fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include <linux/percpu_counter.h>
  28 #include "hash.h"
  29 #include "tree-log.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "volumes.h"
  33 #include "raid56.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36 #include "math.h"
  37 #include "sysfs.h"
  38 #include "qgroup.h"
  39
  40 #undef SCRAMBLE_DELAYED_REFS
  41
  42 /*
  43  * control flags for do_chunk_alloc's force field
  44  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  45  * if we really need one.
  46  *
  47  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  48  * if we have very few chunks already allocated.  This is
  49  * used as part of the clustering code to help make sure
  50  * we have a good pool of storage to cluster in, without
  51  * filling the FS with empty chunks
  52  *
  53  * CHUNK_ALLOC_FORCE means it must try to allocate one
  54  *
  55  */
  56 enum {
  57         CHUNK_ALLOC_NO_FORCE = 0,
  58         CHUNK_ALLOC_LIMITED = 1,
  59         CHUNK_ALLOC_FORCE = 2,
  60 };
  61
  62 /*
  63  * Control how reservations are dealt with.
  64  *
  65  * RESERVE_FREE - freeing a reservation.
  66  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  67  *   ENOSPC accounting
  68  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  69  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  70  */
  71 enum {
  72         RESERVE_FREE = 0,
  73         RESERVE_ALLOC = 1,
  74         RESERVE_ALLOC_NO_ACCOUNT = 2,
  75 };
  76
  77 static int update_block_group(struct btrfs_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                                 int no_quota);
  86 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  87                                     struct extent_buffer *leaf,
  88                                     struct btrfs_extent_item *ei);
  89 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  90                                       struct btrfs_root *root,
  91                                       u64 parent, u64 root_objectid,
  92                                       u64 flags, u64 owner, u64 offset,
  93                                       struct btrfs_key *ins, int ref_mod);
  94 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  95                                      struct btrfs_root *root,
  96                                      u64 parent, u64 root_objectid,
  97                                      u64 flags, struct btrfs_disk_key *key,
  98                                      int level, struct btrfs_key *ins,
  99                                      int no_quota);
 100 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
 101                           struct btrfs_root *extent_root, u64 flags,
 102                           int force);
 103 static int find_next_key(struct btrfs_path *path, int level,
 104                          struct btrfs_key *key);
 105 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 106                             int dump_block_groups);
 107 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 108                                        u64 num_bytes, int reserve);
 109 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 110                                u64 num_bytes);
 111 int btrfs_pin_extent(struct btrfs_root *root,
 112                      u64 bytenr, u64 num_bytes, int reserved);
 113
 114 static noinline int
 115 block_group_cache_done(struct btrfs_block_group_cache *cache)
 116 {
 117         smp_mb();
 118         return cache->cached == BTRFS_CACHE_FINISHED ||
 119                 cache->cached == BTRFS_CACHE_ERROR;
 120 }
 121
 122 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 123 {
 124         return (cache->flags & bits) == bits;
 125 }
 126
 127 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 128 {
 129         atomic_inc(&cache->count);
 130 }
 131
 132 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 133 {
 134         if (atomic_dec_and_test(&cache->count)) {
 135                 WARN_ON(cache->pinned > 0);
 136                 WARN_ON(cache->reserved > 0);
 137                 kfree(cache->free_space_ctl);
 138                 kfree(cache);
 139         }
 140 }
 141
 142 /*
 143  * this adds the block group to the fs_info rb tree for the block group
 144  * cache
 145  */
 146 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 147                                 struct btrfs_block_group_cache *block_group)
 148 {
 149         struct rb_node **p;
 150         struct rb_node *parent = NULL;
 151         struct btrfs_block_group_cache *cache;
 152
 153         spin_lock(&info->block_group_cache_lock);
 154         p = &info->block_group_cache_tree.rb_node;
 155
 156         while (*p) {
 157                 parent = *p;
 158                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 159                                  cache_node);
 160                 if (block_group->key.objectid < cache->key.objectid) {
 161                         p = &(*p)->rb_left;
 162                 } else if (block_group->key.objectid > cache->key.objectid) {
 163                         p = &(*p)->rb_right;
 164                 } else {
 165                         spin_unlock(&info->block_group_cache_lock);
 166                         return -EEXIST;
 167                 }
 168         }
 169
 170         rb_link_node(&block_group->cache_node, parent, p);
 171         rb_insert_color(&block_group->cache_node,
 172                         &info->block_group_cache_tree);
 173
 174         if (info->first_logical_byte > block_group->key.objectid)
 175                 info->first_logical_byte = block_group->key.objectid;
 176
 177         spin_unlock(&info->block_group_cache_lock);
 178
 179         return 0;
 180 }
 181
 182 /*
 183  * This will return the block group at or after bytenr if contains is 0, else
 184  * it will return the block group that contains the bytenr
 185  */
 186 static struct btrfs_block_group_cache *
 187 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 188                               int contains)
 189 {
 190         struct btrfs_block_group_cache *cache, *ret = NULL;
 191         struct rb_node *n;
 192         u64 end, start;
 193
 194         spin_lock(&info->block_group_cache_lock);
 195         n = info->block_group_cache_tree.rb_node;
 196
 197         while (n) {
 198                 cache = rb_entry(n, struct btrfs_block_group_cache,
 199                                  cache_node);
 200                 end = cache->key.objectid + cache->key.offset - 1;
 201                 start = cache->key.objectid;
 202
 203                 if (bytenr < start) {
 204                         if (!contains && (!ret || start < ret->key.objectid))
 205                                 ret = cache;
 206                         n = n->rb_left;
 207                 } else if (bytenr > start) {
 208                         if (contains && bytenr <= end) {
 209                                 ret = cache;
 210                                 break;
 211                         }
 212                         n = n->rb_right;
 213                 } else {
 214                         ret = cache;
 215                         break;
 216                 }
 217         }
 218         if (ret) {
 219                 btrfs_get_block_group(ret);
 220                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 221                         info->first_logical_byte = ret->key.objectid;
 222         }
 223         spin_unlock(&info->block_group_cache_lock);
 224
 225         return ret;
 226 }
 227
 228 static int add_excluded_extent(struct btrfs_root *root,
 229                                u64 start, u64 num_bytes)
 230 {
 231         u64 end = start + num_bytes - 1;
 232         set_extent_bits(&root->fs_info->freed_extents[0],
 233                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 234         set_extent_bits(&root->fs_info->freed_extents[1],
 235                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 236         return 0;
 237 }
 238
 239 static void free_excluded_extents(struct btrfs_root *root,
 240                                   struct btrfs_block_group_cache *cache)
 241 {
 242         u64 start, end;
 243
 244         start = cache->key.objectid;
 245         end = start + cache->key.offset - 1;
 246
 247         clear_extent_bits(&root->fs_info->freed_extents[0],
 248                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 249         clear_extent_bits(&root->fs_info->freed_extents[1],
 250                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 251 }
 252
 253 static int exclude_super_stripes(struct btrfs_root *root,
 254                                  struct btrfs_block_group_cache *cache)
 255 {
 256         u64 bytenr;
 257         u64 *logical;
 258         int stripe_len;
 259         int i, nr, ret;
 260
 261         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 262                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 263                 cache->bytes_super += stripe_len;
 264                 ret = add_excluded_extent(root, cache->key.objectid,
 265                                           stripe_len);
 266                 if (ret)
 267                         return ret;
 268         }
 269
 270         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 271                 bytenr = btrfs_sb_offset(i);
 272                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 273                                        cache->key.objectid, bytenr,
 274                                        0, &logical, &nr, &stripe_len);
 275                 if (ret)
 276                         return ret;
 277
 278                 while (nr--) {
 279                         u64 start, len;
 280
 281                         if (logical[nr] > cache->key.objectid +
 282                             cache->key.offset)
 283                                 continue;
 284
 285                         if (logical[nr] + stripe_len <= cache->key.objectid)
 286                                 continue;
 287
 288                         start = logical[nr];
 289                         if (start < cache->key.objectid) {
 290                                 start = cache->key.objectid;
 291                                 len = (logical[nr] + stripe_len) - start;
 292                         } else {
 293                                 len = min_t(u64, stripe_len,
 294                                             cache->key.objectid +
 295                                             cache->key.offset - start);
 296                         }
 297
 298                         cache->bytes_super += len;
 299                         ret = add_excluded_extent(root, start, len);
 300                         if (ret) {
 301                                 kfree(logical);
 302                                 return ret;
 303                         }
 304                 }
 305
 306                 kfree(logical);
 307         }
 308         return 0;
 309 }
 310
 311 static struct btrfs_caching_control *
 312 get_caching_control(struct btrfs_block_group_cache *cache)
 313 {
 314         struct btrfs_caching_control *ctl;
 315
 316         spin_lock(&cache->lock);
 317         if (cache->cached != BTRFS_CACHE_STARTED) {
 318                 spin_unlock(&cache->lock);
 319                 return NULL;
 320         }
 321
 322         /* We're loading it the fast way, so we don't have a caching_ctl. */
 323         if (!cache->caching_ctl) {
 324                 spin_unlock(&cache->lock);
 325                 return NULL;
 326         }
 327
 328         ctl = cache->caching_ctl;
 329         atomic_inc(&ctl->count);
 330         spin_unlock(&cache->lock);
 331         return ctl;
 332 }
 333
 334 static void put_caching_control(struct btrfs_caching_control *ctl)
 335 {
 336         if (atomic_dec_and_test(&ctl->count))
 337                 kfree(ctl);
 338 }
 339
 340 /*
 341  * this is only called by cache_block_group, since we could have freed extents
 342  * we need to check the pinned_extents for any extents that can't be used yet
 343  * since their free space will be released as soon as the transaction commits.
 344  */
 345 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 346                               struct btrfs_fs_info *info, u64 start, u64 end)
 347 {
 348         u64 extent_start, extent_end, size, total_added = 0;
 349         int ret;
 350
 351         while (start < end) {
 352                 ret = find_first_extent_bit(info->pinned_extents, start,
 353                                             &extent_start, &extent_end,
 354                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 355                                             NULL);
 356                 if (ret)
 357                         break;
 358
 359                 if (extent_start <= start) {
 360                         start = extent_end + 1;
 361                 } else if (extent_start > start && extent_start < end) {
 362                         size = extent_start - start;
 363                         total_added += size;
 364                         ret = btrfs_add_free_space(block_group, start,
 365                                                    size);
 366                         BUG_ON(ret); /* -ENOMEM or logic error */
 367                         start = extent_end + 1;
 368                 } else {
 369                         break;
 370                 }
 371         }
 372
 373         if (start < end) {
 374                 size = end - start;
 375                 total_added += size;
 376                 ret = btrfs_add_free_space(block_group, start, size);
 377                 BUG_ON(ret); /* -ENOMEM or logic error */
 378         }
 379
 380         return total_added;
 381 }
 382
 383 static noinline void caching_thread(struct btrfs_work *work)
 384 {
 385         struct btrfs_block_group_cache *block_group;
 386         struct btrfs_fs_info *fs_info;
 387         struct btrfs_caching_control *caching_ctl;
 388         struct btrfs_root *extent_root;
 389         struct btrfs_path *path;
 390         struct extent_buffer *leaf;
 391         struct btrfs_key key;
 392         u64 total_found = 0;
 393         u64 last = 0;
 394         u32 nritems;
 395         int ret = -ENOMEM;
 396
 397         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 398         block_group = caching_ctl->block_group;
 399         fs_info = block_group->fs_info;
 400         extent_root = fs_info->extent_root;
 401
 402         path = btrfs_alloc_path();
 403         if (!path)
 404                 goto out;
 405
 406         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 407
 408         /*
 409          * We don't want to deadlock with somebody trying to allocate a new
 410          * extent for the extent root while also trying to search the extent
 411          * root to add free space.  So we skip locking and search the commit
 412          * root, since its read-only
 413          */
 414         path->skip_locking = 1;
 415         path->search_commit_root = 1;
 416         path->reada = 1;
 417
 418         key.objectid = last;
 419         key.offset = 0;
 420         key.type = BTRFS_EXTENT_ITEM_KEY;
 421 again:
 422         mutex_lock(&caching_ctl->mutex);
 423         /* need to make sure the commit_root doesn't disappear */
 424         down_read(&fs_info->commit_root_sem);
 425
 426 next:
 427         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 428         if (ret < 0)
 429                 goto err;
 430
 431         leaf = path->nodes[0];
 432         nritems = btrfs_header_nritems(leaf);
 433
 434         while (1) {
 435                 if (btrfs_fs_closing(fs_info) > 1) {
 436                         last = (u64)-1;
 437                         break;
 438                 }
 439
 440                 if (path->slots[0] < nritems) {
 441                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 442                 } else {
 443                         ret = find_next_key(path, 0, &key);
 444                         if (ret)
 445                                 break;
 446
 447                         if (need_resched() ||
 448                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 449                                 caching_ctl->progress = last;
 450                                 btrfs_release_path(path);
 451                                 up_read(&fs_info->commit_root_sem);
 452                                 mutex_unlock(&caching_ctl->mutex);
 453                                 cond_resched();
 454                                 goto again;
 455                         }
 456
 457                         ret = btrfs_next_leaf(extent_root, path);
 458                         if (ret < 0)
 459                                 goto err;
 460                         if (ret)
 461                                 break;
 462                         leaf = path->nodes[0];
 463                         nritems = btrfs_header_nritems(leaf);
 464                         continue;
 465                 }
 466
 467                 if (key.objectid < last) {
 468                         key.objectid = last;
 469                         key.offset = 0;
 470                         key.type = BTRFS_EXTENT_ITEM_KEY;
 471
 472                         caching_ctl->progress = last;
 473                         btrfs_release_path(path);
 474                         goto next;
 475                 }
 476
 477                 if (key.objectid < block_group->key.objectid) {
 478                         path->slots[0]++;
 479                         continue;
 480                 }
 481
 482                 if (key.objectid >= block_group->key.objectid +
 483                     block_group->key.offset)
 484                         break;
 485
 486                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 487                     key.type == BTRFS_METADATA_ITEM_KEY) {
 488                         total_found += add_new_free_space(block_group,
 489                                                           fs_info, last,
 490                                                           key.objectid);
 491                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 492                                 last = key.objectid +
 493                                         fs_info->tree_root->leafsize;
 494                         else
 495                                 last = key.objectid + key.offset;
 496
 497                         if (total_found > (1024 * 1024 * 2)) {
 498                                 total_found = 0;
 499                                 wake_up(&caching_ctl->wait);
 500                         }
 501                 }
 502                 path->slots[0]++;
 503         }
 504         ret = 0;
 505
 506         total_found += add_new_free_space(block_group, fs_info, last,
 507                                           block_group->key.objectid +
 508                                           block_group->key.offset);
 509         caching_ctl->progress = (u64)-1;
 510
 511         spin_lock(&block_group->lock);
 512         block_group->caching_ctl = NULL;
 513         block_group->cached = BTRFS_CACHE_FINISHED;
 514         spin_unlock(&block_group->lock);
 515
 516 err:
 517         btrfs_free_path(path);
 518         up_read(&fs_info->commit_root_sem);
 519
 520         free_excluded_extents(extent_root, block_group);
 521
 522         mutex_unlock(&caching_ctl->mutex);
 523 out:
 524         if (ret) {
 525                 spin_lock(&block_group->lock);
 526                 block_group->caching_ctl = NULL;
 527                 block_group->cached = BTRFS_CACHE_ERROR;
 528                 spin_unlock(&block_group->lock);
 529         }
 530         wake_up(&caching_ctl->wait);
 531
 532         put_caching_control(caching_ctl);
 533         btrfs_put_block_group(block_group);
 534 }
 535
 536 static int cache_block_group(struct btrfs_block_group_cache *cache,
 537                              int load_cache_only)
 538 {
 539         DEFINE_WAIT(wait);
 540         struct btrfs_fs_info *fs_info = cache->fs_info;
 541         struct btrfs_caching_control *caching_ctl;
 542         int ret = 0;
 543
 544         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 545         if (!caching_ctl)
 546                 return -ENOMEM;
 547
 548         INIT_LIST_HEAD(&caching_ctl->list);
 549         mutex_init(&caching_ctl->mutex);
 550         init_waitqueue_head(&caching_ctl->wait);
 551         caching_ctl->block_group = cache;
 552         caching_ctl->progress = cache->key.objectid;
 553         atomic_set(&caching_ctl->count, 1);
 554         btrfs_init_work(&caching_ctl->work, caching_thread, NULL, NULL);
 555
 556         spin_lock(&cache->lock);
 557         /*
 558          * This should be a rare occasion, but this could happen I think in the
 559          * case where one thread starts to load the space cache info, and then
 560          * some other thread starts a transaction commit which tries to do an
 561          * allocation while the other thread is still loading the space cache
 562          * info.  The previous loop should have kept us from choosing this block
 563          * group, but if we've moved to the state where we will wait on caching
 564          * block groups we need to first check if we're doing a fast load here,
 565          * so we can wait for it to finish, otherwise we could end up allocating
 566          * from a block group who's cache gets evicted for one reason or
 567          * another.
 568          */
 569         while (cache->cached == BTRFS_CACHE_FAST) {
 570                 struct btrfs_caching_control *ctl;
 571
 572                 ctl = cache->caching_ctl;
 573                 atomic_inc(&ctl->count);
 574                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 575                 spin_unlock(&cache->lock);
 576
 577                 schedule();
 578
 579                 finish_wait(&ctl->wait, &wait);
 580                 put_caching_control(ctl);
 581                 spin_lock(&cache->lock);
 582         }
 583
 584         if (cache->cached != BTRFS_CACHE_NO) {
 585                 spin_unlock(&cache->lock);
 586                 kfree(caching_ctl);
 587                 return 0;
 588         }
 589         WARN_ON(cache->caching_ctl);
 590         cache->caching_ctl = caching_ctl;
 591         cache->cached = BTRFS_CACHE_FAST;
 592         spin_unlock(&cache->lock);
 593
 594         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 595                 ret = load_free_space_cache(fs_info, cache);
 596
 597                 spin_lock(&cache->lock);
 598                 if (ret == 1) {
 599                         cache->caching_ctl = NULL;
 600                         cache->cached = BTRFS_CACHE_FINISHED;
 601                         cache->last_byte_to_unpin = (u64)-1;
 602                 } else {
 603                         if (load_cache_only) {
 604                                 cache->caching_ctl = NULL;
 605                                 cache->cached = BTRFS_CACHE_NO;
 606                         } else {
 607                                 cache->cached = BTRFS_CACHE_STARTED;
 608                         }
 609                 }
 610                 spin_unlock(&cache->lock);
 611                 wake_up(&caching_ctl->wait);
 612                 if (ret == 1) {
 613                         put_caching_control(caching_ctl);
 614                         free_excluded_extents(fs_info->extent_root, cache);
 615                         return 0;
 616                 }
 617         } else {
 618                 /*
 619                  * We are not going to do the fast caching, set cached to the
 620                  * appropriate value and wakeup any waiters.
 621                  */
 622                 spin_lock(&cache->lock);
 623                 if (load_cache_only) {
 624                         cache->caching_ctl = NULL;
 625                         cache->cached = BTRFS_CACHE_NO;
 626                 } else {
 627                         cache->cached = BTRFS_CACHE_STARTED;
 628                 }
 629                 spin_unlock(&cache->lock);
 630                 wake_up(&caching_ctl->wait);
 631         }
 632
 633         if (load_cache_only) {
 634                 put_caching_control(caching_ctl);
 635                 return 0;
 636         }
 637
 638         down_write(&fs_info->commit_root_sem);
 639         atomic_inc(&caching_ctl->count);
 640         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 641         up_write(&fs_info->commit_root_sem);
 642
 643         btrfs_get_block_group(cache);
 644
 645         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 646
 647         return ret;
 648 }
 649
 650 /*
 651  * return the block group that starts at or after bytenr
 652  */
 653 static struct btrfs_block_group_cache *
 654 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 655 {
 656         struct btrfs_block_group_cache *cache;
 657
 658         cache = block_group_cache_tree_search(info, bytenr, 0);
 659
 660         return cache;
 661 }
 662
 663 /*
 664  * return the block group that contains the given bytenr
 665  */
 666 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 667                                                  struct btrfs_fs_info *info,
 668                                                  u64 bytenr)
 669 {
 670         struct btrfs_block_group_cache *cache;
 671
 672         cache = block_group_cache_tree_search(info, bytenr, 1);
 673
 674         return cache;
 675 }
 676
 677 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 678                                                   u64 flags)
 679 {
 680         struct list_head *head = &info->space_info;
 681         struct btrfs_space_info *found;
 682
 683         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 684
 685         rcu_read_lock();
 686         list_for_each_entry_rcu(found, head, list) {
 687                 if (found->flags & flags) {
 688                         rcu_read_unlock();
 689                         return found;
 690                 }
 691         }
 692         rcu_read_unlock();
 693         return NULL;
 694 }
 695
 696 /*
 697  * after adding space to the filesystem, we need to clear the full flags
 698  * on all the space infos.
 699  */
 700 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 701 {
 702         struct list_head *head = &info->space_info;
 703         struct btrfs_space_info *found;
 704
 705         rcu_read_lock();
 706         list_for_each_entry_rcu(found, head, list)
 707                 found->full = 0;
 708         rcu_read_unlock();
 709 }
 710
 711 /* simple helper to search for an existing extent at a given offset */
 712 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 713 {
 714         int ret;
 715         struct btrfs_key key;
 716         struct btrfs_path *path;
 717
 718         path = btrfs_alloc_path();
 719         if (!path)
 720                 return -ENOMEM;
 721
 722         key.objectid = start;
 723         key.offset = len;
 724         key.type = BTRFS_EXTENT_ITEM_KEY;
 725         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 726                                 0, 0);
 727         if (ret > 0) {
 728                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 729                 if (key.objectid == start &&
 730                     key.type == BTRFS_METADATA_ITEM_KEY)
 731                         ret = 0;
 732         }
 733         btrfs_free_path(path);
 734         return ret;
 735 }
 736
 737 /*
 738  * helper function to lookup reference count and flags of a tree block.
 739  *
 740  * the head node for delayed ref is used to store the sum of all the
 741  * reference count modifications queued up in the rbtree. the head
 742  * node may also store the extent flags to set. This way you can check
 743  * to see what the reference count and extent flags would be if all of
 744  * the delayed refs are not processed.
 745  */
 746 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 747                              struct btrfs_root *root, u64 bytenr,
 748                              u64 offset, int metadata, u64 *refs, u64 *flags)
 749 {
 750         struct btrfs_delayed_ref_head *head;
 751         struct btrfs_delayed_ref_root *delayed_refs;
 752         struct btrfs_path *path;
 753         struct btrfs_extent_item *ei;
 754         struct extent_buffer *leaf;
 755         struct btrfs_key key;
 756         u32 item_size;
 757         u64 num_refs;
 758         u64 extent_flags;
 759         int ret;
 760
 761         /*
 762          * If we don't have skinny metadata, don't bother doing anything
 763          * different
 764          */
 765         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 766                 offset = root->leafsize;
 767                 metadata = 0;
 768         }
 769
 770         path = btrfs_alloc_path();
 771         if (!path)
 772                 return -ENOMEM;
 773
 774         if (!trans) {
 775                 path->skip_locking = 1;
 776                 path->search_commit_root = 1;
 777         }
 778
 779 search_again:
 780         key.objectid = bytenr;
 781         key.offset = offset;
 782         if (metadata)
 783                 key.type = BTRFS_METADATA_ITEM_KEY;
 784         else
 785                 key.type = BTRFS_EXTENT_ITEM_KEY;
 786
 787 again:
 788         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 789                                 &key, path, 0, 0);
 790         if (ret < 0)
 791                 goto out_free;
 792
 793         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 794                 if (path->slots[0]) {
 795                         path->slots[0]--;
 796                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 797                                               path->slots[0]);
 798                         if (key.objectid == bytenr &&
 799                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 800                             key.offset == root->leafsize)
 801                                 ret = 0;
 802                 }
 803                 if (ret) {
 804                         key.objectid = bytenr;
 805                         key.type = BTRFS_EXTENT_ITEM_KEY;
 806                         key.offset = root->leafsize;
 807                         btrfs_release_path(path);
 808                         goto again;
 809                 }
 810         }
 811
 812         if (ret == 0) {
 813                 leaf = path->nodes[0];
 814                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 815                 if (item_size >= sizeof(*ei)) {
 816                         ei = btrfs_item_ptr(leaf, path->slots[0],
 817                                             struct btrfs_extent_item);
 818                         num_refs = btrfs_extent_refs(leaf, ei);
 819                         extent_flags = btrfs_extent_flags(leaf, ei);
 820                 } else {
 821 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 822                         struct btrfs_extent_item_v0 *ei0;
 823                         BUG_ON(item_size != sizeof(*ei0));
 824                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 825                                              struct btrfs_extent_item_v0);
 826                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 827                         /* FIXME: this isn't correct for data */
 828                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 829 #else
 830                         BUG();
 831 #endif
 832                 }
 833                 BUG_ON(num_refs == 0);
 834         } else {
 835                 num_refs = 0;
 836                 extent_flags = 0;
 837                 ret = 0;
 838         }
 839
 840         if (!trans)
 841                 goto out;
 842
 843         delayed_refs = &trans->transaction->delayed_refs;
 844         spin_lock(&delayed_refs->lock);
 845         head = btrfs_find_delayed_ref_head(trans, bytenr);
 846         if (head) {
 847                 if (!mutex_trylock(&head->mutex)) {
 848                         atomic_inc(&head->node.refs);
 849                         spin_unlock(&delayed_refs->lock);
 850
 851                         btrfs_release_path(path);
 852
 853                         /*
 854                          * Mutex was contended, block until it's released and try
 855                          * again
 856                          */
 857                         mutex_lock(&head->mutex);
 858                         mutex_unlock(&head->mutex);
 859                         btrfs_put_delayed_ref(&head->node);
 860                         goto search_again;
 861                 }
 862                 spin_lock(&head->lock);
 863                 if (head->extent_op && head->extent_op->update_flags)
 864                         extent_flags |= head->extent_op->flags_to_set;
 865                 else
 866                         BUG_ON(num_refs == 0);
 867
 868                 num_refs += head->node.ref_mod;
 869                 spin_unlock(&head->lock);
 870                 mutex_unlock(&head->mutex);
 871         }
 872         spin_unlock(&delayed_refs->lock);
 873 out:
 874         WARN_ON(num_refs == 0);
 875         if (refs)
 876                 *refs = num_refs;
 877         if (flags)
 878                 *flags = extent_flags;
 879 out_free:
 880         btrfs_free_path(path);
 881         return ret;
 882 }
 883
 884 /*
 885  * Back reference rules.  Back refs have three main goals:
 886  *
 887  * 1) differentiate between all holders of references to an extent so that
 888  *    when a reference is dropped we can make sure it was a valid reference
 889  *    before freeing the extent.
 890  *
 891  * 2) Provide enough information to quickly find the holders of an extent
 892  *    if we notice a given block is corrupted or bad.
 893  *
 894  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 895  *    maintenance.  This is actually the same as #2, but with a slightly
 896  *    different use case.
 897  *
 898  * There are two kinds of back refs. The implicit back refs is optimized
 899  * for pointers in non-shared tree blocks. For a given pointer in a block,
 900  * back refs of this kind provide information about the block's owner tree
 901  * and the pointer's key. These information allow us to find the block by
 902  * b-tree searching. The full back refs is for pointers in tree blocks not
 903  * referenced by their owner trees. The location of tree block is recorded
 904  * in the back refs. Actually the full back refs is generic, and can be
 905  * used in all cases the implicit back refs is used. The major shortcoming
 906  * of the full back refs is its overhead. Every time a tree block gets
 907  * COWed, we have to update back refs entry for all pointers in it.
 908  *
 909  * For a newly allocated tree block, we use implicit back refs for
 910  * pointers in it. This means most tree related operations only involve
 911  * implicit back refs. For a tree block created in old transaction, the
 912  * only way to drop a reference to it is COW it. So we can detect the
 913  * event that tree block loses its owner tree's reference and do the
 914  * back refs conversion.
 915  *
 916  * When a tree block is COW'd through a tree, there are four cases:
 917  *
 918  * The reference count of the block is one and the tree is the block's
 919  * owner tree. Nothing to do in this case.
 920  *
 921  * The reference count of the block is one and the tree is not the
 922  * block's owner tree. In this case, full back refs is used for pointers
 923  * in the block. Remove these full back refs, add implicit back refs for
 924  * every pointers in the new block.
 925  *
 926  * The reference count of the block is greater than one and the tree is
 927  * the block's owner tree. In this case, implicit back refs is used for
 928  * pointers in the block. Add full back refs for every pointers in the
 929  * block, increase lower level extents' reference counts. The original
 930  * implicit back refs are entailed to the new block.
 931  *
 932  * The reference count of the block is greater than one and the tree is
 933  * not the block's owner tree. Add implicit back refs for every pointer in
 934  * the new block, increase lower level extents' reference count.
 935  *
 936  * Back Reference Key composing:
 937  *
 938  * The key objectid corresponds to the first byte in the extent,
 939  * The key type is used to differentiate between types of back refs.
 940  * There are different meanings of the key offset for different types
 941  * of back refs.
 942  *
 943  * File extents can be referenced by:
 944  *
 945  * - multiple snapshots, subvolumes, or different generations in one subvol
 946  * - different files inside a single subvolume
 947  * - different offsets inside a file (bookend extents in file.c)
 948  *
 949  * The extent ref structure for the implicit back refs has fields for:
 950  *
 951  * - Objectid of the subvolume root
 952  * - objectid of the file holding the reference
 953  * - original offset in the file
 954  * - how many bookend extents
 955  *
 956  * The key offset for the implicit back refs is hash of the first
 957  * three fields.
 958  *
 959  * The extent ref structure for the full back refs has field for:
 960  *
 961  * - number of pointers in the tree leaf
 962  *
 963  * The key offset for the implicit back refs is the first byte of
 964  * the tree leaf
 965  *
 966  * When a file extent is allocated, The implicit back refs is used.
 967  * the fields are filled in:
 968  *
 969  *     (root_key.objectid, inode objectid, offset in file, 1)
 970  *
 971  * When a file extent is removed file truncation, we find the
 972  * corresponding implicit back refs and check the following fields:
 973  *
 974  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 975  *
 976  * Btree extents can be referenced by:
 977  *
 978  * - Different subvolumes
 979  *
 980  * Both the implicit back refs and the full back refs for tree blocks
 981  * only consist of key. The key offset for the implicit back refs is
 982  * objectid of block's owner tree. The key offset for the full back refs
 983  * is the first byte of parent block.
 984  *
 985  * When implicit back refs is used, information about the lowest key and
 986  * level of the tree block are required. These information are stored in
 987  * tree block info structure.
 988  */
 989
 990 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 991 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 992                                   struct btrfs_root *root,
 993                                   struct btrfs_path *path,
 994                                   u64 owner, u32 extra_size)
 995 {
 996         struct btrfs_extent_item *item;
 997         struct btrfs_extent_item_v0 *ei0;
 998         struct btrfs_extent_ref_v0 *ref0;
 999         struct btrfs_tree_block_info *bi;
1000         struct extent_buffer *leaf;
1001         struct btrfs_key key;
1002         struct btrfs_key found_key;
1003         u32 new_size = sizeof(*item);
1004         u64 refs;
1005         int ret;
1006
1007         leaf = path->nodes[0];
1008         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1009
1010         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1011         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1012                              struct btrfs_extent_item_v0);
1013         refs = btrfs_extent_refs_v0(leaf, ei0);
1014
1015         if (owner == (u64)-1) {
1016                 while (1) {
1017                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1018                                 ret = btrfs_next_leaf(root, path);
1019                                 if (ret < 0)
1020                                         return ret;
1021                                 BUG_ON(ret > 0); /* Corruption */
1022                                 leaf = path->nodes[0];
1023                         }
1024                         btrfs_item_key_to_cpu(leaf, &found_key,
1025                                               path->slots[0]);
1026                         BUG_ON(key.objectid != found_key.objectid);
1027                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1028                                 path->slots[0]++;
1029                                 continue;
1030                         }
1031                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1032                                               struct btrfs_extent_ref_v0);
1033                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1034                         break;
1035                 }
1036         }
1037         btrfs_release_path(path);
1038
1039         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1040                 new_size += sizeof(*bi);
1041
1042         new_size -= sizeof(*ei0);
1043         ret = btrfs_search_slot(trans, root, &key, path,
1044                                 new_size + extra_size, 1);
1045         if (ret < 0)
1046                 return ret;
1047         BUG_ON(ret); /* Corruption */
1048
1049         btrfs_extend_item(root, path, new_size);
1050
1051         leaf = path->nodes[0];
1052         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1053         btrfs_set_extent_refs(leaf, item, refs);
1054         /* FIXME: get real generation */
1055         btrfs_set_extent_generation(leaf, item, 0);
1056         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1057                 btrfs_set_extent_flags(leaf, item,
1058                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1059                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1060                 bi = (struct btrfs_tree_block_info *)(item + 1);
1061                 /* FIXME: get first key of the block */
1062                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1063                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1064         } else {
1065                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1066         }
1067         btrfs_mark_buffer_dirty(leaf);
1068         return 0;
1069 }
1070 #endif
1071
1072 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1073 {
1074         u32 high_crc = ~(u32)0;
1075         u32 low_crc = ~(u32)0;
1076         __le64 lenum;
1077
1078         lenum = cpu_to_le64(root_objectid);
1079         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1080         lenum = cpu_to_le64(owner);
1081         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1082         lenum = cpu_to_le64(offset);
1083         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1084
1085         return ((u64)high_crc << 31) ^ (u64)low_crc;
1086 }
1087
1088 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1089                                      struct btrfs_extent_data_ref *ref)
1090 {
1091         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1092                                     btrfs_extent_data_ref_objectid(leaf, ref),
1093                                     btrfs_extent_data_ref_offset(leaf, ref));
1094 }
1095
1096 static int match_extent_data_ref(struct extent_buffer *leaf,
1097                                  struct btrfs_extent_data_ref *ref,
1098                                  u64 root_objectid, u64 owner, u64 offset)
1099 {
1100         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1101             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1102             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1103                 return 0;
1104         return 1;
1105 }
1106
1107 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1108                                            struct btrfs_root *root,
1109                                            struct btrfs_path *path,
1110                                            u64 bytenr, u64 parent,
1111                                            u64 root_objectid,
1112                                            u64 owner, u64 offset)
1113 {
1114         struct btrfs_key key;
1115         struct btrfs_extent_data_ref *ref;
1116         struct extent_buffer *leaf;
1117         u32 nritems;
1118         int ret;
1119         int recow;
1120         int err = -ENOENT;
1121
1122         key.objectid = bytenr;
1123         if (parent) {
1124                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1125                 key.offset = parent;
1126         } else {
1127                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1128                 key.offset = hash_extent_data_ref(root_objectid,
1129                                                   owner, offset);
1130         }
1131 again:
1132         recow = 0;
1133         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1134         if (ret < 0) {
1135                 err = ret;
1136                 goto fail;
1137         }
1138
1139         if (parent) {
1140                 if (!ret)
1141                         return 0;
1142 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1143                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1144                 btrfs_release_path(path);
1145                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1146                 if (ret < 0) {
1147                         err = ret;
1148                         goto fail;
1149                 }
1150                 if (!ret)
1151                         return 0;
1152 #endif
1153                 goto fail;
1154         }
1155
1156         leaf = path->nodes[0];
1157         nritems = btrfs_header_nritems(leaf);
1158         while (1) {
1159                 if (path->slots[0] >= nritems) {
1160                         ret = btrfs_next_leaf(root, path);
1161                         if (ret < 0)
1162                                 err = ret;
1163                         if (ret)
1164                                 goto fail;
1165
1166                         leaf = path->nodes[0];
1167                         nritems = btrfs_header_nritems(leaf);
1168                         recow = 1;
1169                 }
1170
1171                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1172                 if (key.objectid != bytenr ||
1173                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1174                         goto fail;
1175
1176                 ref = btrfs_item_ptr(leaf, path->slots[0],
1177                                      struct btrfs_extent_data_ref);
1178
1179                 if (match_extent_data_ref(leaf, ref, root_objectid,
1180                                           owner, offset)) {
1181                         if (recow) {
1182                                 btrfs_release_path(path);
1183                                 goto again;
1184                         }
1185                         err = 0;
1186                         break;
1187                 }
1188                 path->slots[0]++;
1189         }
1190 fail:
1191         return err;
1192 }
1193
1194 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1195                                            struct btrfs_root *root,
1196                                            struct btrfs_path *path,
1197                                            u64 bytenr, u64 parent,
1198                                            u64 root_objectid, u64 owner,
1199                                            u64 offset, int refs_to_add)
1200 {
1201         struct btrfs_key key;
1202         struct extent_buffer *leaf;
1203         u32 size;
1204         u32 num_refs;
1205         int ret;
1206
1207         key.objectid = bytenr;
1208         if (parent) {
1209                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1210                 key.offset = parent;
1211                 size = sizeof(struct btrfs_shared_data_ref);
1212         } else {
1213                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1214                 key.offset = hash_extent_data_ref(root_objectid,
1215                                                   owner, offset);
1216                 size = sizeof(struct btrfs_extent_data_ref);
1217         }
1218
1219         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1220         if (ret && ret != -EEXIST)
1221                 goto fail;
1222
1223         leaf = path->nodes[0];
1224         if (parent) {
1225                 struct btrfs_shared_data_ref *ref;
1226                 ref = btrfs_item_ptr(leaf, path->slots[0],
1227                                      struct btrfs_shared_data_ref);
1228                 if (ret == 0) {
1229                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1230                 } else {
1231                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1232                         num_refs += refs_to_add;
1233                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1234                 }
1235         } else {
1236                 struct btrfs_extent_data_ref *ref;
1237                 while (ret == -EEXIST) {
1238                         ref = btrfs_item_ptr(leaf, path->slots[0],
1239                                              struct btrfs_extent_data_ref);
1240                         if (match_extent_data_ref(leaf, ref, root_objectid,
1241                                                   owner, offset))
1242                                 break;
1243                         btrfs_release_path(path);
1244                         key.offset++;
1245                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1246                                                       size);
1247                         if (ret && ret != -EEXIST)
1248                                 goto fail;
1249
1250                         leaf = path->nodes[0];
1251                 }
1252                 ref = btrfs_item_ptr(leaf, path->slots[0],
1253                                      struct btrfs_extent_data_ref);
1254                 if (ret == 0) {
1255                         btrfs_set_extent_data_ref_root(leaf, ref,
1256                                                        root_objectid);
1257                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1258                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1259                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1260                 } else {
1261                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1262                         num_refs += refs_to_add;
1263                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1264                 }
1265         }
1266         btrfs_mark_buffer_dirty(leaf);
1267         ret = 0;
1268 fail:
1269         btrfs_release_path(path);
1270         return ret;
1271 }
1272
1273 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1274                                            struct btrfs_root *root,
1275                                            struct btrfs_path *path,
1276                                            int refs_to_drop, int *last_ref)
1277 {
1278         struct btrfs_key key;
1279         struct btrfs_extent_data_ref *ref1 = NULL;
1280         struct btrfs_shared_data_ref *ref2 = NULL;
1281         struct extent_buffer *leaf;
1282         u32 num_refs = 0;
1283         int ret = 0;
1284
1285         leaf = path->nodes[0];
1286         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1287
1288         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1289                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1290                                       struct btrfs_extent_data_ref);
1291                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1292         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1293                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1294                                       struct btrfs_shared_data_ref);
1295                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1296 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1297         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1298                 struct btrfs_extent_ref_v0 *ref0;
1299                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1300                                       struct btrfs_extent_ref_v0);
1301                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1302 #endif
1303         } else {
1304                 BUG();
1305         }
1306
1307         BUG_ON(num_refs < refs_to_drop);
1308         num_refs -= refs_to_drop;
1309
1310         if (num_refs == 0) {
1311                 ret = btrfs_del_item(trans, root, path);
1312                 *last_ref = 1;
1313         } else {
1314                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1315                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1316                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1317                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1318 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1319                 else {
1320                         struct btrfs_extent_ref_v0 *ref0;
1321                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1322                                         struct btrfs_extent_ref_v0);
1323                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1324                 }
1325 #endif
1326                 btrfs_mark_buffer_dirty(leaf);
1327         }
1328         return ret;
1329 }
1330
1331 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1332                                           struct btrfs_path *path,
1333                                           struct btrfs_extent_inline_ref *iref)
1334 {
1335         struct btrfs_key key;
1336         struct extent_buffer *leaf;
1337         struct btrfs_extent_data_ref *ref1;
1338         struct btrfs_shared_data_ref *ref2;
1339         u32 num_refs = 0;
1340
1341         leaf = path->nodes[0];
1342         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1343         if (iref) {
1344                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1345                     BTRFS_EXTENT_DATA_REF_KEY) {
1346                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1347                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1348                 } else {
1349                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1350                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1351                 }
1352         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1353                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1354                                       struct btrfs_extent_data_ref);
1355                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1356         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1357                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1358                                       struct btrfs_shared_data_ref);
1359                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1360 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1361         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1362                 struct btrfs_extent_ref_v0 *ref0;
1363                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1364                                       struct btrfs_extent_ref_v0);
1365                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1366 #endif
1367         } else {
1368                 WARN_ON(1);
1369         }
1370         return num_refs;
1371 }
1372
1373 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1374                                           struct btrfs_root *root,
1375                                           struct btrfs_path *path,
1376                                           u64 bytenr, u64 parent,
1377                                           u64 root_objectid)
1378 {
1379         struct btrfs_key key;
1380         int ret;
1381
1382         key.objectid = bytenr;
1383         if (parent) {
1384                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1385                 key.offset = parent;
1386         } else {
1387                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1388                 key.offset = root_objectid;
1389         }
1390
1391         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1392         if (ret > 0)
1393                 ret = -ENOENT;
1394 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1395         if (ret == -ENOENT && parent) {
1396                 btrfs_release_path(path);
1397                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1398                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1399                 if (ret > 0)
1400                         ret = -ENOENT;
1401         }
1402 #endif
1403         return ret;
1404 }
1405
1406 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1407                                           struct btrfs_root *root,
1408                                           struct btrfs_path *path,
1409                                           u64 bytenr, u64 parent,
1410                                           u64 root_objectid)
1411 {
1412         struct btrfs_key key;
1413         int ret;
1414
1415         key.objectid = bytenr;
1416         if (parent) {
1417                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1418                 key.offset = parent;
1419         } else {
1420                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1421                 key.offset = root_objectid;
1422         }
1423
1424         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1425         btrfs_release_path(path);
1426         return ret;
1427 }
1428
1429 static inline int extent_ref_type(u64 parent, u64 owner)
1430 {
1431         int type;
1432         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1433                 if (parent > 0)
1434                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1435                 else
1436                         type = BTRFS_TREE_BLOCK_REF_KEY;
1437         } else {
1438                 if (parent > 0)
1439                         type = BTRFS_SHARED_DATA_REF_KEY;
1440                 else
1441                         type = BTRFS_EXTENT_DATA_REF_KEY;
1442         }
1443         return type;
1444 }
1445
1446 static int find_next_key(struct btrfs_path *path, int level,
1447                          struct btrfs_key *key)
1448
1449 {
1450         for (; level < BTRFS_MAX_LEVEL; level++) {
1451                 if (!path->nodes[level])
1452                         break;
1453                 if (path->slots[level] + 1 >=
1454                     btrfs_header_nritems(path->nodes[level]))
1455                         continue;
1456                 if (level == 0)
1457                         btrfs_item_key_to_cpu(path->nodes[level], key,
1458                                               path->slots[level] + 1);
1459                 else
1460                         btrfs_node_key_to_cpu(path->nodes[level], key,
1461                                               path->slots[level] + 1);
1462                 return 0;
1463         }
1464         return 1;
1465 }
1466
1467 /*
1468  * look for inline back ref. if back ref is found, *ref_ret is set
1469  * to the address of inline back ref, and 0 is returned.
1470  *
1471  * if back ref isn't found, *ref_ret is set to the address where it
1472  * should be inserted, and -ENOENT is returned.
1473  *
1474  * if insert is true and there are too many inline back refs, the path
1475  * points to the extent item, and -EAGAIN is returned.
1476  *
1477  * NOTE: inline back refs are ordered in the same way that back ref
1478  *       items in the tree are ordered.
1479  */
1480 static noinline_for_stack
1481 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1482                                  struct btrfs_root *root,
1483                                  struct btrfs_path *path,
1484                                  struct btrfs_extent_inline_ref **ref_ret,
1485                                  u64 bytenr, u64 num_bytes,
1486                                  u64 parent, u64 root_objectid,
1487                                  u64 owner, u64 offset, int insert)
1488 {
1489         struct btrfs_key key;
1490         struct extent_buffer *leaf;
1491         struct btrfs_extent_item *ei;
1492         struct btrfs_extent_inline_ref *iref;
1493         u64 flags;
1494         u64 item_size;
1495         unsigned long ptr;
1496         unsigned long end;
1497         int extra_size;
1498         int type;
1499         int want;
1500         int ret;
1501         int err = 0;
1502         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1503                                                  SKINNY_METADATA);
1504
1505         key.objectid = bytenr;
1506         key.type = BTRFS_EXTENT_ITEM_KEY;
1507         key.offset = num_bytes;
1508
1509         want = extent_ref_type(parent, owner);
1510         if (insert) {
1511                 extra_size = btrfs_extent_inline_ref_size(want);
1512                 path->keep_locks = 1;
1513         } else
1514                 extra_size = -1;
1515
1516         /*
1517          * Owner is our parent level, so we can just add one to get the level
1518          * for the block we are interested in.
1519          */
1520         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1521                 key.type = BTRFS_METADATA_ITEM_KEY;
1522                 key.offset = owner;
1523         }
1524
1525 again:
1526         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1527         if (ret < 0) {
1528                 err = ret;
1529                 goto out;
1530         }
1531
1532         /*
1533          * We may be a newly converted file system which still has the old fat
1534          * extent entries for metadata, so try and see if we have one of those.
1535          */
1536         if (ret > 0 && skinny_metadata) {
1537                 skinny_metadata = false;
1538                 if (path->slots[0]) {
1539                         path->slots[0]--;
1540                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1541                                               path->slots[0]);
1542                         if (key.objectid == bytenr &&
1543                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1544                             key.offset == num_bytes)
1545                                 ret = 0;
1546                 }
1547                 if (ret) {
1548                         key.objectid = bytenr;
1549                         key.type = BTRFS_EXTENT_ITEM_KEY;
1550                         key.offset = num_bytes;
1551                         btrfs_release_path(path);
1552                         goto again;
1553                 }
1554         }
1555
1556         if (ret && !insert) {
1557                 err = -ENOENT;
1558                 goto out;
1559         } else if (WARN_ON(ret)) {
1560                 err = -EIO;
1561                 goto out;
1562         }
1563
1564         leaf = path->nodes[0];
1565         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1566 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1567         if (item_size < sizeof(*ei)) {
1568                 if (!insert) {
1569                         err = -ENOENT;
1570                         goto out;
1571                 }
1572                 ret = convert_extent_item_v0(trans, root, path, owner,
1573                                              extra_size);
1574                 if (ret < 0) {
1575                         err = ret;
1576                         goto out;
1577                 }
1578                 leaf = path->nodes[0];
1579                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1580         }
1581 #endif
1582         BUG_ON(item_size < sizeof(*ei));
1583
1584         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1585         flags = btrfs_extent_flags(leaf, ei);
1586
1587         ptr = (unsigned long)(ei + 1);
1588         end = (unsigned long)ei + item_size;
1589
1590         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1591                 ptr += sizeof(struct btrfs_tree_block_info);
1592                 BUG_ON(ptr > end);
1593         }
1594
1595         err = -ENOENT;
1596         while (1) {
1597                 if (ptr >= end) {
1598                         WARN_ON(ptr > end);
1599                         break;
1600                 }
1601                 iref = (struct btrfs_extent_inline_ref *)ptr;
1602                 type = btrfs_extent_inline_ref_type(leaf, iref);
1603                 if (want < type)
1604                         break;
1605                 if (want > type) {
1606                         ptr += btrfs_extent_inline_ref_size(type);
1607                         continue;
1608                 }
1609
1610                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1611                         struct btrfs_extent_data_ref *dref;
1612                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1613                         if (match_extent_data_ref(leaf, dref, root_objectid,
1614                                                   owner, offset)) {
1615                                 err = 0;
1616                                 break;
1617                         }
1618                         if (hash_extent_data_ref_item(leaf, dref) <
1619                             hash_extent_data_ref(root_objectid, owner, offset))
1620                                 break;
1621                 } else {
1622                         u64 ref_offset;
1623                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1624                         if (parent > 0) {
1625                                 if (parent == ref_offset) {
1626                                         err = 0;
1627                                         break;
1628                                 }
1629                                 if (ref_offset < parent)
1630                                         break;
1631                         } else {
1632                                 if (root_objectid == ref_offset) {
1633                                         err = 0;
1634                                         break;
1635                                 }
1636                                 if (ref_offset < root_objectid)
1637                                         break;
1638                         }
1639                 }
1640                 ptr += btrfs_extent_inline_ref_size(type);
1641         }
1642         if (err == -ENOENT && insert) {
1643                 if (item_size + extra_size >=
1644                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1645                         err = -EAGAIN;
1646                         goto out;
1647                 }
1648                 /*
1649                  * To add new inline back ref, we have to make sure
1650                  * there is no corresponding back ref item.
1651                  * For simplicity, we just do not add new inline back
1652                  * ref if there is any kind of item for this block
1653                  */
1654                 if (find_next_key(path, 0, &key) == 0 &&
1655                     key.objectid == bytenr &&
1656                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1657                         err = -EAGAIN;
1658                         goto out;
1659                 }
1660         }
1661         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1662 out:
1663         if (insert) {
1664                 path->keep_locks = 0;
1665                 btrfs_unlock_up_safe(path, 1);
1666         }
1667         return err;
1668 }
1669
1670 /*
1671  * helper to add new inline back ref
1672  */
1673 static noinline_for_stack
1674 void setup_inline_extent_backref(struct btrfs_root *root,
1675                                  struct btrfs_path *path,
1676                                  struct btrfs_extent_inline_ref *iref,
1677                                  u64 parent, u64 root_objectid,
1678                                  u64 owner, u64 offset, int refs_to_add,
1679                                  struct btrfs_delayed_extent_op *extent_op)
1680 {
1681         struct extent_buffer *leaf;
1682         struct btrfs_extent_item *ei;
1683         unsigned long ptr;
1684         unsigned long end;
1685         unsigned long item_offset;
1686         u64 refs;
1687         int size;
1688         int type;
1689
1690         leaf = path->nodes[0];
1691         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1692         item_offset = (unsigned long)iref - (unsigned long)ei;
1693
1694         type = extent_ref_type(parent, owner);
1695         size = btrfs_extent_inline_ref_size(type);
1696
1697         btrfs_extend_item(root, path, size);
1698
1699         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1700         refs = btrfs_extent_refs(leaf, ei);
1701         refs += refs_to_add;
1702         btrfs_set_extent_refs(leaf, ei, refs);
1703         if (extent_op)
1704                 __run_delayed_extent_op(extent_op, leaf, ei);
1705
1706         ptr = (unsigned long)ei + item_offset;
1707         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1708         if (ptr < end - size)
1709                 memmove_extent_buffer(leaf, ptr + size, ptr,
1710                                       end - size - ptr);
1711
1712         iref = (struct btrfs_extent_inline_ref *)ptr;
1713         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1714         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1715                 struct btrfs_extent_data_ref *dref;
1716                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1717                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1718                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1719                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1720                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1721         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1722                 struct btrfs_shared_data_ref *sref;
1723                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1724                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1725                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1726         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1727                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1728         } else {
1729                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1730         }
1731         btrfs_mark_buffer_dirty(leaf);
1732 }
1733
1734 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1735                                  struct btrfs_root *root,
1736                                  struct btrfs_path *path,
1737                                  struct btrfs_extent_inline_ref **ref_ret,
1738                                  u64 bytenr, u64 num_bytes, u64 parent,
1739                                  u64 root_objectid, u64 owner, u64 offset)
1740 {
1741         int ret;
1742
1743         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1744                                            bytenr, num_bytes, parent,
1745                                            root_objectid, owner, offset, 0);
1746         if (ret != -ENOENT)
1747                 return ret;
1748
1749         btrfs_release_path(path);
1750         *ref_ret = NULL;
1751
1752         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1753                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1754                                             root_objectid);
1755         } else {
1756                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1757                                              root_objectid, owner, offset);
1758         }
1759         return ret;
1760 }
1761
1762 /*
1763  * helper to update/remove inline back ref
1764  */
1765 static noinline_for_stack
1766 void update_inline_extent_backref(struct btrfs_root *root,
1767                                   struct btrfs_path *path,
1768                                   struct btrfs_extent_inline_ref *iref,
1769                                   int refs_to_mod,
1770                                   struct btrfs_delayed_extent_op *extent_op,
1771                                   int *last_ref)
1772 {
1773         struct extent_buffer *leaf;
1774         struct btrfs_extent_item *ei;
1775         struct btrfs_extent_data_ref *dref = NULL;
1776         struct btrfs_shared_data_ref *sref = NULL;
1777         unsigned long ptr;
1778         unsigned long end;
1779         u32 item_size;
1780         int size;
1781         int type;
1782         u64 refs;
1783
1784         leaf = path->nodes[0];
1785         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1786         refs = btrfs_extent_refs(leaf, ei);
1787         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1788         refs += refs_to_mod;
1789         btrfs_set_extent_refs(leaf, ei, refs);
1790         if (extent_op)
1791                 __run_delayed_extent_op(extent_op, leaf, ei);
1792
1793         type = btrfs_extent_inline_ref_type(leaf, iref);
1794
1795         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1796                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1797                 refs = btrfs_extent_data_ref_count(leaf, dref);
1798         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1799                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1800                 refs = btrfs_shared_data_ref_count(leaf, sref);
1801         } else {
1802                 refs = 1;
1803                 BUG_ON(refs_to_mod != -1);
1804         }
1805
1806         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1807         refs += refs_to_mod;
1808
1809         if (refs > 0) {
1810                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1811                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1812                 else
1813                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1814         } else {
1815                 *last_ref = 1;
1816                 size =  btrfs_extent_inline_ref_size(type);
1817                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1818                 ptr = (unsigned long)iref;
1819                 end = (unsigned long)ei + item_size;
1820                 if (ptr + size < end)
1821                         memmove_extent_buffer(leaf, ptr, ptr + size,
1822                                               end - ptr - size);
1823                 item_size -= size;
1824                 btrfs_truncate_item(root, path, item_size, 1);
1825         }
1826         btrfs_mark_buffer_dirty(leaf);
1827 }
1828
1829 static noinline_for_stack
1830 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1831                                  struct btrfs_root *root,
1832                                  struct btrfs_path *path,
1833                                  u64 bytenr, u64 num_bytes, u64 parent,
1834                                  u64 root_objectid, u64 owner,
1835                                  u64 offset, int refs_to_add,
1836                                  struct btrfs_delayed_extent_op *extent_op)
1837 {
1838         struct btrfs_extent_inline_ref *iref;
1839         int ret;
1840
1841         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1842                                            bytenr, num_bytes, parent,
1843                                            root_objectid, owner, offset, 1);
1844         if (ret == 0) {
1845                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1846                 update_inline_extent_backref(root, path, iref,
1847                                              refs_to_add, extent_op, NULL);
1848         } else if (ret == -ENOENT) {
1849                 setup_inline_extent_backref(root, path, iref, parent,
1850                                             root_objectid, owner, offset,
1851                                             refs_to_add, extent_op);
1852                 ret = 0;
1853         }
1854         return ret;
1855 }
1856
1857 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1858                                  struct btrfs_root *root,
1859                                  struct btrfs_path *path,
1860                                  u64 bytenr, u64 parent, u64 root_objectid,
1861                                  u64 owner, u64 offset, int refs_to_add)
1862 {
1863         int ret;
1864         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1865                 BUG_ON(refs_to_add != 1);
1866                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1867                                             parent, root_objectid);
1868         } else {
1869                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1870                                              parent, root_objectid,
1871                                              owner, offset, refs_to_add);
1872         }
1873         return ret;
1874 }
1875
1876 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1877                                  struct btrfs_root *root,
1878                                  struct btrfs_path *path,
1879                                  struct btrfs_extent_inline_ref *iref,
1880                                  int refs_to_drop, int is_data, int *last_ref)
1881 {
1882         int ret = 0;
1883
1884         BUG_ON(!is_data && refs_to_drop != 1);
1885         if (iref) {
1886                 update_inline_extent_backref(root, path, iref,
1887                                              -refs_to_drop, NULL, last_ref);
1888         } else if (is_data) {
1889                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1890                                              last_ref);
1891         } else {
1892                 *last_ref = 1;
1893                 ret = btrfs_del_item(trans, root, path);
1894         }
1895         return ret;
1896 }
1897
1898 static int btrfs_issue_discard(struct block_device *bdev,
1899                                 u64 start, u64 len)
1900 {
1901         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1902 }
1903
1904 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1905                                 u64 num_bytes, u64 *actual_bytes)
1906 {
1907         int ret;
1908         u64 discarded_bytes = 0;
1909         struct btrfs_bio *bbio = NULL;
1910
1911
1912         /* Tell the block device(s) that the sectors can be discarded */
1913         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1914                               bytenr, &num_bytes, &bbio, 0);
1915         /* Error condition is -ENOMEM */
1916         if (!ret) {
1917                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1918                 int i;
1919
1920
1921                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1922                         if (!stripe->dev->can_discard)
1923                                 continue;
1924
1925                         ret = btrfs_issue_discard(stripe->dev->bdev,
1926                                                   stripe->physical,
1927                                                   stripe->length);
1928                         if (!ret)
1929                                 discarded_bytes += stripe->length;
1930                         else if (ret != -EOPNOTSUPP)
1931                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1932
1933                         /*
1934                          * Just in case we get back EOPNOTSUPP for some reason,
1935                          * just ignore the return value so we don't screw up
1936                          * people calling discard_extent.
1937                          */
1938                         ret = 0;
1939                 }
1940                 kfree(bbio);
1941         }
1942
1943         if (actual_bytes)
1944                 *actual_bytes = discarded_bytes;
1945
1946
1947         if (ret == -EOPNOTSUPP)
1948                 ret = 0;
1949         return ret;
1950 }
1951
1952 /* Can return -ENOMEM */
1953 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1954                          struct btrfs_root *root,
1955                          u64 bytenr, u64 num_bytes, u64 parent,
1956                          u64 root_objectid, u64 owner, u64 offset,
1957                          int no_quota)
1958 {
1959         int ret;
1960         struct btrfs_fs_info *fs_info = root->fs_info;
1961
1962         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1963                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1964
1965         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1966                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1967                                         num_bytes,
1968                                         parent, root_objectid, (int)owner,
1969                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1970         } else {
1971                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1972                                         num_bytes,
1973                                         parent, root_objectid, owner, offset,
1974                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1975         }
1976         return ret;
1977 }
1978
1979 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1980                                   struct btrfs_root *root,
1981                                   u64 bytenr, u64 num_bytes,
1982                                   u64 parent, u64 root_objectid,
1983                                   u64 owner, u64 offset, int refs_to_add,
1984                                   int no_quota,
1985                                   struct btrfs_delayed_extent_op *extent_op)
1986 {
1987         struct btrfs_fs_info *fs_info = root->fs_info;
1988         struct btrfs_path *path;
1989         struct extent_buffer *leaf;
1990         struct btrfs_extent_item *item;
1991         struct btrfs_key key;
1992         u64 refs;
1993         int ret;
1994         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_ADD_EXCL;
1995
1996         path = btrfs_alloc_path();
1997         if (!path)
1998                 return -ENOMEM;
1999
2000         if (!is_fstree(root_objectid) || !root->fs_info->quota_enabled)
2001                 no_quota = 1;
2002
2003         path->reada = 1;
2004         path->leave_spinning = 1;
2005         /* this will setup the path even if it fails to insert the back ref */
2006         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2007                                            bytenr, num_bytes, parent,
2008                                            root_objectid, owner, offset,
2009                                            refs_to_add, extent_op);
2010         if ((ret < 0 && ret != -EAGAIN) || (!ret && no_quota))
2011                 goto out;
2012         /*
2013          * Ok we were able to insert an inline extent and it appears to be a new
2014          * reference, deal with the qgroup accounting.
2015          */
2016         if (!ret && !no_quota) {
2017                 ASSERT(root->fs_info->quota_enabled);
2018                 leaf = path->nodes[0];
2019                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2020                 item = btrfs_item_ptr(leaf, path->slots[0],
2021                                       struct btrfs_extent_item);
2022                 if (btrfs_extent_refs(leaf, item) > (u64)refs_to_add)
2023                         type = BTRFS_QGROUP_OPER_ADD_SHARED;
2024                 btrfs_release_path(path);
2025
2026                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2027                                               bytenr, num_bytes, type, 0);
2028                 goto out;
2029         }
2030
2031         /*
2032          * Ok we had -EAGAIN which means we didn't have space to insert and
2033          * inline extent ref, so just update the reference count and add a
2034          * normal backref.
2035          */
2036         leaf = path->nodes[0];
2037         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2038         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2039         refs = btrfs_extent_refs(leaf, item);
2040         if (refs)
2041                 type = BTRFS_QGROUP_OPER_ADD_SHARED;
2042         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2043         if (extent_op)
2044                 __run_delayed_extent_op(extent_op, leaf, item);
2045
2046         btrfs_mark_buffer_dirty(leaf);
2047         btrfs_release_path(path);
2048
2049         if (!no_quota) {
2050                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2051                                               bytenr, num_bytes, type, 0);
2052                 if (ret)
2053                         goto out;
2054         }
2055
2056         path->reada = 1;
2057         path->leave_spinning = 1;
2058         /* now insert the actual backref */
2059         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2060                                     path, bytenr, parent, root_objectid,
2061                                     owner, offset, refs_to_add);
2062         if (ret)
2063                 btrfs_abort_transaction(trans, root, ret);
2064 out:
2065         btrfs_free_path(path);
2066         return ret;
2067 }
2068
2069 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2070                                 struct btrfs_root *root,
2071                                 struct btrfs_delayed_ref_node *node,
2072                                 struct btrfs_delayed_extent_op *extent_op,
2073                                 int insert_reserved)
2074 {
2075         int ret = 0;
2076         struct btrfs_delayed_data_ref *ref;
2077         struct btrfs_key ins;
2078         u64 parent = 0;
2079         u64 ref_root = 0;
2080         u64 flags = 0;
2081
2082         ins.objectid = node->bytenr;
2083         ins.offset = node->num_bytes;
2084         ins.type = BTRFS_EXTENT_ITEM_KEY;
2085
2086         ref = btrfs_delayed_node_to_data_ref(node);
2087         trace_run_delayed_data_ref(node, ref, node->action);
2088
2089         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2090                 parent = ref->parent;
2091         ref_root = ref->root;
2092
2093         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2094                 if (extent_op)
2095                         flags |= extent_op->flags_to_set;
2096                 ret = alloc_reserved_file_extent(trans, root,
2097                                                  parent, ref_root, flags,
2098                                                  ref->objectid, ref->offset,
2099                                                  &ins, node->ref_mod);
2100         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2101                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2102                                              node->num_bytes, parent,
2103                                              ref_root, ref->objectid,
2104                                              ref->offset, node->ref_mod,
2105                                              node->no_quota, extent_op);
2106         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2107                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2108                                           node->num_bytes, parent,
2109                                           ref_root, ref->objectid,
2110                                           ref->offset, node->ref_mod,
2111                                           extent_op, node->no_quota);
2112         } else {
2113                 BUG();
2114         }
2115         return ret;
2116 }
2117
2118 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2119                                     struct extent_buffer *leaf,
2120                                     struct btrfs_extent_item *ei)
2121 {
2122         u64 flags = btrfs_extent_flags(leaf, ei);
2123         if (extent_op->update_flags) {
2124                 flags |= extent_op->flags_to_set;
2125                 btrfs_set_extent_flags(leaf, ei, flags);
2126         }
2127
2128         if (extent_op->update_key) {
2129                 struct btrfs_tree_block_info *bi;
2130                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2131                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2132                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2133         }
2134 }
2135
2136 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2137                                  struct btrfs_root *root,
2138                                  struct btrfs_delayed_ref_node *node,
2139                                  struct btrfs_delayed_extent_op *extent_op)
2140 {
2141         struct btrfs_key key;
2142         struct btrfs_path *path;
2143         struct btrfs_extent_item *ei;
2144         struct extent_buffer *leaf;
2145         u32 item_size;
2146         int ret;
2147         int err = 0;
2148         int metadata = !extent_op->is_data;
2149
2150         if (trans->aborted)
2151                 return 0;
2152
2153         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2154                 metadata = 0;
2155
2156         path = btrfs_alloc_path();
2157         if (!path)
2158                 return -ENOMEM;
2159
2160         key.objectid = node->bytenr;
2161
2162         if (metadata) {
2163                 key.type = BTRFS_METADATA_ITEM_KEY;
2164                 key.offset = extent_op->level;
2165         } else {
2166                 key.type = BTRFS_EXTENT_ITEM_KEY;
2167                 key.offset = node->num_bytes;
2168         }
2169
2170 again:
2171         path->reada = 1;
2172         path->leave_spinning = 1;
2173         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2174                                 path, 0, 1);
2175         if (ret < 0) {
2176                 err = ret;
2177                 goto out;
2178         }
2179         if (ret > 0) {
2180                 if (metadata) {
2181                         if (path->slots[0] > 0) {
2182                                 path->slots[0]--;
2183                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2184                                                       path->slots[0]);
2185                                 if (key.objectid == node->bytenr &&
2186                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2187                                     key.offset == node->num_bytes)
2188                                         ret = 0;
2189                         }
2190                         if (ret > 0) {
2191                                 btrfs_release_path(path);
2192                                 metadata = 0;
2193
2194                                 key.objectid = node->bytenr;
2195                                 key.offset = node->num_bytes;
2196                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2197                                 goto again;
2198                         }
2199                 } else {
2200                         err = -EIO;
2201                         goto out;
2202                 }
2203         }
2204
2205         leaf = path->nodes[0];
2206         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2207 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2208         if (item_size < sizeof(*ei)) {
2209                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2210                                              path, (u64)-1, 0);
2211                 if (ret < 0) {
2212                         err = ret;
2213                         goto out;
2214                 }
2215                 leaf = path->nodes[0];
2216                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2217         }
2218 #endif
2219         BUG_ON(item_size < sizeof(*ei));
2220         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2221         __run_delayed_extent_op(extent_op, leaf, ei);
2222
2223         btrfs_mark_buffer_dirty(leaf);
2224 out:
2225         btrfs_free_path(path);
2226         return err;
2227 }
2228
2229 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2230                                 struct btrfs_root *root,
2231                                 struct btrfs_delayed_ref_node *node,
2232                                 struct btrfs_delayed_extent_op *extent_op,
2233                                 int insert_reserved)
2234 {
2235         int ret = 0;
2236         struct btrfs_delayed_tree_ref *ref;
2237         struct btrfs_key ins;
2238         u64 parent = 0;
2239         u64 ref_root = 0;
2240         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2241                                                  SKINNY_METADATA);
2242
2243         ref = btrfs_delayed_node_to_tree_ref(node);
2244         trace_run_delayed_tree_ref(node, ref, node->action);
2245
2246         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2247                 parent = ref->parent;
2248         ref_root = ref->root;
2249
2250         ins.objectid = node->bytenr;
2251         if (skinny_metadata) {
2252                 ins.offset = ref->level;
2253                 ins.type = BTRFS_METADATA_ITEM_KEY;
2254         } else {
2255                 ins.offset = node->num_bytes;
2256                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2257         }
2258
2259         BUG_ON(node->ref_mod != 1);
2260         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2261                 BUG_ON(!extent_op || !extent_op->update_flags);
2262                 ret = alloc_reserved_tree_block(trans, root,
2263                                                 parent, ref_root,
2264                                                 extent_op->flags_to_set,
2265                                                 &extent_op->key,
2266                                                 ref->level, &ins,
2267                                                 node->no_quota);
2268         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2269                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2270                                              node->num_bytes, parent, ref_root,
2271                                              ref->level, 0, 1, node->no_quota,
2272                                              extent_op);
2273         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2274                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2275                                           node->num_bytes, parent, ref_root,
2276                                           ref->level, 0, 1, extent_op,
2277                                           node->no_quota);
2278         } else {
2279                 BUG();
2280         }
2281         return ret;
2282 }
2283
2284 /* helper function to actually process a single delayed ref entry */
2285 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2286                                struct btrfs_root *root,
2287                                struct btrfs_delayed_ref_node *node,
2288                                struct btrfs_delayed_extent_op *extent_op,
2289                                int insert_reserved)
2290 {
2291         int ret = 0;
2292
2293         if (trans->aborted) {
2294                 if (insert_reserved)
2295                         btrfs_pin_extent(root, node->bytenr,
2296                                          node->num_bytes, 1);
2297                 return 0;
2298         }
2299
2300         if (btrfs_delayed_ref_is_head(node)) {
2301                 struct btrfs_delayed_ref_head *head;
2302                 /*
2303                  * we've hit the end of the chain and we were supposed
2304                  * to insert this extent into the tree.  But, it got
2305                  * deleted before we ever needed to insert it, so all
2306                  * we have to do is clean up the accounting
2307                  */
2308                 BUG_ON(extent_op);
2309                 head = btrfs_delayed_node_to_head(node);
2310                 trace_run_delayed_ref_head(node, head, node->action);
2311
2312                 if (insert_reserved) {
2313                         btrfs_pin_extent(root, node->bytenr,
2314                                          node->num_bytes, 1);
2315                         if (head->is_data) {
2316                                 ret = btrfs_del_csums(trans, root,
2317                                                       node->bytenr,
2318                                                       node->num_bytes);
2319                         }
2320                 }
2321                 return ret;
2322         }
2323
2324         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2325             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2326                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2327                                            insert_reserved);
2328         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2329                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2330                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2331                                            insert_reserved);
2332         else
2333                 BUG();
2334         return ret;
2335 }
2336
2337 static noinline struct btrfs_delayed_ref_node *
2338 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2339 {
2340         struct rb_node *node;
2341         struct btrfs_delayed_ref_node *ref, *last = NULL;;
2342
2343         /*
2344          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2345          * this prevents ref count from going down to zero when
2346          * there still are pending delayed ref.
2347          */
2348         node = rb_first(&head->ref_root);
2349         while (node) {
2350                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2351                                 rb_node);
2352                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2353                         return ref;
2354                 else if (last == NULL)
2355                         last = ref;
2356                 node = rb_next(node);
2357         }
2358         return last;
2359 }
2360
2361 /*
2362  * Returns 0 on success or if called with an already aborted transaction.
2363  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2364  */
2365 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2366                                              struct btrfs_root *root,
2367                                              unsigned long nr)
2368 {
2369         struct btrfs_delayed_ref_root *delayed_refs;
2370         struct btrfs_delayed_ref_node *ref;
2371         struct btrfs_delayed_ref_head *locked_ref = NULL;
2372         struct btrfs_delayed_extent_op *extent_op;
2373         struct btrfs_fs_info *fs_info = root->fs_info;
2374         ktime_t start = ktime_get();
2375         int ret;
2376         unsigned long count = 0;
2377         unsigned long actual_count = 0;
2378         int must_insert_reserved = 0;
2379
2380         delayed_refs = &trans->transaction->delayed_refs;
2381         while (1) {
2382                 if (!locked_ref) {
2383                         if (count >= nr)
2384                                 break;
2385
2386                         spin_lock(&delayed_refs->lock);
2387                         locked_ref = btrfs_select_ref_head(trans);
2388                         if (!locked_ref) {
2389                                 spin_unlock(&delayed_refs->lock);
2390                                 break;
2391                         }
2392
2393                         /* grab the lock that says we are going to process
2394                          * all the refs for this head */
2395                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2396                         spin_unlock(&delayed_refs->lock);
2397                         /*
2398                          * we may have dropped the spin lock to get the head
2399                          * mutex lock, and that might have given someone else
2400                          * time to free the head.  If that's true, it has been
2401                          * removed from our list and we can move on.
2402                          */
2403                         if (ret == -EAGAIN) {
2404                                 locked_ref = NULL;
2405                                 count++;
2406                                 continue;
2407                         }
2408                 }
2409
2410                 /*
2411                  * We need to try and merge add/drops of the same ref since we
2412                  * can run into issues with relocate dropping the implicit ref
2413                  * and then it being added back again before the drop can
2414                  * finish.  If we merged anything we need to re-loop so we can
2415                  * get a good ref.
2416                  */
2417                 spin_lock(&locked_ref->lock);
2418                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2419                                          locked_ref);
2420
2421                 /*
2422                  * locked_ref is the head node, so we have to go one
2423                  * node back for any delayed ref updates
2424                  */
2425                 ref = select_delayed_ref(locked_ref);
2426
2427                 if (ref && ref->seq &&
2428                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2429                         spin_unlock(&locked_ref->lock);
2430                         btrfs_delayed_ref_unlock(locked_ref);
2431                         spin_lock(&delayed_refs->lock);
2432                         locked_ref->processing = 0;
2433                         delayed_refs->num_heads_ready++;
2434                         spin_unlock(&delayed_refs->lock);
2435                         locked_ref = NULL;
2436                         cond_resched();
2437                         count++;
2438                         continue;
2439                 }
2440
2441                 /*
2442                  * record the must insert reserved flag before we
2443                  * drop the spin lock.
2444                  */
2445                 must_insert_reserved = locked_ref->must_insert_reserved;
2446                 locked_ref->must_insert_reserved = 0;
2447
2448                 extent_op = locked_ref->extent_op;
2449                 locked_ref->extent_op = NULL;
2450
2451                 if (!ref) {
2452
2453
2454                         /* All delayed refs have been processed, Go ahead
2455                          * and send the head node to run_one_delayed_ref,
2456                          * so that any accounting fixes can happen
2457                          */
2458                         ref = &locked_ref->node;
2459
2460                         if (extent_op && must_insert_reserved) {
2461                                 btrfs_free_delayed_extent_op(extent_op);
2462                                 extent_op = NULL;
2463                         }
2464
2465                         if (extent_op) {
2466                                 spin_unlock(&locked_ref->lock);
2467                                 ret = run_delayed_extent_op(trans, root,
2468                                                             ref, extent_op);
2469                                 btrfs_free_delayed_extent_op(extent_op);
2470
2471                                 if (ret) {
2472                                         /*
2473                                          * Need to reset must_insert_reserved if
2474                                          * there was an error so the abort stuff
2475                                          * can cleanup the reserved space
2476                                          * properly.
2477                                          */
2478                                         if (must_insert_reserved)
2479                                                 locked_ref->must_insert_reserved = 1;
2480                                         locked_ref->processing = 0;
2481                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2482                                         btrfs_delayed_ref_unlock(locked_ref);
2483                                         return ret;
2484                                 }
2485                                 continue;
2486                         }
2487
2488                         /*
2489                          * Need to drop our head ref lock and re-aqcuire the
2490                          * delayed ref lock and then re-check to make sure
2491                          * nobody got added.
2492                          */
2493                         spin_unlock(&locked_ref->lock);
2494                         spin_lock(&delayed_refs->lock);
2495                         spin_lock(&locked_ref->lock);
2496                         if (rb_first(&locked_ref->ref_root) ||
2497                             locked_ref->extent_op) {
2498                                 spin_unlock(&locked_ref->lock);
2499                                 spin_unlock(&delayed_refs->lock);
2500                                 continue;
2501                         }
2502                         ref->in_tree = 0;
2503                         delayed_refs->num_heads--;
2504                         rb_erase(&locked_ref->href_node,
2505                                  &delayed_refs->href_root);
2506                         spin_unlock(&delayed_refs->lock);
2507                 } else {
2508                         actual_count++;
2509                         ref->in_tree = 0;
2510                         rb_erase(&ref->rb_node, &locked_ref->ref_root);
2511                 }
2512                 atomic_dec(&delayed_refs->num_entries);
2513
2514                 if (!btrfs_delayed_ref_is_head(ref)) {
2515                         /*
2516                          * when we play the delayed ref, also correct the
2517                          * ref_mod on head
2518                          */
2519                         switch (ref->action) {
2520                         case BTRFS_ADD_DELAYED_REF:
2521                         case BTRFS_ADD_DELAYED_EXTENT:
2522                                 locked_ref->node.ref_mod -= ref->ref_mod;
2523                                 break;
2524                         case BTRFS_DROP_DELAYED_REF:
2525                                 locked_ref->node.ref_mod += ref->ref_mod;
2526                                 break;
2527                         default:
2528                                 WARN_ON(1);
2529                         }
2530                 }
2531                 spin_unlock(&locked_ref->lock);
2532
2533                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2534                                           must_insert_reserved);
2535
2536                 btrfs_free_delayed_extent_op(extent_op);
2537                 if (ret) {
2538                         locked_ref->processing = 0;
2539                         btrfs_delayed_ref_unlock(locked_ref);
2540                         btrfs_put_delayed_ref(ref);
2541                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2542                         return ret;
2543                 }
2544
2545                 /*
2546                  * If this node is a head, that means all the refs in this head
2547                  * have been dealt with, and we will pick the next head to deal
2548                  * with, so we must unlock the head and drop it from the cluster
2549                  * list before we release it.
2550                  */
2551                 if (btrfs_delayed_ref_is_head(ref)) {
2552                         btrfs_delayed_ref_unlock(locked_ref);
2553                         locked_ref = NULL;
2554                 }
2555                 btrfs_put_delayed_ref(ref);
2556                 count++;
2557                 cond_resched();
2558         }
2559
2560         /*
2561          * We don't want to include ref heads since we can have empty ref heads
2562          * and those will drastically skew our runtime down since we just do
2563          * accounting, no actual extent tree updates.
2564          */
2565         if (actual_count > 0) {
2566                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2567                 u64 avg;
2568
2569                 /*
2570                  * We weigh the current average higher than our current runtime
2571                  * to avoid large swings in the average.
2572                  */
2573                 spin_lock(&delayed_refs->lock);
2574                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2575                 avg = div64_u64(avg, 4);
2576                 fs_info->avg_delayed_ref_runtime = avg;
2577                 spin_unlock(&delayed_refs->lock);
2578         }
2579         return 0;
2580 }
2581
2582 #ifdef SCRAMBLE_DELAYED_REFS
2583 /*
2584  * Normally delayed refs get processed in ascending bytenr order. This
2585  * correlates in most cases to the order added. To expose dependencies on this
2586  * order, we start to process the tree in the middle instead of the beginning
2587  */
2588 static u64 find_middle(struct rb_root *root)
2589 {
2590         struct rb_node *n = root->rb_node;
2591         struct btrfs_delayed_ref_node *entry;
2592         int alt = 1;
2593         u64 middle;
2594         u64 first = 0, last = 0;
2595
2596         n = rb_first(root);
2597         if (n) {
2598                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2599                 first = entry->bytenr;
2600         }
2601         n = rb_last(root);
2602         if (n) {
2603                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2604                 last = entry->bytenr;
2605         }
2606         n = root->rb_node;
2607
2608         while (n) {
2609                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2610                 WARN_ON(!entry->in_tree);
2611
2612                 middle = entry->bytenr;
2613
2614                 if (alt)
2615                         n = n->rb_left;
2616                 else
2617                         n = n->rb_right;
2618
2619                 alt = 1 - alt;
2620         }
2621         return middle;
2622 }
2623 #endif
2624
2625 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2626 {
2627         u64 num_bytes;
2628
2629         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2630                              sizeof(struct btrfs_extent_inline_ref));
2631         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2632                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2633
2634         /*
2635          * We don't ever fill up leaves all the way so multiply by 2 just to be
2636          * closer to what we're really going to want to ouse.
2637          */
2638         return div64_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2639 }
2640
2641 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2642                                        struct btrfs_root *root)
2643 {
2644         struct btrfs_block_rsv *global_rsv;
2645         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2646         u64 num_bytes;
2647         int ret = 0;
2648
2649         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2650         num_heads = heads_to_leaves(root, num_heads);
2651         if (num_heads > 1)
2652                 num_bytes += (num_heads - 1) * root->leafsize;
2653         num_bytes <<= 1;
2654         global_rsv = &root->fs_info->global_block_rsv;
2655
2656         /*
2657          * If we can't allocate any more chunks lets make sure we have _lots_ of
2658          * wiggle room since running delayed refs can create more delayed refs.
2659          */
2660         if (global_rsv->space_info->full)
2661                 num_bytes <<= 1;
2662
2663         spin_lock(&global_rsv->lock);
2664         if (global_rsv->reserved <= num_bytes)
2665                 ret = 1;
2666         spin_unlock(&global_rsv->lock);
2667         return ret;
2668 }
2669
2670 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2671                                        struct btrfs_root *root)
2672 {
2673         struct btrfs_fs_info *fs_info = root->fs_info;
2674         u64 num_entries =
2675                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2676         u64 avg_runtime;
2677         u64 val;
2678
2679         smp_mb();
2680         avg_runtime = fs_info->avg_delayed_ref_runtime;
2681         val = num_entries * avg_runtime;
2682         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2683                 return 1;
2684         if (val >= NSEC_PER_SEC / 2)
2685                 return 2;
2686
2687         return btrfs_check_space_for_delayed_refs(trans, root);
2688 }
2689
2690 struct async_delayed_refs {
2691         struct btrfs_root *root;
2692         int count;
2693         int error;
2694         int sync;
2695         struct completion wait;
2696         struct btrfs_work work;
2697 };
2698
2699 static void delayed_ref_async_start(struct btrfs_work *work)
2700 {
2701         struct async_delayed_refs *async;
2702         struct btrfs_trans_handle *trans;
2703         int ret;
2704
2705         async = container_of(work, struct async_delayed_refs, work);
2706
2707         trans = btrfs_join_transaction(async->root);
2708         if (IS_ERR(trans)) {
2709                 async->error = PTR_ERR(trans);
2710                 goto done;
2711         }
2712
2713         /*
2714          * trans->sync means that when we call end_transaciton, we won't
2715          * wait on delayed refs
2716          */
2717         trans->sync = true;
2718         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2719         if (ret)
2720                 async->error = ret;
2721
2722         ret = btrfs_end_transaction(trans, async->root);
2723         if (ret && !async->error)
2724                 async->error = ret;
2725 done:
2726         if (async->sync)
2727                 complete(&async->wait);
2728         else
2729                 kfree(async);
2730 }
2731
2732 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2733                                  unsigned long count, int wait)
2734 {
2735         struct async_delayed_refs *async;
2736         int ret;
2737
2738         async = kmalloc(sizeof(*async), GFP_NOFS);
2739         if (!async)
2740                 return -ENOMEM;
2741
2742         async->root = root->fs_info->tree_root;
2743         async->count = count;
2744         async->error = 0;
2745         if (wait)
2746                 async->sync = 1;
2747         else
2748                 async->sync = 0;
2749         init_completion(&async->wait);
2750
2751         btrfs_init_work(&async->work, delayed_ref_async_start,
2752                         NULL, NULL);
2753
2754         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2755
2756         if (wait) {
2757                 wait_for_completion(&async->wait);
2758                 ret = async->error;
2759                 kfree(async);
2760                 return ret;
2761         }
2762         return 0;
2763 }
2764
2765 /*
2766  * this starts processing the delayed reference count updates and
2767  * extent insertions we have queued up so far.  count can be
2768  * 0, which means to process everything in the tree at the start
2769  * of the run (but not newly added entries), or it can be some target
2770  * number you'd like to process.
2771  *
2772  * Returns 0 on success or if called with an aborted transaction
2773  * Returns <0 on error and aborts the transaction
2774  */
2775 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2776                            struct btrfs_root *root, unsigned long count)
2777 {
2778         struct rb_node *node;
2779         struct btrfs_delayed_ref_root *delayed_refs;
2780         struct btrfs_delayed_ref_head *head;
2781         int ret;
2782         int run_all = count == (unsigned long)-1;
2783         int run_most = 0;
2784
2785         /* We'll clean this up in btrfs_cleanup_transaction */
2786         if (trans->aborted)
2787                 return 0;
2788
2789         if (root == root->fs_info->extent_root)
2790                 root = root->fs_info->tree_root;
2791
2792         delayed_refs = &trans->transaction->delayed_refs;
2793         if (count == 0) {
2794                 count = atomic_read(&delayed_refs->num_entries) * 2;
2795                 run_most = 1;
2796         }
2797
2798 again:
2799 #ifdef SCRAMBLE_DELAYED_REFS
2800         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2801 #endif
2802         ret = __btrfs_run_delayed_refs(trans, root, count);
2803         if (ret < 0) {
2804                 btrfs_abort_transaction(trans, root, ret);
2805                 return ret;
2806         }
2807
2808         if (run_all) {
2809                 if (!list_empty(&trans->new_bgs))
2810                         btrfs_create_pending_block_groups(trans, root);
2811
2812                 spin_lock(&delayed_refs->lock);
2813                 node = rb_first(&delayed_refs->href_root);
2814                 if (!node) {
2815                         spin_unlock(&delayed_refs->lock);
2816                         goto out;
2817                 }
2818                 count = (unsigned long)-1;
2819
2820                 while (node) {
2821                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2822                                         href_node);
2823                         if (btrfs_delayed_ref_is_head(&head->node)) {
2824                                 struct btrfs_delayed_ref_node *ref;
2825
2826                                 ref = &head->node;
2827                                 atomic_inc(&ref->refs);
2828
2829                                 spin_unlock(&delayed_refs->lock);
2830                                 /*
2831                                  * Mutex was contended, block until it's
2832                                  * released and try again
2833                                  */
2834                                 mutex_lock(&head->mutex);
2835                                 mutex_unlock(&head->mutex);
2836
2837                                 btrfs_put_delayed_ref(ref);
2838                                 cond_resched();
2839                                 goto again;
2840                         } else {
2841                                 WARN_ON(1);
2842                         }
2843                         node = rb_next(node);
2844                 }
2845                 spin_unlock(&delayed_refs->lock);
2846                 cond_resched();
2847                 goto again;
2848         }
2849 out:
2850         ret = btrfs_delayed_qgroup_accounting(trans, root->fs_info);
2851         if (ret)
2852                 return ret;
2853         assert_qgroups_uptodate(trans);
2854         return 0;
2855 }
2856
2857 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2858                                 struct btrfs_root *root,
2859                                 u64 bytenr, u64 num_bytes, u64 flags,
2860                                 int level, int is_data)
2861 {
2862         struct btrfs_delayed_extent_op *extent_op;
2863         int ret;
2864
2865         extent_op = btrfs_alloc_delayed_extent_op();
2866         if (!extent_op)
2867                 return -ENOMEM;
2868
2869         extent_op->flags_to_set = flags;
2870         extent_op->update_flags = 1;
2871         extent_op->update_key = 0;
2872         extent_op->is_data = is_data ? 1 : 0;
2873         extent_op->level = level;
2874
2875         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2876                                           num_bytes, extent_op);
2877         if (ret)
2878                 btrfs_free_delayed_extent_op(extent_op);
2879         return ret;
2880 }
2881
2882 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2883                                       struct btrfs_root *root,
2884                                       struct btrfs_path *path,
2885                                       u64 objectid, u64 offset, u64 bytenr)
2886 {
2887         struct btrfs_delayed_ref_head *head;
2888         struct btrfs_delayed_ref_node *ref;
2889         struct btrfs_delayed_data_ref *data_ref;
2890         struct btrfs_delayed_ref_root *delayed_refs;
2891         struct rb_node *node;
2892         int ret = 0;
2893
2894         delayed_refs = &trans->transaction->delayed_refs;
2895         spin_lock(&delayed_refs->lock);
2896         head = btrfs_find_delayed_ref_head(trans, bytenr);
2897         if (!head) {
2898                 spin_unlock(&delayed_refs->lock);
2899                 return 0;
2900         }
2901
2902         if (!mutex_trylock(&head->mutex)) {
2903                 atomic_inc(&head->node.refs);
2904                 spin_unlock(&delayed_refs->lock);
2905
2906                 btrfs_release_path(path);
2907
2908                 /*
2909                  * Mutex was contended, block until it's released and let
2910                  * caller try again
2911                  */
2912                 mutex_lock(&head->mutex);
2913                 mutex_unlock(&head->mutex);
2914                 btrfs_put_delayed_ref(&head->node);
2915                 return -EAGAIN;
2916         }
2917         spin_unlock(&delayed_refs->lock);
2918
2919         spin_lock(&head->lock);
2920         node = rb_first(&head->ref_root);
2921         while (node) {
2922                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2923                 node = rb_next(node);
2924
2925                 /* If it's a shared ref we know a cross reference exists */
2926                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2927                         ret = 1;
2928                         break;
2929                 }
2930
2931                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2932
2933                 /*
2934                  * If our ref doesn't match the one we're currently looking at
2935                  * then we have a cross reference.
2936                  */
2937                 if (data_ref->root != root->root_key.objectid ||
2938                     data_ref->objectid != objectid ||
2939                     data_ref->offset != offset) {
2940                         ret = 1;
2941                         break;
2942                 }
2943         }
2944         spin_unlock(&head->lock);
2945         mutex_unlock(&head->mutex);
2946         return ret;
2947 }
2948
2949 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2950                                         struct btrfs_root *root,
2951                                         struct btrfs_path *path,
2952                                         u64 objectid, u64 offset, u64 bytenr)
2953 {
2954         struct btrfs_root *extent_root = root->fs_info->extent_root;
2955         struct extent_buffer *leaf;
2956         struct btrfs_extent_data_ref *ref;
2957         struct btrfs_extent_inline_ref *iref;
2958         struct btrfs_extent_item *ei;
2959         struct btrfs_key key;
2960         u32 item_size;
2961         int ret;
2962
2963         key.objectid = bytenr;
2964         key.offset = (u64)-1;
2965         key.type = BTRFS_EXTENT_ITEM_KEY;
2966
2967         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2968         if (ret < 0)
2969                 goto out;
2970         BUG_ON(ret == 0); /* Corruption */
2971
2972         ret = -ENOENT;
2973         if (path->slots[0] == 0)
2974                 goto out;
2975
2976         path->slots[0]--;
2977         leaf = path->nodes[0];
2978         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2979
2980         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2981                 goto out;
2982
2983         ret = 1;
2984         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2985 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2986         if (item_size < sizeof(*ei)) {
2987                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2988                 goto out;
2989         }
2990 #endif
2991         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2992
2993         if (item_size != sizeof(*ei) +
2994             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2995                 goto out;
2996
2997         if (btrfs_extent_generation(leaf, ei) <=
2998             btrfs_root_last_snapshot(&root->root_item))
2999                 goto out;
3000
3001         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3002         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3003             BTRFS_EXTENT_DATA_REF_KEY)
3004                 goto out;
3005
3006         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3007         if (btrfs_extent_refs(leaf, ei) !=
3008             btrfs_extent_data_ref_count(leaf, ref) ||
3009             btrfs_extent_data_ref_root(leaf, ref) !=
3010             root->root_key.objectid ||
3011             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3012             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3013                 goto out;
3014
3015         ret = 0;
3016 out:
3017         return ret;
3018 }
3019
3020 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3021                           struct btrfs_root *root,
3022                           u64 objectid, u64 offset, u64 bytenr)
3023 {
3024         struct btrfs_path *path;
3025         int ret;
3026         int ret2;
3027
3028         path = btrfs_alloc_path();
3029         if (!path)
3030                 return -ENOENT;
3031
3032         do {
3033                 ret = check_committed_ref(trans, root, path, objectid,
3034                                           offset, bytenr);
3035                 if (ret && ret != -ENOENT)
3036                         goto out;
3037
3038                 ret2 = check_delayed_ref(trans, root, path, objectid,
3039                                          offset, bytenr);
3040         } while (ret2 == -EAGAIN);
3041
3042         if (ret2 && ret2 != -ENOENT) {
3043                 ret = ret2;
3044                 goto out;
3045         }
3046
3047         if (ret != -ENOENT || ret2 != -ENOENT)
3048                 ret = 0;
3049 out:
3050         btrfs_free_path(path);
3051         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3052                 WARN_ON(ret > 0);
3053         return ret;
3054 }
3055
3056 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3057                            struct btrfs_root *root,
3058                            struct extent_buffer *buf,
3059                            int full_backref, int inc, int no_quota)
3060 {
3061         u64 bytenr;
3062         u64 num_bytes;
3063         u64 parent;
3064         u64 ref_root;
3065         u32 nritems;
3066         struct btrfs_key key;
3067         struct btrfs_file_extent_item *fi;
3068         int i;
3069         int level;
3070         int ret = 0;
3071         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3072                             u64, u64, u64, u64, u64, u64, int);
3073
3074 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3075         if (unlikely(test_bit(BTRFS_ROOT_DUMMY_ROOT, &root->state)))
3076                 return 0;
3077 #endif
3078         ref_root = btrfs_header_owner(buf);
3079         nritems = btrfs_header_nritems(buf);
3080         level = btrfs_header_level(buf);
3081
3082         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3083                 return 0;
3084
3085         if (inc)
3086                 process_func = btrfs_inc_extent_ref;
3087         else
3088                 process_func = btrfs_free_extent;
3089
3090         if (full_backref)
3091                 parent = buf->start;
3092         else
3093                 parent = 0;
3094
3095         for (i = 0; i < nritems; i++) {
3096                 if (level == 0) {
3097                         btrfs_item_key_to_cpu(buf, &key, i);
3098                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
3099                                 continue;
3100                         fi = btrfs_item_ptr(buf, i,
3101                                             struct btrfs_file_extent_item);
3102                         if (btrfs_file_extent_type(buf, fi) ==
3103                             BTRFS_FILE_EXTENT_INLINE)
3104                                 continue;
3105                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3106                         if (bytenr == 0)
3107                                 continue;
3108
3109                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3110                         key.offset -= btrfs_file_extent_offset(buf, fi);
3111                         ret = process_func(trans, root, bytenr, num_bytes,
3112                                            parent, ref_root, key.objectid,
3113                                            key.offset, no_quota);
3114                         if (ret)
3115                                 goto fail;
3116                 } else {
3117                         bytenr = btrfs_node_blockptr(buf, i);
3118                         num_bytes = btrfs_level_size(root, level - 1);
3119                         ret = process_func(trans, root, bytenr, num_bytes,
3120                                            parent, ref_root, level - 1, 0,
3121                                            no_quota);
3122                         if (ret)
3123                                 goto fail;
3124                 }
3125         }
3126         return 0;
3127 fail:
3128         return ret;
3129 }
3130
3131 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3132                   struct extent_buffer *buf, int full_backref, int no_quota)
3133 {
3134         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, no_quota);
3135 }
3136
3137 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3138                   struct extent_buffer *buf, int full_backref, int no_quota)
3139 {
3140         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, no_quota);
3141 }
3142
3143 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3144                                  struct btrfs_root *root,
3145                                  struct btrfs_path *path,
3146                                  struct btrfs_block_group_cache *cache)
3147 {
3148         int ret;
3149         struct btrfs_root *extent_root = root->fs_info->extent_root;
3150         unsigned long bi;
3151         struct extent_buffer *leaf;
3152
3153         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3154         if (ret < 0)
3155                 goto fail;
3156         BUG_ON(ret); /* Corruption */
3157
3158         leaf = path->nodes[0];
3159         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3160         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3161         btrfs_mark_buffer_dirty(leaf);
3162         btrfs_release_path(path);
3163 fail:
3164         if (ret) {
3165                 btrfs_abort_transaction(trans, root, ret);
3166                 return ret;
3167         }
3168         return 0;
3169
3170 }
3171
3172 static struct btrfs_block_group_cache *
3173 next_block_group(struct btrfs_root *root,
3174                  struct btrfs_block_group_cache *cache)
3175 {
3176         struct rb_node *node;
3177         spin_lock(&root->fs_info->block_group_cache_lock);
3178         node = rb_next(&cache->cache_node);
3179         btrfs_put_block_group(cache);
3180         if (node) {
3181                 cache = rb_entry(node, struct btrfs_block_group_cache,
3182                                  cache_node);
3183                 btrfs_get_block_group(cache);
3184         } else
3185                 cache = NULL;
3186         spin_unlock(&root->fs_info->block_group_cache_lock);
3187         return cache;
3188 }
3189
3190 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3191                             struct btrfs_trans_handle *trans,
3192                             struct btrfs_path *path)
3193 {
3194         struct btrfs_root *root = block_group->fs_info->tree_root;
3195         struct inode *inode = NULL;
3196         u64 alloc_hint = 0;
3197         int dcs = BTRFS_DC_ERROR;
3198         int num_pages = 0;
3199         int retries = 0;
3200         int ret = 0;
3201
3202         /*
3203          * If this block group is smaller than 100 megs don't bother caching the
3204          * block group.
3205          */
3206         if (block_group->key.offset < (100 * 1024 * 1024)) {
3207                 spin_lock(&block_group->lock);
3208                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3209                 spin_unlock(&block_group->lock);
3210                 return 0;
3211         }
3212
3213 again:
3214         inode = lookup_free_space_inode(root, block_group, path);
3215         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3216                 ret = PTR_ERR(inode);
3217                 btrfs_release_path(path);
3218                 goto out;
3219         }
3220
3221         if (IS_ERR(inode)) {
3222                 BUG_ON(retries);
3223                 retries++;
3224
3225                 if (block_group->ro)
3226                         goto out_free;
3227
3228                 ret = create_free_space_inode(root, trans, block_group, path);
3229                 if (ret)
3230                         goto out_free;
3231                 goto again;
3232         }
3233
3234         /* We've already setup this transaction, go ahead and exit */
3235         if (block_group->cache_generation == trans->transid &&
3236             i_size_read(inode)) {
3237                 dcs = BTRFS_DC_SETUP;
3238                 goto out_put;
3239         }
3240
3241         /*
3242          * We want to set the generation to 0, that way if anything goes wrong
3243          * from here on out we know not to trust this cache when we load up next
3244          * time.
3245          */
3246         BTRFS_I(inode)->generation = 0;
3247         ret = btrfs_update_inode(trans, root, inode);
3248         WARN_ON(ret);
3249
3250         if (i_size_read(inode) > 0) {
3251                 ret = btrfs_check_trunc_cache_free_space(root,
3252                                         &root->fs_info->global_block_rsv);
3253                 if (ret)
3254                         goto out_put;
3255
3256                 ret = btrfs_truncate_free_space_cache(root, trans, inode);
3257                 if (ret)
3258                         goto out_put;
3259         }
3260
3261         spin_lock(&block_group->lock);
3262         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3263             !btrfs_test_opt(root, SPACE_CACHE)) {
3264                 /*
3265                  * don't bother trying to write stuff out _if_
3266                  * a) we're not cached,
3267                  * b) we're with nospace_cache mount option.
3268                  */
3269                 dcs = BTRFS_DC_WRITTEN;
3270                 spin_unlock(&block_group->lock);
3271                 goto out_put;
3272         }
3273         spin_unlock(&block_group->lock);
3274
3275         /*
3276          * Try to preallocate enough space based on how big the block group is.
3277          * Keep in mind this has to include any pinned space which could end up
3278          * taking up quite a bit since it's not folded into the other space
3279          * cache.
3280          */
3281         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3282         if (!num_pages)
3283                 num_pages = 1;
3284
3285         num_pages *= 16;
3286         num_pages *= PAGE_CACHE_SIZE;
3287
3288         ret = btrfs_check_data_free_space(inode, num_pages);
3289         if (ret)
3290                 goto out_put;
3291
3292         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3293                                               num_pages, num_pages,
3294                                               &alloc_hint);
3295         if (!ret)
3296                 dcs = BTRFS_DC_SETUP;
3297         btrfs_free_reserved_data_space(inode, num_pages);
3298
3299 out_put:
3300         iput(inode);
3301 out_free:
3302         btrfs_release_path(path);
3303 out:
3304         spin_lock(&block_group->lock);
3305         if (!ret && dcs == BTRFS_DC_SETUP)
3306                 block_group->cache_generation = trans->transid;
3307         block_group->disk_cache_state = dcs;
3308         spin_unlock(&block_group->lock);
3309
3310         return ret;
3311 }
3312
3313 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3314                                    struct btrfs_root *root)
3315 {
3316         struct btrfs_block_group_cache *cache;
3317         int err = 0;
3318         struct btrfs_path *path;
3319         u64 last = 0;
3320
3321         path = btrfs_alloc_path();
3322         if (!path)
3323                 return -ENOMEM;
3324
3325 again:
3326         while (1) {
3327                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3328                 while (cache) {
3329                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3330                                 break;
3331                         cache = next_block_group(root, cache);
3332                 }
3333                 if (!cache) {
3334                         if (last == 0)
3335                                 break;
3336                         last = 0;
3337                         continue;
3338                 }
3339                 err = cache_save_setup(cache, trans, path);
3340                 last = cache->key.objectid + cache->key.offset;
3341                 btrfs_put_block_group(cache);
3342         }
3343
3344         while (1) {
3345                 if (last == 0) {
3346                         err = btrfs_run_delayed_refs(trans, root,
3347                                                      (unsigned long)-1);
3348                         if (err) /* File system offline */
3349                                 goto out;
3350                 }
3351
3352                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3353                 while (cache) {
3354                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3355                                 btrfs_put_block_group(cache);
3356                                 goto again;
3357                         }
3358
3359                         if (cache->dirty)
3360                                 break;
3361                         cache = next_block_group(root, cache);
3362                 }
3363                 if (!cache) {
3364                         if (last == 0)
3365                                 break;
3366                         last = 0;
3367                         continue;
3368                 }
3369
3370                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3371                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3372                 cache->dirty = 0;
3373                 last = cache->key.objectid + cache->key.offset;
3374
3375                 err = write_one_cache_group(trans, root, path, cache);
3376                 btrfs_put_block_group(cache);
3377                 if (err) /* File system offline */
3378                         goto out;
3379         }
3380
3381         while (1) {
3382                 /*
3383                  * I don't think this is needed since we're just marking our
3384                  * preallocated extent as written, but just in case it can't
3385                  * hurt.
3386                  */
3387                 if (last == 0) {
3388                         err = btrfs_run_delayed_refs(trans, root,
3389                                                      (unsigned long)-1);
3390                         if (err) /* File system offline */
3391                                 goto out;
3392                 }
3393
3394                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3395                 while (cache) {
3396                         /*
3397                          * Really this shouldn't happen, but it could if we
3398                          * couldn't write the entire preallocated extent and
3399                          * splitting the extent resulted in a new block.
3400                          */
3401                         if (cache->dirty) {
3402                                 btrfs_put_block_group(cache);
3403                                 goto again;
3404                         }
3405                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3406                                 break;
3407                         cache = next_block_group(root, cache);
3408                 }
3409                 if (!cache) {
3410                         if (last == 0)
3411                                 break;
3412                         last = 0;
3413                         continue;
3414                 }
3415
3416                 err = btrfs_write_out_cache(root, trans, cache, path);
3417
3418                 /*
3419                  * If we didn't have an error then the cache state is still
3420                  * NEED_WRITE, so we can set it to WRITTEN.
3421                  */
3422                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3423                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3424                 last = cache->key.objectid + cache->key.offset;
3425                 btrfs_put_block_group(cache);
3426         }
3427 out:
3428
3429         btrfs_free_path(path);
3430         return err;
3431 }
3432
3433 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3434 {
3435         struct btrfs_block_group_cache *block_group;
3436         int readonly = 0;
3437
3438         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3439         if (!block_group || block_group->ro)
3440                 readonly = 1;
3441         if (block_group)
3442                 btrfs_put_block_group(block_group);
3443         return readonly;
3444 }
3445
3446 static const char *alloc_name(u64 flags)
3447 {
3448         switch (flags) {
3449         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3450                 return "mixed";
3451         case BTRFS_BLOCK_GROUP_METADATA:
3452                 return "metadata";
3453         case BTRFS_BLOCK_GROUP_DATA:
3454                 return "data";
3455         case BTRFS_BLOCK_GROUP_SYSTEM:
3456                 return "system";
3457         default:
3458                 WARN_ON(1);
3459                 return "invalid-combination";
3460         };
3461 }
3462
3463 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3464                              u64 total_bytes, u64 bytes_used,
3465                              struct btrfs_space_info **space_info)
3466 {
3467         struct btrfs_space_info *found;
3468         int i;
3469         int factor;
3470         int ret;
3471
3472         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3473                      BTRFS_BLOCK_GROUP_RAID10))
3474                 factor = 2;
3475         else
3476                 factor = 1;
3477
3478         found = __find_space_info(info, flags);
3479         if (found) {
3480                 spin_lock(&found->lock);
3481                 found->total_bytes += total_bytes;
3482                 found->disk_total += total_bytes * factor;
3483                 found->bytes_used += bytes_used;
3484                 found->disk_used += bytes_used * factor;
3485                 found->full = 0;
3486                 spin_unlock(&found->lock);
3487                 *space_info = found;
3488                 return 0;
3489         }
3490         found = kzalloc(sizeof(*found), GFP_NOFS);
3491         if (!found)
3492                 return -ENOMEM;
3493
3494         ret = percpu_counter_init(&found->total_bytes_pinned, 0);
3495         if (ret) {
3496                 kfree(found);
3497                 return ret;
3498         }
3499
3500         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3501                 INIT_LIST_HEAD(&found->block_groups[i]);
3502         init_rwsem(&found->groups_sem);
3503         spin_lock_init(&found->lock);
3504         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3505         found->total_bytes = total_bytes;
3506         found->disk_total = total_bytes * factor;
3507         found->bytes_used = bytes_used;
3508         found->disk_used = bytes_used * factor;
3509         found->bytes_pinned = 0;
3510         found->bytes_reserved = 0;
3511         found->bytes_readonly = 0;
3512         found->bytes_may_use = 0;
3513         found->full = 0;
3514         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3515         found->chunk_alloc = 0;
3516         found->flush = 0;
3517         init_waitqueue_head(&found->wait);
3518
3519         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3520                                     info->space_info_kobj, "%s",
3521                                     alloc_name(found->flags));
3522         if (ret) {
3523                 kfree(found);
3524                 return ret;
3525         }
3526
3527         *space_info = found;
3528         list_add_rcu(&found->list, &info->space_info);
3529         if (flags & BTRFS_BLOCK_GROUP_DATA)
3530                 info->data_sinfo = found;
3531
3532         return ret;
3533 }
3534
3535 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3536 {
3537         u64 extra_flags = chunk_to_extended(flags) &
3538                                 BTRFS_EXTENDED_PROFILE_MASK;
3539
3540         write_seqlock(&fs_info->profiles_lock);
3541         if (flags & BTRFS_BLOCK_GROUP_DATA)
3542                 fs_info->avail_data_alloc_bits |= extra_flags;
3543         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3544                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3545         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3546                 fs_info->avail_system_alloc_bits |= extra_flags;
3547         write_sequnlock(&fs_info->profiles_lock);
3548 }
3549
3550 /*
3551  * returns target flags in extended format or 0 if restripe for this
3552  * chunk_type is not in progress
3553  *
3554  * should be called with either volume_mutex or balance_lock held
3555  */
3556 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3557 {
3558         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3559         u64 target = 0;
3560
3561         if (!bctl)
3562                 return 0;
3563
3564         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3565             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3566                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3567         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3568                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3569                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3570         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3571                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3572                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3573         }
3574
3575         return target;
3576 }
3577
3578 /*
3579  * @flags: available profiles in extended format (see ctree.h)
3580  *
3581  * Returns reduced profile in chunk format.  If profile changing is in
3582  * progress (either running or paused) picks the target profile (if it's
3583  * already available), otherwise falls back to plain reducing.
3584  */
3585 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3586 {
3587         /*
3588          * we add in the count of missing devices because we want
3589          * to make sure that any RAID levels on a degraded FS
3590          * continue to be honored.
3591          */
3592         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3593                 root->fs_info->fs_devices->missing_devices;
3594         u64 target;
3595         u64 tmp;
3596
3597         /*
3598          * see if restripe for this chunk_type is in progress, if so
3599          * try to reduce to the target profile
3600          */
3601         spin_lock(&root->fs_info->balance_lock);
3602         target = get_restripe_target(root->fs_info, flags);
3603         if (target) {
3604                 /* pick target profile only if it's already available */
3605                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3606                         spin_unlock(&root->fs_info->balance_lock);
3607                         return extended_to_chunk(target);
3608                 }
3609         }
3610         spin_unlock(&root->fs_info->balance_lock);
3611
3612         /* First, mask out the RAID levels which aren't possible */
3613         if (num_devices == 1)
3614                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3615                            BTRFS_BLOCK_GROUP_RAID5);
3616         if (num_devices < 3)
3617                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3618         if (num_devices < 4)
3619                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3620
3621         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3622                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3623                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3624         flags &= ~tmp;
3625
3626         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3627                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3628         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3629                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3630         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3631                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3632         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3633                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3634         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3635                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3636
3637         return extended_to_chunk(flags | tmp);
3638 }
3639
3640 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3641 {
3642         unsigned seq;
3643         u64 flags;
3644
3645         do {
3646                 flags = orig_flags;
3647                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3648
3649                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3650                         flags |= root->fs_info->avail_data_alloc_bits;
3651                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3652                         flags |= root->fs_info->avail_system_alloc_bits;
3653                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3654                         flags |= root->fs_info->avail_metadata_alloc_bits;
3655         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3656
3657         return btrfs_reduce_alloc_profile(root, flags);
3658 }
3659
3660 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3661 {
3662         u64 flags;
3663         u64 ret;
3664
3665         if (data)
3666                 flags = BTRFS_BLOCK_GROUP_DATA;
3667         else if (root == root->fs_info->chunk_root)
3668                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3669         else
3670                 flags = BTRFS_BLOCK_GROUP_METADATA;
3671
3672         ret = get_alloc_profile(root, flags);
3673         return ret;
3674 }
3675
3676 /*
3677  * This will check the space that the inode allocates from to make sure we have
3678  * enough space for bytes.
3679  */
3680 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3681 {
3682         struct btrfs_space_info *data_sinfo;
3683         struct btrfs_root *root = BTRFS_I(inode)->root;
3684         struct btrfs_fs_info *fs_info = root->fs_info;
3685         u64 used;
3686         int ret = 0, committed = 0, alloc_chunk = 1;
3687
3688         /* make sure bytes are sectorsize aligned */
3689         bytes = ALIGN(bytes, root->sectorsize);
3690
3691         if (btrfs_is_free_space_inode(inode)) {
3692                 committed = 1;
3693                 ASSERT(current->journal_info);
3694         }
3695
3696         data_sinfo = fs_info->data_sinfo;
3697         if (!data_sinfo)
3698                 goto alloc;
3699
3700 again:
3701         /* make sure we have enough space to handle the data first */
3702         spin_lock(&data_sinfo->lock);
3703         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3704                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3705                 data_sinfo->bytes_may_use;
3706
3707         if (used + bytes > data_sinfo->total_bytes) {
3708                 struct btrfs_trans_handle *trans;
3709
3710                 /*
3711                  * if we don't have enough free bytes in this space then we need
3712                  * to alloc a new chunk.
3713                  */
3714                 if (!data_sinfo->full && alloc_chunk) {
3715                         u64 alloc_target;
3716
3717                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3718                         spin_unlock(&data_sinfo->lock);
3719 alloc:
3720                         alloc_target = btrfs_get_alloc_profile(root, 1);
3721                         /*
3722                          * It is ugly that we don't call nolock join
3723                          * transaction for the free space inode case here.
3724                          * But it is safe because we only do the data space
3725                          * reservation for the free space cache in the
3726                          * transaction context, the common join transaction
3727                          * just increase the counter of the current transaction
3728                          * handler, doesn't try to acquire the trans_lock of
3729                          * the fs.
3730                          */
3731                         trans = btrfs_join_transaction(root);
3732                         if (IS_ERR(trans))
3733                                 return PTR_ERR(trans);
3734
3735                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3736                                              alloc_target,
3737                                              CHUNK_ALLOC_NO_FORCE);
3738                         btrfs_end_transaction(trans, root);
3739                         if (ret < 0) {
3740                                 if (ret != -ENOSPC)
3741                                         return ret;
3742                                 else
3743                                         goto commit_trans;
3744                         }
3745
3746                         if (!data_sinfo)
3747                                 data_sinfo = fs_info->data_sinfo;
3748
3749                         goto again;
3750                 }
3751
3752                 /*
3753                  * If we don't have enough pinned space to deal with this
3754                  * allocation don't bother committing the transaction.
3755                  */
3756                 if (percpu_counter_compare(&data_sinfo->total_bytes_pinned,
3757                                            bytes) < 0)
3758                         committed = 1;
3759                 spin_unlock(&data_sinfo->lock);
3760
3761                 /* commit the current transaction and try again */
3762 commit_trans:
3763                 if (!committed &&
3764                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3765                         committed = 1;
3766
3767                         trans = btrfs_join_transaction(root);
3768                         if (IS_ERR(trans))
3769                                 return PTR_ERR(trans);
3770                         ret = btrfs_commit_transaction(trans, root);
3771                         if (ret)
3772                                 return ret;
3773                         goto again;
3774                 }
3775
3776                 trace_btrfs_space_reservation(root->fs_info,
3777                                               "space_info:enospc",
3778                                               data_sinfo->flags, bytes, 1);
3779                 return -ENOSPC;
3780         }
3781         data_sinfo->bytes_may_use += bytes;
3782         trace_btrfs_space_reservation(root->fs_info, "space_info",
3783                                       data_sinfo->flags, bytes, 1);
3784         spin_unlock(&data_sinfo->lock);
3785
3786         return 0;
3787 }
3788
3789 /*
3790  * Called if we need to clear a data reservation for this inode.
3791  */
3792 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3793 {
3794         struct btrfs_root *root = BTRFS_I(inode)->root;
3795         struct btrfs_space_info *data_sinfo;
3796
3797         /* make sure bytes are sectorsize aligned */
3798         bytes = ALIGN(bytes, root->sectorsize);
3799
3800         data_sinfo = root->fs_info->data_sinfo;
3801         spin_lock(&data_sinfo->lock);
3802         WARN_ON(data_sinfo->bytes_may_use < bytes);
3803         data_sinfo->bytes_may_use -= bytes;
3804         trace_btrfs_space_reservation(root->fs_info, "space_info",
3805                                       data_sinfo->flags, bytes, 0);
3806         spin_unlock(&data_sinfo->lock);
3807 }
3808
3809 static void force_metadata_allocation(struct btrfs_fs_info *info)
3810 {
3811         struct list_head *head = &info->space_info;
3812         struct btrfs_space_info *found;
3813
3814         rcu_read_lock();
3815         list_for_each_entry_rcu(found, head, list) {
3816                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3817                         found->force_alloc = CHUNK_ALLOC_FORCE;
3818         }
3819         rcu_read_unlock();
3820 }
3821
3822 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3823 {
3824         return (global->size << 1);
3825 }
3826
3827 static int should_alloc_chunk(struct btrfs_root *root,
3828                               struct btrfs_space_info *sinfo, int force)
3829 {
3830         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3831         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3832         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3833         u64 thresh;
3834
3835         if (force == CHUNK_ALLOC_FORCE)
3836                 return 1;
3837
3838         /*
3839          * We need to take into account the global rsv because for all intents
3840          * and purposes it's used space.  Don't worry about locking the
3841          * global_rsv, it doesn't change except when the transaction commits.
3842          */
3843         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3844                 num_allocated += calc_global_rsv_need_space(global_rsv);
3845
3846         /*
3847          * in limited mode, we want to have some free space up to
3848          * about 1% of the FS size.
3849          */
3850         if (force == CHUNK_ALLOC_LIMITED) {
3851                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3852                 thresh = max_t(u64, 64 * 1024 * 1024,
3853                                div_factor_fine(thresh, 1));
3854
3855                 if (num_bytes - num_allocated < thresh)
3856                         return 1;
3857         }
3858
3859         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3860                 return 0;
3861         return 1;
3862 }
3863
3864 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3865 {
3866         u64 num_dev;
3867
3868         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3869                     BTRFS_BLOCK_GROUP_RAID0 |
3870                     BTRFS_BLOCK_GROUP_RAID5 |
3871                     BTRFS_BLOCK_GROUP_RAID6))
3872                 num_dev = root->fs_info->fs_devices->rw_devices;
3873         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3874                 num_dev = 2;
3875         else
3876                 num_dev = 1;    /* DUP or single */
3877
3878         /* metadata for updaing devices and chunk tree */
3879         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3880 }
3881
3882 static void check_system_chunk(struct btrfs_trans_handle *trans,
3883                                struct btrfs_root *root, u64 type)
3884 {
3885         struct btrfs_space_info *info;
3886         u64 left;
3887         u64 thresh;
3888
3889         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3890         spin_lock(&info->lock);
3891         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3892                 info->bytes_reserved - info->bytes_readonly;
3893         spin_unlock(&info->lock);
3894
3895         thresh = get_system_chunk_thresh(root, type);
3896         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3897                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3898                         left, thresh, type);
3899                 dump_space_info(info, 0, 0);
3900         }
3901
3902         if (left < thresh) {
3903                 u64 flags;
3904
3905                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3906                 btrfs_alloc_chunk(trans, root, flags);
3907         }
3908 }
3909
3910 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3911                           struct btrfs_root *extent_root, u64 flags, int force)
3912 {
3913         struct btrfs_space_info *space_info;
3914         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3915         int wait_for_alloc = 0;
3916         int ret = 0;
3917
3918         /* Don't re-enter if we're already allocating a chunk */
3919         if (trans->allocating_chunk)
3920                 return -ENOSPC;
3921
3922         space_info = __find_space_info(extent_root->fs_info, flags);
3923         if (!space_info) {
3924                 ret = update_space_info(extent_root->fs_info, flags,
3925                                         0, 0, &space_info);
3926                 BUG_ON(ret); /* -ENOMEM */
3927         }
3928         BUG_ON(!space_info); /* Logic error */
3929
3930 again:
3931         spin_lock(&space_info->lock);
3932         if (force < space_info->force_alloc)
3933                 force = space_info->force_alloc;
3934         if (space_info->full) {
3935                 if (should_alloc_chunk(extent_root, space_info, force))
3936                         ret = -ENOSPC;
3937                 else
3938                         ret = 0;
3939                 spin_unlock(&space_info->lock);
3940                 return ret;
3941         }
3942
3943         if (!should_alloc_chunk(extent_root, space_info, force)) {
3944                 spin_unlock(&space_info->lock);
3945                 return 0;
3946         } else if (space_info->chunk_alloc) {
3947                 wait_for_alloc = 1;
3948         } else {
3949                 space_info->chunk_alloc = 1;
3950         }
3951
3952         spin_unlock(&space_info->lock);
3953
3954         mutex_lock(&fs_info->chunk_mutex);
3955
3956         /*
3957          * The chunk_mutex is held throughout the entirety of a chunk
3958          * allocation, so once we've acquired the chunk_mutex we know that the
3959          * other guy is done and we need to recheck and see if we should
3960          * allocate.
3961          */
3962         if (wait_for_alloc) {
3963                 mutex_unlock(&fs_info->chunk_mutex);
3964                 wait_for_alloc = 0;
3965                 goto again;
3966         }
3967
3968         trans->allocating_chunk = true;
3969
3970         /*
3971          * If we have mixed data/metadata chunks we want to make sure we keep
3972          * allocating mixed chunks instead of individual chunks.
3973          */
3974         if (btrfs_mixed_space_info(space_info))
3975                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3976
3977         /*
3978          * if we're doing a data chunk, go ahead and make sure that
3979          * we keep a reasonable number of metadata chunks allocated in the
3980          * FS as well.
3981          */
3982         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3983                 fs_info->data_chunk_allocations++;
3984                 if (!(fs_info->data_chunk_allocations %
3985                       fs_info->metadata_ratio))
3986                         force_metadata_allocation(fs_info);
3987         }
3988
3989         /*
3990          * Check if we have enough space in SYSTEM chunk because we may need
3991          * to update devices.
3992          */
3993         check_system_chunk(trans, extent_root, flags);
3994
3995         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3996         trans->allocating_chunk = false;
3997
3998         spin_lock(&space_info->lock);
3999         if (ret < 0 && ret != -ENOSPC)
4000                 goto out;
4001         if (ret)
4002                 space_info->full = 1;
4003         else
4004                 ret = 1;
4005
4006         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4007 out:
4008         space_info->chunk_alloc = 0;
4009         spin_unlock(&space_info->lock);
4010         mutex_unlock(&fs_info->chunk_mutex);
4011         return ret;
4012 }
4013
4014 static int can_overcommit(struct btrfs_root *root,
4015                           struct btrfs_space_info *space_info, u64 bytes,
4016                           enum btrfs_reserve_flush_enum flush)
4017 {
4018         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4019         u64 profile = btrfs_get_alloc_profile(root, 0);
4020         u64 space_size;
4021         u64 avail;
4022         u64 used;
4023
4024         used = space_info->bytes_used + space_info->bytes_reserved +
4025                 space_info->bytes_pinned + space_info->bytes_readonly;
4026
4027         /*
4028          * We only want to allow over committing if we have lots of actual space
4029          * free, but if we don't have enough space to handle the global reserve
4030          * space then we could end up having a real enospc problem when trying
4031          * to allocate a chunk or some other such important allocation.
4032          */
4033         spin_lock(&global_rsv->lock);
4034         space_size = calc_global_rsv_need_space(global_rsv);
4035         spin_unlock(&global_rsv->lock);
4036         if (used + space_size >= space_info->total_bytes)
4037                 return 0;
4038
4039         used += space_info->bytes_may_use;
4040
4041         spin_lock(&root->fs_info->free_chunk_lock);
4042         avail = root->fs_info->free_chunk_space;
4043         spin_unlock(&root->fs_info->free_chunk_lock);
4044
4045         /*
4046          * If we have dup, raid1 or raid10 then only half of the free
4047          * space is actually useable.  For raid56, the space info used
4048          * doesn't include the parity drive, so we don't have to
4049          * change the math
4050          */
4051         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4052                        BTRFS_BLOCK_GROUP_RAID1 |
4053                        BTRFS_BLOCK_GROUP_RAID10))
4054                 avail >>= 1;
4055
4056         /*
4057          * If we aren't flushing all things, let us overcommit up to
4058          * 1/2th of the space. If we can flush, don't let us overcommit
4059          * too much, let it overcommit up to 1/8 of the space.
4060          */
4061         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4062                 avail >>= 3;
4063         else
4064                 avail >>= 1;
4065
4066         if (used + bytes < space_info->total_bytes + avail)
4067                 return 1;
4068         return 0;
4069 }
4070
4071 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4072                                          unsigned long nr_pages, int nr_items)
4073 {
4074         struct super_block *sb = root->fs_info->sb;
4075
4076         if (down_read_trylock(&sb->s_umount)) {
4077                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4078                 up_read(&sb->s_umount);
4079         } else {
4080                 /*
4081                  * We needn't worry the filesystem going from r/w to r/o though
4082                  * we don't acquire ->s_umount mutex, because the filesystem
4083                  * should guarantee the delalloc inodes list be empty after
4084                  * the filesystem is readonly(all dirty pages are written to
4085                  * the disk).
4086                  */
4087                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4088                 if (!current->journal_info)
4089                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4090         }
4091 }
4092
4093 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4094 {
4095         u64 bytes;
4096         int nr;
4097
4098         bytes = btrfs_calc_trans_metadata_size(root, 1);
4099         nr = (int)div64_u64(to_reclaim, bytes);
4100         if (!nr)
4101                 nr = 1;
4102         return nr;
4103 }
4104
4105 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4106
4107 /*
4108  * shrink metadata reservation for delalloc
4109  */
4110 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4111                             bool wait_ordered)
4112 {
4113         struct btrfs_block_rsv *block_rsv;
4114         struct btrfs_space_info *space_info;
4115         struct btrfs_trans_handle *trans;
4116         u64 delalloc_bytes;
4117         u64 max_reclaim;
4118         long time_left;
4119         unsigned long nr_pages;
4120         int loops;
4121         int items;
4122         enum btrfs_reserve_flush_enum flush;
4123
4124         /* Calc the number of the pages we need flush for space reservation */
4125         items = calc_reclaim_items_nr(root, to_reclaim);
4126         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4127
4128         trans = (struct btrfs_trans_handle *)current->journal_info;
4129         block_rsv = &root->fs_info->delalloc_block_rsv;
4130         space_info = block_rsv->space_info;
4131
4132         delalloc_bytes = percpu_counter_sum_positive(
4133                                                 &root->fs_info->delalloc_bytes);
4134         if (delalloc_bytes == 0) {
4135                 if (trans)
4136                         return;
4137                 if (wait_ordered)
4138                         btrfs_wait_ordered_roots(root->fs_info, items);
4139                 return;
4140         }
4141
4142         loops = 0;
4143         while (delalloc_bytes && loops < 3) {
4144                 max_reclaim = min(delalloc_bytes, to_reclaim);
4145                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4146                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4147                 /*
4148                  * We need to wait for the async pages to actually start before
4149                  * we do anything.
4150                  */
4151                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4152                 if (!max_reclaim)
4153                         goto skip_async;
4154
4155                 if (max_reclaim <= nr_pages)
4156                         max_reclaim = 0;
4157                 else
4158                         max_reclaim -= nr_pages;
4159
4160                 wait_event(root->fs_info->async_submit_wait,
4161                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4162                            (int)max_reclaim);
4163 skip_async:
4164                 if (!trans)
4165                         flush = BTRFS_RESERVE_FLUSH_ALL;
4166                 else
4167                         flush = BTRFS_RESERVE_NO_FLUSH;
4168                 spin_lock(&space_info->lock);
4169                 if (can_overcommit(root, space_info, orig, flush)) {
4170                         spin_unlock(&space_info->lock);
4171                         break;
4172                 }
4173                 spin_unlock(&space_info->lock);
4174
4175                 loops++;
4176                 if (wait_ordered && !trans) {
4177                         btrfs_wait_ordered_roots(root->fs_info, items);
4178                 } else {
4179                         time_left = schedule_timeout_killable(1);
4180                         if (time_left)
4181                                 break;
4182                 }
4183                 delalloc_bytes = percpu_counter_sum_positive(
4184                                                 &root->fs_info->delalloc_bytes);
4185         }
4186 }
4187
4188 /**
4189  * maybe_commit_transaction - possibly commit the transaction if its ok to
4190  * @root - the root we're allocating for
4191  * @bytes - the number of bytes we want to reserve
4192  * @force - force the commit
4193  *
4194  * This will check to make sure that committing the transaction will actually
4195  * get us somewhere and then commit the transaction if it does.  Otherwise it
4196  * will return -ENOSPC.
4197  */
4198 static int may_commit_transaction(struct btrfs_root *root,
4199                                   struct btrfs_space_info *space_info,
4200                                   u64 bytes, int force)
4201 {
4202         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4203         struct btrfs_trans_handle *trans;
4204
4205         trans = (struct btrfs_trans_handle *)current->journal_info;
4206         if (trans)
4207                 return -EAGAIN;
4208
4209         if (force)
4210                 goto commit;
4211
4212         /* See if there is enough pinned space to make this reservation */
4213         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4214                                    bytes) >= 0)
4215                 goto commit;
4216
4217         /*
4218          * See if there is some space in the delayed insertion reservation for
4219          * this reservation.
4220          */
4221         if (space_info != delayed_rsv->space_info)
4222                 return -ENOSPC;
4223
4224         spin_lock(&delayed_rsv->lock);
4225         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4226                                    bytes - delayed_rsv->size) >= 0) {
4227                 spin_unlock(&delayed_rsv->lock);
4228                 return -ENOSPC;
4229         }
4230         spin_unlock(&delayed_rsv->lock);
4231
4232 commit:
4233         trans = btrfs_join_transaction(root);
4234         if (IS_ERR(trans))
4235                 return -ENOSPC;
4236
4237         return btrfs_commit_transaction(trans, root);
4238 }
4239
4240 enum flush_state {
4241         FLUSH_DELAYED_ITEMS_NR  =       1,
4242         FLUSH_DELAYED_ITEMS     =       2,
4243         FLUSH_DELALLOC          =       3,
4244         FLUSH_DELALLOC_WAIT     =       4,
4245         ALLOC_CHUNK             =       5,
4246         COMMIT_TRANS            =       6,
4247 };
4248
4249 static int flush_space(struct btrfs_root *root,
4250                        struct btrfs_space_info *space_info, u64 num_bytes,
4251                        u64 orig_bytes, int state)
4252 {
4253         struct btrfs_trans_handle *trans;
4254         int nr;
4255         int ret = 0;
4256
4257         switch (state) {
4258         case FLUSH_DELAYED_ITEMS_NR:
4259         case FLUSH_DELAYED_ITEMS:
4260                 if (state == FLUSH_DELAYED_ITEMS_NR)
4261                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4262                 else
4263                         nr = -1;
4264
4265                 trans = btrfs_join_transaction(root);
4266                 if (IS_ERR(trans)) {
4267                         ret = PTR_ERR(trans);
4268                         break;
4269                 }
4270                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4271                 btrfs_end_transaction(trans, root);
4272                 break;
4273         case FLUSH_DELALLOC:
4274         case FLUSH_DELALLOC_WAIT:
4275                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4276                                 state == FLUSH_DELALLOC_WAIT);
4277                 break;
4278         case ALLOC_CHUNK:
4279                 trans = btrfs_join_transaction(root);
4280                 if (IS_ERR(trans)) {
4281                         ret = PTR_ERR(trans);
4282                         break;
4283                 }
4284                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4285                                      btrfs_get_alloc_profile(root, 0),
4286                                      CHUNK_ALLOC_NO_FORCE);
4287                 btrfs_end_transaction(trans, root);
4288                 if (ret == -ENOSPC)
4289                         ret = 0;
4290                 break;
4291         case COMMIT_TRANS:
4292                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4293                 break;
4294         default:
4295                 ret = -ENOSPC;
4296                 break;
4297         }
4298
4299         return ret;
4300 }
4301
4302 static inline u64
4303 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4304                                  struct btrfs_space_info *space_info)
4305 {
4306         u64 used;
4307         u64 expected;
4308         u64 to_reclaim;
4309
4310         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4311                                 16 * 1024 * 1024);
4312         spin_lock(&space_info->lock);
4313         if (can_overcommit(root, space_info, to_reclaim,
4314                            BTRFS_RESERVE_FLUSH_ALL)) {
4315                 to_reclaim = 0;
4316                 goto out;
4317         }
4318
4319         used = space_info->bytes_used + space_info->bytes_reserved +
4320                space_info->bytes_pinned + space_info->bytes_readonly +
4321                space_info->bytes_may_use;
4322         if (can_overcommit(root, space_info, 1024 * 1024,
4323                            BTRFS_RESERVE_FLUSH_ALL))
4324                 expected = div_factor_fine(space_info->total_bytes, 95);
4325         else
4326                 expected = div_factor_fine(space_info->total_bytes, 90);
4327
4328         if (used > expected)
4329                 to_reclaim = used - expected;
4330         else
4331                 to_reclaim = 0;
4332         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4333                                      space_info->bytes_reserved);
4334 out:
4335         spin_unlock(&space_info->lock);
4336
4337         return to_reclaim;
4338 }
4339
4340 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4341                                         struct btrfs_fs_info *fs_info, u64 used)
4342 {
4343         return (used >= div_factor_fine(space_info->total_bytes, 98) &&
4344                 !btrfs_fs_closing(fs_info) &&
4345                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4346 }
4347
4348 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4349                                        struct btrfs_fs_info *fs_info)
4350 {
4351         u64 used;
4352
4353         spin_lock(&space_info->lock);
4354         used = space_info->bytes_used + space_info->bytes_reserved +
4355                space_info->bytes_pinned + space_info->bytes_readonly +
4356                space_info->bytes_may_use;
4357         if (need_do_async_reclaim(space_info, fs_info, used)) {
4358                 spin_unlock(&space_info->lock);
4359                 return 1;
4360         }
4361         spin_unlock(&space_info->lock);
4362
4363         return 0;
4364 }
4365
4366 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4367 {
4368         struct btrfs_fs_info *fs_info;
4369         struct btrfs_space_info *space_info;
4370         u64 to_reclaim;
4371         int flush_state;
4372
4373         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4374         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4375
4376         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4377                                                       space_info);
4378         if (!to_reclaim)
4379                 return;
4380
4381         flush_state = FLUSH_DELAYED_ITEMS_NR;
4382         do {
4383                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4384                             to_reclaim, flush_state);
4385                 flush_state++;
4386                 if (!btrfs_need_do_async_reclaim(space_info, fs_info))
4387                         return;
4388         } while (flush_state <= COMMIT_TRANS);
4389
4390         if (btrfs_need_do_async_reclaim(space_info, fs_info))
4391                 queue_work(system_unbound_wq, work);
4392 }
4393
4394 void btrfs_init_async_reclaim_work(struct work_struct *work)
4395 {
4396         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4397 }
4398
4399 /**
4400  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4401  * @root - the root we're allocating for
4402  * @block_rsv - the block_rsv we're allocating for
4403  * @orig_bytes - the number of bytes we want
4404  * @flush - whether or not we can flush to make our reservation
4405  *
4406  * This will reserve orgi_bytes number of bytes from the space info associated
4407  * with the block_rsv.  If there is not enough space it will make an attempt to
4408  * flush out space to make room.  It will do this by flushing delalloc if
4409  * possible or committing the transaction.  If flush is 0 then no attempts to
4410  * regain reservations will be made and this will fail if there is not enough
4411  * space already.
4412  */
4413 static int reserve_metadata_bytes(struct btrfs_root *root,
4414                                   struct btrfs_block_rsv *block_rsv,
4415                                   u64 orig_bytes,
4416                                   enum btrfs_reserve_flush_enum flush)
4417 {
4418         struct btrfs_space_info *space_info = block_rsv->space_info;
4419         u64 used;
4420         u64 num_bytes = orig_bytes;
4421         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4422         int ret = 0;
4423         bool flushing = false;
4424
4425 again:
4426         ret = 0;
4427         spin_lock(&space_info->lock);
4428         /*
4429          * We only want to wait if somebody other than us is flushing and we
4430          * are actually allowed to flush all things.
4431          */
4432         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4433                space_info->flush) {
4434                 spin_unlock(&space_info->lock);
4435                 /*
4436                  * If we have a trans handle we can't wait because the flusher
4437                  * may have to commit the transaction, which would mean we would
4438                  * deadlock since we are waiting for the flusher to finish, but
4439                  * hold the current transaction open.
4440                  */
4441                 if (current->journal_info)
4442                         return -EAGAIN;
4443                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4444                 /* Must have been killed, return */
4445                 if (ret)
4446                         return -EINTR;
4447
4448                 spin_lock(&space_info->lock);
4449         }
4450
4451         ret = -ENOSPC;
4452         used = space_info->bytes_used + space_info->bytes_reserved +
4453                 space_info->bytes_pinned + space_info->bytes_readonly +
4454                 space_info->bytes_may_use;
4455
4456         /*
4457          * The idea here is that we've not already over-reserved the block group
4458          * then we can go ahead and save our reservation first and then start
4459          * flushing if we need to.  Otherwise if we've already overcommitted
4460          * lets start flushing stuff first and then come back and try to make
4461          * our reservation.
4462          */
4463         if (used <= space_info->total_bytes) {
4464                 if (used + orig_bytes <= space_info->total_bytes) {
4465                         space_info->bytes_may_use += orig_bytes;
4466                         trace_btrfs_space_reservation(root->fs_info,
4467                                 "space_info", space_info->flags, orig_bytes, 1);
4468                         ret = 0;
4469                 } else {
4470                         /*
4471                          * Ok set num_bytes to orig_bytes since we aren't
4472                          * overocmmitted, this way we only try and reclaim what
4473                          * we need.
4474                          */
4475                         num_bytes = orig_bytes;
4476                 }
4477         } else {
4478                 /*
4479                  * Ok we're over committed, set num_bytes to the overcommitted
4480                  * amount plus the amount of bytes that we need for this
4481                  * reservation.
4482                  */
4483                 num_bytes = used - space_info->total_bytes +
4484                         (orig_bytes * 2);
4485         }
4486
4487         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4488                 space_info->bytes_may_use += orig_bytes;
4489                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4490                                               space_info->flags, orig_bytes,
4491                                               1);
4492                 ret = 0;
4493         }
4494
4495         /*
4496          * Couldn't make our reservation, save our place so while we're trying
4497          * to reclaim space we can actually use it instead of somebody else
4498          * stealing it from us.
4499          *
4500          * We make the other tasks wait for the flush only when we can flush
4501          * all things.
4502          */
4503         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4504                 flushing = true;
4505                 space_info->flush = 1;
4506         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4507                 used += orig_bytes;
4508                 if (need_do_async_reclaim(space_info, root->fs_info, used) &&
4509                     !work_busy(&root->fs_info->async_reclaim_work))
4510                         queue_work(system_unbound_wq,
4511                                    &root->fs_info->async_reclaim_work);
4512         }
4513         spin_unlock(&space_info->lock);
4514
4515         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4516                 goto out;
4517
4518         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4519                           flush_state);
4520         flush_state++;
4521
4522         /*
4523          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4524          * would happen. So skip delalloc flush.
4525          */
4526         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4527             (flush_state == FLUSH_DELALLOC ||
4528              flush_state == FLUSH_DELALLOC_WAIT))
4529                 flush_state = ALLOC_CHUNK;
4530
4531         if (!ret)
4532                 goto again;
4533         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4534                  flush_state < COMMIT_TRANS)
4535                 goto again;
4536         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4537                  flush_state <= COMMIT_TRANS)
4538                 goto again;
4539
4540 out:
4541         if (ret == -ENOSPC &&
4542             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4543                 struct btrfs_block_rsv *global_rsv =
4544                         &root->fs_info->global_block_rsv;
4545
4546                 if (block_rsv != global_rsv &&
4547                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4548                         ret = 0;
4549         }
4550         if (ret == -ENOSPC)
4551                 trace_btrfs_space_reservation(root->fs_info,
4552                                               "space_info:enospc",
4553                                               space_info->flags, orig_bytes, 1);
4554         if (flushing) {
4555                 spin_lock(&space_info->lock);
4556                 space_info->flush = 0;
4557                 wake_up_all(&space_info->wait);
4558                 spin_unlock(&space_info->lock);
4559         }
4560         return ret;
4561 }
4562
4563 static struct btrfs_block_rsv *get_block_rsv(
4564                                         const struct btrfs_trans_handle *trans,
4565                                         const struct btrfs_root *root)
4566 {
4567         struct btrfs_block_rsv *block_rsv = NULL;
4568
4569         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4570                 block_rsv = trans->block_rsv;
4571
4572         if (root == root->fs_info->csum_root && trans->adding_csums)
4573                 block_rsv = trans->block_rsv;
4574
4575         if (root == root->fs_info->uuid_root)
4576                 block_rsv = trans->block_rsv;
4577
4578         if (!block_rsv)
4579                 block_rsv = root->block_rsv;
4580
4581         if (!block_rsv)
4582                 block_rsv = &root->fs_info->empty_block_rsv;
4583
4584         return block_rsv;
4585 }
4586
4587 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4588                                u64 num_bytes)
4589 {
4590         int ret = -ENOSPC;
4591         spin_lock(&block_rsv->lock);
4592         if (block_rsv->reserved >= num_bytes) {
4593                 block_rsv->reserved -= num_bytes;
4594                 if (block_rsv->reserved < block_rsv->size)
4595                         block_rsv->full = 0;
4596                 ret = 0;
4597         }
4598         spin_unlock(&block_rsv->lock);
4599         return ret;
4600 }
4601
4602 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4603                                 u64 num_bytes, int update_size)
4604 {
4605         spin_lock(&block_rsv->lock);
4606         block_rsv->reserved += num_bytes;
4607         if (update_size)
4608                 block_rsv->size += num_bytes;
4609         else if (block_rsv->reserved >= block_rsv->size)
4610                 block_rsv->full = 1;
4611         spin_unlock(&block_rsv->lock);
4612 }
4613
4614 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4615                              struct btrfs_block_rsv *dest, u64 num_bytes,
4616                              int min_factor)
4617 {
4618         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4619         u64 min_bytes;
4620
4621         if (global_rsv->space_info != dest->space_info)
4622                 return -ENOSPC;
4623
4624         spin_lock(&global_rsv->lock);
4625         min_bytes = div_factor(global_rsv->size, min_factor);
4626         if (global_rsv->reserved < min_bytes + num_bytes) {
4627                 spin_unlock(&global_rsv->lock);
4628                 return -ENOSPC;
4629         }
4630         global_rsv->reserved -= num_bytes;
4631         if (global_rsv->reserved < global_rsv->size)
4632                 global_rsv->full = 0;
4633         spin_unlock(&global_rsv->lock);
4634
4635         block_rsv_add_bytes(dest, num_bytes, 1);
4636         return 0;
4637 }
4638
4639 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4640                                     struct btrfs_block_rsv *block_rsv,
4641                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4642 {
4643         struct btrfs_space_info *space_info = block_rsv->space_info;
4644
4645         spin_lock(&block_rsv->lock);
4646         if (num_bytes == (u64)-1)
4647                 num_bytes = block_rsv->size;
4648         block_rsv->size -= num_bytes;
4649         if (block_rsv->reserved >= block_rsv->size) {
4650                 num_bytes = block_rsv->reserved - block_rsv->size;
4651                 block_rsv->reserved = block_rsv->size;
4652                 block_rsv->full = 1;
4653         } else {
4654                 num_bytes = 0;
4655         }
4656         spin_unlock(&block_rsv->lock);
4657
4658         if (num_bytes > 0) {
4659                 if (dest) {
4660                         spin_lock(&dest->lock);
4661                         if (!dest->full) {
4662                                 u64 bytes_to_add;
4663
4664                                 bytes_to_add = dest->size - dest->reserved;
4665                                 bytes_to_add = min(num_bytes, bytes_to_add);
4666                                 dest->reserved += bytes_to_add;
4667                                 if (dest->reserved >= dest->size)
4668                                         dest->full = 1;
4669                                 num_bytes -= bytes_to_add;
4670                         }
4671                         spin_unlock(&dest->lock);
4672                 }
4673                 if (num_bytes) {
4674                         spin_lock(&space_info->lock);
4675                         space_info->bytes_may_use -= num_bytes;
4676                         trace_btrfs_space_reservation(fs_info, "space_info",
4677                                         space_info->flags, num_bytes, 0);
4678                         spin_unlock(&space_info->lock);
4679                 }
4680         }
4681 }
4682
4683 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4684                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4685 {
4686         int ret;
4687
4688         ret = block_rsv_use_bytes(src, num_bytes);
4689         if (ret)
4690                 return ret;
4691
4692         block_rsv_add_bytes(dst, num_bytes, 1);
4693         return 0;
4694 }
4695
4696 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4697 {
4698         memset(rsv, 0, sizeof(*rsv));
4699         spin_lock_init(&rsv->lock);
4700         rsv->type = type;
4701 }
4702
4703 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4704                                               unsigned short type)
4705 {
4706         struct btrfs_block_rsv *block_rsv;
4707         struct btrfs_fs_info *fs_info = root->fs_info;
4708
4709         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4710         if (!block_rsv)
4711                 return NULL;
4712
4713         btrfs_init_block_rsv(block_rsv, type);
4714         block_rsv->space_info = __find_space_info(fs_info,
4715                                                   BTRFS_BLOCK_GROUP_METADATA);
4716         return block_rsv;
4717 }
4718
4719 void btrfs_free_block_rsv(struct btrfs_root *root,
4720                           struct btrfs_block_rsv *rsv)
4721 {
4722         if (!rsv)
4723                 return;
4724         btrfs_block_rsv_release(root, rsv, (u64)-1);
4725         kfree(rsv);
4726 }
4727
4728 int btrfs_block_rsv_add(struct btrfs_root *root,
4729                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4730                         enum btrfs_reserve_flush_enum flush)
4731 {
4732         int ret;
4733
4734         if (num_bytes == 0)
4735                 return 0;
4736
4737         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4738         if (!ret) {
4739                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4740                 return 0;
4741         }
4742
4743         return ret;
4744 }
4745
4746 int btrfs_block_rsv_check(struct btrfs_root *root,
4747                           struct btrfs_block_rsv *block_rsv, int min_factor)
4748 {
4749         u64 num_bytes = 0;
4750         int ret = -ENOSPC;
4751
4752         if (!block_rsv)
4753                 return 0;
4754
4755         spin_lock(&block_rsv->lock);
4756         num_bytes = div_factor(block_rsv->size, min_factor);
4757         if (block_rsv->reserved >= num_bytes)
4758                 ret = 0;
4759         spin_unlock(&block_rsv->lock);
4760
4761         return ret;
4762 }
4763
4764 int btrfs_block_rsv_refill(struct btrfs_root *root,
4765                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4766                            enum btrfs_reserve_flush_enum flush)
4767 {
4768         u64 num_bytes = 0;
4769         int ret = -ENOSPC;
4770
4771         if (!block_rsv)
4772                 return 0;
4773
4774         spin_lock(&block_rsv->lock);
4775         num_bytes = min_reserved;
4776         if (block_rsv->reserved >= num_bytes)
4777                 ret = 0;
4778         else
4779                 num_bytes -= block_rsv->reserved;
4780         spin_unlock(&block_rsv->lock);
4781
4782         if (!ret)
4783                 return 0;
4784
4785         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4786         if (!ret) {
4787                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4788                 return 0;
4789         }
4790
4791         return ret;
4792 }
4793
4794 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4795                             struct btrfs_block_rsv *dst_rsv,
4796                             u64 num_bytes)
4797 {
4798         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4799 }
4800
4801 void btrfs_block_rsv_release(struct btrfs_root *root,
4802                              struct btrfs_block_rsv *block_rsv,
4803                              u64 num_bytes)
4804 {
4805         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4806         if (global_rsv == block_rsv ||
4807             block_rsv->space_info != global_rsv->space_info)
4808                 global_rsv = NULL;
4809         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4810                                 num_bytes);
4811 }
4812
4813 /*
4814  * helper to calculate size of global block reservation.
4815  * the desired value is sum of space used by extent tree,
4816  * checksum tree and root tree
4817  */
4818 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4819 {
4820         struct btrfs_space_info *sinfo;
4821         u64 num_bytes;
4822         u64 meta_used;
4823         u64 data_used;
4824         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4825
4826         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4827         spin_lock(&sinfo->lock);
4828         data_used = sinfo->bytes_used;
4829         spin_unlock(&sinfo->lock);
4830
4831         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4832         spin_lock(&sinfo->lock);
4833         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4834                 data_used = 0;
4835         meta_used = sinfo->bytes_used;
4836         spin_unlock(&sinfo->lock);
4837
4838         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4839                     csum_size * 2;
4840         num_bytes += div64_u64(data_used + meta_used, 50);
4841
4842         if (num_bytes * 3 > meta_used)
4843                 num_bytes = div64_u64(meta_used, 3);
4844
4845         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4846 }
4847
4848 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4849 {
4850         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4851         struct btrfs_space_info *sinfo = block_rsv->space_info;
4852         u64 num_bytes;
4853
4854         num_bytes = calc_global_metadata_size(fs_info);
4855
4856         spin_lock(&sinfo->lock);
4857         spin_lock(&block_rsv->lock);
4858
4859         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4860
4861         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4862                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4863                     sinfo->bytes_may_use;
4864
4865         if (sinfo->total_bytes > num_bytes) {
4866                 num_bytes = sinfo->total_bytes - num_bytes;
4867                 block_rsv->reserved += num_bytes;
4868                 sinfo->bytes_may_use += num_bytes;
4869                 trace_btrfs_space_reservation(fs_info, "space_info",
4870                                       sinfo->flags, num_bytes, 1);
4871         }
4872
4873         if (block_rsv->reserved >= block_rsv->size) {
4874                 num_bytes = block_rsv->reserved - block_rsv->size;
4875                 sinfo->bytes_may_use -= num_bytes;
4876                 trace_btrfs_space_reservation(fs_info, "space_info",
4877                                       sinfo->flags, num_bytes, 0);
4878                 block_rsv->reserved = block_rsv->size;
4879                 block_rsv->full = 1;
4880         }
4881
4882         spin_unlock(&block_rsv->lock);
4883         spin_unlock(&sinfo->lock);
4884 }
4885
4886 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4887 {
4888         struct btrfs_space_info *space_info;
4889
4890         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4891         fs_info->chunk_block_rsv.space_info = space_info;
4892
4893         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4894         fs_info->global_block_rsv.space_info = space_info;
4895         fs_info->delalloc_block_rsv.space_info = space_info;
4896         fs_info->trans_block_rsv.space_info = space_info;
4897         fs_info->empty_block_rsv.space_info = space_info;
4898         fs_info->delayed_block_rsv.space_info = space_info;
4899
4900         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4901         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4902         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4903         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4904         if (fs_info->quota_root)
4905                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
4906         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4907
4908         update_global_block_rsv(fs_info);
4909 }
4910
4911 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4912 {
4913         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4914                                 (u64)-1);
4915         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4916         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4917         WARN_ON(fs_info->trans_block_rsv.size > 0);
4918         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4919         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4920         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4921         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4922         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4923 }
4924
4925 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4926                                   struct btrfs_root *root)
4927 {
4928         if (!trans->block_rsv)
4929                 return;
4930
4931         if (!trans->bytes_reserved)
4932                 return;
4933
4934         trace_btrfs_space_reservation(root->fs_info, "transaction",
4935                                       trans->transid, trans->bytes_reserved, 0);
4936         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4937         trans->bytes_reserved = 0;
4938 }
4939
4940 /* Can only return 0 or -ENOSPC */
4941 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4942                                   struct inode *inode)
4943 {
4944         struct btrfs_root *root = BTRFS_I(inode)->root;
4945         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4946         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4947
4948         /*
4949          * We need to hold space in order to delete our orphan item once we've
4950          * added it, so this takes the reservation so we can release it later
4951          * when we are truly done with the orphan item.
4952          */
4953         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4954         trace_btrfs_space_reservation(root->fs_info, "orphan",
4955                                       btrfs_ino(inode), num_bytes, 1);
4956         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4957 }
4958
4959 void btrfs_orphan_release_metadata(struct inode *inode)
4960 {
4961         struct btrfs_root *root = BTRFS_I(inode)->root;
4962         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4963         trace_btrfs_space_reservation(root->fs_info, "orphan",
4964                                       btrfs_ino(inode), num_bytes, 0);
4965         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4966 }
4967
4968 /*
4969  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4970  * root: the root of the parent directory
4971  * rsv: block reservation
4972  * items: the number of items that we need do reservation
4973  * qgroup_reserved: used to return the reserved size in qgroup
4974  *
4975  * This function is used to reserve the space for snapshot/subvolume
4976  * creation and deletion. Those operations are different with the
4977  * common file/directory operations, they change two fs/file trees
4978  * and root tree, the number of items that the qgroup reserves is
4979  * different with the free space reservation. So we can not use
4980  * the space reseravtion mechanism in start_transaction().
4981  */
4982 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4983                                      struct btrfs_block_rsv *rsv,
4984                                      int items,
4985                                      u64 *qgroup_reserved,
4986                                      bool use_global_rsv)
4987 {
4988         u64 num_bytes;
4989         int ret;
4990         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4991
4992         if (root->fs_info->quota_enabled) {
4993                 /* One for parent inode, two for dir entries */
4994                 num_bytes = 3 * root->leafsize;
4995                 ret = btrfs_qgroup_reserve(root, num_bytes);
4996                 if (ret)
4997                         return ret;
4998         } else {
4999                 num_bytes = 0;
5000         }
5001
5002         *qgroup_reserved = num_bytes;
5003
5004         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5005         rsv->space_info = __find_space_info(root->fs_info,
5006                                             BTRFS_BLOCK_GROUP_METADATA);
5007         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5008                                   BTRFS_RESERVE_FLUSH_ALL);
5009
5010         if (ret == -ENOSPC && use_global_rsv)
5011                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5012
5013         if (ret) {
5014                 if (*qgroup_reserved)
5015                         btrfs_qgroup_free(root, *qgroup_reserved);
5016         }
5017
5018         return ret;
5019 }
5020
5021 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5022                                       struct btrfs_block_rsv *rsv,
5023                                       u64 qgroup_reserved)
5024 {
5025         btrfs_block_rsv_release(root, rsv, (u64)-1);
5026         if (qgroup_reserved)
5027                 btrfs_qgroup_free(root, qgroup_reserved);
5028 }
5029
5030 /**
5031  * drop_outstanding_extent - drop an outstanding extent
5032  * @inode: the inode we're dropping the extent for
5033  *
5034  * This is called when we are freeing up an outstanding extent, either called
5035  * after an error or after an extent is written.  This will return the number of
5036  * reserved extents that need to be freed.  This must be called with
5037  * BTRFS_I(inode)->lock held.
5038  */
5039 static unsigned drop_outstanding_extent(struct inode *inode)
5040 {
5041         unsigned drop_inode_space = 0;
5042         unsigned dropped_extents = 0;
5043
5044         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
5045         BTRFS_I(inode)->outstanding_extents--;
5046
5047         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5048             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5049                                &BTRFS_I(inode)->runtime_flags))
5050                 drop_inode_space = 1;
5051
5052         /*
5053          * If we have more or the same amount of outsanding extents than we have
5054          * reserved then we need to leave the reserved extents count alone.
5055          */
5056         if (BTRFS_I(inode)->outstanding_extents >=
5057             BTRFS_I(inode)->reserved_extents)
5058                 return drop_inode_space;
5059
5060         dropped_extents = BTRFS_I(inode)->reserved_extents -
5061                 BTRFS_I(inode)->outstanding_extents;
5062         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5063         return dropped_extents + drop_inode_space;
5064 }
5065
5066 /**
5067  * calc_csum_metadata_size - return the amount of metada space that must be
5068  *      reserved/free'd for the given bytes.
5069  * @inode: the inode we're manipulating
5070  * @num_bytes: the number of bytes in question
5071  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5072  *
5073  * This adjusts the number of csum_bytes in the inode and then returns the
5074  * correct amount of metadata that must either be reserved or freed.  We
5075  * calculate how many checksums we can fit into one leaf and then divide the
5076  * number of bytes that will need to be checksumed by this value to figure out
5077  * how many checksums will be required.  If we are adding bytes then the number
5078  * may go up and we will return the number of additional bytes that must be
5079  * reserved.  If it is going down we will return the number of bytes that must
5080  * be freed.
5081  *
5082  * This must be called with BTRFS_I(inode)->lock held.
5083  */
5084 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5085                                    int reserve)
5086 {
5087         struct btrfs_root *root = BTRFS_I(inode)->root;
5088         u64 csum_size;
5089         int num_csums_per_leaf;
5090         int num_csums;
5091         int old_csums;
5092
5093         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5094             BTRFS_I(inode)->csum_bytes == 0)
5095                 return 0;
5096
5097         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
5098         if (reserve)
5099                 BTRFS_I(inode)->csum_bytes += num_bytes;
5100         else
5101                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5102         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
5103         num_csums_per_leaf = (int)div64_u64(csum_size,
5104                                             sizeof(struct btrfs_csum_item) +
5105                                             sizeof(struct btrfs_disk_key));
5106         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
5107         num_csums = num_csums + num_csums_per_leaf - 1;
5108         num_csums = num_csums / num_csums_per_leaf;
5109
5110         old_csums = old_csums + num_csums_per_leaf - 1;
5111         old_csums = old_csums / num_csums_per_leaf;
5112
5113         /* No change, no need to reserve more */
5114         if (old_csums == num_csums)
5115                 return 0;
5116
5117         if (reserve)
5118                 return btrfs_calc_trans_metadata_size(root,
5119                                                       num_csums - old_csums);
5120
5121         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5122 }
5123
5124 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5125 {
5126         struct btrfs_root *root = BTRFS_I(inode)->root;
5127         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5128         u64 to_reserve = 0;
5129         u64 csum_bytes;
5130         unsigned nr_extents = 0;
5131         int extra_reserve = 0;
5132         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5133         int ret = 0;
5134         bool delalloc_lock = true;
5135         u64 to_free = 0;
5136         unsigned dropped;
5137
5138         /* If we are a free space inode we need to not flush since we will be in
5139          * the middle of a transaction commit.  We also don't need the delalloc
5140          * mutex since we won't race with anybody.  We need this mostly to make
5141          * lockdep shut its filthy mouth.
5142          */
5143         if (btrfs_is_free_space_inode(inode)) {
5144                 flush = BTRFS_RESERVE_NO_FLUSH;
5145                 delalloc_lock = false;
5146         }
5147
5148         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5149             btrfs_transaction_in_commit(root->fs_info))
5150                 schedule_timeout(1);
5151
5152         if (delalloc_lock)
5153                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5154
5155         num_bytes = ALIGN(num_bytes, root->sectorsize);
5156
5157         spin_lock(&BTRFS_I(inode)->lock);
5158         BTRFS_I(inode)->outstanding_extents++;
5159
5160         if (BTRFS_I(inode)->outstanding_extents >
5161             BTRFS_I(inode)->reserved_extents)
5162                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5163                         BTRFS_I(inode)->reserved_extents;
5164
5165         /*
5166          * Add an item to reserve for updating the inode when we complete the
5167          * delalloc io.
5168          */
5169         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5170                       &BTRFS_I(inode)->runtime_flags)) {
5171                 nr_extents++;
5172                 extra_reserve = 1;
5173         }
5174
5175         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5176         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5177         csum_bytes = BTRFS_I(inode)->csum_bytes;
5178         spin_unlock(&BTRFS_I(inode)->lock);
5179
5180         if (root->fs_info->quota_enabled) {
5181                 ret = btrfs_qgroup_reserve(root, num_bytes +
5182                                            nr_extents * root->leafsize);
5183                 if (ret)
5184                         goto out_fail;
5185         }
5186
5187         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5188         if (unlikely(ret)) {
5189                 if (root->fs_info->quota_enabled)
5190                         btrfs_qgroup_free(root, num_bytes +
5191                                                 nr_extents * root->leafsize);
5192                 goto out_fail;
5193         }
5194
5195         spin_lock(&BTRFS_I(inode)->lock);
5196         if (extra_reserve) {
5197                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5198                         &BTRFS_I(inode)->runtime_flags);
5199                 nr_extents--;
5200         }
5201         BTRFS_I(inode)->reserved_extents += nr_extents;
5202         spin_unlock(&BTRFS_I(inode)->lock);
5203
5204         if (delalloc_lock)
5205                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5206
5207         if (to_reserve)
5208                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5209                                               btrfs_ino(inode), to_reserve, 1);
5210         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5211
5212         return 0;
5213
5214 out_fail:
5215         spin_lock(&BTRFS_I(inode)->lock);
5216         dropped = drop_outstanding_extent(inode);
5217         /*
5218          * If the inodes csum_bytes is the same as the original
5219          * csum_bytes then we know we haven't raced with any free()ers
5220          * so we can just reduce our inodes csum bytes and carry on.
5221          */
5222         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5223                 calc_csum_metadata_size(inode, num_bytes, 0);
5224         } else {
5225                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5226                 u64 bytes;
5227
5228                 /*
5229                  * This is tricky, but first we need to figure out how much we
5230                  * free'd from any free-ers that occured during this
5231                  * reservation, so we reset ->csum_bytes to the csum_bytes
5232                  * before we dropped our lock, and then call the free for the
5233                  * number of bytes that were freed while we were trying our
5234                  * reservation.
5235                  */
5236                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5237                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5238                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5239
5240
5241                 /*
5242                  * Now we need to see how much we would have freed had we not
5243                  * been making this reservation and our ->csum_bytes were not
5244                  * artificially inflated.
5245                  */
5246                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5247                 bytes = csum_bytes - orig_csum_bytes;
5248                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5249
5250                 /*
5251                  * Now reset ->csum_bytes to what it should be.  If bytes is
5252                  * more than to_free then we would have free'd more space had we
5253                  * not had an artificially high ->csum_bytes, so we need to free
5254                  * the remainder.  If bytes is the same or less then we don't
5255                  * need to do anything, the other free-ers did the correct
5256                  * thing.
5257                  */
5258                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5259                 if (bytes > to_free)
5260                         to_free = bytes - to_free;
5261                 else
5262                         to_free = 0;
5263         }
5264         spin_unlock(&BTRFS_I(inode)->lock);
5265         if (dropped)
5266                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5267
5268         if (to_free) {
5269                 btrfs_block_rsv_release(root, block_rsv, to_free);
5270                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5271                                               btrfs_ino(inode), to_free, 0);
5272         }
5273         if (delalloc_lock)
5274                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5275         return ret;
5276 }
5277
5278 /**
5279  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5280  * @inode: the inode to release the reservation for
5281  * @num_bytes: the number of bytes we're releasing
5282  *
5283  * This will release the metadata reservation for an inode.  This can be called
5284  * once we complete IO for a given set of bytes to release their metadata
5285  * reservations.
5286  */
5287 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5288 {
5289         struct btrfs_root *root = BTRFS_I(inode)->root;
5290         u64 to_free = 0;
5291         unsigned dropped;
5292
5293         num_bytes = ALIGN(num_bytes, root->sectorsize);
5294         spin_lock(&BTRFS_I(inode)->lock);
5295         dropped = drop_outstanding_extent(inode);
5296
5297         if (num_bytes)
5298                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5299         spin_unlock(&BTRFS_I(inode)->lock);
5300         if (dropped > 0)
5301                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5302
5303         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5304                                       btrfs_ino(inode), to_free, 0);
5305         if (root->fs_info->quota_enabled) {
5306                 btrfs_qgroup_free(root, num_bytes +
5307                                         dropped * root->leafsize);
5308         }
5309
5310         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5311                                 to_free);
5312 }
5313
5314 /**
5315  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5316  * @inode: inode we're writing to
5317  * @num_bytes: the number of bytes we want to allocate
5318  *
5319  * This will do the following things
5320  *
5321  * o reserve space in the data space info for num_bytes
5322  * o reserve space in the metadata space info based on number of outstanding
5323  *   extents and how much csums will be needed
5324  * o add to the inodes ->delalloc_bytes
5325  * o add it to the fs_info's delalloc inodes list.
5326  *
5327  * This will return 0 for success and -ENOSPC if there is no space left.
5328  */
5329 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5330 {
5331         int ret;
5332
5333         ret = btrfs_check_data_free_space(inode, num_bytes);
5334         if (ret)
5335                 return ret;
5336
5337         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5338         if (ret) {
5339                 btrfs_free_reserved_data_space(inode, num_bytes);
5340                 return ret;
5341         }
5342
5343         return 0;
5344 }
5345
5346 /**
5347  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5348  * @inode: inode we're releasing space for
5349  * @num_bytes: the number of bytes we want to free up
5350  *
5351  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5352  * called in the case that we don't need the metadata AND data reservations
5353  * anymore.  So if there is an error or we insert an inline extent.
5354  *
5355  * This function will release the metadata space that was not used and will
5356  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5357  * list if there are no delalloc bytes left.
5358  */
5359 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5360 {
5361         btrfs_delalloc_release_metadata(inode, num_bytes);
5362         btrfs_free_reserved_data_space(inode, num_bytes);
5363 }
5364
5365 static int update_block_group(struct btrfs_root *root,
5366                               u64 bytenr, u64 num_bytes, int alloc)
5367 {
5368         struct btrfs_block_group_cache *cache = NULL;
5369         struct btrfs_fs_info *info = root->fs_info;
5370         u64 total = num_bytes;
5371         u64 old_val;
5372         u64 byte_in_group;
5373         int factor;
5374
5375         /* block accounting for super block */
5376         spin_lock(&info->delalloc_root_lock);
5377         old_val = btrfs_super_bytes_used(info->super_copy);
5378         if (alloc)
5379                 old_val += num_bytes;
5380         else
5381                 old_val -= num_bytes;
5382         btrfs_set_super_bytes_used(info->super_copy, old_val);
5383         spin_unlock(&info->delalloc_root_lock);
5384
5385         while (total) {
5386                 cache = btrfs_lookup_block_group(info, bytenr);
5387                 if (!cache)
5388                         return -ENOENT;
5389                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5390                                     BTRFS_BLOCK_GROUP_RAID1 |
5391                                     BTRFS_BLOCK_GROUP_RAID10))
5392                         factor = 2;
5393                 else
5394                         factor = 1;
5395                 /*
5396                  * If this block group has free space cache written out, we
5397                  * need to make sure to load it if we are removing space.  This
5398                  * is because we need the unpinning stage to actually add the
5399                  * space back to the block group, otherwise we will leak space.
5400                  */
5401                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5402                         cache_block_group(cache, 1);
5403
5404                 byte_in_group = bytenr - cache->key.objectid;
5405                 WARN_ON(byte_in_group > cache->key.offset);
5406
5407                 spin_lock(&cache->space_info->lock);
5408                 spin_lock(&cache->lock);
5409
5410                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5411                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5412                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5413
5414                 cache->dirty = 1;
5415                 old_val = btrfs_block_group_used(&cache->item);
5416                 num_bytes = min(total, cache->key.offset - byte_in_group);
5417                 if (alloc) {
5418                         old_val += num_bytes;
5419                         btrfs_set_block_group_used(&cache->item, old_val);
5420                         cache->reserved -= num_bytes;
5421                         cache->space_info->bytes_reserved -= num_bytes;
5422                         cache->space_info->bytes_used += num_bytes;
5423                         cache->space_info->disk_used += num_bytes * factor;
5424                         spin_unlock(&cache->lock);
5425                         spin_unlock(&cache->space_info->lock);
5426                 } else {
5427                         old_val -= num_bytes;
5428                         btrfs_set_block_group_used(&cache->item, old_val);
5429                         cache->pinned += num_bytes;
5430                         cache->space_info->bytes_pinned += num_bytes;
5431                         cache->space_info->bytes_used -= num_bytes;
5432                         cache->space_info->disk_used -= num_bytes * factor;
5433                         spin_unlock(&cache->lock);
5434                         spin_unlock(&cache->space_info->lock);
5435
5436                         set_extent_dirty(info->pinned_extents,
5437                                          bytenr, bytenr + num_bytes - 1,
5438                                          GFP_NOFS | __GFP_NOFAIL);
5439                 }
5440                 btrfs_put_block_group(cache);
5441                 total -= num_bytes;
5442                 bytenr += num_bytes;
5443         }
5444         return 0;
5445 }
5446
5447 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5448 {
5449         struct btrfs_block_group_cache *cache;
5450         u64 bytenr;
5451
5452         spin_lock(&root->fs_info->block_group_cache_lock);
5453         bytenr = root->fs_info->first_logical_byte;
5454         spin_unlock(&root->fs_info->block_group_cache_lock);
5455
5456         if (bytenr < (u64)-1)
5457                 return bytenr;
5458
5459         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5460         if (!cache)
5461                 return 0;
5462
5463         bytenr = cache->key.objectid;
5464         btrfs_put_block_group(cache);
5465
5466         return bytenr;
5467 }
5468
5469 static int pin_down_extent(struct btrfs_root *root,
5470                            struct btrfs_block_group_cache *cache,
5471                            u64 bytenr, u64 num_bytes, int reserved)
5472 {
5473         spin_lock(&cache->space_info->lock);
5474         spin_lock(&cache->lock);
5475         cache->pinned += num_bytes;
5476         cache->space_info->bytes_pinned += num_bytes;
5477         if (reserved) {
5478                 cache->reserved -= num_bytes;
5479                 cache->space_info->bytes_reserved -= num_bytes;
5480         }
5481         spin_unlock(&cache->lock);
5482         spin_unlock(&cache->space_info->lock);
5483
5484         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5485                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5486         if (reserved)
5487                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5488         return 0;
5489 }
5490
5491 /*
5492  * this function must be called within transaction
5493  */
5494 int btrfs_pin_extent(struct btrfs_root *root,
5495                      u64 bytenr, u64 num_bytes, int reserved)
5496 {
5497         struct btrfs_block_group_cache *cache;
5498
5499         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5500         BUG_ON(!cache); /* Logic error */
5501
5502         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5503
5504         btrfs_put_block_group(cache);
5505         return 0;
5506 }
5507
5508 /*
5509  * this function must be called within transaction
5510  */
5511 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5512                                     u64 bytenr, u64 num_bytes)
5513 {
5514         struct btrfs_block_group_cache *cache;
5515         int ret;
5516
5517         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5518         if (!cache)
5519                 return -EINVAL;
5520
5521         /*
5522          * pull in the free space cache (if any) so that our pin
5523          * removes the free space from the cache.  We have load_only set
5524          * to one because the slow code to read in the free extents does check
5525          * the pinned extents.
5526          */
5527         cache_block_group(cache, 1);
5528
5529         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5530
5531         /* remove us from the free space cache (if we're there at all) */
5532         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5533         btrfs_put_block_group(cache);
5534         return ret;
5535 }
5536
5537 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5538 {
5539         int ret;
5540         struct btrfs_block_group_cache *block_group;
5541         struct btrfs_caching_control *caching_ctl;
5542
5543         block_group = btrfs_lookup_block_group(root->fs_info, start);
5544         if (!block_group)
5545                 return -EINVAL;
5546
5547         cache_block_group(block_group, 0);
5548         caching_ctl = get_caching_control(block_group);
5549
5550         if (!caching_ctl) {
5551                 /* Logic error */
5552                 BUG_ON(!block_group_cache_done(block_group));
5553                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5554         } else {
5555                 mutex_lock(&caching_ctl->mutex);
5556
5557                 if (start >= caching_ctl->progress) {
5558                         ret = add_excluded_extent(root, start, num_bytes);
5559                 } else if (start + num_bytes <= caching_ctl->progress) {
5560                         ret = btrfs_remove_free_space(block_group,
5561                                                       start, num_bytes);
5562                 } else {
5563                         num_bytes = caching_ctl->progress - start;
5564                         ret = btrfs_remove_free_space(block_group,
5565                                                       start, num_bytes);
5566                         if (ret)
5567                                 goto out_lock;
5568
5569                         num_bytes = (start + num_bytes) -
5570                                 caching_ctl->progress;
5571                         start = caching_ctl->progress;
5572                         ret = add_excluded_extent(root, start, num_bytes);
5573                 }
5574 out_lock:
5575                 mutex_unlock(&caching_ctl->mutex);
5576                 put_caching_control(caching_ctl);
5577         }
5578         btrfs_put_block_group(block_group);
5579         return ret;
5580 }
5581
5582 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5583                                  struct extent_buffer *eb)
5584 {
5585         struct btrfs_file_extent_item *item;
5586         struct btrfs_key key;
5587         int found_type;
5588         int i;
5589
5590         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5591                 return 0;
5592
5593         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5594                 btrfs_item_key_to_cpu(eb, &key, i);
5595                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5596                         continue;
5597                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5598                 found_type = btrfs_file_extent_type(eb, item);
5599                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5600                         continue;
5601                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5602                         continue;
5603                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5604                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5605                 __exclude_logged_extent(log, key.objectid, key.offset);
5606         }
5607
5608         return 0;
5609 }
5610
5611 /**
5612  * btrfs_update_reserved_bytes - update the block_group and space info counters
5613  * @cache:      The cache we are manipulating
5614  * @num_bytes:  The number of bytes in question
5615  * @reserve:    One of the reservation enums
5616  *
5617  * This is called by the allocator when it reserves space, or by somebody who is
5618  * freeing space that was never actually used on disk.  For example if you
5619  * reserve some space for a new leaf in transaction A and before transaction A
5620  * commits you free that leaf, you call this with reserve set to 0 in order to
5621  * clear the reservation.
5622  *
5623  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5624  * ENOSPC accounting.  For data we handle the reservation through clearing the
5625  * delalloc bits in the io_tree.  We have to do this since we could end up
5626  * allocating less disk space for the amount of data we have reserved in the
5627  * case of compression.
5628  *
5629  * If this is a reservation and the block group has become read only we cannot
5630  * make the reservation and return -EAGAIN, otherwise this function always
5631  * succeeds.
5632  */
5633 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5634                                        u64 num_bytes, int reserve)
5635 {
5636         struct btrfs_space_info *space_info = cache->space_info;
5637         int ret = 0;
5638
5639         spin_lock(&space_info->lock);
5640         spin_lock(&cache->lock);
5641         if (reserve != RESERVE_FREE) {
5642                 if (cache->ro) {
5643                         ret = -EAGAIN;
5644                 } else {
5645                         cache->reserved += num_bytes;
5646                         space_info->bytes_reserved += num_bytes;
5647                         if (reserve == RESERVE_ALLOC) {
5648                                 trace_btrfs_space_reservation(cache->fs_info,
5649                                                 "space_info", space_info->flags,
5650                                                 num_bytes, 0);
5651                                 space_info->bytes_may_use -= num_bytes;
5652                         }
5653                 }
5654         } else {
5655                 if (cache->ro)
5656                         space_info->bytes_readonly += num_bytes;
5657                 cache->reserved -= num_bytes;
5658                 space_info->bytes_reserved -= num_bytes;
5659         }
5660         spin_unlock(&cache->lock);
5661         spin_unlock(&space_info->lock);
5662         return ret;
5663 }
5664
5665 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5666                                 struct btrfs_root *root)
5667 {
5668         struct btrfs_fs_info *fs_info = root->fs_info;
5669         struct btrfs_caching_control *next;
5670         struct btrfs_caching_control *caching_ctl;
5671         struct btrfs_block_group_cache *cache;
5672         struct btrfs_space_info *space_info;
5673
5674         down_write(&fs_info->commit_root_sem);
5675
5676         list_for_each_entry_safe(caching_ctl, next,
5677                                  &fs_info->caching_block_groups, list) {
5678                 cache = caching_ctl->block_group;
5679                 if (block_group_cache_done(cache)) {
5680                         cache->last_byte_to_unpin = (u64)-1;
5681                         list_del_init(&caching_ctl->list);
5682                         put_caching_control(caching_ctl);
5683                 } else {
5684                         cache->last_byte_to_unpin = caching_ctl->progress;
5685                 }
5686         }
5687
5688         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5689                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5690         else
5691                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5692
5693         up_write(&fs_info->commit_root_sem);
5694
5695         list_for_each_entry_rcu(space_info, &fs_info->space_info, list)
5696                 percpu_counter_set(&space_info->total_bytes_pinned, 0);
5697
5698         update_global_block_rsv(fs_info);
5699 }
5700
5701 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5702 {
5703         struct btrfs_fs_info *fs_info = root->fs_info;
5704         struct btrfs_block_group_cache *cache = NULL;
5705         struct btrfs_space_info *space_info;
5706         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5707         u64 len;
5708         bool readonly;
5709
5710         while (start <= end) {
5711                 readonly = false;
5712                 if (!cache ||
5713                     start >= cache->key.objectid + cache->key.offset) {
5714                         if (cache)
5715                                 btrfs_put_block_group(cache);
5716                         cache = btrfs_lookup_block_group(fs_info, start);
5717                         BUG_ON(!cache); /* Logic error */
5718                 }
5719
5720                 len = cache->key.objectid + cache->key.offset - start;
5721                 len = min(len, end + 1 - start);
5722
5723                 if (start < cache->last_byte_to_unpin) {
5724                         len = min(len, cache->last_byte_to_unpin - start);
5725                         btrfs_add_free_space(cache, start, len);
5726                 }
5727
5728                 start += len;
5729                 space_info = cache->space_info;
5730
5731                 spin_lock(&space_info->lock);
5732                 spin_lock(&cache->lock);
5733                 cache->pinned -= len;
5734                 space_info->bytes_pinned -= len;
5735                 if (cache->ro) {
5736                         space_info->bytes_readonly += len;
5737                         readonly = true;
5738                 }
5739                 spin_unlock(&cache->lock);
5740                 if (!readonly && global_rsv->space_info == space_info) {
5741                         spin_lock(&global_rsv->lock);
5742                         if (!global_rsv->full) {
5743                                 len = min(len, global_rsv->size -
5744                                           global_rsv->reserved);
5745                                 global_rsv->reserved += len;
5746                                 space_info->bytes_may_use += len;
5747                                 if (global_rsv->reserved >= global_rsv->size)
5748                                         global_rsv->full = 1;
5749                         }
5750                         spin_unlock(&global_rsv->lock);
5751                 }
5752                 spin_unlock(&space_info->lock);
5753         }
5754
5755         if (cache)
5756                 btrfs_put_block_group(cache);
5757         return 0;
5758 }
5759
5760 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5761                                struct btrfs_root *root)
5762 {
5763         struct btrfs_fs_info *fs_info = root->fs_info;
5764         struct extent_io_tree *unpin;
5765         u64 start;
5766         u64 end;
5767         int ret;
5768
5769         if (trans->aborted)
5770                 return 0;
5771
5772         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5773                 unpin = &fs_info->freed_extents[1];
5774         else
5775                 unpin = &fs_info->freed_extents[0];
5776
5777         while (1) {
5778                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5779                                             EXTENT_DIRTY, NULL);
5780                 if (ret)
5781                         break;
5782
5783                 if (btrfs_test_opt(root, DISCARD))
5784                         ret = btrfs_discard_extent(root, start,
5785                                                    end + 1 - start, NULL);
5786
5787                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5788                 unpin_extent_range(root, start, end);
5789                 cond_resched();
5790         }
5791
5792         return 0;
5793 }
5794
5795 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
5796                              u64 owner, u64 root_objectid)
5797 {
5798         struct btrfs_space_info *space_info;
5799         u64 flags;
5800
5801         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5802                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
5803                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
5804                 else
5805                         flags = BTRFS_BLOCK_GROUP_METADATA;
5806         } else {
5807                 flags = BTRFS_BLOCK_GROUP_DATA;
5808         }
5809
5810         space_info = __find_space_info(fs_info, flags);
5811         BUG_ON(!space_info); /* Logic bug */
5812         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
5813 }
5814
5815
5816 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5817                                 struct btrfs_root *root,
5818                                 u64 bytenr, u64 num_bytes, u64 parent,
5819                                 u64 root_objectid, u64 owner_objectid,
5820                                 u64 owner_offset, int refs_to_drop,
5821                                 struct btrfs_delayed_extent_op *extent_op,
5822                                 int no_quota)
5823 {
5824         struct btrfs_key key;
5825         struct btrfs_path *path;
5826         struct btrfs_fs_info *info = root->fs_info;
5827         struct btrfs_root *extent_root = info->extent_root;
5828         struct extent_buffer *leaf;
5829         struct btrfs_extent_item *ei;
5830         struct btrfs_extent_inline_ref *iref;
5831         int ret;
5832         int is_data;
5833         int extent_slot = 0;
5834         int found_extent = 0;
5835         int num_to_del = 1;
5836         u32 item_size;
5837         u64 refs;
5838         int last_ref = 0;
5839         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_SUB_EXCL;
5840         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5841                                                  SKINNY_METADATA);
5842
5843         if (!info->quota_enabled || !is_fstree(root_objectid))
5844                 no_quota = 1;
5845
5846         path = btrfs_alloc_path();
5847         if (!path)
5848                 return -ENOMEM;
5849
5850         path->reada = 1;
5851         path->leave_spinning = 1;
5852
5853         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5854         BUG_ON(!is_data && refs_to_drop != 1);
5855
5856         if (is_data)
5857                 skinny_metadata = 0;
5858
5859         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5860                                     bytenr, num_bytes, parent,
5861                                     root_objectid, owner_objectid,
5862                                     owner_offset);
5863         if (ret == 0) {
5864                 extent_slot = path->slots[0];
5865                 while (extent_slot >= 0) {
5866                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5867                                               extent_slot);
5868                         if (key.objectid != bytenr)
5869                                 break;
5870                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5871                             key.offset == num_bytes) {
5872                                 found_extent = 1;
5873                                 break;
5874                         }
5875                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5876                             key.offset == owner_objectid) {
5877                                 found_extent = 1;
5878                                 break;
5879                         }
5880                         if (path->slots[0] - extent_slot > 5)
5881                                 break;
5882                         extent_slot--;
5883                 }
5884 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5885                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5886                 if (found_extent && item_size < sizeof(*ei))
5887                         found_extent = 0;
5888 #endif
5889                 if (!found_extent) {
5890                         BUG_ON(iref);
5891                         ret = remove_extent_backref(trans, extent_root, path,
5892                                                     NULL, refs_to_drop,
5893                                                     is_data, &last_ref);
5894                         if (ret) {
5895                                 btrfs_abort_transaction(trans, extent_root, ret);
5896                                 goto out;
5897                         }
5898                         btrfs_release_path(path);
5899                         path->leave_spinning = 1;
5900
5901                         key.objectid = bytenr;
5902                         key.type = BTRFS_EXTENT_ITEM_KEY;
5903                         key.offset = num_bytes;
5904
5905                         if (!is_data && skinny_metadata) {
5906                                 key.type = BTRFS_METADATA_ITEM_KEY;
5907                                 key.offset = owner_objectid;
5908                         }
5909
5910                         ret = btrfs_search_slot(trans, extent_root,
5911                                                 &key, path, -1, 1);
5912                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5913                                 /*
5914                                  * Couldn't find our skinny metadata item,
5915                                  * see if we have ye olde extent item.
5916                                  */
5917                                 path->slots[0]--;
5918                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5919                                                       path->slots[0]);
5920                                 if (key.objectid == bytenr &&
5921                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5922                                     key.offset == num_bytes)
5923                                         ret = 0;
5924                         }
5925
5926                         if (ret > 0 && skinny_metadata) {
5927                                 skinny_metadata = false;
5928                                 key.objectid = bytenr;
5929                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5930                                 key.offset = num_bytes;
5931                                 btrfs_release_path(path);
5932                                 ret = btrfs_search_slot(trans, extent_root,
5933                                                         &key, path, -1, 1);
5934                         }
5935
5936                         if (ret) {
5937                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5938                                         ret, bytenr);
5939                                 if (ret > 0)
5940                                         btrfs_print_leaf(extent_root,
5941                                                          path->nodes[0]);
5942                         }
5943                         if (ret < 0) {
5944                                 btrfs_abort_transaction(trans, extent_root, ret);
5945                                 goto out;
5946                         }
5947                         extent_slot = path->slots[0];
5948                 }
5949         } else if (WARN_ON(ret == -ENOENT)) {
5950                 btrfs_print_leaf(extent_root, path->nodes[0]);
5951                 btrfs_err(info,
5952                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5953                         bytenr, parent, root_objectid, owner_objectid,
5954                         owner_offset);
5955                 btrfs_abort_transaction(trans, extent_root, ret);
5956                 goto out;
5957         } else {
5958                 btrfs_abort_transaction(trans, extent_root, ret);
5959                 goto out;
5960         }
5961
5962         leaf = path->nodes[0];
5963         item_size = btrfs_item_size_nr(leaf, extent_slot);
5964 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5965         if (item_size < sizeof(*ei)) {
5966                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5967                 ret = convert_extent_item_v0(trans, extent_root, path,
5968                                              owner_objectid, 0);
5969                 if (ret < 0) {
5970                         btrfs_abort_transaction(trans, extent_root, ret);
5971                         goto out;
5972                 }
5973
5974                 btrfs_release_path(path);
5975                 path->leave_spinning = 1;
5976
5977                 key.objectid = bytenr;
5978                 key.type = BTRFS_EXTENT_ITEM_KEY;
5979                 key.offset = num_bytes;
5980
5981                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5982                                         -1, 1);
5983                 if (ret) {
5984                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5985                                 ret, bytenr);
5986                         btrfs_print_leaf(extent_root, path->nodes[0]);
5987                 }
5988                 if (ret < 0) {
5989                         btrfs_abort_transaction(trans, extent_root, ret);
5990                         goto out;
5991                 }
5992
5993                 extent_slot = path->slots[0];
5994                 leaf = path->nodes[0];
5995                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5996         }
5997 #endif
5998         BUG_ON(item_size < sizeof(*ei));
5999         ei = btrfs_item_ptr(leaf, extent_slot,
6000                             struct btrfs_extent_item);
6001         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6002             key.type == BTRFS_EXTENT_ITEM_KEY) {
6003                 struct btrfs_tree_block_info *bi;
6004                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6005                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6006                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6007         }
6008
6009         refs = btrfs_extent_refs(leaf, ei);
6010         if (refs < refs_to_drop) {
6011                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6012                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6013                 ret = -EINVAL;
6014                 btrfs_abort_transaction(trans, extent_root, ret);
6015                 goto out;
6016         }
6017         refs -= refs_to_drop;
6018
6019         if (refs > 0) {
6020                 type = BTRFS_QGROUP_OPER_SUB_SHARED;
6021                 if (extent_op)
6022                         __run_delayed_extent_op(extent_op, leaf, ei);
6023                 /*
6024                  * In the case of inline back ref, reference count will
6025                  * be updated by remove_extent_backref
6026                  */
6027                 if (iref) {
6028                         BUG_ON(!found_extent);
6029                 } else {
6030                         btrfs_set_extent_refs(leaf, ei, refs);
6031                         btrfs_mark_buffer_dirty(leaf);
6032                 }
6033                 if (found_extent) {
6034                         ret = remove_extent_backref(trans, extent_root, path,
6035                                                     iref, refs_to_drop,
6036                                                     is_data, &last_ref);
6037                         if (ret) {
6038                                 btrfs_abort_transaction(trans, extent_root, ret);
6039                                 goto out;
6040                         }
6041                 }
6042                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6043                                  root_objectid);
6044         } else {
6045                 if (found_extent) {
6046                         BUG_ON(is_data && refs_to_drop !=
6047                                extent_data_ref_count(root, path, iref));
6048                         if (iref) {
6049                                 BUG_ON(path->slots[0] != extent_slot);
6050                         } else {
6051                                 BUG_ON(path->slots[0] != extent_slot + 1);
6052                                 path->slots[0] = extent_slot;
6053                                 num_to_del = 2;
6054                         }
6055                 }
6056
6057                 last_ref = 1;
6058                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6059                                       num_to_del);
6060                 if (ret) {
6061                         btrfs_abort_transaction(trans, extent_root, ret);
6062                         goto out;
6063                 }
6064                 btrfs_release_path(path);
6065
6066                 if (is_data) {
6067                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6068                         if (ret) {
6069                                 btrfs_abort_transaction(trans, extent_root, ret);
6070                                 goto out;
6071                         }
6072                 }
6073
6074                 ret = update_block_group(root, bytenr, num_bytes, 0);
6075                 if (ret) {
6076                         btrfs_abort_transaction(trans, extent_root, ret);
6077                         goto out;
6078                 }
6079         }
6080         btrfs_release_path(path);
6081
6082         /* Deal with the quota accounting */
6083         if (!ret && last_ref && !no_quota) {
6084                 int mod_seq = 0;
6085
6086                 if (owner_objectid >= BTRFS_FIRST_FREE_OBJECTID &&
6087                     type == BTRFS_QGROUP_OPER_SUB_SHARED)
6088                         mod_seq = 1;
6089
6090                 ret = btrfs_qgroup_record_ref(trans, info, root_objectid,
6091                                               bytenr, num_bytes, type,
6092                                               mod_seq);
6093         }
6094 out:
6095         btrfs_free_path(path);
6096         return ret;
6097 }
6098
6099 /*
6100  * when we free an block, it is possible (and likely) that we free the last
6101  * delayed ref for that extent as well.  This searches the delayed ref tree for
6102  * a given extent, and if there are no other delayed refs to be processed, it
6103  * removes it from the tree.
6104  */
6105 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6106                                       struct btrfs_root *root, u64 bytenr)
6107 {
6108         struct btrfs_delayed_ref_head *head;
6109         struct btrfs_delayed_ref_root *delayed_refs;
6110         int ret = 0;
6111
6112         delayed_refs = &trans->transaction->delayed_refs;
6113         spin_lock(&delayed_refs->lock);
6114         head = btrfs_find_delayed_ref_head(trans, bytenr);
6115         if (!head)
6116                 goto out_delayed_unlock;
6117
6118         spin_lock(&head->lock);
6119         if (rb_first(&head->ref_root))
6120                 goto out;
6121
6122         if (head->extent_op) {
6123                 if (!head->must_insert_reserved)
6124                         goto out;
6125                 btrfs_free_delayed_extent_op(head->extent_op);
6126                 head->extent_op = NULL;
6127         }
6128
6129         /*
6130          * waiting for the lock here would deadlock.  If someone else has it
6131          * locked they are already in the process of dropping it anyway
6132          */
6133         if (!mutex_trylock(&head->mutex))
6134                 goto out;
6135
6136         /*
6137          * at this point we have a head with no other entries.  Go
6138          * ahead and process it.
6139          */
6140         head->node.in_tree = 0;
6141         rb_erase(&head->href_node, &delayed_refs->href_root);
6142
6143         atomic_dec(&delayed_refs->num_entries);
6144
6145         /*
6146          * we don't take a ref on the node because we're removing it from the
6147          * tree, so we just steal the ref the tree was holding.
6148          */
6149         delayed_refs->num_heads--;
6150         if (head->processing == 0)
6151                 delayed_refs->num_heads_ready--;
6152         head->processing = 0;
6153         spin_unlock(&head->lock);
6154         spin_unlock(&delayed_refs->lock);
6155
6156         BUG_ON(head->extent_op);
6157         if (head->must_insert_reserved)
6158                 ret = 1;
6159
6160         mutex_unlock(&head->mutex);
6161         btrfs_put_delayed_ref(&head->node);
6162         return ret;
6163 out:
6164         spin_unlock(&head->lock);
6165
6166 out_delayed_unlock:
6167         spin_unlock(&delayed_refs->lock);
6168         return 0;
6169 }
6170
6171 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6172                            struct btrfs_root *root,
6173                            struct extent_buffer *buf,
6174                            u64 parent, int last_ref)
6175 {
6176         struct btrfs_block_group_cache *cache = NULL;
6177         int pin = 1;
6178         int ret;
6179
6180         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6181                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6182                                         buf->start, buf->len,
6183                                         parent, root->root_key.objectid,
6184                                         btrfs_header_level(buf),
6185                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6186                 BUG_ON(ret); /* -ENOMEM */
6187         }
6188
6189         if (!last_ref)
6190                 return;
6191
6192         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6193
6194         if (btrfs_header_generation(buf) == trans->transid) {
6195                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6196                         ret = check_ref_cleanup(trans, root, buf->start);
6197                         if (!ret)
6198                                 goto out;
6199                 }
6200
6201                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6202                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6203                         goto out;
6204                 }
6205
6206                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6207
6208                 btrfs_add_free_space(cache, buf->start, buf->len);
6209                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
6210                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6211                 pin = 0;
6212         }
6213 out:
6214         if (pin)
6215                 add_pinned_bytes(root->fs_info, buf->len,
6216                                  btrfs_header_level(buf),
6217                                  root->root_key.objectid);
6218
6219         /*
6220          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6221          * anymore.
6222          */
6223         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6224         btrfs_put_block_group(cache);
6225 }
6226
6227 /* Can return -ENOMEM */
6228 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6229                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6230                       u64 owner, u64 offset, int no_quota)
6231 {
6232         int ret;
6233         struct btrfs_fs_info *fs_info = root->fs_info;
6234
6235 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
6236         if (unlikely(test_bit(BTRFS_ROOT_DUMMY_ROOT, &root->state)))
6237                 return 0;
6238 #endif
6239         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6240
6241         /*
6242          * tree log blocks never actually go into the extent allocation
6243          * tree, just update pinning info and exit early.
6244          */
6245         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6246                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6247                 /* unlocks the pinned mutex */
6248                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6249                 ret = 0;
6250         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6251                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6252                                         num_bytes,
6253                                         parent, root_objectid, (int)owner,
6254                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6255         } else {
6256                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6257                                                 num_bytes,
6258                                                 parent, root_objectid, owner,
6259                                                 offset, BTRFS_DROP_DELAYED_REF,
6260                                                 NULL, no_quota);
6261         }
6262         return ret;
6263 }
6264
6265 static u64 stripe_align(struct btrfs_root *root,
6266                         struct btrfs_block_group_cache *cache,
6267                         u64 val, u64 num_bytes)
6268 {
6269         u64 ret = ALIGN(val, root->stripesize);
6270         return ret;
6271 }
6272
6273 /*
6274  * when we wait for progress in the block group caching, its because
6275  * our allocation attempt failed at least once.  So, we must sleep
6276  * and let some progress happen before we try again.
6277  *
6278  * This function will sleep at least once waiting for new free space to
6279  * show up, and then it will check the block group free space numbers
6280  * for our min num_bytes.  Another option is to have it go ahead
6281  * and look in the rbtree for a free extent of a given size, but this
6282  * is a good start.
6283  *
6284  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6285  * any of the information in this block group.
6286  */
6287 static noinline void
6288 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6289                                 u64 num_bytes)
6290 {
6291         struct btrfs_caching_control *caching_ctl;
6292
6293         caching_ctl = get_caching_control(cache);
6294         if (!caching_ctl)
6295                 return;
6296
6297         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6298                    (cache->free_space_ctl->free_space >= num_bytes));
6299
6300         put_caching_control(caching_ctl);
6301 }
6302
6303 static noinline int
6304 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6305 {
6306         struct btrfs_caching_control *caching_ctl;
6307         int ret = 0;
6308
6309         caching_ctl = get_caching_control(cache);
6310         if (!caching_ctl)
6311                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6312
6313         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6314         if (cache->cached == BTRFS_CACHE_ERROR)
6315                 ret = -EIO;
6316         put_caching_control(caching_ctl);
6317         return ret;
6318 }
6319
6320 int __get_raid_index(u64 flags)
6321 {
6322         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6323                 return BTRFS_RAID_RAID10;
6324         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6325                 return BTRFS_RAID_RAID1;
6326         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6327                 return BTRFS_RAID_DUP;
6328         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6329                 return BTRFS_RAID_RAID0;
6330         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6331                 return BTRFS_RAID_RAID5;
6332         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6333                 return BTRFS_RAID_RAID6;
6334
6335         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6336 }
6337
6338 int get_block_group_index(struct btrfs_block_group_cache *cache)
6339 {
6340         return __get_raid_index(cache->flags);
6341 }
6342
6343 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6344         [BTRFS_RAID_RAID10]     = "raid10",
6345         [BTRFS_RAID_RAID1]      = "raid1",
6346         [BTRFS_RAID_DUP]        = "dup",
6347         [BTRFS_RAID_RAID0]      = "raid0",
6348         [BTRFS_RAID_SINGLE]     = "single",
6349         [BTRFS_RAID_RAID5]      = "raid5",
6350         [BTRFS_RAID_RAID6]      = "raid6",
6351 };
6352
6353 static const char *get_raid_name(enum btrfs_raid_types type)
6354 {
6355         if (type >= BTRFS_NR_RAID_TYPES)
6356                 return NULL;
6357
6358         return btrfs_raid_type_names[type];
6359 }
6360
6361 enum btrfs_loop_type {
6362         LOOP_CACHING_NOWAIT = 0,
6363         LOOP_CACHING_WAIT = 1,
6364         LOOP_ALLOC_CHUNK = 2,
6365         LOOP_NO_EMPTY_SIZE = 3,
6366 };
6367
6368 /*
6369  * walks the btree of allocated extents and find a hole of a given size.
6370  * The key ins is changed to record the hole:
6371  * ins->objectid == start position
6372  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6373  * ins->offset == the size of the hole.
6374  * Any available blocks before search_start are skipped.
6375  *
6376  * If there is no suitable free space, we will record the max size of
6377  * the free space extent currently.
6378  */
6379 static noinline int find_free_extent(struct btrfs_root *orig_root,
6380                                      u64 num_bytes, u64 empty_size,
6381                                      u64 hint_byte, struct btrfs_key *ins,
6382                                      u64 flags)
6383 {
6384         int ret = 0;
6385         struct btrfs_root *root = orig_root->fs_info->extent_root;
6386         struct btrfs_free_cluster *last_ptr = NULL;
6387         struct btrfs_block_group_cache *block_group = NULL;
6388         u64 search_start = 0;
6389         u64 max_extent_size = 0;
6390         int empty_cluster = 2 * 1024 * 1024;
6391         struct btrfs_space_info *space_info;
6392         int loop = 0;
6393         int index = __get_raid_index(flags);
6394         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6395                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6396         bool failed_cluster_refill = false;
6397         bool failed_alloc = false;
6398         bool use_cluster = true;
6399         bool have_caching_bg = false;
6400
6401         WARN_ON(num_bytes < root->sectorsize);
6402         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
6403         ins->objectid = 0;
6404         ins->offset = 0;
6405
6406         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6407
6408         space_info = __find_space_info(root->fs_info, flags);
6409         if (!space_info) {
6410                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6411                 return -ENOSPC;
6412         }
6413
6414         /*
6415          * If the space info is for both data and metadata it means we have a
6416          * small filesystem and we can't use the clustering stuff.
6417          */
6418         if (btrfs_mixed_space_info(space_info))
6419                 use_cluster = false;
6420
6421         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6422                 last_ptr = &root->fs_info->meta_alloc_cluster;
6423                 if (!btrfs_test_opt(root, SSD))
6424                         empty_cluster = 64 * 1024;
6425         }
6426
6427         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6428             btrfs_test_opt(root, SSD)) {
6429                 last_ptr = &root->fs_info->data_alloc_cluster;
6430         }
6431
6432         if (last_ptr) {
6433                 spin_lock(&last_ptr->lock);
6434                 if (last_ptr->block_group)
6435                         hint_byte = last_ptr->window_start;
6436                 spin_unlock(&last_ptr->lock);
6437         }
6438
6439         search_start = max(search_start, first_logical_byte(root, 0));
6440         search_start = max(search_start, hint_byte);
6441
6442         if (!last_ptr)
6443                 empty_cluster = 0;
6444
6445         if (search_start == hint_byte) {
6446                 block_group = btrfs_lookup_block_group(root->fs_info,
6447                                                        search_start);
6448                 /*
6449                  * we don't want to use the block group if it doesn't match our
6450                  * allocation bits, or if its not cached.
6451                  *
6452                  * However if we are re-searching with an ideal block group
6453                  * picked out then we don't care that the block group is cached.
6454                  */
6455                 if (block_group && block_group_bits(block_group, flags) &&
6456                     block_group->cached != BTRFS_CACHE_NO) {
6457                         down_read(&space_info->groups_sem);
6458                         if (list_empty(&block_group->list) ||
6459                             block_group->ro) {
6460                                 /*
6461                                  * someone is removing this block group,
6462                                  * we can't jump into the have_block_group
6463                                  * target because our list pointers are not
6464                                  * valid
6465                                  */
6466                                 btrfs_put_block_group(block_group);
6467                                 up_read(&space_info->groups_sem);
6468                         } else {
6469                                 index = get_block_group_index(block_group);
6470                                 goto have_block_group;
6471                         }
6472                 } else if (block_group) {
6473                         btrfs_put_block_group(block_group);
6474                 }
6475         }
6476 search:
6477         have_caching_bg = false;
6478         down_read(&space_info->groups_sem);
6479         list_for_each_entry(block_group, &space_info->block_groups[index],
6480                             list) {
6481                 u64 offset;
6482                 int cached;
6483
6484                 btrfs_get_block_group(block_group);
6485                 search_start = block_group->key.objectid;
6486
6487                 /*
6488                  * this can happen if we end up cycling through all the
6489                  * raid types, but we want to make sure we only allocate
6490                  * for the proper type.
6491                  */
6492                 if (!block_group_bits(block_group, flags)) {
6493                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6494                                 BTRFS_BLOCK_GROUP_RAID1 |
6495                                 BTRFS_BLOCK_GROUP_RAID5 |
6496                                 BTRFS_BLOCK_GROUP_RAID6 |
6497                                 BTRFS_BLOCK_GROUP_RAID10;
6498
6499                         /*
6500                          * if they asked for extra copies and this block group
6501                          * doesn't provide them, bail.  This does allow us to
6502                          * fill raid0 from raid1.
6503                          */
6504                         if ((flags & extra) && !(block_group->flags & extra))
6505                                 goto loop;
6506                 }
6507
6508 have_block_group:
6509                 cached = block_group_cache_done(block_group);
6510                 if (unlikely(!cached)) {
6511                         ret = cache_block_group(block_group, 0);
6512                         BUG_ON(ret < 0);
6513                         ret = 0;
6514                 }
6515
6516                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6517                         goto loop;
6518                 if (unlikely(block_group->ro))
6519                         goto loop;
6520
6521                 /*
6522                  * Ok we want to try and use the cluster allocator, so
6523                  * lets look there
6524                  */
6525                 if (last_ptr) {
6526                         struct btrfs_block_group_cache *used_block_group;
6527                         unsigned long aligned_cluster;
6528                         /*
6529                          * the refill lock keeps out other
6530                          * people trying to start a new cluster
6531                          */
6532                         spin_lock(&last_ptr->refill_lock);
6533                         used_block_group = last_ptr->block_group;
6534                         if (used_block_group != block_group &&
6535                             (!used_block_group ||
6536                              used_block_group->ro ||
6537                              !block_group_bits(used_block_group, flags)))
6538                                 goto refill_cluster;
6539
6540                         if (used_block_group != block_group)
6541                                 btrfs_get_block_group(used_block_group);
6542
6543                         offset = btrfs_alloc_from_cluster(used_block_group,
6544                                                 last_ptr,
6545                                                 num_bytes,
6546                                                 used_block_group->key.objectid,
6547                                                 &max_extent_size);
6548                         if (offset) {
6549                                 /* we have a block, we're done */
6550                                 spin_unlock(&last_ptr->refill_lock);
6551                                 trace_btrfs_reserve_extent_cluster(root,
6552                                                 used_block_group,
6553                                                 search_start, num_bytes);
6554                                 if (used_block_group != block_group) {
6555                                         btrfs_put_block_group(block_group);
6556                                         block_group = used_block_group;
6557                                 }
6558                                 goto checks;
6559                         }
6560
6561                         WARN_ON(last_ptr->block_group != used_block_group);
6562                         if (used_block_group != block_group)
6563                                 btrfs_put_block_group(used_block_group);
6564 refill_cluster:
6565                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6566                          * set up a new clusters, so lets just skip it
6567                          * and let the allocator find whatever block
6568                          * it can find.  If we reach this point, we
6569                          * will have tried the cluster allocator
6570                          * plenty of times and not have found
6571                          * anything, so we are likely way too
6572                          * fragmented for the clustering stuff to find
6573                          * anything.
6574                          *
6575                          * However, if the cluster is taken from the
6576                          * current block group, release the cluster
6577                          * first, so that we stand a better chance of
6578                          * succeeding in the unclustered
6579                          * allocation.  */
6580                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6581                             last_ptr->block_group != block_group) {
6582                                 spin_unlock(&last_ptr->refill_lock);
6583                                 goto unclustered_alloc;
6584                         }
6585
6586                         /*
6587                          * this cluster didn't work out, free it and
6588                          * start over
6589                          */
6590                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6591
6592                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6593                                 spin_unlock(&last_ptr->refill_lock);
6594                                 goto unclustered_alloc;
6595                         }
6596
6597                         aligned_cluster = max_t(unsigned long,
6598                                                 empty_cluster + empty_size,
6599                                               block_group->full_stripe_len);
6600
6601                         /* allocate a cluster in this block group */
6602                         ret = btrfs_find_space_cluster(root, block_group,
6603                                                        last_ptr, search_start,
6604                                                        num_bytes,
6605                                                        aligned_cluster);
6606                         if (ret == 0) {
6607                                 /*
6608                                  * now pull our allocation out of this
6609                                  * cluster
6610                                  */
6611                                 offset = btrfs_alloc_from_cluster(block_group,
6612                                                         last_ptr,
6613                                                         num_bytes,
6614                                                         search_start,
6615                                                         &max_extent_size);
6616                                 if (offset) {
6617                                         /* we found one, proceed */
6618                                         spin_unlock(&last_ptr->refill_lock);
6619                                         trace_btrfs_reserve_extent_cluster(root,
6620                                                 block_group, search_start,
6621                                                 num_bytes);
6622                                         goto checks;
6623                                 }
6624                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6625                                    && !failed_cluster_refill) {
6626                                 spin_unlock(&last_ptr->refill_lock);
6627
6628                                 failed_cluster_refill = true;
6629                                 wait_block_group_cache_progress(block_group,
6630                                        num_bytes + empty_cluster + empty_size);
6631                                 goto have_block_group;
6632                         }
6633
6634                         /*
6635                          * at this point we either didn't find a cluster
6636                          * or we weren't able to allocate a block from our
6637                          * cluster.  Free the cluster we've been trying
6638                          * to use, and go to the next block group
6639                          */
6640                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6641                         spin_unlock(&last_ptr->refill_lock);
6642                         goto loop;
6643                 }
6644
6645 unclustered_alloc:
6646                 spin_lock(&block_group->free_space_ctl->tree_lock);
6647                 if (cached &&
6648                     block_group->free_space_ctl->free_space <
6649                     num_bytes + empty_cluster + empty_size) {
6650                         if (block_group->free_space_ctl->free_space >
6651                             max_extent_size)
6652                                 max_extent_size =
6653                                         block_group->free_space_ctl->free_space;
6654                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6655                         goto loop;
6656                 }
6657                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6658
6659                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6660                                                     num_bytes, empty_size,
6661                                                     &max_extent_size);
6662                 /*
6663                  * If we didn't find a chunk, and we haven't failed on this
6664                  * block group before, and this block group is in the middle of
6665                  * caching and we are ok with waiting, then go ahead and wait
6666                  * for progress to be made, and set failed_alloc to true.
6667                  *
6668                  * If failed_alloc is true then we've already waited on this
6669                  * block group once and should move on to the next block group.
6670                  */
6671                 if (!offset && !failed_alloc && !cached &&
6672                     loop > LOOP_CACHING_NOWAIT) {
6673                         wait_block_group_cache_progress(block_group,
6674                                                 num_bytes + empty_size);
6675                         failed_alloc = true;
6676                         goto have_block_group;
6677                 } else if (!offset) {
6678                         if (!cached)
6679                                 have_caching_bg = true;
6680                         goto loop;
6681                 }
6682 checks:
6683                 search_start = stripe_align(root, block_group,
6684                                             offset, num_bytes);
6685
6686                 /* move on to the next group */
6687                 if (search_start + num_bytes >
6688                     block_group->key.objectid + block_group->key.offset) {
6689                         btrfs_add_free_space(block_group, offset, num_bytes);
6690                         goto loop;
6691                 }
6692
6693                 if (offset < search_start)
6694                         btrfs_add_free_space(block_group, offset,
6695                                              search_start - offset);
6696                 BUG_ON(offset > search_start);
6697
6698                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
6699                                                   alloc_type);
6700                 if (ret == -EAGAIN) {
6701                         btrfs_add_free_space(block_group, offset, num_bytes);
6702                         goto loop;
6703                 }
6704
6705                 /* we are all good, lets return */
6706                 ins->objectid = search_start;
6707                 ins->offset = num_bytes;
6708
6709                 trace_btrfs_reserve_extent(orig_root, block_group,
6710                                            search_start, num_bytes);
6711                 btrfs_put_block_group(block_group);
6712                 break;
6713 loop:
6714                 failed_cluster_refill = false;
6715                 failed_alloc = false;
6716                 BUG_ON(index != get_block_group_index(block_group));
6717                 btrfs_put_block_group(block_group);
6718         }
6719         up_read(&space_info->groups_sem);
6720
6721         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6722                 goto search;
6723
6724         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6725                 goto search;
6726
6727         /*
6728          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6729          *                      caching kthreads as we move along
6730          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6731          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6732          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6733          *                      again
6734          */
6735         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6736                 index = 0;
6737                 loop++;
6738                 if (loop == LOOP_ALLOC_CHUNK) {
6739                         struct btrfs_trans_handle *trans;
6740                         int exist = 0;
6741
6742                         trans = current->journal_info;
6743                         if (trans)
6744                                 exist = 1;
6745                         else
6746                                 trans = btrfs_join_transaction(root);
6747
6748                         if (IS_ERR(trans)) {
6749                                 ret = PTR_ERR(trans);
6750                                 goto out;
6751                         }
6752
6753                         ret = do_chunk_alloc(trans, root, flags,
6754                                              CHUNK_ALLOC_FORCE);
6755                         /*
6756                          * Do not bail out on ENOSPC since we
6757                          * can do more things.
6758                          */
6759                         if (ret < 0 && ret != -ENOSPC)
6760                                 btrfs_abort_transaction(trans,
6761                                                         root, ret);
6762                         else
6763                                 ret = 0;
6764                         if (!exist)
6765                                 btrfs_end_transaction(trans, root);
6766                         if (ret)
6767                                 goto out;
6768                 }
6769
6770                 if (loop == LOOP_NO_EMPTY_SIZE) {
6771                         empty_size = 0;
6772                         empty_cluster = 0;
6773                 }
6774
6775                 goto search;
6776         } else if (!ins->objectid) {
6777                 ret = -ENOSPC;
6778         } else if (ins->objectid) {
6779                 ret = 0;
6780         }
6781 out:
6782         if (ret == -ENOSPC)
6783                 ins->offset = max_extent_size;
6784         return ret;
6785 }
6786
6787 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6788                             int dump_block_groups)
6789 {
6790         struct btrfs_block_group_cache *cache;
6791         int index = 0;
6792
6793         spin_lock(&info->lock);
6794         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
6795                info->flags,
6796                info->total_bytes - info->bytes_used - info->bytes_pinned -
6797                info->bytes_reserved - info->bytes_readonly,
6798                (info->full) ? "" : "not ");
6799         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
6800                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6801                info->total_bytes, info->bytes_used, info->bytes_pinned,
6802                info->bytes_reserved, info->bytes_may_use,
6803                info->bytes_readonly);
6804         spin_unlock(&info->lock);
6805
6806         if (!dump_block_groups)
6807                 return;
6808
6809         down_read(&info->groups_sem);
6810 again:
6811         list_for_each_entry(cache, &info->block_groups[index], list) {
6812                 spin_lock(&cache->lock);
6813                 printk(KERN_INFO "BTRFS: "
6814                            "block group %llu has %llu bytes, "
6815                            "%llu used %llu pinned %llu reserved %s\n",
6816                        cache->key.objectid, cache->key.offset,
6817                        btrfs_block_group_used(&cache->item), cache->pinned,
6818                        cache->reserved, cache->ro ? "[readonly]" : "");
6819                 btrfs_dump_free_space(cache, bytes);
6820                 spin_unlock(&cache->lock);
6821         }
6822         if (++index < BTRFS_NR_RAID_TYPES)
6823                 goto again;
6824         up_read(&info->groups_sem);
6825 }
6826
6827 int btrfs_reserve_extent(struct btrfs_root *root,
6828                          u64 num_bytes, u64 min_alloc_size,
6829                          u64 empty_size, u64 hint_byte,
6830                          struct btrfs_key *ins, int is_data)
6831 {
6832         bool final_tried = false;
6833         u64 flags;
6834         int ret;
6835
6836         flags = btrfs_get_alloc_profile(root, is_data);
6837 again:
6838         WARN_ON(num_bytes < root->sectorsize);
6839         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
6840                                flags);
6841
6842         if (ret == -ENOSPC) {
6843                 if (!final_tried && ins->offset) {
6844                         num_bytes = min(num_bytes >> 1, ins->offset);
6845                         num_bytes = round_down(num_bytes, root->sectorsize);
6846                         num_bytes = max(num_bytes, min_alloc_size);
6847                         if (num_bytes == min_alloc_size)
6848                                 final_tried = true;
6849                         goto again;
6850                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6851                         struct btrfs_space_info *sinfo;
6852
6853                         sinfo = __find_space_info(root->fs_info, flags);
6854                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6855                                 flags, num_bytes);
6856                         if (sinfo)
6857                                 dump_space_info(sinfo, num_bytes, 1);
6858                 }
6859         }
6860
6861         return ret;
6862 }
6863
6864 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6865                                         u64 start, u64 len, int pin)
6866 {
6867         struct btrfs_block_group_cache *cache;
6868         int ret = 0;
6869
6870         cache = btrfs_lookup_block_group(root->fs_info, start);
6871         if (!cache) {
6872                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6873                         start);
6874                 return -ENOSPC;
6875         }
6876
6877         if (btrfs_test_opt(root, DISCARD))
6878                 ret = btrfs_discard_extent(root, start, len, NULL);
6879
6880         if (pin)
6881                 pin_down_extent(root, cache, start, len, 1);
6882         else {
6883                 btrfs_add_free_space(cache, start, len);
6884                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6885         }
6886         btrfs_put_block_group(cache);
6887
6888         trace_btrfs_reserved_extent_free(root, start, len);
6889
6890         return ret;
6891 }
6892
6893 int btrfs_free_reserved_extent(struct btrfs_root *root,
6894                                         u64 start, u64 len)
6895 {
6896         return __btrfs_free_reserved_extent(root, start, len, 0);
6897 }
6898
6899 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6900                                        u64 start, u64 len)
6901 {
6902         return __btrfs_free_reserved_extent(root, start, len, 1);
6903 }
6904
6905 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6906                                       struct btrfs_root *root,
6907                                       u64 parent, u64 root_objectid,
6908                                       u64 flags, u64 owner, u64 offset,
6909                                       struct btrfs_key *ins, int ref_mod)
6910 {
6911         int ret;
6912         struct btrfs_fs_info *fs_info = root->fs_info;
6913         struct btrfs_extent_item *extent_item;
6914         struct btrfs_extent_inline_ref *iref;
6915         struct btrfs_path *path;
6916         struct extent_buffer *leaf;
6917         int type;
6918         u32 size;
6919
6920         if (parent > 0)
6921                 type = BTRFS_SHARED_DATA_REF_KEY;
6922         else
6923                 type = BTRFS_EXTENT_DATA_REF_KEY;
6924
6925         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6926
6927         path = btrfs_alloc_path();
6928         if (!path)
6929                 return -ENOMEM;
6930
6931         path->leave_spinning = 1;
6932         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6933                                       ins, size);
6934         if (ret) {
6935                 btrfs_free_path(path);
6936                 return ret;
6937         }
6938
6939         leaf = path->nodes[0];
6940         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6941                                      struct btrfs_extent_item);
6942         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6943         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6944         btrfs_set_extent_flags(leaf, extent_item,
6945                                flags | BTRFS_EXTENT_FLAG_DATA);
6946
6947         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6948         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6949         if (parent > 0) {
6950                 struct btrfs_shared_data_ref *ref;
6951                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6952                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6953                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6954         } else {
6955                 struct btrfs_extent_data_ref *ref;
6956                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6957                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6958                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6959                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6960                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6961         }
6962
6963         btrfs_mark_buffer_dirty(path->nodes[0]);
6964         btrfs_free_path(path);
6965
6966         /* Always set parent to 0 here since its exclusive anyway. */
6967         ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
6968                                       ins->objectid, ins->offset,
6969                                       BTRFS_QGROUP_OPER_ADD_EXCL, 0);
6970         if (ret)
6971                 return ret;
6972
6973         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6974         if (ret) { /* -ENOENT, logic error */
6975                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6976                         ins->objectid, ins->offset);
6977                 BUG();
6978         }
6979         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6980         return ret;
6981 }
6982
6983 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6984                                      struct btrfs_root *root,
6985                                      u64 parent, u64 root_objectid,
6986                                      u64 flags, struct btrfs_disk_key *key,
6987                                      int level, struct btrfs_key *ins,
6988                                      int no_quota)
6989 {
6990         int ret;
6991         struct btrfs_fs_info *fs_info = root->fs_info;
6992         struct btrfs_extent_item *extent_item;
6993         struct btrfs_tree_block_info *block_info;
6994         struct btrfs_extent_inline_ref *iref;
6995         struct btrfs_path *path;
6996         struct extent_buffer *leaf;
6997         u32 size = sizeof(*extent_item) + sizeof(*iref);
6998         u64 num_bytes = ins->offset;
6999         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7000                                                  SKINNY_METADATA);
7001
7002         if (!skinny_metadata)
7003                 size += sizeof(*block_info);
7004
7005         path = btrfs_alloc_path();
7006         if (!path) {
7007                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7008                                                    root->leafsize);
7009                 return -ENOMEM;
7010         }
7011
7012         path->leave_spinning = 1;
7013         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7014                                       ins, size);
7015         if (ret) {
7016                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7017                                                    root->leafsize);
7018                 btrfs_free_path(path);
7019                 return ret;
7020         }
7021
7022         leaf = path->nodes[0];
7023         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7024                                      struct btrfs_extent_item);
7025         btrfs_set_extent_refs(leaf, extent_item, 1);
7026         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7027         btrfs_set_extent_flags(leaf, extent_item,
7028                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7029
7030         if (skinny_metadata) {
7031                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7032                 num_bytes = root->leafsize;
7033         } else {
7034                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7035                 btrfs_set_tree_block_key(leaf, block_info, key);
7036                 btrfs_set_tree_block_level(leaf, block_info, level);
7037                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7038         }
7039
7040         if (parent > 0) {
7041                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7042                 btrfs_set_extent_inline_ref_type(leaf, iref,
7043                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7044                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7045         } else {
7046                 btrfs_set_extent_inline_ref_type(leaf, iref,
7047                                                  BTRFS_TREE_BLOCK_REF_KEY);
7048                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7049         }
7050
7051         btrfs_mark_buffer_dirty(leaf);
7052         btrfs_free_path(path);
7053
7054         if (!no_quota) {
7055                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7056                                               ins->objectid, num_bytes,
7057                                               BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7058                 if (ret)
7059                         return ret;
7060         }
7061
7062         ret = update_block_group(root, ins->objectid, root->leafsize, 1);
7063         if (ret) { /* -ENOENT, logic error */
7064                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7065                         ins->objectid, ins->offset);
7066                 BUG();
7067         }
7068
7069         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->leafsize);
7070         return ret;
7071 }
7072
7073 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7074                                      struct btrfs_root *root,
7075                                      u64 root_objectid, u64 owner,
7076                                      u64 offset, struct btrfs_key *ins)
7077 {
7078         int ret;
7079
7080         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7081
7082         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7083                                          ins->offset, 0,
7084                                          root_objectid, owner, offset,
7085                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7086         return ret;
7087 }
7088
7089 /*
7090  * this is used by the tree logging recovery code.  It records that
7091  * an extent has been allocated and makes sure to clear the free
7092  * space cache bits as well
7093  */
7094 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7095                                    struct btrfs_root *root,
7096                                    u64 root_objectid, u64 owner, u64 offset,
7097                                    struct btrfs_key *ins)
7098 {
7099         int ret;
7100         struct btrfs_block_group_cache *block_group;
7101
7102         /*
7103          * Mixed block groups will exclude before processing the log so we only
7104          * need to do the exlude dance if this fs isn't mixed.
7105          */
7106         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7107                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7108                 if (ret)
7109                         return ret;
7110         }
7111
7112         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7113         if (!block_group)
7114                 return -EINVAL;
7115
7116         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7117                                           RESERVE_ALLOC_NO_ACCOUNT);
7118         BUG_ON(ret); /* logic error */
7119         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7120                                          0, owner, offset, ins, 1);
7121         btrfs_put_block_group(block_group);
7122         return ret;
7123 }
7124
7125 static struct extent_buffer *
7126 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7127                       u64 bytenr, u32 blocksize, int level)
7128 {
7129         struct extent_buffer *buf;
7130
7131         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
7132         if (!buf)
7133                 return ERR_PTR(-ENOMEM);
7134         btrfs_set_header_generation(buf, trans->transid);
7135         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7136         btrfs_tree_lock(buf);
7137         clean_tree_block(trans, root, buf);
7138         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7139
7140         btrfs_set_lock_blocking(buf);
7141         btrfs_set_buffer_uptodate(buf);
7142
7143         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7144                 /*
7145                  * we allow two log transactions at a time, use different
7146                  * EXENT bit to differentiate dirty pages.
7147                  */
7148                 if (root->log_transid % 2 == 0)
7149                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7150                                         buf->start + buf->len - 1, GFP_NOFS);
7151                 else
7152                         set_extent_new(&root->dirty_log_pages, buf->start,
7153                                         buf->start + buf->len - 1, GFP_NOFS);
7154         } else {
7155                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7156                          buf->start + buf->len - 1, GFP_NOFS);
7157         }
7158         trans->blocks_used++;
7159         /* this returns a buffer locked for blocking */
7160         return buf;
7161 }
7162
7163 static struct btrfs_block_rsv *
7164 use_block_rsv(struct btrfs_trans_handle *trans,
7165               struct btrfs_root *root, u32 blocksize)
7166 {
7167         struct btrfs_block_rsv *block_rsv;
7168         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7169         int ret;
7170         bool global_updated = false;
7171
7172         block_rsv = get_block_rsv(trans, root);
7173
7174         if (unlikely(block_rsv->size == 0))
7175                 goto try_reserve;
7176 again:
7177         ret = block_rsv_use_bytes(block_rsv, blocksize);
7178         if (!ret)
7179                 return block_rsv;
7180
7181         if (block_rsv->failfast)
7182                 return ERR_PTR(ret);
7183
7184         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7185                 global_updated = true;
7186                 update_global_block_rsv(root->fs_info);
7187                 goto again;
7188         }
7189
7190         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7191                 static DEFINE_RATELIMIT_STATE(_rs,
7192                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7193                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7194                 if (__ratelimit(&_rs))
7195                         WARN(1, KERN_DEBUG
7196                                 "BTRFS: block rsv returned %d\n", ret);
7197         }
7198 try_reserve:
7199         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7200                                      BTRFS_RESERVE_NO_FLUSH);
7201         if (!ret)
7202                 return block_rsv;
7203         /*
7204          * If we couldn't reserve metadata bytes try and use some from
7205          * the global reserve if its space type is the same as the global
7206          * reservation.
7207          */
7208         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7209             block_rsv->space_info == global_rsv->space_info) {
7210                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7211                 if (!ret)
7212                         return global_rsv;
7213         }
7214         return ERR_PTR(ret);
7215 }
7216
7217 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7218                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7219 {
7220         block_rsv_add_bytes(block_rsv, blocksize, 0);
7221         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7222 }
7223
7224 /*
7225  * finds a free extent and does all the dirty work required for allocation
7226  * returns the key for the extent through ins, and a tree buffer for
7227  * the first block of the extent through buf.
7228  *
7229  * returns the tree buffer or NULL.
7230  */
7231 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
7232                                         struct btrfs_root *root, u32 blocksize,
7233                                         u64 parent, u64 root_objectid,
7234                                         struct btrfs_disk_key *key, int level,
7235                                         u64 hint, u64 empty_size)
7236 {
7237         struct btrfs_key ins;
7238         struct btrfs_block_rsv *block_rsv;
7239         struct extent_buffer *buf;
7240         u64 flags = 0;
7241         int ret;
7242         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7243                                                  SKINNY_METADATA);
7244
7245 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
7246         if (unlikely(test_bit(BTRFS_ROOT_DUMMY_ROOT, &root->state))) {
7247                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7248                                             blocksize, level);
7249                 if (!IS_ERR(buf))
7250                         root->alloc_bytenr += blocksize;
7251                 return buf;
7252         }
7253 #endif
7254         block_rsv = use_block_rsv(trans, root, blocksize);
7255         if (IS_ERR(block_rsv))
7256                 return ERR_CAST(block_rsv);
7257
7258         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7259                                    empty_size, hint, &ins, 0);
7260         if (ret) {
7261                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7262                 return ERR_PTR(ret);
7263         }
7264
7265         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
7266                                     blocksize, level);
7267         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
7268
7269         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7270                 if (parent == 0)
7271                         parent = ins.objectid;
7272                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7273         } else
7274                 BUG_ON(parent > 0);
7275
7276         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7277                 struct btrfs_delayed_extent_op *extent_op;
7278                 extent_op = btrfs_alloc_delayed_extent_op();
7279                 BUG_ON(!extent_op); /* -ENOMEM */
7280                 if (key)
7281                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7282                 else
7283                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7284                 extent_op->flags_to_set = flags;
7285                 if (skinny_metadata)
7286                         extent_op->update_key = 0;
7287                 else
7288                         extent_op->update_key = 1;
7289                 extent_op->update_flags = 1;
7290                 extent_op->is_data = 0;
7291                 extent_op->level = level;
7292
7293                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7294                                         ins.objectid,
7295                                         ins.offset, parent, root_objectid,
7296                                         level, BTRFS_ADD_DELAYED_EXTENT,
7297                                         extent_op, 0);
7298                 BUG_ON(ret); /* -ENOMEM */
7299         }
7300         return buf;
7301 }
7302
7303 struct walk_control {
7304         u64 refs[BTRFS_MAX_LEVEL];
7305         u64 flags[BTRFS_MAX_LEVEL];
7306         struct btrfs_key update_progress;
7307         int stage;
7308         int level;
7309         int shared_level;
7310         int update_ref;
7311         int keep_locks;
7312         int reada_slot;
7313         int reada_count;
7314         int for_reloc;
7315 };
7316
7317 #define DROP_REFERENCE  1
7318 #define UPDATE_BACKREF  2
7319
7320 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7321                                      struct btrfs_root *root,
7322                                      struct walk_control *wc,
7323                                      struct btrfs_path *path)
7324 {
7325         u64 bytenr;
7326         u64 generation;
7327         u64 refs;
7328         u64 flags;
7329         u32 nritems;
7330         u32 blocksize;
7331         struct btrfs_key key;
7332         struct extent_buffer *eb;
7333         int ret;
7334         int slot;
7335         int nread = 0;
7336
7337         if (path->slots[wc->level] < wc->reada_slot) {
7338                 wc->reada_count = wc->reada_count * 2 / 3;
7339                 wc->reada_count = max(wc->reada_count, 2);
7340         } else {
7341                 wc->reada_count = wc->reada_count * 3 / 2;
7342                 wc->reada_count = min_t(int, wc->reada_count,
7343                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7344         }
7345
7346         eb = path->nodes[wc->level];
7347         nritems = btrfs_header_nritems(eb);
7348         blocksize = btrfs_level_size(root, wc->level - 1);
7349
7350         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7351                 if (nread >= wc->reada_count)
7352                         break;
7353
7354                 cond_resched();
7355                 bytenr = btrfs_node_blockptr(eb, slot);
7356                 generation = btrfs_node_ptr_generation(eb, slot);
7357
7358                 if (slot == path->slots[wc->level])
7359                         goto reada;
7360
7361                 if (wc->stage == UPDATE_BACKREF &&
7362                     generation <= root->root_key.offset)
7363                         continue;
7364
7365                 /* We don't lock the tree block, it's OK to be racy here */
7366                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7367                                                wc->level - 1, 1, &refs,
7368                                                &flags);
7369                 /* We don't care about errors in readahead. */
7370                 if (ret < 0)
7371                         continue;
7372                 BUG_ON(refs == 0);
7373
7374                 if (wc->stage == DROP_REFERENCE) {
7375                         if (refs == 1)
7376                                 goto reada;
7377
7378                         if (wc->level == 1 &&
7379                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7380                                 continue;
7381                         if (!wc->update_ref ||
7382                             generation <= root->root_key.offset)
7383                                 continue;
7384                         btrfs_node_key_to_cpu(eb, &key, slot);
7385                         ret = btrfs_comp_cpu_keys(&key,
7386                                                   &wc->update_progress);
7387                         if (ret < 0)
7388                                 continue;
7389                 } else {
7390                         if (wc->level == 1 &&
7391                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7392                                 continue;
7393                 }
7394 reada:
7395                 ret = readahead_tree_block(root, bytenr, blocksize,
7396                                            generation);
7397                 if (ret)
7398                         break;
7399                 nread++;
7400         }
7401         wc->reada_slot = slot;
7402 }
7403
7404 /*
7405  * helper to process tree block while walking down the tree.
7406  *
7407  * when wc->stage == UPDATE_BACKREF, this function updates
7408  * back refs for pointers in the block.
7409  *
7410  * NOTE: return value 1 means we should stop walking down.
7411  */
7412 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7413                                    struct btrfs_root *root,
7414                                    struct btrfs_path *path,
7415                                    struct walk_control *wc, int lookup_info)
7416 {
7417         int level = wc->level;
7418         struct extent_buffer *eb = path->nodes[level];
7419         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7420         int ret;
7421
7422         if (wc->stage == UPDATE_BACKREF &&
7423             btrfs_header_owner(eb) != root->root_key.objectid)
7424                 return 1;
7425
7426         /*
7427          * when reference count of tree block is 1, it won't increase
7428          * again. once full backref flag is set, we never clear it.
7429          */
7430         if (lookup_info &&
7431             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7432              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7433                 BUG_ON(!path->locks[level]);
7434                 ret = btrfs_lookup_extent_info(trans, root,
7435                                                eb->start, level, 1,
7436                                                &wc->refs[level],
7437                                                &wc->flags[level]);
7438                 BUG_ON(ret == -ENOMEM);
7439                 if (ret)
7440                         return ret;
7441                 BUG_ON(wc->refs[level] == 0);
7442         }
7443
7444         if (wc->stage == DROP_REFERENCE) {
7445                 if (wc->refs[level] > 1)
7446                         return 1;
7447
7448                 if (path->locks[level] && !wc->keep_locks) {
7449                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7450                         path->locks[level] = 0;
7451                 }
7452                 return 0;
7453         }
7454
7455         /* wc->stage == UPDATE_BACKREF */
7456         if (!(wc->flags[level] & flag)) {
7457                 BUG_ON(!path->locks[level]);
7458                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
7459                 BUG_ON(ret); /* -ENOMEM */
7460                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
7461                 BUG_ON(ret); /* -ENOMEM */
7462                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7463                                                   eb->len, flag,
7464                                                   btrfs_header_level(eb), 0);
7465                 BUG_ON(ret); /* -ENOMEM */
7466                 wc->flags[level] |= flag;
7467         }
7468
7469         /*
7470          * the block is shared by multiple trees, so it's not good to
7471          * keep the tree lock
7472          */
7473         if (path->locks[level] && level > 0) {
7474                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7475                 path->locks[level] = 0;
7476         }
7477         return 0;
7478 }
7479
7480 /*
7481  * helper to process tree block pointer.
7482  *
7483  * when wc->stage == DROP_REFERENCE, this function checks
7484  * reference count of the block pointed to. if the block
7485  * is shared and we need update back refs for the subtree
7486  * rooted at the block, this function changes wc->stage to
7487  * UPDATE_BACKREF. if the block is shared and there is no
7488  * need to update back, this function drops the reference
7489  * to the block.
7490  *
7491  * NOTE: return value 1 means we should stop walking down.
7492  */
7493 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7494                                  struct btrfs_root *root,
7495                                  struct btrfs_path *path,
7496                                  struct walk_control *wc, int *lookup_info)
7497 {
7498         u64 bytenr;
7499         u64 generation;
7500         u64 parent;
7501         u32 blocksize;
7502         struct btrfs_key key;
7503         struct extent_buffer *next;
7504         int level = wc->level;
7505         int reada = 0;
7506         int ret = 0;
7507
7508         generation = btrfs_node_ptr_generation(path->nodes[level],
7509                                                path->slots[level]);
7510         /*
7511          * if the lower level block was created before the snapshot
7512          * was created, we know there is no need to update back refs
7513          * for the subtree
7514          */
7515         if (wc->stage == UPDATE_BACKREF &&
7516             generation <= root->root_key.offset) {
7517                 *lookup_info = 1;
7518                 return 1;
7519         }
7520
7521         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7522         blocksize = btrfs_level_size(root, level - 1);
7523
7524         next = btrfs_find_tree_block(root, bytenr, blocksize);
7525         if (!next) {
7526                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7527                 if (!next)
7528                         return -ENOMEM;
7529                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
7530                                                level - 1);
7531                 reada = 1;
7532         }
7533         btrfs_tree_lock(next);
7534         btrfs_set_lock_blocking(next);
7535
7536         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7537                                        &wc->refs[level - 1],
7538                                        &wc->flags[level - 1]);
7539         if (ret < 0) {
7540                 btrfs_tree_unlock(next);
7541                 return ret;
7542         }
7543
7544         if (unlikely(wc->refs[level - 1] == 0)) {
7545                 btrfs_err(root->fs_info, "Missing references.");
7546                 BUG();
7547         }
7548         *lookup_info = 0;
7549
7550         if (wc->stage == DROP_REFERENCE) {
7551                 if (wc->refs[level - 1] > 1) {
7552                         if (level == 1 &&
7553                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7554                                 goto skip;
7555
7556                         if (!wc->update_ref ||
7557                             generation <= root->root_key.offset)
7558                                 goto skip;
7559
7560                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7561                                               path->slots[level]);
7562                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7563                         if (ret < 0)
7564                                 goto skip;
7565
7566                         wc->stage = UPDATE_BACKREF;
7567                         wc->shared_level = level - 1;
7568                 }
7569         } else {
7570                 if (level == 1 &&
7571                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7572                         goto skip;
7573         }
7574
7575         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7576                 btrfs_tree_unlock(next);
7577                 free_extent_buffer(next);
7578                 next = NULL;
7579                 *lookup_info = 1;
7580         }
7581
7582         if (!next) {
7583                 if (reada && level == 1)
7584                         reada_walk_down(trans, root, wc, path);
7585                 next = read_tree_block(root, bytenr, blocksize, generation);
7586                 if (!next || !extent_buffer_uptodate(next)) {
7587                         free_extent_buffer(next);
7588                         return -EIO;
7589                 }
7590                 btrfs_tree_lock(next);
7591                 btrfs_set_lock_blocking(next);
7592         }
7593
7594         level--;
7595         BUG_ON(level != btrfs_header_level(next));
7596         path->nodes[level] = next;
7597         path->slots[level] = 0;
7598         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7599         wc->level = level;
7600         if (wc->level == 1)
7601                 wc->reada_slot = 0;
7602         return 0;
7603 skip:
7604         wc->refs[level - 1] = 0;
7605         wc->flags[level - 1] = 0;
7606         if (wc->stage == DROP_REFERENCE) {
7607                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7608                         parent = path->nodes[level]->start;
7609                 } else {
7610                         BUG_ON(root->root_key.objectid !=
7611                                btrfs_header_owner(path->nodes[level]));
7612                         parent = 0;
7613                 }
7614
7615                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7616                                 root->root_key.objectid, level - 1, 0, 0);
7617                 BUG_ON(ret); /* -ENOMEM */
7618         }
7619         btrfs_tree_unlock(next);
7620         free_extent_buffer(next);
7621         *lookup_info = 1;
7622         return 1;
7623 }
7624
7625 /*
7626  * helper to process tree block while walking up the tree.
7627  *
7628  * when wc->stage == DROP_REFERENCE, this function drops
7629  * reference count on the block.
7630  *
7631  * when wc->stage == UPDATE_BACKREF, this function changes
7632  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7633  * to UPDATE_BACKREF previously while processing the block.
7634  *
7635  * NOTE: return value 1 means we should stop walking up.
7636  */
7637 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7638                                  struct btrfs_root *root,
7639                                  struct btrfs_path *path,
7640                                  struct walk_control *wc)
7641 {
7642         int ret;
7643         int level = wc->level;
7644         struct extent_buffer *eb = path->nodes[level];
7645         u64 parent = 0;
7646
7647         if (wc->stage == UPDATE_BACKREF) {
7648                 BUG_ON(wc->shared_level < level);
7649                 if (level < wc->shared_level)
7650                         goto out;
7651
7652                 ret = find_next_key(path, level + 1, &wc->update_progress);
7653                 if (ret > 0)
7654                         wc->update_ref = 0;
7655
7656                 wc->stage = DROP_REFERENCE;
7657                 wc->shared_level = -1;
7658                 path->slots[level] = 0;
7659
7660                 /*
7661                  * check reference count again if the block isn't locked.
7662                  * we should start walking down the tree again if reference
7663                  * count is one.
7664                  */
7665                 if (!path->locks[level]) {
7666                         BUG_ON(level == 0);
7667                         btrfs_tree_lock(eb);
7668                         btrfs_set_lock_blocking(eb);
7669                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7670
7671                         ret = btrfs_lookup_extent_info(trans, root,
7672                                                        eb->start, level, 1,
7673                                                        &wc->refs[level],
7674                                                        &wc->flags[level]);
7675                         if (ret < 0) {
7676                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7677                                 path->locks[level] = 0;
7678                                 return ret;
7679                         }
7680                         BUG_ON(wc->refs[level] == 0);
7681                         if (wc->refs[level] == 1) {
7682                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7683                                 path->locks[level] = 0;
7684                                 return 1;
7685                         }
7686                 }
7687         }
7688
7689         /* wc->stage == DROP_REFERENCE */
7690         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7691
7692         if (wc->refs[level] == 1) {
7693                 if (level == 0) {
7694                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7695                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7696                                                     wc->for_reloc);
7697                         else
7698                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7699                                                     wc->for_reloc);
7700                         BUG_ON(ret); /* -ENOMEM */
7701                 }
7702                 /* make block locked assertion in clean_tree_block happy */
7703                 if (!path->locks[level] &&
7704                     btrfs_header_generation(eb) == trans->transid) {
7705                         btrfs_tree_lock(eb);
7706                         btrfs_set_lock_blocking(eb);
7707                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7708                 }
7709                 clean_tree_block(trans, root, eb);
7710         }
7711
7712         if (eb == root->node) {
7713                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7714                         parent = eb->start;
7715                 else
7716                         BUG_ON(root->root_key.objectid !=
7717                                btrfs_header_owner(eb));
7718         } else {
7719                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7720                         parent = path->nodes[level + 1]->start;
7721                 else
7722                         BUG_ON(root->root_key.objectid !=
7723                                btrfs_header_owner(path->nodes[level + 1]));
7724         }
7725
7726         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7727 out:
7728         wc->refs[level] = 0;
7729         wc->flags[level] = 0;
7730         return 0;
7731 }
7732
7733 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7734                                    struct btrfs_root *root,
7735                                    struct btrfs_path *path,
7736                                    struct walk_control *wc)
7737 {
7738         int level = wc->level;
7739         int lookup_info = 1;
7740         int ret;
7741
7742         while (level >= 0) {
7743                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7744                 if (ret > 0)
7745                         break;
7746
7747                 if (level == 0)
7748                         break;
7749
7750                 if (path->slots[level] >=
7751                     btrfs_header_nritems(path->nodes[level]))
7752                         break;
7753
7754                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7755                 if (ret > 0) {
7756                         path->slots[level]++;
7757                         continue;
7758                 } else if (ret < 0)
7759                         return ret;
7760                 level = wc->level;
7761         }
7762         return 0;
7763 }
7764
7765 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7766                                  struct btrfs_root *root,
7767                                  struct btrfs_path *path,
7768                                  struct walk_control *wc, int max_level)
7769 {
7770         int level = wc->level;
7771         int ret;
7772
7773         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7774         while (level < max_level && path->nodes[level]) {
7775                 wc->level = level;
7776                 if (path->slots[level] + 1 <
7777                     btrfs_header_nritems(path->nodes[level])) {
7778                         path->slots[level]++;
7779                         return 0;
7780                 } else {
7781                         ret = walk_up_proc(trans, root, path, wc);
7782                         if (ret > 0)
7783                                 return 0;
7784
7785                         if (path->locks[level]) {
7786                                 btrfs_tree_unlock_rw(path->nodes[level],
7787                                                      path->locks[level]);
7788                                 path->locks[level] = 0;
7789                         }
7790                         free_extent_buffer(path->nodes[level]);
7791                         path->nodes[level] = NULL;
7792                         level++;
7793                 }
7794         }
7795         return 1;
7796 }
7797
7798 /*
7799  * drop a subvolume tree.
7800  *
7801  * this function traverses the tree freeing any blocks that only
7802  * referenced by the tree.
7803  *
7804  * when a shared tree block is found. this function decreases its
7805  * reference count by one. if update_ref is true, this function
7806  * also make sure backrefs for the shared block and all lower level
7807  * blocks are properly updated.
7808  *
7809  * If called with for_reloc == 0, may exit early with -EAGAIN
7810  */
7811 int btrfs_drop_snapshot(struct btrfs_root *root,
7812                          struct btrfs_block_rsv *block_rsv, int update_ref,
7813                          int for_reloc)
7814 {
7815         struct btrfs_path *path;
7816         struct btrfs_trans_handle *trans;
7817         struct btrfs_root *tree_root = root->fs_info->tree_root;
7818         struct btrfs_root_item *root_item = &root->root_item;
7819         struct walk_control *wc;
7820         struct btrfs_key key;
7821         int err = 0;
7822         int ret;
7823         int level;
7824         bool root_dropped = false;
7825
7826         path = btrfs_alloc_path();
7827         if (!path) {
7828                 err = -ENOMEM;
7829                 goto out;
7830         }
7831
7832         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7833         if (!wc) {
7834                 btrfs_free_path(path);
7835                 err = -ENOMEM;
7836                 goto out;
7837         }
7838
7839         trans = btrfs_start_transaction(tree_root, 0);
7840         if (IS_ERR(trans)) {
7841                 err = PTR_ERR(trans);
7842                 goto out_free;
7843         }
7844
7845         if (block_rsv)
7846                 trans->block_rsv = block_rsv;
7847
7848         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7849                 level = btrfs_header_level(root->node);
7850                 path->nodes[level] = btrfs_lock_root_node(root);
7851                 btrfs_set_lock_blocking(path->nodes[level]);
7852                 path->slots[level] = 0;
7853                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7854                 memset(&wc->update_progress, 0,
7855                        sizeof(wc->update_progress));
7856         } else {
7857                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7858                 memcpy(&wc->update_progress, &key,
7859                        sizeof(wc->update_progress));
7860
7861                 level = root_item->drop_level;
7862                 BUG_ON(level == 0);
7863                 path->lowest_level = level;
7864                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7865                 path->lowest_level = 0;
7866                 if (ret < 0) {
7867                         err = ret;
7868                         goto out_end_trans;
7869                 }
7870                 WARN_ON(ret > 0);
7871
7872                 /*
7873                  * unlock our path, this is safe because only this
7874                  * function is allowed to delete this snapshot
7875                  */
7876                 btrfs_unlock_up_safe(path, 0);
7877
7878                 level = btrfs_header_level(root->node);
7879                 while (1) {
7880                         btrfs_tree_lock(path->nodes[level]);
7881                         btrfs_set_lock_blocking(path->nodes[level]);
7882                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7883
7884                         ret = btrfs_lookup_extent_info(trans, root,
7885                                                 path->nodes[level]->start,
7886                                                 level, 1, &wc->refs[level],
7887                                                 &wc->flags[level]);
7888                         if (ret < 0) {
7889                                 err = ret;
7890                                 goto out_end_trans;
7891                         }
7892                         BUG_ON(wc->refs[level] == 0);
7893
7894                         if (level == root_item->drop_level)
7895                                 break;
7896
7897                         btrfs_tree_unlock(path->nodes[level]);
7898                         path->locks[level] = 0;
7899                         WARN_ON(wc->refs[level] != 1);
7900                         level--;
7901                 }
7902         }
7903
7904         wc->level = level;
7905         wc->shared_level = -1;
7906         wc->stage = DROP_REFERENCE;
7907         wc->update_ref = update_ref;
7908         wc->keep_locks = 0;
7909         wc->for_reloc = for_reloc;
7910         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7911
7912         while (1) {
7913
7914                 ret = walk_down_tree(trans, root, path, wc);
7915                 if (ret < 0) {
7916                         err = ret;
7917                         break;
7918                 }
7919
7920                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7921                 if (ret < 0) {
7922                         err = ret;
7923                         break;
7924                 }
7925
7926                 if (ret > 0) {
7927                         BUG_ON(wc->stage != DROP_REFERENCE);
7928                         break;
7929                 }
7930
7931                 if (wc->stage == DROP_REFERENCE) {
7932                         level = wc->level;
7933                         btrfs_node_key(path->nodes[level],
7934                                        &root_item->drop_progress,
7935                                        path->slots[level]);
7936                         root_item->drop_level = level;
7937                 }
7938
7939                 BUG_ON(wc->level == 0);
7940                 if (btrfs_should_end_transaction(trans, tree_root) ||
7941                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
7942                         ret = btrfs_update_root(trans, tree_root,
7943                                                 &root->root_key,
7944                                                 root_item);
7945                         if (ret) {
7946                                 btrfs_abort_transaction(trans, tree_root, ret);
7947                                 err = ret;
7948                                 goto out_end_trans;
7949                         }
7950
7951                         btrfs_end_transaction_throttle(trans, tree_root);
7952                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
7953                                 pr_debug("BTRFS: drop snapshot early exit\n");
7954                                 err = -EAGAIN;
7955                                 goto out_free;
7956                         }
7957
7958                         trans = btrfs_start_transaction(tree_root, 0);
7959                         if (IS_ERR(trans)) {
7960                                 err = PTR_ERR(trans);
7961                                 goto out_free;
7962                         }
7963                         if (block_rsv)
7964                                 trans->block_rsv = block_rsv;
7965                 }
7966         }
7967         btrfs_release_path(path);
7968         if (err)
7969                 goto out_end_trans;
7970
7971         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7972         if (ret) {
7973                 btrfs_abort_transaction(trans, tree_root, ret);
7974                 goto out_end_trans;
7975         }
7976
7977         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7978                 ret = btrfs_find_root(tree_root, &root->root_key, path,
7979                                       NULL, NULL);
7980                 if (ret < 0) {
7981                         btrfs_abort_transaction(trans, tree_root, ret);
7982                         err = ret;
7983                         goto out_end_trans;
7984                 } else if (ret > 0) {
7985                         /* if we fail to delete the orphan item this time
7986                          * around, it'll get picked up the next time.
7987                          *
7988                          * The most common failure here is just -ENOENT.
7989                          */
7990                         btrfs_del_orphan_item(trans, tree_root,
7991                                               root->root_key.objectid);
7992                 }
7993         }
7994
7995         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
7996                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
7997         } else {
7998                 free_extent_buffer(root->node);
7999                 free_extent_buffer(root->commit_root);
8000                 btrfs_put_fs_root(root);
8001         }
8002         root_dropped = true;
8003 out_end_trans:
8004         btrfs_end_transaction_throttle(trans, tree_root);
8005 out_free:
8006         kfree(wc);
8007         btrfs_free_path(path);
8008 out:
8009         /*
8010          * So if we need to stop dropping the snapshot for whatever reason we
8011          * need to make sure to add it back to the dead root list so that we
8012          * keep trying to do the work later.  This also cleans up roots if we
8013          * don't have it in the radix (like when we recover after a power fail
8014          * or unmount) so we don't leak memory.
8015          */
8016         if (!for_reloc && root_dropped == false)
8017                 btrfs_add_dead_root(root);
8018         if (err && err != -EAGAIN)
8019                 btrfs_std_error(root->fs_info, err);
8020         return err;
8021 }
8022
8023 /*
8024  * drop subtree rooted at tree block 'node'.
8025  *
8026  * NOTE: this function will unlock and release tree block 'node'
8027  * only used by relocation code
8028  */
8029 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8030                         struct btrfs_root *root,
8031                         struct extent_buffer *node,
8032                         struct extent_buffer *parent)
8033 {
8034         struct btrfs_path *path;
8035         struct walk_control *wc;
8036         int level;
8037         int parent_level;
8038         int ret = 0;
8039         int wret;
8040
8041         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8042
8043         path = btrfs_alloc_path();
8044         if (!path)
8045                 return -ENOMEM;
8046
8047         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8048         if (!wc) {
8049                 btrfs_free_path(path);
8050                 return -ENOMEM;
8051         }
8052
8053         btrfs_assert_tree_locked(parent);
8054         parent_level = btrfs_header_level(parent);
8055         extent_buffer_get(parent);
8056         path->nodes[parent_level] = parent;
8057         path->slots[parent_level] = btrfs_header_nritems(parent);
8058
8059         btrfs_assert_tree_locked(node);
8060         level = btrfs_header_level(node);
8061         path->nodes[level] = node;
8062         path->slots[level] = 0;
8063         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8064
8065         wc->refs[parent_level] = 1;
8066         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8067         wc->level = level;
8068         wc->shared_level = -1;
8069         wc->stage = DROP_REFERENCE;
8070         wc->update_ref = 0;
8071         wc->keep_locks = 1;
8072         wc->for_reloc = 1;
8073         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8074
8075         while (1) {
8076                 wret = walk_down_tree(trans, root, path, wc);
8077                 if (wret < 0) {
8078                         ret = wret;
8079                         break;
8080                 }
8081
8082                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8083                 if (wret < 0)
8084                         ret = wret;
8085                 if (wret != 0)
8086                         break;
8087         }
8088
8089         kfree(wc);
8090         btrfs_free_path(path);
8091         return ret;
8092 }
8093
8094 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8095 {
8096         u64 num_devices;
8097         u64 stripped;
8098
8099         /*
8100          * if restripe for this chunk_type is on pick target profile and
8101          * return, otherwise do the usual balance
8102          */
8103         stripped = get_restripe_target(root->fs_info, flags);
8104         if (stripped)
8105                 return extended_to_chunk(stripped);
8106
8107         /*
8108          * we add in the count of missing devices because we want
8109          * to make sure that any RAID levels on a degraded FS
8110          * continue to be honored.
8111          */
8112         num_devices = root->fs_info->fs_devices->rw_devices +
8113                 root->fs_info->fs_devices->missing_devices;
8114
8115         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8116                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8117                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8118
8119         if (num_devices == 1) {
8120                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8121                 stripped = flags & ~stripped;
8122
8123                 /* turn raid0 into single device chunks */
8124                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8125                         return stripped;
8126
8127                 /* turn mirroring into duplication */
8128                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8129                              BTRFS_BLOCK_GROUP_RAID10))
8130                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8131         } else {
8132                 /* they already had raid on here, just return */
8133                 if (flags & stripped)
8134                         return flags;
8135
8136                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8137                 stripped = flags & ~stripped;
8138
8139                 /* switch duplicated blocks with raid1 */
8140                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8141                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8142
8143                 /* this is drive concat, leave it alone */
8144         }
8145
8146         return flags;
8147 }
8148
8149 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8150 {
8151         struct btrfs_space_info *sinfo = cache->space_info;
8152         u64 num_bytes;
8153         u64 min_allocable_bytes;
8154         int ret = -ENOSPC;
8155
8156
8157         /*
8158          * We need some metadata space and system metadata space for
8159          * allocating chunks in some corner cases until we force to set
8160          * it to be readonly.
8161          */
8162         if ((sinfo->flags &
8163              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8164             !force)
8165                 min_allocable_bytes = 1 * 1024 * 1024;
8166         else
8167                 min_allocable_bytes = 0;
8168
8169         spin_lock(&sinfo->lock);
8170         spin_lock(&cache->lock);
8171
8172         if (cache->ro) {
8173                 ret = 0;
8174                 goto out;
8175         }
8176
8177         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8178                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8179
8180         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8181             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8182             min_allocable_bytes <= sinfo->total_bytes) {
8183                 sinfo->bytes_readonly += num_bytes;
8184                 cache->ro = 1;
8185                 ret = 0;
8186         }
8187 out:
8188         spin_unlock(&cache->lock);
8189         spin_unlock(&sinfo->lock);
8190         return ret;
8191 }
8192
8193 int btrfs_set_block_group_ro(struct btrfs_root *root,
8194                              struct btrfs_block_group_cache *cache)
8195
8196 {
8197         struct btrfs_trans_handle *trans;
8198         u64 alloc_flags;
8199         int ret;
8200
8201         BUG_ON(cache->ro);
8202
8203         trans = btrfs_join_transaction(root);
8204         if (IS_ERR(trans))
8205                 return PTR_ERR(trans);
8206
8207         alloc_flags = update_block_group_flags(root, cache->flags);
8208         if (alloc_flags != cache->flags) {
8209                 ret = do_chunk_alloc(trans, root, alloc_flags,
8210                                      CHUNK_ALLOC_FORCE);
8211                 if (ret < 0)
8212                         goto out;
8213         }
8214
8215         ret = set_block_group_ro(cache, 0);
8216         if (!ret)
8217                 goto out;
8218         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8219         ret = do_chunk_alloc(trans, root, alloc_flags,
8220                              CHUNK_ALLOC_FORCE);
8221         if (ret < 0)
8222                 goto out;
8223         ret = set_block_group_ro(cache, 0);
8224 out:
8225         btrfs_end_transaction(trans, root);
8226         return ret;
8227 }
8228
8229 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8230                             struct btrfs_root *root, u64 type)
8231 {
8232         u64 alloc_flags = get_alloc_profile(root, type);
8233         return do_chunk_alloc(trans, root, alloc_flags,
8234                               CHUNK_ALLOC_FORCE);
8235 }
8236
8237 /*
8238  * helper to account the unused space of all the readonly block group in the
8239  * list. takes mirrors into account.
8240  */
8241 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
8242 {
8243         struct btrfs_block_group_cache *block_group;
8244         u64 free_bytes = 0;
8245         int factor;
8246
8247         list_for_each_entry(block_group, groups_list, list) {
8248                 spin_lock(&block_group->lock);
8249
8250                 if (!block_group->ro) {
8251                         spin_unlock(&block_group->lock);
8252                         continue;
8253                 }
8254
8255                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8256                                           BTRFS_BLOCK_GROUP_RAID10 |
8257                                           BTRFS_BLOCK_GROUP_DUP))
8258                         factor = 2;
8259                 else
8260                         factor = 1;
8261
8262                 free_bytes += (block_group->key.offset -
8263                                btrfs_block_group_used(&block_group->item)) *
8264                                factor;
8265
8266                 spin_unlock(&block_group->lock);
8267         }
8268
8269         return free_bytes;
8270 }
8271
8272 /*
8273  * helper to account the unused space of all the readonly block group in the
8274  * space_info. takes mirrors into account.
8275  */
8276 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8277 {
8278         int i;
8279         u64 free_bytes = 0;
8280
8281         spin_lock(&sinfo->lock);
8282
8283         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
8284                 if (!list_empty(&sinfo->block_groups[i]))
8285                         free_bytes += __btrfs_get_ro_block_group_free_space(
8286                                                 &sinfo->block_groups[i]);
8287
8288         spin_unlock(&sinfo->lock);
8289
8290         return free_bytes;
8291 }
8292
8293 void btrfs_set_block_group_rw(struct btrfs_root *root,
8294                               struct btrfs_block_group_cache *cache)
8295 {
8296         struct btrfs_space_info *sinfo = cache->space_info;
8297         u64 num_bytes;
8298
8299         BUG_ON(!cache->ro);
8300
8301         spin_lock(&sinfo->lock);
8302         spin_lock(&cache->lock);
8303         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8304                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8305         sinfo->bytes_readonly -= num_bytes;
8306         cache->ro = 0;
8307         spin_unlock(&cache->lock);
8308         spin_unlock(&sinfo->lock);
8309 }
8310
8311 /*
8312  * checks to see if its even possible to relocate this block group.
8313  *
8314  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8315  * ok to go ahead and try.
8316  */
8317 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8318 {
8319         struct btrfs_block_group_cache *block_group;
8320         struct btrfs_space_info *space_info;
8321         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8322         struct btrfs_device *device;
8323         struct btrfs_trans_handle *trans;
8324         u64 min_free;
8325         u64 dev_min = 1;
8326         u64 dev_nr = 0;
8327         u64 target;
8328         int index;
8329         int full = 0;
8330         int ret = 0;
8331
8332         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8333
8334         /* odd, couldn't find the block group, leave it alone */
8335         if (!block_group)
8336                 return -1;
8337
8338         min_free = btrfs_block_group_used(&block_group->item);
8339
8340         /* no bytes used, we're good */
8341         if (!min_free)
8342                 goto out;
8343
8344         space_info = block_group->space_info;
8345         spin_lock(&space_info->lock);
8346
8347         full = space_info->full;
8348
8349         /*
8350          * if this is the last block group we have in this space, we can't
8351          * relocate it unless we're able to allocate a new chunk below.
8352          *
8353          * Otherwise, we need to make sure we have room in the space to handle
8354          * all of the extents from this block group.  If we can, we're good
8355          */
8356         if ((space_info->total_bytes != block_group->key.offset) &&
8357             (space_info->bytes_used + space_info->bytes_reserved +
8358              space_info->bytes_pinned + space_info->bytes_readonly +
8359              min_free < space_info->total_bytes)) {
8360                 spin_unlock(&space_info->lock);
8361                 goto out;
8362         }
8363         spin_unlock(&space_info->lock);
8364
8365         /*
8366          * ok we don't have enough space, but maybe we have free space on our
8367          * devices to allocate new chunks for relocation, so loop through our
8368          * alloc devices and guess if we have enough space.  if this block
8369          * group is going to be restriped, run checks against the target
8370          * profile instead of the current one.
8371          */
8372         ret = -1;
8373
8374         /*
8375          * index:
8376          *      0: raid10
8377          *      1: raid1
8378          *      2: dup
8379          *      3: raid0
8380          *      4: single
8381          */
8382         target = get_restripe_target(root->fs_info, block_group->flags);
8383         if (target) {
8384                 index = __get_raid_index(extended_to_chunk(target));
8385         } else {
8386                 /*
8387                  * this is just a balance, so if we were marked as full
8388                  * we know there is no space for a new chunk
8389                  */
8390                 if (full)
8391                         goto out;
8392
8393                 index = get_block_group_index(block_group);
8394         }
8395
8396         if (index == BTRFS_RAID_RAID10) {
8397                 dev_min = 4;
8398                 /* Divide by 2 */
8399                 min_free >>= 1;
8400         } else if (index == BTRFS_RAID_RAID1) {
8401                 dev_min = 2;
8402         } else if (index == BTRFS_RAID_DUP) {
8403                 /* Multiply by 2 */
8404                 min_free <<= 1;
8405         } else if (index == BTRFS_RAID_RAID0) {
8406                 dev_min = fs_devices->rw_devices;
8407                 do_div(min_free, dev_min);
8408         }
8409
8410         /* We need to do this so that we can look at pending chunks */
8411         trans = btrfs_join_transaction(root);
8412         if (IS_ERR(trans)) {
8413                 ret = PTR_ERR(trans);
8414                 goto out;
8415         }
8416
8417         mutex_lock(&root->fs_info->chunk_mutex);
8418         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8419                 u64 dev_offset;
8420
8421                 /*
8422                  * check to make sure we can actually find a chunk with enough
8423                  * space to fit our block group in.
8424                  */
8425                 if (device->total_bytes > device->bytes_used + min_free &&
8426                     !device->is_tgtdev_for_dev_replace) {
8427                         ret = find_free_dev_extent(trans, device, min_free,
8428                                                    &dev_offset, NULL);
8429                         if (!ret)
8430                                 dev_nr++;
8431
8432                         if (dev_nr >= dev_min)
8433                                 break;
8434
8435                         ret = -1;
8436                 }
8437         }
8438         mutex_unlock(&root->fs_info->chunk_mutex);
8439         btrfs_end_transaction(trans, root);
8440 out:
8441         btrfs_put_block_group(block_group);
8442         return ret;
8443 }
8444
8445 static int find_first_block_group(struct btrfs_root *root,
8446                 struct btrfs_path *path, struct btrfs_key *key)
8447 {
8448         int ret = 0;
8449         struct btrfs_key found_key;
8450         struct extent_buffer *leaf;
8451         int slot;
8452
8453         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8454         if (ret < 0)
8455                 goto out;
8456
8457         while (1) {
8458                 slot = path->slots[0];
8459                 leaf = path->nodes[0];
8460                 if (slot >= btrfs_header_nritems(leaf)) {
8461                         ret = btrfs_next_leaf(root, path);
8462                         if (ret == 0)
8463                                 continue;
8464                         if (ret < 0)
8465                                 goto out;
8466                         break;
8467                 }
8468                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8469
8470                 if (found_key.objectid >= key->objectid &&
8471                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8472                         ret = 0;
8473                         goto out;
8474                 }
8475                 path->slots[0]++;
8476         }
8477 out:
8478         return ret;
8479 }
8480
8481 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8482 {
8483         struct btrfs_block_group_cache *block_group;
8484         u64 last = 0;
8485
8486         while (1) {
8487                 struct inode *inode;
8488
8489                 block_group = btrfs_lookup_first_block_group(info, last);
8490                 while (block_group) {
8491                         spin_lock(&block_group->lock);
8492                         if (block_group->iref)
8493                                 break;
8494                         spin_unlock(&block_group->lock);
8495                         block_group = next_block_group(info->tree_root,
8496                                                        block_group);
8497                 }
8498                 if (!block_group) {
8499                         if (last == 0)
8500                                 break;
8501                         last = 0;
8502                         continue;
8503                 }
8504
8505                 inode = block_group->inode;
8506                 block_group->iref = 0;
8507                 block_group->inode = NULL;
8508                 spin_unlock(&block_group->lock);
8509                 iput(inode);
8510                 last = block_group->key.objectid + block_group->key.offset;
8511                 btrfs_put_block_group(block_group);
8512         }
8513 }
8514
8515 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8516 {
8517         struct btrfs_block_group_cache *block_group;
8518         struct btrfs_space_info *space_info;
8519         struct btrfs_caching_control *caching_ctl;
8520         struct rb_node *n;
8521
8522         down_write(&info->commit_root_sem);
8523         while (!list_empty(&info->caching_block_groups)) {
8524                 caching_ctl = list_entry(info->caching_block_groups.next,
8525                                          struct btrfs_caching_control, list);
8526                 list_del(&caching_ctl->list);
8527                 put_caching_control(caching_ctl);
8528         }
8529         up_write(&info->commit_root_sem);
8530
8531         spin_lock(&info->block_group_cache_lock);
8532         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8533                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8534                                        cache_node);
8535                 rb_erase(&block_group->cache_node,
8536                          &info->block_group_cache_tree);
8537                 spin_unlock(&info->block_group_cache_lock);
8538
8539                 down_write(&block_group->space_info->groups_sem);
8540                 list_del(&block_group->list);
8541                 up_write(&block_group->space_info->groups_sem);
8542
8543                 if (block_group->cached == BTRFS_CACHE_STARTED)
8544                         wait_block_group_cache_done(block_group);
8545
8546                 /*
8547                  * We haven't cached this block group, which means we could
8548                  * possibly have excluded extents on this block group.
8549                  */
8550                 if (block_group->cached == BTRFS_CACHE_NO ||
8551                     block_group->cached == BTRFS_CACHE_ERROR)
8552                         free_excluded_extents(info->extent_root, block_group);
8553
8554                 btrfs_remove_free_space_cache(block_group);
8555                 btrfs_put_block_group(block_group);
8556
8557                 spin_lock(&info->block_group_cache_lock);
8558         }
8559         spin_unlock(&info->block_group_cache_lock);
8560
8561         /* now that all the block groups are freed, go through and
8562          * free all the space_info structs.  This is only called during
8563          * the final stages of unmount, and so we know nobody is
8564          * using them.  We call synchronize_rcu() once before we start,
8565          * just to be on the safe side.
8566          */
8567         synchronize_rcu();
8568
8569         release_global_block_rsv(info);
8570
8571         while (!list_empty(&info->space_info)) {
8572                 int i;
8573
8574                 space_info = list_entry(info->space_info.next,
8575                                         struct btrfs_space_info,
8576                                         list);
8577                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8578                         if (WARN_ON(space_info->bytes_pinned > 0 ||
8579                             space_info->bytes_reserved > 0 ||
8580                             space_info->bytes_may_use > 0)) {
8581                                 dump_space_info(space_info, 0, 0);
8582                         }
8583                 }
8584                 list_del(&space_info->list);
8585                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
8586                         struct kobject *kobj;
8587                         kobj = space_info->block_group_kobjs[i];
8588                         space_info->block_group_kobjs[i] = NULL;
8589                         if (kobj) {
8590                                 kobject_del(kobj);
8591                                 kobject_put(kobj);
8592                         }
8593                 }
8594                 kobject_del(&space_info->kobj);
8595                 kobject_put(&space_info->kobj);
8596         }
8597         return 0;
8598 }
8599
8600 static void __link_block_group(struct btrfs_space_info *space_info,
8601                                struct btrfs_block_group_cache *cache)
8602 {
8603         int index = get_block_group_index(cache);
8604         bool first = false;
8605
8606         down_write(&space_info->groups_sem);
8607         if (list_empty(&space_info->block_groups[index]))
8608                 first = true;
8609         list_add_tail(&cache->list, &space_info->block_groups[index]);
8610         up_write(&space_info->groups_sem);
8611
8612         if (first) {
8613                 struct raid_kobject *rkobj;
8614                 int ret;
8615
8616                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
8617                 if (!rkobj)
8618                         goto out_err;
8619                 rkobj->raid_type = index;
8620                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
8621                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
8622                                   "%s", get_raid_name(index));
8623                 if (ret) {
8624                         kobject_put(&rkobj->kobj);
8625                         goto out_err;
8626                 }
8627                 space_info->block_group_kobjs[index] = &rkobj->kobj;
8628         }
8629
8630         return;
8631 out_err:
8632         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
8633 }
8634
8635 static struct btrfs_block_group_cache *
8636 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
8637 {
8638         struct btrfs_block_group_cache *cache;
8639
8640         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8641         if (!cache)
8642                 return NULL;
8643
8644         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8645                                         GFP_NOFS);
8646         if (!cache->free_space_ctl) {
8647                 kfree(cache);
8648                 return NULL;
8649         }
8650
8651         cache->key.objectid = start;
8652         cache->key.offset = size;
8653         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8654
8655         cache->sectorsize = root->sectorsize;
8656         cache->fs_info = root->fs_info;
8657         cache->full_stripe_len = btrfs_full_stripe_len(root,
8658                                                &root->fs_info->mapping_tree,
8659                                                start);
8660         atomic_set(&cache->count, 1);
8661         spin_lock_init(&cache->lock);
8662         INIT_LIST_HEAD(&cache->list);
8663         INIT_LIST_HEAD(&cache->cluster_list);
8664         INIT_LIST_HEAD(&cache->new_bg_list);
8665         btrfs_init_free_space_ctl(cache);
8666
8667         return cache;
8668 }
8669
8670 int btrfs_read_block_groups(struct btrfs_root *root)
8671 {
8672         struct btrfs_path *path;
8673         int ret;
8674         struct btrfs_block_group_cache *cache;
8675         struct btrfs_fs_info *info = root->fs_info;
8676         struct btrfs_space_info *space_info;
8677         struct btrfs_key key;
8678         struct btrfs_key found_key;
8679         struct extent_buffer *leaf;
8680         int need_clear = 0;
8681         u64 cache_gen;
8682
8683         root = info->extent_root;
8684         key.objectid = 0;
8685         key.offset = 0;
8686         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8687         path = btrfs_alloc_path();
8688         if (!path)
8689                 return -ENOMEM;
8690         path->reada = 1;
8691
8692         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
8693         if (btrfs_test_opt(root, SPACE_CACHE) &&
8694             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
8695                 need_clear = 1;
8696         if (btrfs_test_opt(root, CLEAR_CACHE))
8697                 need_clear = 1;
8698
8699         while (1) {
8700                 ret = find_first_block_group(root, path, &key);
8701                 if (ret > 0)
8702                         break;
8703                 if (ret != 0)
8704                         goto error;
8705
8706                 leaf = path->nodes[0];
8707                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8708
8709                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
8710                                                        found_key.offset);
8711                 if (!cache) {
8712                         ret = -ENOMEM;
8713                         goto error;
8714                 }
8715
8716                 if (need_clear) {
8717                         /*
8718                          * When we mount with old space cache, we need to
8719                          * set BTRFS_DC_CLEAR and set dirty flag.
8720                          *
8721                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8722                          *    truncate the old free space cache inode and
8723                          *    setup a new one.
8724                          * b) Setting 'dirty flag' makes sure that we flush
8725                          *    the new space cache info onto disk.
8726                          */
8727                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8728                         if (btrfs_test_opt(root, SPACE_CACHE))
8729                                 cache->dirty = 1;
8730                 }
8731
8732                 read_extent_buffer(leaf, &cache->item,
8733                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8734                                    sizeof(cache->item));
8735                 cache->flags = btrfs_block_group_flags(&cache->item);
8736
8737                 key.objectid = found_key.objectid + found_key.offset;
8738                 btrfs_release_path(path);
8739
8740                 /*
8741                  * We need to exclude the super stripes now so that the space
8742                  * info has super bytes accounted for, otherwise we'll think
8743                  * we have more space than we actually do.
8744                  */
8745                 ret = exclude_super_stripes(root, cache);
8746                 if (ret) {
8747                         /*
8748                          * We may have excluded something, so call this just in
8749                          * case.
8750                          */
8751                         free_excluded_extents(root, cache);
8752                         btrfs_put_block_group(cache);
8753                         goto error;
8754                 }
8755
8756                 /*
8757                  * check for two cases, either we are full, and therefore
8758                  * don't need to bother with the caching work since we won't
8759                  * find any space, or we are empty, and we can just add all
8760                  * the space in and be done with it.  This saves us _alot_ of
8761                  * time, particularly in the full case.
8762                  */
8763                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8764                         cache->last_byte_to_unpin = (u64)-1;
8765                         cache->cached = BTRFS_CACHE_FINISHED;
8766                         free_excluded_extents(root, cache);
8767                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8768                         cache->last_byte_to_unpin = (u64)-1;
8769                         cache->cached = BTRFS_CACHE_FINISHED;
8770                         add_new_free_space(cache, root->fs_info,
8771                                            found_key.objectid,
8772                                            found_key.objectid +
8773                                            found_key.offset);
8774                         free_excluded_extents(root, cache);
8775                 }
8776
8777                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8778                 if (ret) {
8779                         btrfs_remove_free_space_cache(cache);
8780                         btrfs_put_block_group(cache);
8781                         goto error;
8782                 }
8783
8784                 ret = update_space_info(info, cache->flags, found_key.offset,
8785                                         btrfs_block_group_used(&cache->item),
8786                                         &space_info);
8787                 if (ret) {
8788                         btrfs_remove_free_space_cache(cache);
8789                         spin_lock(&info->block_group_cache_lock);
8790                         rb_erase(&cache->cache_node,
8791                                  &info->block_group_cache_tree);
8792                         spin_unlock(&info->block_group_cache_lock);
8793                         btrfs_put_block_group(cache);
8794                         goto error;
8795                 }
8796
8797                 cache->space_info = space_info;
8798                 spin_lock(&cache->space_info->lock);
8799                 cache->space_info->bytes_readonly += cache->bytes_super;
8800                 spin_unlock(&cache->space_info->lock);
8801
8802                 __link_block_group(space_info, cache);
8803
8804                 set_avail_alloc_bits(root->fs_info, cache->flags);
8805                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8806                         set_block_group_ro(cache, 1);
8807         }
8808
8809         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8810                 if (!(get_alloc_profile(root, space_info->flags) &
8811                       (BTRFS_BLOCK_GROUP_RAID10 |
8812                        BTRFS_BLOCK_GROUP_RAID1 |
8813                        BTRFS_BLOCK_GROUP_RAID5 |
8814                        BTRFS_BLOCK_GROUP_RAID6 |
8815                        BTRFS_BLOCK_GROUP_DUP)))
8816                         continue;
8817                 /*
8818                  * avoid allocating from un-mirrored block group if there are
8819                  * mirrored block groups.
8820                  */
8821                 list_for_each_entry(cache,
8822                                 &space_info->block_groups[BTRFS_RAID_RAID0],
8823                                 list)
8824                         set_block_group_ro(cache, 1);
8825                 list_for_each_entry(cache,
8826                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
8827                                 list)
8828                         set_block_group_ro(cache, 1);
8829         }
8830
8831         init_global_block_rsv(info);
8832         ret = 0;
8833 error:
8834         btrfs_free_path(path);
8835         return ret;
8836 }
8837
8838 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8839                                        struct btrfs_root *root)
8840 {
8841         struct btrfs_block_group_cache *block_group, *tmp;
8842         struct btrfs_root *extent_root = root->fs_info->extent_root;
8843         struct btrfs_block_group_item item;
8844         struct btrfs_key key;
8845         int ret = 0;
8846
8847         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8848                                  new_bg_list) {
8849                 list_del_init(&block_group->new_bg_list);
8850
8851                 if (ret)
8852                         continue;
8853
8854                 spin_lock(&block_group->lock);
8855                 memcpy(&item, &block_group->item, sizeof(item));
8856                 memcpy(&key, &block_group->key, sizeof(key));
8857                 spin_unlock(&block_group->lock);
8858
8859                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8860                                         sizeof(item));
8861                 if (ret)
8862                         btrfs_abort_transaction(trans, extent_root, ret);
8863                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
8864                                                key.objectid, key.offset);
8865                 if (ret)
8866                         btrfs_abort_transaction(trans, extent_root, ret);
8867         }
8868 }
8869
8870 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8871                            struct btrfs_root *root, u64 bytes_used,
8872                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8873                            u64 size)
8874 {
8875         int ret;
8876         struct btrfs_root *extent_root;
8877         struct btrfs_block_group_cache *cache;
8878
8879         extent_root = root->fs_info->extent_root;
8880
8881         btrfs_set_log_full_commit(root->fs_info, trans);
8882
8883         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
8884         if (!cache)
8885                 return -ENOMEM;
8886
8887         btrfs_set_block_group_used(&cache->item, bytes_used);
8888         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8889         btrfs_set_block_group_flags(&cache->item, type);
8890
8891         cache->flags = type;
8892         cache->last_byte_to_unpin = (u64)-1;
8893         cache->cached = BTRFS_CACHE_FINISHED;
8894         ret = exclude_super_stripes(root, cache);
8895         if (ret) {
8896                 /*
8897                  * We may have excluded something, so call this just in
8898                  * case.
8899                  */
8900                 free_excluded_extents(root, cache);
8901                 btrfs_put_block_group(cache);
8902                 return ret;
8903         }
8904
8905         add_new_free_space(cache, root->fs_info, chunk_offset,
8906                            chunk_offset + size);
8907
8908         free_excluded_extents(root, cache);
8909
8910         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8911         if (ret) {
8912                 btrfs_remove_free_space_cache(cache);
8913                 btrfs_put_block_group(cache);
8914                 return ret;
8915         }
8916
8917         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8918                                 &cache->space_info);
8919         if (ret) {
8920                 btrfs_remove_free_space_cache(cache);
8921                 spin_lock(&root->fs_info->block_group_cache_lock);
8922                 rb_erase(&cache->cache_node,
8923                          &root->fs_info->block_group_cache_tree);
8924                 spin_unlock(&root->fs_info->block_group_cache_lock);
8925                 btrfs_put_block_group(cache);
8926                 return ret;
8927         }
8928         update_global_block_rsv(root->fs_info);
8929
8930         spin_lock(&cache->space_info->lock);
8931         cache->space_info->bytes_readonly += cache->bytes_super;
8932         spin_unlock(&cache->space_info->lock);
8933
8934         __link_block_group(cache->space_info, cache);
8935
8936         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8937
8938         set_avail_alloc_bits(extent_root->fs_info, type);
8939
8940         return 0;
8941 }
8942
8943 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8944 {
8945         u64 extra_flags = chunk_to_extended(flags) &
8946                                 BTRFS_EXTENDED_PROFILE_MASK;
8947
8948         write_seqlock(&fs_info->profiles_lock);
8949         if (flags & BTRFS_BLOCK_GROUP_DATA)
8950                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8951         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8952                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8953         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8954                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8955         write_sequnlock(&fs_info->profiles_lock);
8956 }
8957
8958 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8959                              struct btrfs_root *root, u64 group_start)
8960 {
8961         struct btrfs_path *path;
8962         struct btrfs_block_group_cache *block_group;
8963         struct btrfs_free_cluster *cluster;
8964         struct btrfs_root *tree_root = root->fs_info->tree_root;
8965         struct btrfs_key key;
8966         struct inode *inode;
8967         struct kobject *kobj = NULL;
8968         int ret;
8969         int index;
8970         int factor;
8971
8972         root = root->fs_info->extent_root;
8973
8974         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8975         BUG_ON(!block_group);
8976         BUG_ON(!block_group->ro);
8977
8978         /*
8979          * Free the reserved super bytes from this block group before
8980          * remove it.
8981          */
8982         free_excluded_extents(root, block_group);
8983
8984         memcpy(&key, &block_group->key, sizeof(key));
8985         index = get_block_group_index(block_group);
8986         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8987                                   BTRFS_BLOCK_GROUP_RAID1 |
8988                                   BTRFS_BLOCK_GROUP_RAID10))
8989                 factor = 2;
8990         else
8991                 factor = 1;
8992
8993         /* make sure this block group isn't part of an allocation cluster */
8994         cluster = &root->fs_info->data_alloc_cluster;
8995         spin_lock(&cluster->refill_lock);
8996         btrfs_return_cluster_to_free_space(block_group, cluster);
8997         spin_unlock(&cluster->refill_lock);
8998
8999         /*
9000          * make sure this block group isn't part of a metadata
9001          * allocation cluster
9002          */
9003         cluster = &root->fs_info->meta_alloc_cluster;
9004         spin_lock(&cluster->refill_lock);
9005         btrfs_return_cluster_to_free_space(block_group, cluster);
9006         spin_unlock(&cluster->refill_lock);
9007
9008         path = btrfs_alloc_path();
9009         if (!path) {
9010                 ret = -ENOMEM;
9011                 goto out;
9012         }
9013
9014         inode = lookup_free_space_inode(tree_root, block_group, path);
9015         if (!IS_ERR(inode)) {
9016                 ret = btrfs_orphan_add(trans, inode);
9017                 if (ret) {
9018                         btrfs_add_delayed_iput(inode);
9019                         goto out;
9020                 }
9021                 clear_nlink(inode);
9022                 /* One for the block groups ref */
9023                 spin_lock(&block_group->lock);
9024                 if (block_group->iref) {
9025                         block_group->iref = 0;
9026                         block_group->inode = NULL;
9027                         spin_unlock(&block_group->lock);
9028                         iput(inode);
9029                 } else {
9030                         spin_unlock(&block_group->lock);
9031                 }
9032                 /* One for our lookup ref */
9033                 btrfs_add_delayed_iput(inode);
9034         }
9035
9036         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9037         key.offset = block_group->key.objectid;
9038         key.type = 0;
9039
9040         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9041         if (ret < 0)
9042                 goto out;
9043         if (ret > 0)
9044                 btrfs_release_path(path);
9045         if (ret == 0) {
9046                 ret = btrfs_del_item(trans, tree_root, path);
9047                 if (ret)
9048                         goto out;
9049                 btrfs_release_path(path);
9050         }
9051
9052         spin_lock(&root->fs_info->block_group_cache_lock);
9053         rb_erase(&block_group->cache_node,
9054                  &root->fs_info->block_group_cache_tree);
9055
9056         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9057                 root->fs_info->first_logical_byte = (u64)-1;
9058         spin_unlock(&root->fs_info->block_group_cache_lock);
9059
9060         down_write(&block_group->space_info->groups_sem);
9061         /*
9062          * we must use list_del_init so people can check to see if they
9063          * are still on the list after taking the semaphore
9064          */
9065         list_del_init(&block_group->list);
9066         if (list_empty(&block_group->space_info->block_groups[index])) {
9067                 kobj = block_group->space_info->block_group_kobjs[index];
9068                 block_group->space_info->block_group_kobjs[index] = NULL;
9069                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9070         }
9071         up_write(&block_group->space_info->groups_sem);
9072         if (kobj) {
9073                 kobject_del(kobj);
9074                 kobject_put(kobj);
9075         }
9076
9077         if (block_group->cached == BTRFS_CACHE_STARTED)
9078                 wait_block_group_cache_done(block_group);
9079
9080         btrfs_remove_free_space_cache(block_group);
9081
9082         spin_lock(&block_group->space_info->lock);
9083         block_group->space_info->total_bytes -= block_group->key.offset;
9084         block_group->space_info->bytes_readonly -= block_group->key.offset;
9085         block_group->space_info->disk_total -= block_group->key.offset * factor;
9086         spin_unlock(&block_group->space_info->lock);
9087
9088         memcpy(&key, &block_group->key, sizeof(key));
9089
9090         btrfs_clear_space_info_full(root->fs_info);
9091
9092         btrfs_put_block_group(block_group);
9093         btrfs_put_block_group(block_group);
9094
9095         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9096         if (ret > 0)
9097                 ret = -EIO;
9098         if (ret < 0)
9099                 goto out;
9100
9101         ret = btrfs_del_item(trans, root, path);
9102 out:
9103         btrfs_free_path(path);
9104         return ret;
9105 }
9106
9107 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
9108 {
9109         struct btrfs_space_info *space_info;
9110         struct btrfs_super_block *disk_super;
9111         u64 features;
9112         u64 flags;
9113         int mixed = 0;
9114         int ret;
9115
9116         disk_super = fs_info->super_copy;
9117         if (!btrfs_super_root(disk_super))
9118                 return 1;
9119
9120         features = btrfs_super_incompat_flags(disk_super);
9121         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
9122                 mixed = 1;
9123
9124         flags = BTRFS_BLOCK_GROUP_SYSTEM;
9125         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9126         if (ret)
9127                 goto out;
9128
9129         if (mixed) {
9130                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
9131                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9132         } else {
9133                 flags = BTRFS_BLOCK_GROUP_METADATA;
9134                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9135                 if (ret)
9136                         goto out;
9137
9138                 flags = BTRFS_BLOCK_GROUP_DATA;
9139                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9140         }
9141 out:
9142         return ret;
9143 }
9144
9145 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
9146 {
9147         return unpin_extent_range(root, start, end);
9148 }
9149
9150 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
9151                                u64 num_bytes, u64 *actual_bytes)
9152 {
9153         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
9154 }
9155
9156 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
9157 {
9158         struct btrfs_fs_info *fs_info = root->fs_info;
9159         struct btrfs_block_group_cache *cache = NULL;
9160         u64 group_trimmed;
9161         u64 start;
9162         u64 end;
9163         u64 trimmed = 0;
9164         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
9165         int ret = 0;
9166
9167         /*
9168          * try to trim all FS space, our block group may start from non-zero.
9169          */
9170         if (range->len == total_bytes)
9171                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
9172         else
9173                 cache = btrfs_lookup_block_group(fs_info, range->start);
9174
9175         while (cache) {
9176                 if (cache->key.objectid >= (range->start + range->len)) {
9177                         btrfs_put_block_group(cache);
9178                         break;
9179                 }
9180
9181                 start = max(range->start, cache->key.objectid);
9182                 end = min(range->start + range->len,
9183                                 cache->key.objectid + cache->key.offset);
9184
9185                 if (end - start >= range->minlen) {
9186                         if (!block_group_cache_done(cache)) {
9187                                 ret = cache_block_group(cache, 0);
9188                                 if (ret) {
9189                                         btrfs_put_block_group(cache);
9190                                         break;
9191                                 }
9192                                 ret = wait_block_group_cache_done(cache);
9193                                 if (ret) {
9194                                         btrfs_put_block_group(cache);
9195                                         break;
9196                                 }
9197                         }
9198                         ret = btrfs_trim_block_group(cache,
9199                                                      &group_trimmed,
9200                                                      start,
9201                                                      end,
9202                                                      range->minlen);
9203
9204                         trimmed += group_trimmed;
9205                         if (ret) {
9206                                 btrfs_put_block_group(cache);
9207                                 break;
9208                         }
9209                 }
9210
9211                 cache = next_block_group(fs_info->tree_root, cache);
9212         }
9213
9214         range->len = trimmed;
9215         return ret;
9216 }
9217
9218 /*
9219  * btrfs_{start,end}_write() is similar to mnt_{want, drop}_write(),
9220  * they are used to prevent the some tasks writing data into the page cache
9221  * by nocow before the subvolume is snapshoted, but flush the data into
9222  * the disk after the snapshot creation.
9223  */
9224 void btrfs_end_nocow_write(struct btrfs_root *root)
9225 {
9226         percpu_counter_dec(&root->subv_writers->counter);
9227         /*
9228          * Make sure counter is updated before we wake up
9229          * waiters.
9230          */
9231         smp_mb();
9232         if (waitqueue_active(&root->subv_writers->wait))
9233                 wake_up(&root->subv_writers->wait);
9234 }
9235
9236 int btrfs_start_nocow_write(struct btrfs_root *root)
9237 {
9238         if (unlikely(atomic_read(&root->will_be_snapshoted)))
9239                 return 0;
9240
9241         percpu_counter_inc(&root->subv_writers->counter);
9242         /*
9243          * Make sure counter is updated before we check for snapshot creation.
9244          */
9245         smp_mb();
9246         if (unlikely(atomic_read(&root->will_be_snapshoted))) {
9247                 btrfs_end_nocow_write(root);
9248                 return 0;
9249         }
9250         return 1;
9251 }