last_ptr = &root->fs_info->data_alloc_cluster;
6176         }
6177
6178         if (last_ptr) {
6179                 spin_lock(&last_ptr->lock);
6180                 if (last_ptr->block_group)
6181                         hint_byte = last_ptr->window_start;
6182                 spin_unlock(&last_ptr->lock);
6183         }
6184
6185         search_start = max(search_start, first_logical_byte(root, 0));
6186         search_start = max(search_start, hint_byte);
6187
6188         if (!last_ptr)
6189                 empty_cluster = 0;
6190
6191         if (search_start == hint_byte) {
6192                 block_group = btrfs_lookup_block_group(root->fs_info,
6193                                                        search_start);
6194                 used_block_group = block_group;
6195                 /*
6196                  * we don't want to use the block group if it doesn't match our
6197                  * allocation bits, or if its not cached.
6198                  *
6199                  * However if we are re-searching with an ideal block group
6200                  * picked out then we don't care that the block group is cached.
6201                  */
6202                 if (block_group && block_group_bits(block_group, flags) &&
6203                     block_group->cached != BTRFS_CACHE_NO) {
6204                         down_read(&space_info->groups_sem);
6205                         if (list_empty(&block_group->list) ||
6206                             block_group->ro) {
6207                                 /*
6208                                  * someone is removing this block group,
6209                                  * we can't jump into the have_block_group
6210                                  * target because our list pointers are not
6211                                  * valid
6212                                  */
6213                                 btrfs_put_block_group(block_group);
6214                                 up_read(&space_info->groups_sem);
6215                         } else {
6216                                 index = get_block_group_index(block_group);
6217                                 goto have_block_group;
6218                         }
6219                 } else if (block_group) {
6220                         btrfs_put_block_group(block_group);
6221                 }
6222         }
6223 search:
6224         have_caching_bg = false;
6225         down_read(&space_info->groups_sem);
6226         list_for_each_entry(block_group, &space_info->block_groups[index],
6227                             list) {
6228                 u64 offset;
6229                 int cached;
6230
6231                 used_block_group = block_group;
6232                 btrfs_get_block_group(block_group);
6233                 search_start = block_group->key.objectid;
6234
6235                 /*
6236                  * this can happen if we end up cycling through all the
6237                  * raid types, but we want to make sure we only allocate
6238                  * for the proper type.
6239                  */
6240                 if (!block_group_bits(block_group, flags)) {
6241                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6242                                 BTRFS_BLOCK_GROUP_RAID1 |
6243                                 BTRFS_BLOCK_GROUP_RAID5 |
6244                                 BTRFS_BLOCK_GROUP_RAID6 |
6245                                 BTRFS_BLOCK_GROUP_RAID10;
6246
6247                         /*
6248                          * if they asked for extra copies and this block group
6249                          * doesn't provide them, bail.  This does allow us to
6250                          * fill raid0 from raid1.
6251                          */
6252                         if ((flags & extra) && !(block_group->flags & extra))
6253                                 goto loop;
6254                 }
6255
6256 have_block_group:
6257                 cached = block_group_cache_done(block_group);
6258                 if (unlikely(!cached)) {
6259                         found_uncached_bg = true;
6260                         ret = cache_block_group(block_group, 0);
6261                         BUG_ON(ret < 0);
6262                         ret = 0;
6263                 }
6264
6265                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6266                         goto loop;
6267                 if (unlikely(block_group->ro))
6268                         goto loop;
6269
6270                 /*
6271                  * Ok we want to try and use the cluster allocator, so
6272                  * lets look there
6273                  */
6274                 if (last_ptr) {
6275                         unsigned long aligned_cluster;
6276                         /*
6277                          * the refill lock keeps out other
6278                          * people trying to start a new cluster
6279                          */
6280                         spin_lock(&last_ptr->refill_lock);
6281                         used_block_group = last_ptr->block_group;
6282                         if (used_block_group != block_group &&
6283                             (!used_block_group ||
6284                              used_block_group->ro ||
6285                              !block_group_bits(used_block_group, flags))) {
6286                                 used_block_group = block_group;
6287                                 goto refill_cluster;
6288                         }
6289
6290                         if (used_block_group != block_group)
6291                                 btrfs_get_block_group(used_block_group);
6292
6293                         offset = btrfs_alloc_from_cluster(used_block_group,
6294                                                 last_ptr,
6295                                                 num_bytes,
6296                                                 used_block_group->key.objectid,
6297                                                 &max_extent_size);
6298                         if (offset) {
6299                                 /* we have a block, we're done */
6300                                 spin_unlock(&last_ptr->refill_lock);
6301                                 trace_btrfs_reserve_extent_cluster(root,
6302                                         block_group, search_start, num_bytes);
6303                                 goto checks;
6304                         }
6305
6306                         WARN_ON(last_ptr->block_group != used_block_group);
6307                         if (used_block_group != block_group) {
6308                                 btrfs_put_block_group(used_block_group);
6309                                 used_block_group = block_group;
6310                         }
6311 refill_cluster:
6312                         BUG_ON(used_block_group != block_group);
6313                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6314                          * set up a new clusters, so lets just skip it
6315                          * and let the allocator find whatever block
6316                          * it can find.  If we reach this point, we
6317                          * will have tried the cluster allocator
6318                          * plenty of times and not have found
6319                          * anything, so we are likely way too
6320                          * fragmented for the clustering stuff to find
6321                          * anything.
6322                          *
6323                          * However, if the cluster is taken from the
6324                          * current block group, release the cluster
6325                          * first, so that we stand a better chance of
6326                          * succeeding in the unclustered
6327                          * allocation.  */
6328                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6329                             last_ptr->block_group != block_group) {
6330                                 spin_unlock(&last_ptr->refill_lock);
6331                                 goto unclustered_alloc;
6332                         }
6333
6334                         /*
6335                          * this cluster didn't work out, free it and
6336                          * start over
6337                          */
6338                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6339
6340                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6341                                 spin_unlock(&last_ptr->refill_lock);
6342                                 goto unclustered_alloc;
6343                         }
6344
6345                         aligned_cluster = max_t(unsigned long,
6346                                                 empty_cluster + empty_size,
6347                                               block_group->full_stripe_len);
6348
6349                         /* allocate a cluster in this block group */
6350                         ret = btrfs_find_space_cluster(root, block_group,
6351                                                        last_ptr, search_start,
6352                                                        num_bytes,
6353                                                        aligned_cluster);
6354                         if (ret == 0) {
6355                                 /*
6356                                  * now pull our allocation out of this
6357                                  * cluster
6358                                  */
6359                                 offset = btrfs_alloc_from_cluster(block_group,
6360                                                         last_ptr,
6361                                                         num_bytes,
6362                                                         search_start,
6363                                                         &max_extent_size);
6364                                 if (offset) {
6365                                         /* we found one, proceed */
6366                                         spin_unlock(&last_ptr->refill_lock);
6367                                         trace_btrfs_reserve_extent_cluster(root,
6368                                                 block_group, search_start,
6369                                                 num_bytes);
6370                                         goto checks;
6371                                 }
6372                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6373                                    && !failed_cluster_refill) {
6374                                 spin_unlock(&last_ptr->refill_lock);
6375
6376                                 failed_cluster_refill = true;
6377                                 wait_block_group_cache_progress(block_group,
6378                                        num_bytes + empty_cluster + empty_size);
6379                                 goto have_block_group;
6380                         }
6381
6382                         /*
6383                          * at this point we either didn't find a cluster
6384                          * or we weren't able to allocate a block from our
6385                          * cluster.  Free the cluster we've been trying
6386                          * to use, and go to the next block group
6387                          */
6388                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6389                         spin_unlock(&last_ptr->refill_lock);
6390                         goto loop;
6391                 }
6392
6393 unclustered_alloc:
6394                 spin_lock(&block_group->free_space_ctl->tree_lock);
6395                 if (cached &&
6396                     block_group->free_space_ctl->free_space <
6397                     num_bytes + empty_cluster + empty_size) {
6398                         if (block_group->free_space_ctl->free_space >
6399                             max_extent_size)
6400                                 max_extent_size =
6401                                         block_group->free_space_ctl->free_space;
6402                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6403                         goto loop;
6404                 }
6405                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6406
6407                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6408                                                     num_bytes, empty_size,
6409                                                     &max_extent_size);
6410                 /*
6411                  * If we didn't find a chunk, and we haven't failed on this
6412                  * block group before, and this block group is in the middle of
6413                  * caching and we are ok with waiting, then go ahead and wait
6414                  * for progress to be made, and set failed_alloc to true.
6415                  *
6416                  * If failed_alloc is true then we've already waited on this
6417                  * block group once and should move on to the next block group.
6418                  */
6419                 if (!offset && !failed_alloc && !cached &&
6420                     loop > LOOP_CACHING_NOWAIT) {
6421                         wait_block_group_cache_progress(block_group,
6422                                                 num_bytes + empty_size);
6423                         failed_alloc = true;
6424                         goto have_block_group;
6425                 } else if (!offset) {
6426                         if (!cached)
6427                                 have_caching_bg = true;
6428                         goto loop;
6429                 }
6430 checks:
6431                 search_start = stripe_align(root, used_block_group,
6432                                             offset, num_bytes);
6433
6434                 /* move on to the next group */
6435                 if (search_start + num_bytes >
6436                     used_block_group->key.objectid + used_block_group->key.offset) {
6437                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6438                         goto loop;
6439                 }
6440
6441                 if (offset < search_start)
6442                         btrfs_add_free_space(used_block_group, offset,
6443                                              search_start - offset);
6444                 BUG_ON(offset > search_start);
6445
6446                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
6447                                                   alloc_type);
6448                 if (ret == -EAGAIN) {
6449                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6450                         goto loop;
6451                 }
6452
6453                 /* we are all good, lets return */
6454                 ins->objectid = search_start;
6455                 ins->offset = num_bytes;
6456
6457                 trace_btrfs_reserve_extent(orig_root, block_group,
6458                                            search_start, num_bytes);
6459                 if (used_block_group != block_group)
6460                         btrfs_put_block_group(used_block_group);
6461                 btrfs_put_block_group(block_group);
6462                 break;
6463 loop:
6464                 failed_cluster_refill = false;
6465                 failed_alloc = false;
6466                 BUG_ON(index != get_block_group_index(block_group));
6467                 if (used_block_group != block_group)
6468                         btrfs_put_block_group(used_block_group);
6469                 btrfs_put_block_group(block_group);
6470         }
6471         up_read(&space_info->groups_sem);
6472
6473         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6474                 goto search;
6475
6476         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6477                 goto search;
6478
6479         /*
6480          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6481          *                      caching kthreads as we move along
6482          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6483          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6484          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6485          *                      again
6486          */
6487         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6488                 index = 0;
6489                 loop++;
6490                 if (loop == LOOP_ALLOC_CHUNK) {
6491                         struct btrfs_trans_handle *trans;
6492
6493                         trans = btrfs_join_transaction(root);
6494                         if (IS_ERR(trans)) {
6495                                 ret = PTR_ERR(trans);
6496                                 goto out;
6497                         }
6498
6499                         ret = do_chunk_alloc(trans, root, flags,
6500                                              CHUNK_ALLOC_FORCE);
6501                         /*
6502                          * Do not bail out on ENOSPC since we
6503                          * can do more things.
6504                          */
6505                         if (ret < 0 && ret != -ENOSPC)
6506                                 btrfs_abort_transaction(trans,
6507                                                         root, ret);
6508                         else
6509                                 ret = 0;
6510                         btrfs_end_transaction(trans, root);
6511                         if (ret)
6512                                 goto out;
6513                 }
6514
6515                 if (loop == LOOP_NO_EMPTY_SIZE) {
6516                         empty_size = 0;
6517                         empty_cluster = 0;
6518                 }
6519
6520                 goto search;
6521         } else if (!ins->objectid) {
6522                 ret = -ENOSPC;
6523         } else if (ins->objectid) {
6524                 ret = 0;
6525         }
6526 out:
6527         if (ret == -ENOSPC)
6528                 ins->offset = max_extent_size;
6529         return ret;
6530 }
6531
6532 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6533                             int dump_block_groups)
6534 {
6535         struct btrfs_block_group_cache *cache;
6536         int index = 0;
6537
6538         spin_lock(&info->lock);
6539         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
6540                info->flags,
6541                info->total_bytes - info->bytes_used - info->bytes_pinned -
6542                info->bytes_reserved - info->bytes_readonly,
6543                (info->full) ? "" : "not ");
6544         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
6545                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6546                info->total_bytes, info->bytes_used, info->bytes_pinned,
6547                info->bytes_reserved, info->bytes_may_use,
6548                info->bytes_readonly);
6549         spin_unlock(&info->lock);
6550
6551         if (!dump_block_groups)
6552                 return;
6553
6554         down_read(&info->groups_sem);
6555 again:
6556         list_for_each_entry(cache, &info->block_groups[index], list) {
6557                 spin_lock(&cache->lock);
6558                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
6559                        cache->key.objectid, cache->key.offset,
6560                        btrfs_block_group_used(&cache->item), cache->pinned,
6561                        cache->reserved, cache->ro ? "[readonly]" : "");
6562                 btrfs_dump_free_space(cache, bytes);
6563                 spin_unlock(&cache->lock);
6564         }
6565         if (++index < BTRFS_NR_RAID_TYPES)
6566                 goto again;
6567         up_read(&info->groups_sem);
6568 }
6569
6570 int btrfs_reserve_extent(struct btrfs_root *root,
6571                          u64 num_bytes, u64 min_alloc_size,
6572                          u64 empty_size, u64 hint_byte,
6573                          struct btrfs_key *ins, int is_data)
6574 {
6575         bool final_tried = false;
6576         u64 flags;
6577         int ret;
6578
6579         flags = btrfs_get_alloc_profile(root, is_data);
6580 again:
6581         WARN_ON(num_bytes < root->sectorsize);
6582         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
6583                                flags);
6584
6585         if (ret == -ENOSPC) {
6586                 if (!final_tried && ins->offset) {
6587                         num_bytes = min(num_bytes >> 1, ins->offset);
6588                         num_bytes = round_down(num_bytes, root->sectorsize);
6589                         num_bytes = max(num_bytes, min_alloc_size);
6590                         if (num_bytes == min_alloc_size)
6591                                 final_tried = true;
6592                         goto again;
6593                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6594                         struct btrfs_space_info *sinfo;
6595
6596                         sinfo = __find_space_info(root->fs_info, flags);
6597                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6598                                 flags, num_bytes);
6599                         if (sinfo)
6600                                 dump_space_info(sinfo, num_bytes, 1);
6601                 }
6602         }
6603
6604         return ret;
6605 }
6606
6607 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6608                                         u64 start, u64 len, int pin)
6609 {
6610         struct btrfs_block_group_cache *cache;
6611         int ret = 0;
6612
6613         cache = btrfs_lookup_block_group(root->fs_info, start);
6614         if (!cache) {
6615                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6616                         start);
6617                 return -ENOSPC;
6618         }
6619
6620         if (btrfs_test_opt(root, DISCARD))
6621                 ret = btrfs_discard_extent(root, start, len, NULL);
6622
6623         if (pin)
6624                 pin_down_extent(root, cache, start, len, 1);
6625         else {
6626                 btrfs_add_free_space(cache, start, len);
6627                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6628         }
6629         btrfs_put_block_group(cache);
6630
6631         trace_btrfs_reserved_extent_free(root, start, len);
6632
6633         return ret;
6634 }
6635
6636 int btrfs_free_reserved_extent(struct btrfs_root *root,
6637                                         u64 start, u64 len)
6638 {
6639         return __btrfs_free_reserved_extent(root, start, len, 0);
6640 }
6641
6642 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6643                                        u64 start, u64 len)
6644 {
6645         return __btrfs_free_reserved_extent(root, start, len, 1);
6646 }
6647
6648 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6649                                       struct btrfs_root *root,
6650                                       u64 parent, u64 root_objectid,
6651                                       u64 flags, u64 owner, u64 offset,
6652                                       struct btrfs_key *ins, int ref_mod)
6653 {
6654         int ret;
6655         struct btrfs_fs_info *fs_info = root->fs_info;
6656         struct btrfs_extent_item *extent_item;
6657         struct btrfs_extent_inline_ref *iref;
6658         struct btrfs_path *path;
6659         struct extent_buffer *leaf;
6660         int type;
6661         u32 size;
6662
6663         if (parent > 0)
6664                 type = BTRFS_SHARED_DATA_REF_KEY;
6665         else
6666                 type = BTRFS_EXTENT_DATA_REF_KEY;
6667
6668         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6669
6670         path = btrfs_alloc_path();
6671         if (!path)
6672                 return -ENOMEM;
6673
6674         path->leave_spinning = 1;
6675         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6676                                       ins, size);
6677         if (ret) {
6678                 btrfs_free_path(path);
6679                 return ret;
6680         }
6681
6682         leaf = path->nodes[0];
6683         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6684                                      struct btrfs_extent_item);
6685         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6686         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6687         btrfs_set_extent_flags(leaf, extent_item,
6688                                flags | BTRFS_EXTENT_FLAG_DATA);
6689
6690         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6691         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6692         if (parent > 0) {
6693                 struct btrfs_shared_data_ref *ref;
6694                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6695                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6696                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6697         } else {
6698                 struct btrfs_extent_data_ref *ref;
6699                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6700                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6701                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6702                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6703                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6704         }
6705
6706         btrfs_mark_buffer_dirty(path->nodes[0]);
6707         btrfs_free_path(path);
6708
6709         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6710         if (ret) { /* -ENOENT, logic error */
6711                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6712                         ins->objectid, ins->offset);
6713                 BUG();
6714         }
6715         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6716         return ret;
6717 }
6718
6719 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6720                                      struct btrfs_root *root,
6721                                      u64 parent, u64 root_objectid,
6722                                      u64 flags, struct btrfs_disk_key *key,
6723                                      int level, struct btrfs_key *ins)
6724 {
6725         int ret;
6726         struct btrfs_fs_info *fs_info = root->fs_info;
6727         struct btrfs_extent_item *extent_item;
6728         struct btrfs_tree_block_info *block_info;
6729         struct btrfs_extent_inline_ref *iref;
6730         struct btrfs_path *path;
6731         struct extent_buffer *leaf;
6732         u32 size = sizeof(*extent_item) + sizeof(*iref);
6733         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6734                                                  SKINNY_METADATA);
6735
6736         if (!skinny_metadata)
6737                 size += sizeof(*block_info);
6738
6739         path = btrfs_alloc_path();
6740         if (!path) {
6741                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
6742                                                    root->leafsize);
6743                 return -ENOMEM;
6744         }
6745
6746         path->leave_spinning = 1;
6747         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6748                                       ins, size);
6749         if (ret) {
6750                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
6751                                                    root->leafsize);
6752                 btrfs_free_path(path);
6753                 return ret;
6754         }
6755
6756         leaf = path->nodes[0];
6757         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6758                                      struct btrfs_extent_item);
6759         btrfs_set_extent_refs(leaf, extent_item, 1);
6760         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6761         btrfs_set_extent_flags(leaf, extent_item,
6762                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6763
6764         if (skinny_metadata) {
6765                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6766         } else {
6767                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6768                 btrfs_set_tree_block_key(leaf, block_info, key);
6769                 btrfs_set_tree_block_level(leaf, block_info, level);
6770                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6771         }
6772
6773         if (parent > 0) {
6774                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6775                 btrfs_set_extent_inline_ref_type(leaf, iref,
6776                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6777                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6778         } else {
6779                 btrfs_set_extent_inline_ref_type(leaf, iref,
6780                                                  BTRFS_TREE_BLOCK_REF_KEY);
6781                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6782         }
6783
6784         btrfs_mark_buffer_dirty(leaf);
6785         btrfs_free_path(path);
6786
6787         ret = update_block_group(root, ins->objectid, root->leafsize, 1);
6788         if (ret) { /* -ENOENT, logic error */
6789                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6790                         ins->objectid, ins->offset);
6791                 BUG();
6792         }
6793
6794         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->leafsize);
6795         return ret;
6796 }
6797
6798 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6799                                      struct btrfs_root *root,
6800                                      u64 root_objectid, u64 owner,
6801                                      u64 offset, struct btrfs_key *ins)
6802 {
6803         int ret;
6804
6805         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6806
6807         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6808                                          ins->offset, 0,
6809                                          root_objectid, owner, offset,
6810                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6811         return ret;
6812 }
6813
6814 /*
6815  * this is used by the tree logging recovery code.  It records that
6816  * an extent has been allocated and makes sure to clear the free
6817  * space cache bits as well
6818  */
6819 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6820                                    struct btrfs_root *root,
6821                                    u64 root_objectid, u64 owner, u64 offset,
6822                                    struct btrfs_key *ins)
6823 {
6824         int ret;
6825         struct btrfs_block_group_cache *block_group;
6826
6827         /*
6828          * Mixed block groups will exclude before processing the log so we only
6829          * need to do the exlude dance if this fs isn't mixed.
6830          */
6831         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
6832                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
6833                 if (ret)
6834                         return ret;
6835         }
6836
6837         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6838         if (!block_group)
6839                 return -EINVAL;
6840
6841         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6842                                           RESERVE_ALLOC_NO_ACCOUNT);
6843         BUG_ON(ret); /* logic error */
6844         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6845                                          0, owner, offset, ins, 1);
6846         btrfs_put_block_group(block_group);
6847         return ret;
6848 }
6849
6850 static struct extent_buffer *
6851 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6852                       u64 bytenr, u32 blocksize, int level)
6853 {
6854         struct extent_buffer *buf;
6855
6856         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6857         if (!buf)
6858                 return ERR_PTR(-ENOMEM);
6859         btrfs_set_header_generation(buf, trans->transid);
6860         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6861         btrfs_tree_lock(buf);
6862         clean_tree_block(trans, root, buf);
6863         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6864
6865         btrfs_set_lock_blocking(buf);
6866         btrfs_set_buffer_uptodate(buf);
6867
6868         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6869                 /*
6870                  * we allow two log transactions at a time, use different
6871                  * EXENT bit to differentiate dirty pages.
6872                  */
6873                 if (root->log_transid % 2 == 0)
6874                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6875                                         buf->start + buf->len - 1, GFP_NOFS);
6876                 else
6877                         set_extent_new(&root->dirty_log_pages, buf->start,
6878                                         buf->start + buf->len - 1, GFP_NOFS);
6879         } else {
6880                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6881                          buf->start + buf->len - 1, GFP_NOFS);
6882         }
6883         trans->blocks_used++;
6884         /* this returns a buffer locked for blocking */
6885         return buf;
6886 }
6887
6888 static struct btrfs_block_rsv *
6889 use_block_rsv(struct btrfs_trans_handle *trans,
6890               struct btrfs_root *root, u32 blocksize)
6891 {
6892         struct btrfs_block_rsv *block_rsv;
6893         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6894         int ret;
6895         bool global_updated = false;
6896
6897         block_rsv = get_block_rsv(trans, root);
6898
6899         if (unlikely(block_rsv->size == 0))
6900                 goto try_reserve;
6901 again:
6902         ret = block_rsv_use_bytes(block_rsv, blocksize);
6903         if (!ret)
6904                 return block_rsv;
6905
6906         if (block_rsv->failfast)
6907                 return ERR_PTR(ret);
6908
6909         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
6910                 global_updated = true;
6911                 update_global_block_rsv(root->fs_info);
6912                 goto again;
6913         }
6914
6915         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6916                 static DEFINE_RATELIMIT_STATE(_rs,
6917                                 DEFAULT_RATELIMIT_INTERVAL * 10,
6918                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
6919                 if (__ratelimit(&_rs))
6920                         WARN(1, KERN_DEBUG
6921                                 "btrfs: block rsv returned %d\n", ret);
6922         }
6923 try_reserve:
6924         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6925                                      BTRFS_RESERVE_NO_FLUSH);
6926         if (!ret)
6927                 return block_rsv;
6928         /*
6929          * If we couldn't reserve metadata bytes try and use some from
6930          * the global reserve if its space type is the same as the global
6931          * reservation.
6932          */
6933         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
6934             block_rsv->space_info == global_rsv->space_info) {
6935                 ret = block_rsv_use_bytes(global_rsv, blocksize);
6936                 if (!ret)
6937                         return global_rsv;
6938         }
6939         return ERR_PTR(ret);
6940 }
6941
6942 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6943                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6944 {
6945         block_rsv_add_bytes(block_rsv, blocksize, 0);
6946         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6947 }
6948
6949 /*
6950  * finds a free extent and does all the dirty work required for allocation
6951  * returns the key for the extent through ins, and a tree buffer for
6952  * the first block of the extent through buf.
6953  *
6954  * returns the tree buffer or NULL.
6955  */
6956 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6957                                         struct btrfs_root *root, u32 blocksize,
6958                                         u64 parent, u64 root_objectid,
6959                                         struct btrfs_disk_key *key, int level,
6960                                         u64 hint, u64 empty_size)
6961 {
6962         struct btrfs_key ins;
6963         struct btrfs_block_rsv *block_rsv;
6964         struct extent_buffer *buf;
6965         u64 flags = 0;
6966         int ret;
6967         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6968                                                  SKINNY_METADATA);
6969
6970         block_rsv = use_block_rsv(trans, root, blocksize);
6971         if (IS_ERR(block_rsv))
6972                 return ERR_CAST(block_rsv);
6973
6974         ret = btrfs_reserve_extent(root, blocksize, blocksize,
6975                                    empty_size, hint, &ins, 0);
6976         if (ret) {
6977                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6978                 return ERR_PTR(ret);
6979         }
6980
6981         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6982                                     blocksize, level);
6983         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6984
6985         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6986                 if (parent == 0)
6987                         parent = ins.objectid;
6988                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6989         } else
6990                 BUG_ON(parent > 0);
6991
6992         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6993                 struct btrfs_delayed_extent_op *extent_op;
6994                 extent_op = btrfs_alloc_delayed_extent_op();
6995                 BUG_ON(!extent_op); /* -ENOMEM */
6996                 if (key)
6997                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6998                 else
6999                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7000                 extent_op->flags_to_set = flags;
7001                 if (skinny_metadata)
7002                         extent_op->update_key = 0;
7003                 else
7004                         extent_op->update_key = 1;
7005                 extent_op->update_flags = 1;
7006                 extent_op->is_data = 0;
7007                 extent_op->level = level;
7008
7009                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7010                                         ins.objectid,
7011                                         ins.offset, parent, root_objectid,
7012                                         level, BTRFS_ADD_DELAYED_EXTENT,
7013                                         extent_op, 0);
7014                 BUG_ON(ret); /* -ENOMEM */
7015         }
7016         return buf;
7017 }
7018
7019 struct walk_control {
7020         u64 refs[BTRFS_MAX_LEVEL];
7021         u64 flags[BTRFS_MAX_LEVEL];
7022         struct btrfs_key update_progress;
7023         int stage;
7024         int level;
7025         int shared_level;
7026         int update_ref;
7027         int keep_locks;
7028         int reada_slot;
7029         int reada_count;
7030         int for_reloc;
7031 };
7032
7033 #define DROP_REFERENCE  1
7034 #define UPDATE_BACKREF  2
7035
7036 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7037                                      struct btrfs_root *root,
7038                                      struct walk_control *wc,
7039                                      struct btrfs_path *path)
7040 {
7041         u64 bytenr;
7042         u64 generation;
7043         u64 refs;
7044         u64 flags;
7045         u32 nritems;
7046         u32 blocksize;
7047         struct btrfs_key key;
7048         struct extent_buffer *eb;
7049         int ret;
7050         int slot;
7051         int nread = 0;
7052
7053         if (path->slots[wc->level] < wc->reada_slot) {
7054                 wc->reada_count = wc->reada_count * 2 / 3;
7055                 wc->reada_count = max(wc->reada_count, 2);
7056         } else {
7057                 wc->reada_count = wc->reada_count * 3 / 2;
7058                 wc->reada_count = min_t(int, wc->reada_count,
7059                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7060         }
7061
7062         eb = path->nodes[wc->level];
7063         nritems = btrfs_header_nritems(eb);
7064         blocksize = btrfs_level_size(root, wc->level - 1);
7065
7066         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7067                 if (nread >= wc->reada_count)
7068                         break;
7069
7070                 cond_resched();
7071                 bytenr = btrfs_node_blockptr(eb, slot);
7072                 generation = btrfs_node_ptr_generation(eb, slot);
7073
7074                 if (slot == path->slots[wc->level])
7075                         goto reada;
7076
7077                 if (wc->stage == UPDATE_BACKREF &&
7078                     generation <= root->root_key.offset)
7079                         continue;
7080
7081                 /* We don't lock the tree block, it's OK to be racy here */
7082                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7083                                                wc->level - 1, 1, &refs,
7084                                                &flags);
7085                 /* We don't care about errors in readahead. */
7086                 if (ret < 0)
7087                         continue;
7088                 BUG_ON(refs == 0);
7089
7090                 if (wc->stage == DROP_REFERENCE) {
7091                         if (refs == 1)
7092                                 goto reada;
7093
7094                         if (wc->level == 1 &&
7095                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7096                                 continue;
7097                         if (!wc->update_ref ||
7098                             generation <= root->root_key.offset)
7099                                 continue;
7100                         btrfs_node_key_to_cpu(eb, &key, slot);
7101                         ret = btrfs_comp_cpu_keys(&key,
7102                                                   &wc->update_progress);
7103                         if (ret < 0)
7104                                 continue;
7105                 } else {
7106                         if (wc->level == 1 &&
7107                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7108                                 continue;
7109                 }
7110 reada:
7111                 ret = readahead_tree_block(root, bytenr, blocksize,
7112                                            generation);
7113                 if (ret)
7114                         break;
7115                 nread++;
7116         }
7117         wc->reada_slot = slot;
7118 }
7119
7120 /*
7121  * helper to process tree block while walking down the tree.
7122  *
7123  * when wc->stage == UPDATE_BACKREF, this function updates
7124  * back refs for pointers in the block.
7125  *
7126  * NOTE: return value 1 means we should stop walking down.
7127  */
7128 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7129                                    struct btrfs_root *root,
7130                                    struct btrfs_path *path,
7131                                    struct walk_control *wc, int lookup_info)
7132 {
7133         int level = wc->level;
7134         struct extent_buffer *eb = path->nodes[level];
7135         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7136         int ret;
7137
7138         if (wc->stage == UPDATE_BACKREF &&
7139             btrfs_header_owner(eb) != root->root_key.objectid)
7140                 return 1;
7141
7142         /*
7143          * when reference count of tree block is 1, it won't increase
7144          * again. once full backref flag is set, we never clear it.
7145          */
7146         if (lookup_info &&
7147             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7148              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7149                 BUG_ON(!path->locks[level]);
7150                 ret = btrfs_lookup_extent_info(trans, root,
7151                                                eb->start, level, 1,
7152                                                &wc->refs[level],
7153                                                &wc->flags[level]);
7154                 BUG_ON(ret == -ENOMEM);
7155                 if (ret)
7156                         return ret;
7157                 BUG_ON(wc->refs[level] == 0);
7158         }
7159
7160         if (wc->stage == DROP_REFERENCE) {
7161                 if (wc->refs[level] > 1)
7162                         return 1;
7163
7164                 if (path->locks[level] && !wc->keep_locks) {
7165                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7166                         path->locks[level] = 0;
7167                 }
7168                 return 0;
7169         }
7170
7171         /* wc->stage == UPDATE_BACKREF */
7172         if (!(wc->flags[level] & flag)) {
7173                 BUG_ON(!path->locks[level]);
7174                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
7175                 BUG_ON(ret); /* -ENOMEM */
7176                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
7177                 BUG_ON(ret); /* -ENOMEM */
7178                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7179                                                   eb->len, flag,
7180                                                   btrfs_header_level(eb), 0);
7181                 BUG_ON(ret); /* -ENOMEM */
7182                 wc->flags[level] |= flag;
7183         }
7184
7185         /*
7186          * the block is shared by multiple trees, so it's not good to
7187          * keep the tree lock
7188          */
7189         if (path->locks[level] && level > 0) {
7190                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7191                 path->locks[level] = 0;
7192         }
7193         return 0;
7194 }
7195
7196 /*
7197  * helper to process tree block pointer.
7198  *
7199  * when wc->stage == DROP_REFERENCE, this function checks
7200  * reference count of the block pointed to. if the block
7201  * is shared and we need update back refs for the subtree
7202  * rooted at the block, this function changes wc->stage to
7203  * UPDATE_BACKREF. if the block is shared and there is no
7204  * need to update back, this function drops the reference
7205  * to the block.
7206  *
7207  * NOTE: return value 1 means we should stop walking down.
7208  */
7209 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7210                                  struct btrfs_root *root,
7211                                  struct btrfs_path *path,
7212                                  struct walk_control *wc, int *lookup_info)
7213 {
7214         u64 bytenr;
7215         u64 generation;
7216         u64 parent;
7217         u32 blocksize;
7218         struct btrfs_key key;
7219         struct extent_buffer *next;
7220         int level = wc->level;
7221         int reada = 0;
7222         int ret = 0;
7223
7224         generation = btrfs_node_ptr_generation(path->nodes[level],
7225                                                path->slots[level]);
7226         /*
7227          * if the lower level block was created before the snapshot
7228          * was created, we know there is no need to update back refs
7229          * for the subtree
7230          */
7231         if (wc->stage == UPDATE_BACKREF &&
7232             generation <= root->root_key.offset) {
7233                 *lookup_info = 1;
7234                 return 1;
7235         }
7236
7237         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7238         blocksize = btrfs_level_size(root, level - 1);
7239
7240         next = btrfs_find_tree_block(root, bytenr, blocksize);
7241         if (!next) {
7242                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7243                 if (!next)
7244                         return -ENOMEM;
7245                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
7246                                                level - 1);
7247                 reada = 1;
7248         }
7249         btrfs_tree_lock(next);
7250         btrfs_set_lock_blocking(next);
7251
7252         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7253                                        &wc->refs[level - 1],
7254                                        &wc->flags[level - 1]);
7255         if (ret < 0) {
7256                 btrfs_tree_unlock(next);
7257                 return ret;
7258         }
7259
7260         if (unlikely(wc->refs[level - 1] == 0)) {
7261                 btrfs_err(root->fs_info, "Missing references.");
7262                 BUG();
7263         }
7264         *lookup_info = 0;
7265
7266         if (wc->stage == DROP_REFERENCE) {
7267                 if (wc->refs[level - 1] > 1) {
7268                         if (level == 1 &&
7269                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7270                                 goto skip;
7271
7272                         if (!wc->update_ref ||
7273                             generation <= root->root_key.offset)
7274                                 goto skip;
7275
7276                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7277                                               path->slots[level]);
7278                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7279                         if (ret < 0)
7280                                 goto skip;
7281
7282                         wc->stage = UPDATE_BACKREF;
7283                         wc->shared_level = level - 1;
7284                 }
7285         } else {
7286                 if (level == 1 &&
7287                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7288                         goto skip;
7289         }
7290
7291         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7292                 btrfs_tree_unlock(next);
7293                 free_extent_buffer(next);
7294                 next = NULL;
7295                 *lookup_info = 1;
7296         }
7297
7298         if (!next) {
7299                 if (reada && level == 1)
7300                         reada_walk_down(trans, root, wc, path);
7301                 next = read_tree_block(root, bytenr, blocksize, generation);
7302                 if (!next || !extent_buffer_uptodate(next)) {
7303                         free_extent_buffer(next);
7304                         return -EIO;
7305                 }
7306                 btrfs_tree_lock(next);
7307                 btrfs_set_lock_blocking(next);
7308         }
7309
7310         level--;
7311         BUG_ON(level != btrfs_header_level(next));
7312         path->nodes[level] = next;
7313         path->slots[level] = 0;
7314         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7315         wc->level = level;
7316         if (wc->level == 1)
7317                 wc->reada_slot = 0;
7318         return 0;
7319 skip:
7320         wc->refs[level - 1] = 0;
7321         wc->flags[level - 1] = 0;
7322         if (wc->stage == DROP_REFERENCE) {
7323                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7324                         parent = path->nodes[level]->start;
7325                 } else {
7326                         BUG_ON(root->root_key.objectid !=
7327                                btrfs_header_owner(path->nodes[level]));
7328                         parent = 0;
7329                 }
7330
7331                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7332                                 root->root_key.objectid, level - 1, 0, 0);
7333                 BUG_ON(ret); /* -ENOMEM */
7334         }
7335         btrfs_tree_unlock(next);
7336         free_extent_buffer(next);
7337         *lookup_info = 1;
7338         return 1;
7339 }
7340
7341 /*
7342  * helper to process tree block while walking up the tree.
7343  *
7344  * when wc->stage == DROP_REFERENCE, this function drops
7345  * reference count on the block.
7346  *
7347  * when wc->stage == UPDATE_BACKREF, this function changes
7348  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7349  * to UPDATE_BACKREF previously while processing the block.
7350  *
7351  * NOTE: return value 1 means we should stop walking up.
7352  */
7353 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7354                                  struct btrfs_root *root,
7355                                  struct btrfs_path *path,
7356                                  struct walk_control *wc)
7357 {
7358         int ret;
7359         int level = wc->level;
7360         struct extent_buffer *eb = path->nodes[level];
7361         u64 parent = 0;
7362
7363         if (wc->stage == UPDATE_BACKREF) {
7364                 BUG_ON(wc->shared_level < level);
7365                 if (level < wc->shared_level)
7366                         goto out;
7367
7368                 ret = find_next_key(path, level + 1, &wc->update_progress);
7369                 if (ret > 0)
7370                         wc->update_ref = 0;
7371
7372                 wc->stage = DROP_REFERENCE;
7373                 wc->shared_level = -1;
7374                 path->slots[level] = 0;
7375
7376                 /*
7377                  * check reference count again if the block isn't locked.
7378                  * we should start walking down the tree again if reference
7379                  * count is one.
7380                  */
7381                 if (!path->locks[level]) {
7382                         BUG_ON(level == 0);
7383                         btrfs_tree_lock(eb);
7384                         btrfs_set_lock_blocking(eb);
7385                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7386
7387                         ret = btrfs_lookup_extent_info(trans, root,
7388                                                        eb->start, level, 1,
7389                                                        &wc->refs[level],
7390                                                        &wc->flags[level]);
7391                         if (ret < 0) {
7392                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7393                                 path->locks[level] = 0;
7394                                 return ret;
7395                         }
7396                         BUG_ON(wc->refs[level] == 0);
7397                         if (wc->refs[level] == 1) {
7398                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7399                                 path->locks[level] = 0;
7400                                 return 1;
7401                         }
7402                 }
7403         }
7404
7405         /* wc->stage == DROP_REFERENCE */
7406         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7407
7408         if (wc->refs[level] == 1) {
7409                 if (level == 0) {
7410                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7411                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7412                                                     wc->for_reloc);
7413                         else
7414                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7415                                                     wc->for_reloc);
7416                         BUG_ON(ret); /* -ENOMEM */
7417                 }
7418                 /* make block locked assertion in clean_tree_block happy */
7419                 if (!path->locks[level] &&
7420                     btrfs_header_generation(eb) == trans->transid) {
7421                         btrfs_tree_lock(eb);
7422                         btrfs_set_lock_blocking(eb);
7423                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7424                 }
7425                 clean_tree_block(trans, root, eb);
7426         }
7427
7428         if (eb == root->node) {
7429                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7430                         parent = eb->start;
7431                 else
7432                         BUG_ON(root->root_key.objectid !=
7433                                btrfs_header_owner(eb));
7434         } else {
7435                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7436                         parent = path->nodes[level + 1]->start;
7437                 else
7438                         BUG_ON(root->root_key.objectid !=
7439                                btrfs_header_owner(path->nodes[level + 1]));
7440         }
7441
7442         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7443 out:
7444         wc->refs[level] = 0;
7445         wc->flags[level] = 0;
7446         return 0;
7447 }
7448
7449 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7450                                    struct btrfs_root *root,
7451                                    struct btrfs_path *path,
7452                                    struct walk_control *wc)
7453 {
7454         int level = wc->level;
7455         int lookup_info = 1;
7456         int ret;
7457
7458         while (level >= 0) {
7459                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7460                 if (ret > 0)
7461                         break;
7462
7463                 if (level == 0)
7464                         break;
7465
7466                 if (path->slots[level] >=
7467                     btrfs_header_nritems(path->nodes[level]))
7468                         break;
7469
7470                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7471                 if (ret > 0) {
7472                         path->slots[level]++;
7473                         continue;
7474                 } else if (ret < 0)
7475                         return ret;
7476                 level = wc->level;
7477         }
7478         return 0;
7479 }
7480
7481 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7482                                  struct btrfs_root *root,
7483                                  struct btrfs_path *path,
7484                                  struct walk_control *wc, int max_level)
7485 {
7486         int level = wc->level;
7487         int ret;
7488
7489         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7490         while (level < max_level && path->nodes[level]) {
7491                 wc->level = level;
7492                 if (path->slots[level] + 1 <
7493                     btrfs_header_nritems(path->nodes[level])) {
7494                         path->slots[level]++;
7495                         return 0;
7496                 } else {
7497                         ret = walk_up_proc(trans, root, path, wc);
7498                         if (ret > 0)
7499                                 return 0;
7500
7501                         if (path->locks[level]) {
7502                                 btrfs_tree_unlock_rw(path->nodes[level],
7503                                                      path->locks[level]);
7504                                 path->locks[level] = 0;
7505                         }
7506                         free_extent_buffer(path->nodes[level]);
7507                         path->nodes[level] = NULL;
7508                         level++;
7509                 }
7510         }
7511         return 1;
7512 }
7513
7514 /*
7515  * drop a subvolume tree.
7516  *
7517  * this function traverses the tree freeing any blocks that only
7518  * referenced by the tree.
7519  *
7520  * when a shared tree block is found. this function decreases its
7521  * reference count by one. if update_ref is true, this function
7522  * also make sure backrefs for the shared block and all lower level
7523  * blocks are properly updated.
7524  *
7525  * If called with for_reloc == 0, may exit early with -EAGAIN
7526  */
7527 int btrfs_drop_snapshot(struct btrfs_root *root,
7528                          struct btrfs_block_rsv *block_rsv, int update_ref,
7529                          int for_reloc)
7530 {
7531         struct btrfs_path *path;
7532         struct btrfs_trans_handle *trans;
7533         struct btrfs_root *tree_root = root->fs_info->tree_root;
7534         struct btrfs_root_item *root_item = &root->root_item;
7535         struct walk_control *wc;
7536         struct btrfs_key key;
7537         int err = 0;
7538         int ret;
7539         int level;
7540         bool root_dropped = false;
7541
7542         path = btrfs_alloc_path();
7543         if (!path) {
7544                 err = -ENOMEM;
7545                 goto out;
7546         }
7547
7548         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7549         if (!wc) {
7550                 btrfs_free_path(path);
7551                 err = -ENOMEM;
7552                 goto out;
7553         }
7554
7555         trans = btrfs_start_transaction(tree_root, 0);
7556         if (IS_ERR(trans)) {
7557                 err = PTR_ERR(trans);
7558                 goto out_free;
7559         }
7560
7561         if (block_rsv)
7562                 trans->block_rsv = block_rsv;
7563
7564         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7565                 level = btrfs_header_level(root->node);
7566                 path->nodes[level] = btrfs_lock_root_node(root);
7567                 btrfs_set_lock_blocking(path->nodes[level]);
7568                 path->slots[level] = 0;
7569                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7570                 memset(&wc->update_progress, 0,
7571                        sizeof(wc->update_progress));
7572         } else {
7573                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7574                 memcpy(&wc->update_progress, &key,
7575                        sizeof(wc->update_progress));
7576
7577                 level = root_item->drop_level;
7578                 BUG_ON(level == 0);
7579                 path->lowest_level = level;
7580                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7581                 path->lowest_level = 0;
7582                 if (ret < 0) {
7583                         err = ret;
7584                         goto out_end_trans;
7585                 }
7586                 WARN_ON(ret > 0);
7587
7588                 /*
7589                  * unlock our path, this is safe because only this
7590                  * function is allowed to delete this snapshot
7591                  */
7592                 btrfs_unlock_up_safe(path, 0);
7593
7594                 level = btrfs_header_level(root->node);
7595                 while (1) {
7596                         btrfs_tree_lock(path->nodes[level]);
7597                         btrfs_set_lock_blocking(path->nodes[level]);
7598                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7599
7600                         ret = btrfs_lookup_extent_info(trans, root,
7601                                                 path->nodes[level]->start,
7602                                                 level, 1, &wc->refs[level],
7603                                                 &wc->flags[level]);
7604                         if (ret < 0) {
7605                                 err = ret;
7606                                 goto out_end_trans;
7607                         }
7608                         BUG_ON(wc->refs[level] == 0);
7609
7610                         if (level == root_item->drop_level)
7611                                 break;
7612
7613                         btrfs_tree_unlock(path->nodes[level]);
7614                         path->locks[level] = 0;
7615                         WARN_ON(wc->refs[level] != 1);
7616                         level--;
7617                 }
7618         }
7619
7620         wc->level = level;
7621         wc->shared_level = -1;
7622         wc->stage = DROP_REFERENCE;
7623         wc->update_ref = update_ref;
7624         wc->keep_locks = 0;
7625         wc->for_reloc = for_reloc;
7626         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7627
7628         while (1) {
7629
7630                 ret = walk_down_tree(trans, root, path, wc);
7631                 if (ret < 0) {
7632                         err = ret;
7633                         break;
7634                 }
7635
7636                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7637                 if (ret < 0) {
7638                         err = ret;
7639                         break;
7640                 }
7641
7642                 if (ret > 0) {
7643                         BUG_ON(wc->stage != DROP_REFERENCE);
7644                         break;
7645                 }
7646
7647                 if (wc->stage == DROP_REFERENCE) {
7648                         level = wc->level;
7649                         btrfs_node_key(path->nodes[level],
7650                                        &root_item->drop_progress,
7651                                        path->slots[level]);
7652                         root_item->drop_level = level;
7653                 }
7654
7655                 BUG_ON(wc->level == 0);
7656                 if (btrfs_should_end_transaction(trans, tree_root) ||
7657                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
7658                         ret = btrfs_update_root(trans, tree_root,
7659                                                 &root->root_key,
7660                                                 root_item);
7661                         if (ret) {
7662                                 btrfs_abort_transaction(trans, tree_root, ret);
7663                                 err = ret;
7664                                 goto out_end_trans;
7665                         }
7666
7667                         btrfs_end_transaction_throttle(trans, tree_root);
7668                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
7669                                 pr_debug("btrfs: drop snapshot early exit\n");
7670                                 err = -EAGAIN;
7671                                 goto out_free;
7672                         }
7673
7674                         trans = btrfs_start_transaction(tree_root, 0);
7675                         if (IS_ERR(trans)) {
7676                                 err = PTR_ERR(trans);
7677                                 goto out_free;
7678                         }
7679                         if (block_rsv)
7680                                 trans->block_rsv = block_rsv;
7681                 }
7682         }
7683         btrfs_release_path(path);
7684         if (err)
7685                 goto out_end_trans;
7686
7687         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7688         if (ret) {
7689                 btrfs_abort_transaction(trans, tree_root, ret);
7690                 goto out_end_trans;
7691         }
7692
7693         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7694                 ret = btrfs_find_root(tree_root, &root->root_key, path,
7695                                       NULL, NULL);
7696                 if (ret < 0) {
7697                         btrfs_abort_transaction(trans, tree_root, ret);
7698                         err = ret;
7699                         goto out_end_trans;
7700                 } else if (ret > 0) {
7701                         /* if we fail to delete the orphan item this time
7702                          * around, it'll get picked up the next time.
7703                          *
7704                          * The most common failure here is just -ENOENT.
7705                          */
7706                         btrfs_del_orphan_item(trans, tree_root,
7707                                               root->root_key.objectid);
7708                 }
7709         }
7710
7711         if (root->in_radix) {
7712                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
7713         } else {
7714                 free_extent_buffer(root->node);
7715                 free_extent_buffer(root->commit_root);
7716                 btrfs_put_fs_root(root);
7717         }
7718         root_dropped = true;
7719 out_end_trans:
7720         btrfs_end_transaction_throttle(trans, tree_root);
7721 out_free:
7722         kfree(wc);
7723         btrfs_free_path(path);
7724 out:
7725         /*
7726          * So if we need to stop dropping the snapshot for whatever reason we
7727          * need to make sure to add it back to the dead root list so that we
7728          * keep trying to do the work later.  This also cleans up roots if we
7729          * don't have it in the radix (like when we recover after a power fail
7730          * or unmount) so we don't leak memory.
7731          */
7732         if (!for_reloc && root_dropped == false)
7733                 btrfs_add_dead_root(root);
7734         if (err)
7735                 btrfs_std_error(root->fs_info, err);
7736         return err;
7737 }
7738
7739 /*
7740  * drop subtree rooted at tree block 'node'.
7741  *
7742  * NOTE: this function will unlock and release tree block 'node'
7743  * only used by relocation code
7744  */
7745 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7746                         struct btrfs_root *root,
7747                         struct extent_buffer *node,
7748                         struct extent_buffer *parent)
7749 {
7750         struct btrfs_path *path;
7751         struct walk_control *wc;
7752         int level;
7753         int parent_level;
7754         int ret = 0;
7755         int wret;
7756
7757         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7758
7759         path = btrfs_alloc_path();
7760         if (!path)
7761                 return -ENOMEM;
7762
7763         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7764         if (!wc) {
7765                 btrfs_free_path(path);
7766                 return -ENOMEM;
7767         }
7768
7769         btrfs_assert_tree_locked(parent);
7770         parent_level = btrfs_header_level(parent);
7771         extent_buffer_get(parent);
7772         path->nodes[parent_level] = parent;
7773         path->slots[parent_level] = btrfs_header_nritems(parent);
7774
7775         btrfs_assert_tree_locked(node);
7776         level = btrfs_header_level(node);
7777         path->nodes[level] = node;
7778         path->slots[level] = 0;
7779         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7780
7781         wc->refs[parent_level] = 1;
7782         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7783         wc->level = level;
7784         wc->shared_level = -1;
7785         wc->stage = DROP_REFERENCE;
7786         wc->update_ref = 0;
7787         wc->keep_locks = 1;
7788         wc->for_reloc = 1;
7789         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7790
7791         while (1) {
7792                 wret = walk_down_tree(trans, root, path, wc);
7793                 if (wret < 0) {
7794                         ret = wret;
7795                         break;
7796                 }
7797
7798                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7799                 if (wret < 0)
7800                         ret = wret;
7801                 if (wret != 0)
7802                         break;
7803         }
7804
7805         kfree(wc);
7806         btrfs_free_path(path);
7807         return ret;
7808 }
7809
7810 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7811 {
7812         u64 num_devices;
7813         u64 stripped;
7814
7815         /*
7816          * if restripe for this chunk_type is on pick target profile and
7817          * return, otherwise do the usual balance
7818          */
7819         stripped = get_restripe_target(root->fs_info, flags);
7820         if (stripped)
7821                 return extended_to_chunk(stripped);
7822
7823         /*
7824          * we add in the count of missing devices because we want
7825          * to make sure that any RAID levels on a degraded FS
7826          * continue to be honored.
7827          */
7828         num_devices = root->fs_info->fs_devices->rw_devices +
7829                 root->fs_info->fs_devices->missing_devices;
7830
7831         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7832                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7833                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7834
7835         if (num_devices == 1) {
7836                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7837                 stripped = flags & ~stripped;
7838
7839                 /* turn raid0 into single device chunks */
7840                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7841                         return stripped;
7842
7843                 /* turn mirroring into duplication */
7844                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7845                              BTRFS_BLOCK_GROUP_RAID10))
7846                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7847         } else {
7848                 /* they already had raid on here, just return */
7849                 if (flags & stripped)
7850                         return flags;
7851
7852                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7853                 stripped = flags & ~stripped;
7854
7855                 /* switch duplicated blocks with raid1 */
7856                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7857                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7858
7859                 /* this is drive concat, leave it alone */
7860         }
7861
7862         return flags;
7863 }
7864
7865 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7866 {
7867         struct btrfs_space_info *sinfo = cache->space_info;
7868         u64 num_bytes;
7869         u64 min_allocable_bytes;
7870         int ret = -ENOSPC;
7871
7872
7873         /*
7874          * We need some metadata space and system metadata space for
7875          * allocating chunks in some corner cases until we force to set
7876          * it to be readonly.
7877          */
7878         if ((sinfo->flags &
7879              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7880             !force)
7881                 min_allocable_bytes = 1 * 1024 * 1024;
7882         else
7883                 min_allocable_bytes = 0;
7884
7885         spin_lock(&sinfo->lock);
7886         spin_lock(&cache->lock);
7887
7888         if (cache->ro) {
7889                 ret = 0;
7890                 goto out;
7891         }
7892
7893         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7894                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7895
7896         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7897             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7898             min_allocable_bytes <= sinfo->total_bytes) {
7899                 sinfo->bytes_readonly += num_bytes;
7900                 cache->ro = 1;
7901                 ret = 0;
7902         }
7903 out:
7904         spin_unlock(&cache->lock);
7905         spin_unlock(&sinfo->lock);
7906         return ret;
7907 }
7908
7909 int btrfs_set_block_group_ro(struct btrfs_root *root,
7910                              struct btrfs_block_group_cache *cache)
7911
7912 {
7913         struct btrfs_trans_handle *trans;
7914         u64 alloc_flags;
7915         int ret;
7916
7917         BUG_ON(cache->ro);
7918
7919         trans = btrfs_join_transaction(root);
7920         if (IS_ERR(trans))
7921                 return PTR_ERR(trans);
7922
7923         alloc_flags = update_block_group_flags(root, cache->flags);
7924         if (alloc_flags != cache->flags) {
7925                 ret = do_chunk_alloc(trans, root, alloc_flags,
7926                                      CHUNK_ALLOC_FORCE);
7927                 if (ret < 0)
7928                         goto out;
7929         }
7930
7931         ret = set_block_group_ro(cache, 0);
7932         if (!ret)
7933                 goto out;
7934         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7935         ret = do_chunk_alloc(trans, root, alloc_flags,
7936                              CHUNK_ALLOC_FORCE);
7937         if (ret < 0)
7938                 goto out;
7939         ret = set_block_group_ro(cache, 0);
7940 out:
7941         btrfs_end_transaction(trans, root);
7942         return ret;
7943 }
7944
7945 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7946                             struct btrfs_root *root, u64 type)
7947 {
7948         u64 alloc_flags = get_alloc_profile(root, type);
7949         return do_chunk_alloc(trans, root, alloc_flags,
7950                               CHUNK_ALLOC_FORCE);
7951 }
7952
7953 /*
7954  * helper to account the unused space of all the readonly block group in the
7955  * list. takes mirrors into account.
7956  */
7957 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7958 {
7959         struct btrfs_block_group_cache *block_group;
7960         u64 free_bytes = 0;
7961         int factor;
7962
7963         list_for_each_entry(block_group, groups_list, list) {
7964                 spin_lock(&block_group->lock);
7965
7966                 if (!block_group->ro) {
7967                         spin_unlock(&block_group->lock);
7968                         continue;
7969                 }
7970
7971                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7972                                           BTRFS_BLOCK_GROUP_RAID10 |
7973                                           BTRFS_BLOCK_GROUP_DUP))
7974                         factor = 2;
7975                 else
7976                         factor = 1;
7977
7978                 free_bytes += (block_group->key.offset -
7979                                btrfs_block_group_used(&block_group->item)) *
7980                                factor;
7981
7982                 spin_unlock(&block_group->lock);
7983         }
7984
7985         return free_bytes;
7986 }
7987
7988 /*
7989  * helper to account the unused space of all the readonly block group in the
7990  * space_info. takes mirrors into account.
7991  */
7992 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7993 {
7994         int i;
7995         u64 free_bytes = 0;
7996
7997         spin_lock(&sinfo->lock);
7998
7999         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
8000                 if (!list_empty(&sinfo->block_groups[i]))
8001                         free_bytes += __btrfs_get_ro_block_group_free_space(
8002                                                 &sinfo->block_groups[i]);
8003
8004         spin_unlock(&sinfo->lock);
8005
8006         return free_bytes;
8007 }
8008
8009 void btrfs_set_block_group_rw(struct btrfs_root *root,
8010                               struct btrfs_block_group_cache *cache)
8011 {
8012         struct btrfs_space_info *sinfo = cache->space_info;
8013         u64 num_bytes;
8014
8015         BUG_ON(!cache->ro);
8016
8017         spin_lock(&sinfo->lock);
8018         spin_lock(&cache->lock);
8019         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8020                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8021         sinfo->bytes_readonly -= num_bytes;
8022         cache->ro = 0;
8023         spin_unlock(&cache->lock);
8024         spin_unlock(&sinfo->lock);
8025 }
8026
8027 /*
8028  * checks to see if its even possible to relocate this block group.
8029  *
8030  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8031  * ok to go ahead and try.
8032  */
8033 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8034 {
8035         struct btrfs_block_group_cache *block_group;
8036         struct btrfs_space_info *space_info;
8037         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8038         struct btrfs_device *device;
8039         struct btrfs_trans_handle *trans;
8040         u64 min_free;
8041         u64 dev_min = 1;
8042         u64 dev_nr = 0;
8043         u64 target;
8044         int index;
8045         int full = 0;
8046         int ret = 0;
8047
8048         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8049
8050         /* odd, couldn't find the block group, leave it alone */
8051         if (!block_group)
8052                 return -1;
8053
8054         min_free = btrfs_block_group_used(&block_group->item);
8055
8056         /* no bytes used, we're good */
8057         if (!min_free)
8058                 goto out;
8059
8060         space_info = block_group->space_info;
8061         spin_lock(&space_info->lock);
8062
8063         full = space_info->full;
8064
8065         /*
8066          * if this is the last block group we have in this space, we can't
8067          * relocate it unless we're able to allocate a new chunk below.
8068          *
8069          * Otherwise, we need to make sure we have room in the space to handle
8070          * all of the extents from this block group.  If we can, we're good
8071          */
8072         if ((space_info->total_bytes != block_group->key.offset) &&
8073             (space_info->bytes_used + space_info->bytes_reserved +
8074              space_info->bytes_pinned + space_info->bytes_readonly +
8075              min_free < space_info->total_bytes)) {
8076                 spin_unlock(&space_info->lock);
8077                 goto out;
8078         }
8079         spin_unlock(&space_info->lock);
8080
8081         /*
8082          * ok we don't have enough space, but maybe we have free space on our
8083          * devices to allocate new chunks for relocation, so loop through our
8084          * alloc devices and guess if we have enough space.  if this block
8085          * group is going to be restriped, run checks against the target
8086          * profile instead of the current one.
8087          */
8088         ret = -1;
8089
8090         /*
8091          * index:
8092          *      0: raid10
8093          *      1: raid1
8094          *      2: dup
8095          *      3: raid0
8096          *      4: single
8097          */
8098         target = get_restripe_target(root->fs_info, block_group->flags);
8099         if (target) {
8100                 index = __get_raid_index(extended_to_chunk(target));
8101         } else {
8102                 /*
8103                  * this is just a balance, so if we were marked as full
8104                  * we know there is no space for a new chunk
8105                  */
8106                 if (full)
8107                         goto out;
8108
8109                 index = get_block_group_index(block_group);
8110         }
8111
8112         if (index == BTRFS_RAID_RAID10) {
8113                 dev_min = 4;
8114                 /* Divide by 2 */
8115                 min_free >>= 1;
8116         } else if (index == BTRFS_RAID_RAID1) {
8117                 dev_min = 2;
8118         } else if (index == BTRFS_RAID_DUP) {
8119                 /* Multiply by 2 */
8120                 min_free <<= 1;
8121         } else if (index == BTRFS_RAID_RAID0) {
8122                 dev_min = fs_devices->rw_devices;
8123                 do_div(min_free, dev_min);
8124         }
8125
8126         /* We need to do this so that we can look at pending chunks */
8127         trans = btrfs_join_transaction(root);
8128         if (IS_ERR(trans)) {
8129                 ret = PTR_ERR(trans);
8130                 goto out;
8131         }
8132
8133         mutex_lock(&root->fs_info->chunk_mutex);
8134         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8135                 u64 dev_offset;
8136
8137                 /*
8138                  * check to make sure we can actually find a chunk with enough
8139                  * space to fit our block group in.
8140                  */
8141                 if (device->total_bytes > device->bytes_used + min_free &&
8142                     !device->is_tgtdev_for_dev_replace) {
8143                         ret = find_free_dev_extent(trans, device, min_free,
8144                                                    &dev_offset, NULL);
8145                         if (!ret)
8146                                 dev_nr++;
8147
8148                         if (dev_nr >= dev_min)
8149                                 break;
8150
8151                         ret = -1;
8152                 }
8153         }
8154         mutex_unlock(&root->fs_info->chunk_mutex);
8155         btrfs_end_transaction(trans, root);
8156 out:
8157         btrfs_put_block_group(block_group);
8158         return ret;
8159 }
8160
8161 static int find_first_block_group(struct btrfs_root *root,
8162                 struct btrfs_path *path, struct btrfs_key *key)
8163 {
8164         int ret = 0;
8165         struct btrfs_key found_key;
8166         struct extent_buffer *leaf;
8167         int slot;
8168
8169         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8170         if (ret < 0)
8171                 goto out;
8172
8173         while (1) {
8174                 slot = path->slots[0];
8175                 leaf = path->nodes[0];
8176                 if (slot >= btrfs_header_nritems(leaf)) {
8177                         ret = btrfs_next_leaf(root, path);
8178                         if (ret == 0)
8179                                 continue;
8180                         if (ret < 0)
8181                                 goto out;
8182                         break;
8183                 }
8184                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8185
8186                 if (found_key.objectid >= key->objectid &&
8187                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8188                         ret = 0;
8189                         goto out;
8190                 }
8191                 path->slots[0]++;
8192         }
8193 out:
8194         return ret;
8195 }
8196
8197 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8198 {
8199         struct btrfs_block_group_cache *block_group;
8200         u64 last = 0;
8201
8202         while (1) {
8203                 struct inode *inode;
8204
8205                 block_group = btrfs_lookup_first_block_group(info, last);
8206                 while (block_group) {
8207                         spin_lock(&block_group->lock);
8208                         if (block_group->iref)
8209                                 break;
8210                         spin_unlock(&block_group->lock);
8211                         block_group = next_block_group(info->tree_root,
8212                                                        block_group);
8213                 }
8214                 if (!block_group) {
8215                         if (last == 0)
8216                                 break;
8217                         last = 0;
8218                         continue;
8219                 }
8220
8221                 inode = block_group->inode;
8222                 block_group->iref = 0;
8223                 block_group->inode = NULL;
8224                 spin_unlock(&block_group->lock);
8225                 iput(inode);
8226                 last = block_group->key.objectid + block_group->key.offset;
8227                 btrfs_put_block_group(block_group);
8228         }
8229 }
8230
8231 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8232 {
8233         struct btrfs_block_group_cache *block_group;
8234         struct btrfs_space_info *space_info;
8235         struct btrfs_caching_control *caching_ctl;
8236         struct rb_node *n;
8237
8238         down_write(&info->extent_commit_sem);
8239         while (!list_empty(&info->caching_block_groups)) {
8240                 caching_ctl = list_entry(info->caching_block_groups.next,
8241                                          struct btrfs_caching_control, list);
8242                 list_del(&caching_ctl->list);
8243                 put_caching_control(caching_ctl);
8244         }
8245         up_write(&info->extent_commit_sem);
8246
8247         spin_lock(&info->block_group_cache_lock);
8248         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8249                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8250                                        cache_node);
8251                 rb_erase(&block_group->cache_node,
8252                          &info->block_group_cache_tree);
8253                 spin_unlock(&info->block_group_cache_lock);
8254
8255                 down_write(&block_group->space_info->groups_sem);
8256                 list_del(&block_group->list);
8257                 up_write(&block_group->space_info->groups_sem);
8258
8259                 if (block_group->cached == BTRFS_CACHE_STARTED)
8260                         wait_block_group_cache_done(block_group);
8261
8262                 /*
8263                  * We haven't cached this block group, which means we could
8264                  * possibly have excluded extents on this block group.
8265                  */
8266                 if (block_group->cached == BTRFS_CACHE_NO ||
8267                     block_group->cached == BTRFS_CACHE_ERROR)
8268                         free_excluded_extents(info->extent_root, block_group);
8269
8270                 btrfs_remove_free_space_cache(block_group);
8271                 btrfs_put_block_group(block_group);
8272
8273                 spin_lock(&info->block_group_cache_lock);
8274         }
8275         spin_unlock(&info->block_group_cache_lock);
8276
8277         /* now that all the block groups are freed, go through and
8278          * free all the space_info structs.  This is only called during
8279          * the final stages of unmount, and so we know nobody is
8280          * using them.  We call synchronize_rcu() once before we start,
8281          * just to be on the safe side.
8282          */
8283         synchronize_rcu();
8284
8285         release_global_block_rsv(info);
8286
8287         while(!list_empty(&info->space_info)) {
8288                 space_info = list_entry(info->space_info.next,
8289                                         struct btrfs_space_info,
8290                                         list);
8291                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8292                         if (space_info->bytes_pinned > 0 ||
8293                             space_info->bytes_reserved > 0 ||
8294                             space_info->bytes_may_use > 0) {
8295                                 WARN_ON(1);
8296                                 dump_space_info(space_info, 0, 0);
8297                         }
8298                 }
8299                 percpu_counter_destroy(&space_info->total_bytes_pinned);
8300                 list_del(&space_info->list);
8301                 kfree(space_info);
8302         }
8303         return 0;
8304 }
8305
8306 static void __link_block_group(struct btrfs_space_info *space_info,
8307                                struct btrfs_block_group_cache *cache)
8308 {
8309         int index = get_block_group_index(cache);
8310
8311         down_write(&space_info->groups_sem);
8312         list_add_tail(&cache->list, &space_info->block_groups[index]);
8313         up_write(&space_info->groups_sem);
8314 }
8315
8316 int btrfs_read_block_groups(struct btrfs_root *root)
8317 {
8318         struct btrfs_path *path;
8319         int ret;
8320         struct btrfs_block_group_cache *cache;
8321         struct btrfs_fs_info *info = root->fs_info;
8322         struct btrfs_space_info *space_info;
8323         struct btrfs_key key;
8324         struct btrfs_key found_key;
8325         struct extent_buffer *leaf;
8326         int need_clear = 0;
8327         u64 cache_gen;
8328
8329         root = info->extent_root;
8330         key.objectid = 0;
8331         key.offset = 0;
8332         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8333         path = btrfs_alloc_path();
8334         if (!path)
8335                 return -ENOMEM;
8336         path->reada = 1;
8337
8338         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
8339         if (btrfs_test_opt(root, SPACE_CACHE) &&
8340             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
8341                 need_clear = 1;
8342         if (btrfs_test_opt(root, CLEAR_CACHE))
8343                 need_clear = 1;
8344
8345         while (1) {
8346                 ret = find_first_block_group(root, path, &key);
8347                 if (ret > 0)
8348                         break;
8349                 if (ret != 0)
8350                         goto error;
8351                 leaf = path->nodes[0];
8352                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8353                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
8354                 if (!cache) {
8355                         ret = -ENOMEM;
8356                         goto error;
8357                 }
8358                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8359                                                 GFP_NOFS);
8360                 if (!cache->free_space_ctl) {
8361                         kfree(cache);
8362                         ret = -ENOMEM;
8363                         goto error;
8364                 }
8365
8366                 atomic_set(&cache->count, 1);
8367                 spin_lock_init(&cache->lock);
8368                 cache->fs_info = info;
8369                 INIT_LIST_HEAD(&cache->list);
8370                 INIT_LIST_HEAD(&cache->cluster_list);
8371
8372                 if (need_clear) {
8373                         /*
8374                          * When we mount with old space cache, we need to
8375                          * set BTRFS_DC_CLEAR and set dirty flag.
8376                          *
8377                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8378                          *    truncate the old free space cache inode and
8379                          *    setup a new one.
8380                          * b) Setting 'dirty flag' makes sure that we flush
8381                          *    the new space cache info onto disk.
8382                          */
8383                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8384                         if (btrfs_test_opt(root, SPACE_CACHE))
8385                                 cache->dirty = 1;
8386                 }
8387
8388                 read_extent_buffer(leaf, &cache->item,
8389                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8390                                    sizeof(cache->item));
8391                 memcpy(&cache->key, &found_key, sizeof(found_key));
8392
8393                 key.objectid = found_key.objectid + found_key.offset;
8394                 btrfs_release_path(path);
8395                 cache->flags = btrfs_block_group_flags(&cache->item);
8396                 cache->sectorsize = root->sectorsize;
8397                 cache->full_stripe_len = btrfs_full_stripe_len(root,
8398                                                &root->fs_info->mapping_tree,
8399                                                found_key.objectid);
8400                 btrfs_init_free_space_ctl(cache);
8401
8402                 /*
8403                  * We need to exclude the super stripes now so that the space
8404                  * info has super bytes accounted for, otherwise we'll think
8405                  * we have more space than we actually do.
8406                  */
8407                 ret = exclude_super_stripes(root, cache);
8408                 if (ret) {
8409                         /*
8410                          * We may have excluded something, so call this just in
8411                          * case.
8412                          */
8413                         free_excluded_extents(root, cache);
8414                         kfree(cache->free_space_ctl);
8415                         kfree(cache);
8416                         goto error;
8417                 }
8418
8419                 /*
8420                  * check for two cases, either we are full, and therefore
8421                  * don't need to bother with the caching work since we won't
8422                  * find any space, or we are empty, and we can just add all
8423                  * the space in and be done with it.  This saves us _alot_ of
8424                  * time, particularly in the full case.
8425                  */
8426                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8427                         cache->last_byte_to_unpin = (u64)-1;
8428                         cache->cached = BTRFS_CACHE_FINISHED;
8429                         free_excluded_extents(root, cache);
8430                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8431                         cache->last_byte_to_unpin = (u64)-1;
8432                         cache->cached = BTRFS_CACHE_FINISHED;
8433                         add_new_free_space(cache, root->fs_info,
8434                                            found_key.objectid,
8435                                            found_key.objectid +
8436                                            found_key.offset);
8437                         free_excluded_extents(root, cache);
8438                 }
8439
8440                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8441                 if (ret) {
8442                         btrfs_remove_free_space_cache(cache);
8443                         btrfs_put_block_group(cache);
8444                         goto error;
8445                 }
8446
8447                 ret = update_space_info(info, cache->flags, found_key.offset,
8448                                         btrfs_block_group_used(&cache->item),
8449                                         &space_info);
8450                 if (ret) {
8451                         btrfs_remove_free_space_cache(cache);
8452                         spin_lock(&info->block_group_cache_lock);
8453                         rb_erase(&cache->cache_node,
8454                                  &info->block_group_cache_tree);
8455                         spin_unlock(&info->block_group_cache_lock);
8456                         btrfs_put_block_group(cache);
8457                         goto error;
8458                 }
8459
8460                 cache->space_info = space_info;
8461                 spin_lock(&cache->space_info->lock);
8462                 cache->space_info->bytes_readonly += cache->bytes_super;
8463                 spin_unlock(&cache->space_info->lock);
8464
8465                 __link_block_group(space_info, cache);
8466
8467                 set_avail_alloc_bits(root->fs_info, cache->flags);
8468                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8469                         set_block_group_ro(cache, 1);
8470         }
8471
8472         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8473                 if (!(get_alloc_profile(root, space_info->flags) &
8474                       (BTRFS_BLOCK_GROUP_RAID10 |
8475                        BTRFS_BLOCK_GROUP_RAID1 |
8476                        BTRFS_BLOCK_GROUP_RAID5 |
8477                        BTRFS_BLOCK_GROUP_RAID6 |
8478                        BTRFS_BLOCK_GROUP_DUP)))
8479                         continue;
8480                 /*
8481                  * avoid allocating from un-mirrored block group if there are
8482                  * mirrored block groups.
8483                  */
8484                 list_for_each_entry(cache,
8485                                 &space_info->block_groups[BTRFS_RAID_RAID0],
8486                                 list)
8487                         set_block_group_ro(cache, 1);
8488                 list_for_each_entry(cache,
8489                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
8490                                 list)
8491                         set_block_group_ro(cache, 1);
8492         }
8493
8494         init_global_block_rsv(info);
8495         ret = 0;
8496 error:
8497         btrfs_free_path(path);
8498         return ret;
8499 }
8500
8501 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8502                                        struct btrfs_root *root)
8503 {
8504         struct btrfs_block_group_cache *block_group, *tmp;
8505         struct btrfs_root *extent_root = root->fs_info->extent_root;
8506         struct btrfs_block_group_item item;
8507         struct btrfs_key key;
8508         int ret = 0;
8509
8510         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8511                                  new_bg_list) {
8512                 list_del_init(&block_group->new_bg_list);
8513
8514                 if (ret)
8515                         continue;
8516
8517                 spin_lock(&block_group->lock);
8518                 memcpy(&item, &block_group->item, sizeof(item));
8519                 memcpy(&key, &block_group->key, sizeof(key));
8520                 spin_unlock(&block_group->lock);
8521
8522                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8523                                         sizeof(item));
8524                 if (ret)
8525                         btrfs_abort_transaction(trans, extent_root, ret);
8526                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
8527                                                key.objectid, key.offset);
8528                 if (ret)
8529                         btrfs_abort_transaction(trans, extent_root, ret);
8530         }
8531 }
8532
8533 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8534                            struct btrfs_root *root, u64 bytes_used,
8535                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8536                            u64 size)
8537 {
8538         int ret;
8539         struct btrfs_root *extent_root;
8540         struct btrfs_block_group_cache *cache;
8541
8542         extent_root = root->fs_info->extent_root;
8543
8544         root->fs_info->last_trans_log_full_commit = trans->transid;
8545
8546         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8547         if (!cache)
8548                 return -ENOMEM;
8549         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8550                                         GFP_NOFS);
8551         if (!cache->free_space_ctl) {
8552                 kfree(cache);
8553                 return -ENOMEM;
8554         }
8555
8556         cache->key.objectid = chunk_offset;
8557         cache->key.offset = size;
8558         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8559         cache->sectorsize = root->sectorsize;
8560         cache->fs_info = root->fs_info;
8561         cache->full_stripe_len = btrfs_full_stripe_len(root,
8562                                                &root->fs_info->mapping_tree,
8563                                                chunk_offset);
8564
8565         atomic_set(&cache->count, 1);
8566         spin_lock_init(&cache->lock);
8567         INIT_LIST_HEAD(&cache->list);
8568         INIT_LIST_HEAD(&cache->cluster_list);
8569         INIT_LIST_HEAD(&cache->new_bg_list);
8570
8571         btrfs_init_free_space_ctl(cache);
8572
8573         btrfs_set_block_group_used(&cache->item, bytes_used);
8574         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8575         cache->flags = type;
8576         btrfs_set_block_group_flags(&cache->item, type);
8577
8578         cache->last_byte_to_unpin = (u64)-1;
8579         cache->cached = BTRFS_CACHE_FINISHED;
8580         ret = exclude_super_stripes(root, cache);
8581         if (ret) {
8582                 /*
8583                  * We may have excluded something, so call this just in
8584                  * case.
8585                  */
8586                 free_excluded_extents(root, cache);
8587                 kfree(cache->free_space_ctl);
8588                 kfree(cache);
8589                 return ret;
8590         }
8591
8592         add_new_free_space(cache, root->fs_info, chunk_offset,
8593                            chunk_offset + size);
8594
8595         free_excluded_extents(root, cache);
8596
8597         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8598         if (ret) {
8599                 btrfs_remove_free_space_cache(cache);
8600                 btrfs_put_block_group(cache);
8601                 return ret;
8602         }
8603
8604         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8605                                 &cache->space_info);
8606         if (ret) {
8607                 btrfs_remove_free_space_cache(cache);
8608                 spin_lock(&root->fs_info->block_group_cache_lock);
8609                 rb_erase(&cache->cache_node,
8610                          &root->fs_info->block_group_cache_tree);
8611                 spin_unlock(&root->fs_info->block_group_cache_lock);
8612                 btrfs_put_block_group(cache);
8613                 return ret;
8614         }
8615         update_global_block_rsv(root->fs_info);
8616
8617         spin_lock(&cache->space_info->lock);
8618         cache->space_info->bytes_readonly += cache->bytes_super;
8619         spin_unlock(&cache->space_info->lock);
8620
8621         __link_block_group(cache->space_info, cache);
8622
8623         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8624
8625         set_avail_alloc_bits(extent_root->fs_info, type);
8626
8627         return 0;
8628 }
8629
8630 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8631 {
8632         u64 extra_flags = chunk_to_extended(flags) &
8633                                 BTRFS_EXTENDED_PROFILE_MASK;
8634
8635         write_seqlock(&fs_info->profiles_lock);
8636         if (flags & BTRFS_BLOCK_GROUP_DATA)
8637                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8638         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8639                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8640         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8641                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8642         write_sequnlock(&fs_info->profiles_lock);
8643 }
8644
8645 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8646                              struct btrfs_root *root, u64 group_start)
8647 {
8648         struct btrfs_path *path;
8649         struct btrfs_block_group_cache *block_group;
8650         struct btrfs_free_cluster *cluster;
8651         struct btrfs_root *tree_root = root->fs_info->tree_root;
8652         struct btrfs_key key;
8653         struct inode *inode;
8654         int ret;
8655         int index;
8656         int factor;
8657
8658         root = root->fs_info->extent_root;
8659
8660         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8661         BUG_ON(!block_group);
8662         BUG_ON(!block_group->ro);
8663
8664         /*
8665          * Free the reserved super bytes from this block group before
8666          * remove it.
8667          */
8668         free_excluded_extents(root, block_group);
8669
8670         memcpy(&key, &block_group->key, sizeof(key));
8671         index = get_block_group_index(block_group);
8672         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8673                                   BTRFS_BLOCK_GROUP_RAID1 |
8674                                   BTRFS_BLOCK_GROUP_RAID10))
8675                 factor = 2;
8676         else
8677                 factor = 1;
8678
8679         /* make sure this block group isn't part of an allocation cluster */
8680         cluster = &root->fs_info->data_alloc_cluster;
8681         spin_lock(&cluster->refill_lock);
8682         btrfs_return_cluster_to_free_space(block_group, cluster);
8683         spin_unlock(&cluster->refill_lock);
8684
8685         /*
8686          * make sure this block group isn't part of a metadata
8687          * allocation cluster
8688          */
8689         cluster = &root->fs_info->meta_alloc_cluster;
8690         spin_lock(&cluster->refill_lock);
8691         btrfs_return_cluster_to_free_space(block_group, cluster);
8692         spin_unlock(&cluster->refill_lock);
8693
8694         path = btrfs_alloc_path();
8695         if (!path) {
8696                 ret = -ENOMEM;
8697                 goto out;
8698         }
8699
8700         inode = lookup_free_space_inode(tree_root, block_group, path);
8701         if (!IS_ERR(inode)) {
8702                 ret = btrfs_orphan_add(trans, inode);
8703                 if (ret) {
8704                         btrfs_add_delayed_iput(inode);
8705                         goto out;
8706                 }
8707                 clear_nlink(inode);
8708                 /* One for the block groups ref */
8709                 spin_lock(&block_group->lock);
8710                 if (block_group->iref) {
8711                         block_group->iref = 0;
8712                         block_group->inode = NULL;
8713                         spin_unlock(&block_group->lock);
8714                         iput(inode);
8715                 } else {
8716                         spin_unlock(&block_group->lock);
8717                 }
8718                 /* One for our lookup ref */
8719                 btrfs_add_delayed_iput(inode);
8720         }
8721
8722         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8723         key.offset = block_group->key.objectid;
8724         key.type = 0;
8725
8726         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8727         if (ret < 0)
8728                 goto out;
8729         if (ret > 0)
8730                 btrfs_release_path(path);
8731         if (ret == 0) {
8732                 ret = btrfs_del_item(trans, tree_root, path);
8733                 if (ret)
8734                         goto out;
8735                 btrfs_release_path(path);
8736         }
8737
8738         spin_lock(&root->fs_info->block_group_cache_lock);
8739         rb_erase(&block_group->cache_node,
8740                  &root->fs_info->block_group_cache_tree);
8741
8742         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8743                 root->fs_info->first_logical_byte = (u64)-1;
8744         spin_unlock(&root->fs_info->block_group_cache_lock);
8745
8746         down_write(&block_group->space_info->groups_sem);
8747         /*
8748          * we must use list_del_init so people can check to see if they
8749          * are still on the list after taking the semaphore
8750          */
8751         list_del_init(&block_group->list);
8752         if (list_empty(&block_group->space_info->block_groups[index]))
8753                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8754         up_write(&block_group->space_info->groups_sem);
8755
8756         if (block_group->cached == BTRFS_CACHE_STARTED)
8757                 wait_block_group_cache_done(block_group);
8758
8759         btrfs_remove_free_space_cache(block_group);
8760
8761         spin_lock(&block_group->space_info->lock);
8762         block_group->space_info->total_bytes -= block_group->key.offset;
8763         block_group->space_info->bytes_readonly -= block_group->key.offset;
8764         block_group->space_info->disk_total -= block_group->key.offset * factor;
8765         spin_unlock(&block_group->space_info->lock);
8766
8767         memcpy(&key, &block_group->key, sizeof(key));
8768
8769         btrfs_clear_space_info_full(root->fs_info);
8770
8771         btrfs_put_block_group(block_group);
8772         btrfs_put_block_group(block_group);
8773
8774         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8775         if (ret > 0)
8776                 ret = -EIO;
8777         if (ret < 0)
8778                 goto out;
8779
8780         ret = btrfs_del_item(trans, root, path);
8781 out:
8782         btrfs_free_path(path);
8783         return ret;
8784 }
8785
8786 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8787 {
8788         struct btrfs_space_info *space_info;
8789         struct btrfs_super_block *disk_super;
8790         u64 features;
8791         u64 flags;
8792         int mixed = 0;
8793         int ret;
8794
8795         disk_super = fs_info->super_copy;
8796         if (!btrfs_super_root(disk_super))
8797                 return 1;
8798
8799         features = btrfs_super_incompat_flags(disk_super);
8800         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8801                 mixed = 1;
8802
8803         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8804         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8805         if (ret)
8806                 goto out;
8807
8808         if (mixed) {
8809                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8810                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8811         } else {
8812                 flags = BTRFS_BLOCK_GROUP_METADATA;
8813                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8814                 if (ret)
8815                         goto out;
8816
8817                 flags = BTRFS_BLOCK_GROUP_DATA;
8818                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8819         }
8820 out:
8821         return ret;
8822 }
8823
8824 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8825 {
8826         return unpin_extent_range(root, start, end);
8827 }
8828
8829 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8830                                u64 num_bytes, u64 *actual_bytes)
8831 {
8832         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8833 }
8834
8835 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8836 {
8837         struct btrfs_fs_info *fs_info = root->fs_info;
8838         struct btrfs_block_group_cache *cache = NULL;
8839         u64 group_trimmed;
8840         u64 start;
8841         u64 end;
8842         u64 trimmed = 0;
8843         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8844         int ret = 0;
8845
8846         /*
8847          * try to trim all FS space, our block group may start from non-zero.
8848          */
8849         if (range->len == total_bytes)
8850                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8851         else
8852                 cache = btrfs_lookup_block_group(fs_info, range->start);
8853
8854         while (cache) {
8855                 if (cache->key.objectid >= (range->start + range->len)) {
8856                         btrfs_put_block_group(cache);
8857                         break;
8858                 }
8859
8860                 start = max(range->start, cache->key.objectid);
8861                 end = min(range->start + range->len,
8862                                 cache->key.objectid + cache->key.offset);
8863
8864                 if (end - start >= range->minlen) {
8865                         if (!block_group_cache_done(cache)) {
8866                                 ret = cache_block_group(cache, 0);
8867                                 if (ret) {
8868                                         btrfs_put_block_group(cache);
8869                                         break;
8870                                 }
8871                                 ret = wait_block_group_cache_done(cache);
8872                                 if (ret) {
8873                                         btrfs_put_block_group(cache);
8874                                         break;
8875                                 }
8876                         }
8877                         ret = btrfs_trim_block_group(cache,
8878                                                      &group_trimmed,
8879                                                      start,
8880                                                      end,
8881                                                      range->minlen);
8882
8883                         trimmed += group_trimmed;
8884                         if (ret) {
8885                                 btrfs_put_block_group(cache);
8886                                 break;
8887                         }
8888                 }
8889
8890                 cache = next_block_group(fs_info->tree_root, cache);
8891         }
8892
8893         range->len = trimmed;
8894         return ret;
8895 }