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 "compat.h"
  28 #include "hash.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "raid56.h"
  35 #include "locking.h"
  36 #include "free-space-cache.h"
  37 #include "math.h"
  38
  39 #undef SCRAMBLE_DELAYED_REFS
  40
  41 /*
  42  * control flags for do_chunk_alloc's force field
  43  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  44  * if we really need one.
  45  *
  46  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  47  * if we have very few chunks already allocated.  This is
  48  * used as part of the clustering code to help make sure
  49  * we have a good pool of storage to cluster in, without
  50  * filling the FS with empty chunks
  51  *
  52  * CHUNK_ALLOC_FORCE means it must try to allocate one
  53  *
  54  */
  55 enum {
  56         CHUNK_ALLOC_NO_FORCE = 0,
  57         CHUNK_ALLOC_LIMITED = 1,
  58         CHUNK_ALLOC_FORCE = 2,
  59 };
  60
  61 /*
  62  * Control how reservations are dealt with.
  63  *
  64  * RESERVE_FREE - freeing a reservation.
  65  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  66  *   ENOSPC accounting
  67  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  68  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  69  */
  70 enum {
  71         RESERVE_FREE = 0,
  72         RESERVE_ALLOC = 1,
  73         RESERVE_ALLOC_NO_ACCOUNT = 2,
  74 };
  75
  76 static int update_block_group(struct btrfs_root *root,
  77                               u64 bytenr, u64 num_bytes, int alloc);
  78 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  79                                 struct btrfs_root *root,
  80                                 u64 bytenr, u64 num_bytes, u64 parent,
  81                                 u64 root_objectid, u64 owner_objectid,
  82                                 u64 owner_offset, int refs_to_drop,
  83                                 struct btrfs_delayed_extent_op *extra_op);
  84 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  85                                     struct extent_buffer *leaf,
  86                                     struct btrfs_extent_item *ei);
  87 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  88                                       struct btrfs_root *root,
  89                                       u64 parent, u64 root_objectid,
  90                                       u64 flags, u64 owner, u64 offset,
  91                                       struct btrfs_key *ins, int ref_mod);
  92 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  93                                      struct btrfs_root *root,
  94                                      u64 parent, u64 root_objectid,
  95                                      u64 flags, struct btrfs_disk_key *key,
  96                                      int level, struct btrfs_key *ins);
  97 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  98                           struct btrfs_root *extent_root, u64 flags,
  99                           int force);
 100 static int find_next_key(struct btrfs_path *path, int level,
 101                          struct btrfs_key *key);
 102 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 103                             int dump_block_groups);
 104 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 105                                        u64 num_bytes, int reserve);
 106 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 107                                u64 num_bytes);
 108
 109 static noinline int
 110 block_group_cache_done(struct btrfs_block_group_cache *cache)
 111 {
 112         smp_mb();
 113         return cache->cached == BTRFS_CACHE_FINISHED;
 114 }
 115
 116 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 117 {
 118         return (cache->flags & bits) == bits;
 119 }
 120
 121 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         atomic_inc(&cache->count);
 124 }
 125
 126 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 127 {
 128         if (atomic_dec_and_test(&cache->count)) {
 129                 WARN_ON(cache->pinned > 0);
 130                 WARN_ON(cache->reserved > 0);
 131                 kfree(cache->free_space_ctl);
 132                 kfree(cache);
 133         }
 134 }
 135
 136 /*
 137  * this adds the block group to the fs_info rb tree for the block group
 138  * cache
 139  */
 140 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 141                                 struct btrfs_block_group_cache *block_group)
 142 {
 143         struct rb_node **p;
 144         struct rb_node *parent = NULL;
 145         struct btrfs_block_group_cache *cache;
 146
 147         spin_lock(&info->block_group_cache_lock);
 148         p = &info->block_group_cache_tree.rb_node;
 149
 150         while (*p) {
 151                 parent = *p;
 152                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 153                                  cache_node);
 154                 if (block_group->key.objectid < cache->key.objectid) {
 155                         p = &(*p)->rb_left;
 156                 } else if (block_group->key.objectid > cache->key.objectid) {
 157                         p = &(*p)->rb_right;
 158                 } else {
 159                         spin_unlock(&info->block_group_cache_lock);
 160                         return -EEXIST;
 161                 }
 162         }
 163
 164         rb_link_node(&block_group->cache_node, parent, p);
 165         rb_insert_color(&block_group->cache_node,
 166                         &info->block_group_cache_tree);
 167
 168         if (info->first_logical_byte > block_group->key.objectid)
 169                 info->first_logical_byte = block_group->key.objectid;
 170
 171         spin_unlock(&info->block_group_cache_lock);
 172
 173         return 0;
 174 }
 175
 176 /*
 177  * This will return the block group at or after bytenr if contains is 0, else
 178  * it will return the block group that contains the bytenr
 179  */
 180 static struct btrfs_block_group_cache *
 181 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 182                               int contains)
 183 {
 184         struct btrfs_block_group_cache *cache, *ret = NULL;
 185         struct rb_node *n;
 186         u64 end, start;
 187
 188         spin_lock(&info->block_group_cache_lock);
 189         n = info->block_group_cache_tree.rb_node;
 190
 191         while (n) {
 192                 cache = rb_entry(n, struct btrfs_block_group_cache,
 193                                  cache_node);
 194                 end = cache->key.objectid + cache->key.offset - 1;
 195                 start = cache->key.objectid;
 196
 197                 if (bytenr < start) {
 198                         if (!contains && (!ret || start < ret->key.objectid))
 199                                 ret = cache;
 200                         n = n->rb_left;
 201                 } else if (bytenr > start) {
 202                         if (contains && bytenr <= end) {
 203                                 ret = cache;
 204                                 break;
 205                         }
 206                         n = n->rb_right;
 207                 } else {
 208                         ret = cache;
 209                         break;
 210                 }
 211         }
 212         if (ret) {
 213                 btrfs_get_block_group(ret);
 214                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 215                         info->first_logical_byte = ret->key.objectid;
 216         }
 217         spin_unlock(&info->block_group_cache_lock);
 218
 219         return ret;
 220 }
 221
 222 static int add_excluded_extent(struct btrfs_root *root,
 223                                u64 start, u64 num_bytes)
 224 {
 225         u64 end = start + num_bytes - 1;
 226         set_extent_bits(&root->fs_info->freed_extents[0],
 227                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 228         set_extent_bits(&root->fs_info->freed_extents[1],
 229                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 230         return 0;
 231 }
 232
 233 static void free_excluded_extents(struct btrfs_root *root,
 234                                   struct btrfs_block_group_cache *cache)
 235 {
 236         u64 start, end;
 237
 238         start = cache->key.objectid;
 239         end = start + cache->key.offset - 1;
 240
 241         clear_extent_bits(&root->fs_info->freed_extents[0],
 242                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 243         clear_extent_bits(&root->fs_info->freed_extents[1],
 244                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 245 }
 246
 247 static int exclude_super_stripes(struct btrfs_root *root,
 248                                  struct btrfs_block_group_cache *cache)
 249 {
 250         u64 bytenr;
 251         u64 *logical;
 252         int stripe_len;
 253         int i, nr, ret;
 254
 255         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 256                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 257                 cache->bytes_super += stripe_len;
 258                 ret = add_excluded_extent(root, cache->key.objectid,
 259                                           stripe_len);
 260                 if (ret)
 261                         return ret;
 262         }
 263
 264         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 265                 bytenr = btrfs_sb_offset(i);
 266                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 267                                        cache->key.objectid, bytenr,
 268                                        0, &logical, &nr, &stripe_len);
 269                 if (ret)
 270                         return ret;
 271
 272                 while (nr--) {
 273                         cache->bytes_super += stripe_len;
 274                         ret = add_excluded_extent(root, logical[nr],
 275                                                   stripe_len);
 276                         if (ret) {
 277                                 kfree(logical);
 278                                 return ret;
 279                         }
 280                 }
 281
 282                 kfree(logical);
 283         }
 284         return 0;
 285 }
 286
 287 static struct btrfs_caching_control *
 288 get_caching_control(struct btrfs_block_group_cache *cache)
 289 {
 290         struct btrfs_caching_control *ctl;
 291
 292         spin_lock(&cache->lock);
 293         if (cache->cached != BTRFS_CACHE_STARTED) {
 294                 spin_unlock(&cache->lock);
 295                 return NULL;
 296         }
 297
 298         /* We're loading it the fast way, so we don't have a caching_ctl. */
 299         if (!cache->caching_ctl) {
 300                 spin_unlock(&cache->lock);
 301                 return NULL;
 302         }
 303
 304         ctl = cache->caching_ctl;
 305         atomic_inc(&ctl->count);
 306         spin_unlock(&cache->lock);
 307         return ctl;
 308 }
 309
 310 static void put_caching_control(struct btrfs_caching_control *ctl)
 311 {
 312         if (atomic_dec_and_test(&ctl->count))
 313                 kfree(ctl);
 314 }
 315
 316 /*
 317  * this is only called by cache_block_group, since we could have freed extents
 318  * we need to check the pinned_extents for any extents that can't be used yet
 319  * since their free space will be released as soon as the transaction commits.
 320  */
 321 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 322                               struct btrfs_fs_info *info, u64 start, u64 end)
 323 {
 324         u64 extent_start, extent_end, size, total_added = 0;
 325         int ret;
 326
 327         while (start < end) {
 328                 ret = find_first_extent_bit(info->pinned_extents, start,
 329                                             &extent_start, &extent_end,
 330                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 331                                             NULL);
 332                 if (ret)
 333                         break;
 334
 335                 if (extent_start <= start) {
 336                         start = extent_end + 1;
 337                 } else if (extent_start > start && extent_start < end) {
 338                         size = extent_start - start;
 339                         total_added += size;
 340                         ret = btrfs_add_free_space(block_group, start,
 341                                                    size);
 342                         BUG_ON(ret); /* -ENOMEM or logic error */
 343                         start = extent_end + 1;
 344                 } else {
 345                         break;
 346                 }
 347         }
 348
 349         if (start < end) {
 350                 size = end - start;
 351                 total_added += size;
 352                 ret = btrfs_add_free_space(block_group, start, size);
 353                 BUG_ON(ret); /* -ENOMEM or logic error */
 354         }
 355
 356         return total_added;
 357 }
 358
 359 static noinline void caching_thread(struct btrfs_work *work)
 360 {
 361         struct btrfs_block_group_cache *block_group;
 362         struct btrfs_fs_info *fs_info;
 363         struct btrfs_caching_control *caching_ctl;
 364         struct btrfs_root *extent_root;
 365         struct btrfs_path *path;
 366         struct extent_buffer *leaf;
 367         struct btrfs_key key;
 368         u64 total_found = 0;
 369         u64 last = 0;
 370         u32 nritems;
 371         int ret = 0;
 372
 373         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 374         block_group = caching_ctl->block_group;
 375         fs_info = block_group->fs_info;
 376         extent_root = fs_info->extent_root;
 377
 378         path = btrfs_alloc_path();
 379         if (!path)
 380                 goto out;
 381
 382         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 383
 384         /*
 385          * We don't want to deadlock with somebody trying to allocate a new
 386          * extent for the extent root while also trying to search the extent
 387          * root to add free space.  So we skip locking and search the commit
 388          * root, since its read-only
 389          */
 390         path->skip_locking = 1;
 391         path->search_commit_root = 1;
 392         path->reada = 1;
 393
 394         key.objectid = last;
 395         key.offset = 0;
 396         key.type = BTRFS_EXTENT_ITEM_KEY;
 397 again:
 398         mutex_lock(&caching_ctl->mutex);
 399         /* need to make sure the commit_root doesn't disappear */
 400         down_read(&fs_info->extent_commit_sem);
 401
 402         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 403         if (ret < 0)
 404                 goto err;
 405
 406         leaf = path->nodes[0];
 407         nritems = btrfs_header_nritems(leaf);
 408
 409         while (1) {
 410                 if (btrfs_fs_closing(fs_info) > 1) {
 411                         last = (u64)-1;
 412                         break;
 413                 }
 414
 415                 if (path->slots[0] < nritems) {
 416                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 417                 } else {
 418                         ret = find_next_key(path, 0, &key);
 419                         if (ret)
 420                                 break;
 421
 422                         if (need_resched() ||
 423                             btrfs_next_leaf(extent_root, path)) {
 424                                 caching_ctl->progress = last;
 425                                 btrfs_release_path(path);
 426                                 up_read(&fs_info->extent_commit_sem);
 427                                 mutex_unlock(&caching_ctl->mutex);
 428                                 cond_resched();
 429                                 goto again;
 430                         }
 431                         leaf = path->nodes[0];
 432                         nritems = btrfs_header_nritems(leaf);
 433                         continue;
 434                 }
 435
 436                 if (key.objectid < block_group->key.objectid) {
 437                         path->slots[0]++;
 438                         continue;
 439                 }
 440
 441                 if (key.objectid >= block_group->key.objectid +
 442                     block_group->key.offset)
 443                         break;
 444
 445                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 446                     key.type == BTRFS_METADATA_ITEM_KEY) {
 447                         total_found += add_new_free_space(block_group,
 448                                                           fs_info, last,
 449                                                           key.objectid);
 450                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 451                                 last = key.objectid +
 452                                         fs_info->tree_root->leafsize;
 453                         else
 454                                 last = key.objectid + key.offset;
 455
 456                         if (total_found > (1024 * 1024 * 2)) {
 457                                 total_found = 0;
 458                                 wake_up(&caching_ctl->wait);
 459                         }
 460                 }
 461                 path->slots[0]++;
 462         }
 463         ret = 0;
 464
 465         total_found += add_new_free_space(block_group, fs_info, last,
 466                                           block_group->key.objectid +
 467                                           block_group->key.offset);
 468         caching_ctl->progress = (u64)-1;
 469
 470         spin_lock(&block_group->lock);
 471         block_group->caching_ctl = NULL;
 472         block_group->cached = BTRFS_CACHE_FINISHED;
 473         spin_unlock(&block_group->lock);
 474
 475 err:
 476         btrfs_free_path(path);
 477         up_read(&fs_info->extent_commit_sem);
 478
 479         free_excluded_extents(extent_root, block_group);
 480
 481         mutex_unlock(&caching_ctl->mutex);
 482 out:
 483         wake_up(&caching_ctl->wait);
 484
 485         put_caching_control(caching_ctl);
 486         btrfs_put_block_group(block_group);
 487 }
 488
 489 static int cache_block_group(struct btrfs_block_group_cache *cache,
 490                              int load_cache_only)
 491 {
 492         DEFINE_WAIT(wait);
 493         struct btrfs_fs_info *fs_info = cache->fs_info;
 494         struct btrfs_caching_control *caching_ctl;
 495         int ret = 0;
 496
 497         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 498         if (!caching_ctl)
 499                 return -ENOMEM;
 500
 501         INIT_LIST_HEAD(&caching_ctl->list);
 502         mutex_init(&caching_ctl->mutex);
 503         init_waitqueue_head(&caching_ctl->wait);
 504         caching_ctl->block_group = cache;
 505         caching_ctl->progress = cache->key.objectid;
 506         atomic_set(&caching_ctl->count, 1);
 507         caching_ctl->work.func = caching_thread;
 508
 509         spin_lock(&cache->lock);
 510         /*
 511          * This should be a rare occasion, but this could happen I think in the
 512          * case where one thread starts to load the space cache info, and then
 513          * some other thread starts a transaction commit which tries to do an
 514          * allocation while the other thread is still loading the space cache
 515          * info.  The previous loop should have kept us from choosing this block
 516          * group, but if we've moved to the state where we will wait on caching
 517          * block groups we need to first check if we're doing a fast load here,
 518          * so we can wait for it to finish, otherwise we could end up allocating
 519          * from a block group who's cache gets evicted for one reason or
 520          * another.
 521          */
 522         while (cache->cached == BTRFS_CACHE_FAST) {
 523                 struct btrfs_caching_control *ctl;
 524
 525                 ctl = cache->caching_ctl;
 526                 atomic_inc(&ctl->count);
 527                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 528                 spin_unlock(&cache->lock);
 529
 530                 schedule();
 531
 532                 finish_wait(&ctl->wait, &wait);
 533                 put_caching_control(ctl);
 534                 spin_lock(&cache->lock);
 535         }
 536
 537         if (cache->cached != BTRFS_CACHE_NO) {
 538                 spin_unlock(&cache->lock);
 539                 kfree(caching_ctl);
 540                 return 0;
 541         }
 542         WARN_ON(cache->caching_ctl);
 543         cache->caching_ctl = caching_ctl;
 544         cache->cached = BTRFS_CACHE_FAST;
 545         spin_unlock(&cache->lock);
 546
 547         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 548                 ret = load_free_space_cache(fs_info, cache);
 549
 550                 spin_lock(&cache->lock);
 551                 if (ret == 1) {
 552                         cache->caching_ctl = NULL;
 553                         cache->cached = BTRFS_CACHE_FINISHED;
 554                         cache->last_byte_to_unpin = (u64)-1;
 555                 } else {
 556                         if (load_cache_only) {
 557                                 cache->caching_ctl = NULL;
 558                                 cache->cached = BTRFS_CACHE_NO;
 559                         } else {
 560                                 cache->cached = BTRFS_CACHE_STARTED;
 561                         }
 562                 }
 563                 spin_unlock(&cache->lock);
 564                 wake_up(&caching_ctl->wait);
 565                 if (ret == 1) {
 566                         put_caching_control(caching_ctl);
 567                         free_excluded_extents(fs_info->extent_root, cache);
 568                         return 0;
 569                 }
 570         } else {
 571                 /*
 572                  * We are not going to do the fast caching, set cached to the
 573                  * appropriate value and wakeup any waiters.
 574                  */
 575                 spin_lock(&cache->lock);
 576                 if (load_cache_only) {
 577                         cache->caching_ctl = NULL;
 578                         cache->cached = BTRFS_CACHE_NO;
 579                 } else {
 580                         cache->cached = BTRFS_CACHE_STARTED;
 581                 }
 582                 spin_unlock(&cache->lock);
 583                 wake_up(&caching_ctl->wait);
 584         }
 585
 586         if (load_cache_only) {
 587                 put_caching_control(caching_ctl);
 588                 return 0;
 589         }
 590
 591         down_write(&fs_info->extent_commit_sem);
 592         atomic_inc(&caching_ctl->count);
 593         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 594         up_write(&fs_info->extent_commit_sem);
 595
 596         btrfs_get_block_group(cache);
 597
 598         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 599
 600         return ret;
 601 }
 602
 603 /*
 604  * return the block group that starts at or after bytenr
 605  */
 606 static struct btrfs_block_group_cache *
 607 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 608 {
 609         struct btrfs_block_group_cache *cache;
 610
 611         cache = block_group_cache_tree_search(info, bytenr, 0);
 612
 613         return cache;
 614 }
 615
 616 /*
 617  * return the block group that contains the given bytenr
 618  */
 619 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 620                                                  struct btrfs_fs_info *info,
 621                                                  u64 bytenr)
 622 {
 623         struct btrfs_block_group_cache *cache;
 624
 625         cache = block_group_cache_tree_search(info, bytenr, 1);
 626
 627         return cache;
 628 }
 629
 630 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 631                                                   u64 flags)
 632 {
 633         struct list_head *head = &info->space_info;
 634         struct btrfs_space_info *found;
 635
 636         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 637
 638         rcu_read_lock();
 639         list_for_each_entry_rcu(found, head, list) {
 640                 if (found->flags & flags) {
 641                         rcu_read_unlock();
 642                         return found;
 643                 }
 644         }
 645         rcu_read_unlock();
 646         return NULL;
 647 }
 648
 649 /*
 650  * after adding space to the filesystem, we need to clear the full flags
 651  * on all the space infos.
 652  */
 653 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 654 {
 655         struct list_head *head = &info->space_info;
 656         struct btrfs_space_info *found;
 657
 658         rcu_read_lock();
 659         list_for_each_entry_rcu(found, head, list)
 660                 found->full = 0;
 661         rcu_read_unlock();
 662 }
 663
 664 u64 btrfs_find_block_group(struct btrfs_root *root,
 665                            u64 search_start, u64 search_hint, int owner)
 666 {
 667         struct btrfs_block_group_cache *cache;
 668         u64 used;
 669         u64 last = max(search_hint, search_start);
 670         u64 group_start = 0;
 671         int full_search = 0;
 672         int factor = 9;
 673         int wrapped = 0;
 674 again:
 675         while (1) {
 676                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 677                 if (!cache)
 678                         break;
 679
 680                 spin_lock(&cache->lock);
 681                 last = cache->key.objectid + cache->key.offset;
 682                 used = btrfs_block_group_used(&cache->item);
 683
 684                 if ((full_search || !cache->ro) &&
 685                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 686                         if (used + cache->pinned + cache->reserved <
 687                             div_factor(cache->key.offset, factor)) {
 688                                 group_start = cache->key.objectid;
 689                                 spin_unlock(&cache->lock);
 690                                 btrfs_put_block_group(cache);
 691                                 goto found;
 692                         }
 693                 }
 694                 spin_unlock(&cache->lock);
 695                 btrfs_put_block_group(cache);
 696                 cond_resched();
 697         }
 698         if (!wrapped) {
 699                 last = search_start;
 700                 wrapped = 1;
 701                 goto again;
 702         }
 703         if (!full_search && factor < 10) {
 704                 last = search_start;
 705                 full_search = 1;
 706                 factor = 10;
 707                 goto again;
 708         }
 709 found:
 710         return group_start;
 711 }
 712
 713 /* simple helper to search for an existing extent at a given offset */
 714 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 715 {
 716         int ret;
 717         struct btrfs_key key;
 718         struct btrfs_path *path;
 719
 720         path = btrfs_alloc_path();
 721         if (!path)
 722                 return -ENOMEM;
 723
 724         key.objectid = start;
 725         key.offset = len;
 726         key.type = BTRFS_EXTENT_ITEM_KEY;
 727         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 728                                 0, 0);
 729         if (ret > 0) {
 730                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 731                 if (key.objectid == start &&
 732                     key.type == BTRFS_METADATA_ITEM_KEY)
 733                         ret = 0;
 734         }
 735         btrfs_free_path(path);
 736         return ret;
 737 }
 738
 739 /*
 740  * helper function to lookup reference count and flags of a tree block.
 741  *
 742  * the head node for delayed ref is used to store the sum of all the
 743  * reference count modifications queued up in the rbtree. the head
 744  * node may also store the extent flags to set. This way you can check
 745  * to see what the reference count and extent flags would be if all of
 746  * the delayed refs are not processed.
 747  */
 748 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 749                              struct btrfs_root *root, u64 bytenr,
 750                              u64 offset, int metadata, u64 *refs, u64 *flags)
 751 {
 752         struct btrfs_delayed_ref_head *head;
 753         struct btrfs_delayed_ref_root *delayed_refs;
 754         struct btrfs_path *path;
 755         struct btrfs_extent_item *ei;
 756         struct extent_buffer *leaf;
 757         struct btrfs_key key;
 758         u32 item_size;
 759         u64 num_refs;
 760         u64 extent_flags;
 761         int ret;
 762
 763         /*
 764          * If we don't have skinny metadata, don't bother doing anything
 765          * different
 766          */
 767         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 768                 offset = root->leafsize;
 769                 metadata = 0;
 770         }
 771
 772         path = btrfs_alloc_path();
 773         if (!path)
 774                 return -ENOMEM;
 775
 776         if (metadata) {
 777                 key.objectid = bytenr;
 778                 key.type = BTRFS_METADATA_ITEM_KEY;
 779                 key.offset = offset;
 780         } else {
 781                 key.objectid = bytenr;
 782                 key.type = BTRFS_EXTENT_ITEM_KEY;
 783                 key.offset = offset;
 784         }
 785
 786         if (!trans) {
 787                 path->skip_locking = 1;
 788                 path->search_commit_root = 1;
 789         }
 790 again:
 791         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 792                                 &key, path, 0, 0);
 793         if (ret < 0)
 794                 goto out_free;
 795
 796         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 797                 key.type = BTRFS_EXTENT_ITEM_KEY;
 798                 key.offset = root->leafsize;
 799                 btrfs_release_path(path);
 800                 goto again;
 801         }
 802
 803         if (ret == 0) {
 804                 leaf = path->nodes[0];
 805                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 806                 if (item_size >= sizeof(*ei)) {
 807                         ei = btrfs_item_ptr(leaf, path->slots[0],
 808                                             struct btrfs_extent_item);
 809                         num_refs = btrfs_extent_refs(leaf, ei);
 810                         extent_flags = btrfs_extent_flags(leaf, ei);
 811                 } else {
 812 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 813                         struct btrfs_extent_item_v0 *ei0;
 814                         BUG_ON(item_size != sizeof(*ei0));
 815                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 816                                              struct btrfs_extent_item_v0);
 817                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 818                         /* FIXME: this isn't correct for data */
 819                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 820 #else
 821                         BUG();
 822 #endif
 823                 }
 824                 BUG_ON(num_refs == 0);
 825         } else {
 826                 num_refs = 0;
 827                 extent_flags = 0;
 828                 ret = 0;
 829         }
 830
 831         if (!trans)
 832                 goto out;
 833
 834         delayed_refs = &trans->transaction->delayed_refs;
 835         spin_lock(&delayed_refs->lock);
 836         head = btrfs_find_delayed_ref_head(trans, bytenr);
 837         if (head) {
 838                 if (!mutex_trylock(&head->mutex)) {
 839                         atomic_inc(&head->node.refs);
 840                         spin_unlock(&delayed_refs->lock);
 841
 842                         btrfs_release_path(path);
 843
 844                         /*
 845                          * Mutex was contended, block until it's released and try
 846                          * again
 847                          */
 848                         mutex_lock(&head->mutex);
 849                         mutex_unlock(&head->mutex);
 850                         btrfs_put_delayed_ref(&head->node);
 851                         goto again;
 852                 }
 853                 if (head->extent_op && head->extent_op->update_flags)
 854                         extent_flags |= head->extent_op->flags_to_set;
 855                 else
 856                         BUG_ON(num_refs == 0);
 857
 858                 num_refs += head->node.ref_mod;
 859                 mutex_unlock(&head->mutex);
 860         }
 861         spin_unlock(&delayed_refs->lock);
 862 out:
 863         WARN_ON(num_refs == 0);
 864         if (refs)
 865                 *refs = num_refs;
 866         if (flags)
 867                 *flags = extent_flags;
 868 out_free:
 869         btrfs_free_path(path);
 870         return ret;
 871 }
 872
 873 /*
 874  * Back reference rules.  Back refs have three main goals:
 875  *
 876  * 1) differentiate between all holders of references to an extent so that
 877  *    when a reference is dropped we can make sure it was a valid reference
 878  *    before freeing the extent.
 879  *
 880  * 2) Provide enough information to quickly find the holders of an extent
 881  *    if we notice a given block is corrupted or bad.
 882  *
 883  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 884  *    maintenance.  This is actually the same as #2, but with a slightly
 885  *    different use case.
 886  *
 887  * There are two kinds of back refs. The implicit back refs is optimized
 888  * for pointers in non-shared tree blocks. For a given pointer in a block,
 889  * back refs of this kind provide information about the block's owner tree
 890  * and the pointer's key. These information allow us to find the block by
 891  * b-tree searching. The full back refs is for pointers in tree blocks not
 892  * referenced by their owner trees. The location of tree block is recorded
 893  * in the back refs. Actually the full back refs is generic, and can be
 894  * used in all cases the implicit back refs is used. The major shortcoming
 895  * of the full back refs is its overhead. Every time a tree block gets
 896  * COWed, we have to update back refs entry for all pointers in it.
 897  *
 898  * For a newly allocated tree block, we use implicit back refs for
 899  * pointers in it. This means most tree related operations only involve
 900  * implicit back refs. For a tree block created in old transaction, the
 901  * only way to drop a reference to it is COW it. So we can detect the
 902  * event that tree block loses its owner tree's reference and do the
 903  * back refs conversion.
 904  *
 905  * When a tree block is COW'd through a tree, there are four cases:
 906  *
 907  * The reference count of the block is one and the tree is the block's
 908  * owner tree. Nothing to do in this case.
 909  *
 910  * The reference count of the block is one and the tree is not the
 911  * block's owner tree. In this case, full back refs is used for pointers
 912  * in the block. Remove these full back refs, add implicit back refs for
 913  * every pointers in the new block.
 914  *
 915  * The reference count of the block is greater than one and the tree is
 916  * the block's owner tree. In this case, implicit back refs is used for
 917  * pointers in the block. Add full back refs for every pointers in the
 918  * block, increase lower level extents' reference counts. The original
 919  * implicit back refs are entailed to the new block.
 920  *
 921  * The reference count of the block is greater than one and the tree is
 922  * not the block's owner tree. Add implicit back refs for every pointer in
 923  * the new block, increase lower level extents' reference count.
 924  *
 925  * Back Reference Key composing:
 926  *
 927  * The key objectid corresponds to the first byte in the extent,
 928  * The key type is used to differentiate between types of back refs.
 929  * There are different meanings of the key offset for different types
 930  * of back refs.
 931  *
 932  * File extents can be referenced by:
 933  *
 934  * - multiple snapshots, subvolumes, or different generations in one subvol
 935  * - different files inside a single subvolume
 936  * - different offsets inside a file (bookend extents in file.c)
 937  *
 938  * The extent ref structure for the implicit back refs has fields for:
 939  *
 940  * - Objectid of the subvolume root
 941  * - objectid of the file holding the reference
 942  * - original offset in the file
 943  * - how many bookend extents
 944  *
 945  * The key offset for the implicit back refs is hash of the first
 946  * three fields.
 947  *
 948  * The extent ref structure for the full back refs has field for:
 949  *
 950  * - number of pointers in the tree leaf
 951  *
 952  * The key offset for the implicit back refs is the first byte of
 953  * the tree leaf
 954  *
 955  * When a file extent is allocated, The implicit back refs is used.
 956  * the fields are filled in:
 957  *
 958  *     (root_key.objectid, inode objectid, offset in file, 1)
 959  *
 960  * When a file extent is removed file truncation, we find the
 961  * corresponding implicit back refs and check the following fields:
 962  *
 963  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 964  *
 965  * Btree extents can be referenced by:
 966  *
 967  * - Different subvolumes
 968  *
 969  * Both the implicit back refs and the full back refs for tree blocks
 970  * only consist of key. The key offset for the implicit back refs is
 971  * objectid of block's owner tree. The key offset for the full back refs
 972  * is the first byte of parent block.
 973  *
 974  * When implicit back refs is used, information about the lowest key and
 975  * level of the tree block are required. These information are stored in
 976  * tree block info structure.
 977  */
 978
 979 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 980 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 981                                   struct btrfs_root *root,
 982                                   struct btrfs_path *path,
 983                                   u64 owner, u32 extra_size)
 984 {
 985         struct btrfs_extent_item *item;
 986         struct btrfs_extent_item_v0 *ei0;
 987         struct btrfs_extent_ref_v0 *ref0;
 988         struct btrfs_tree_block_info *bi;
 989         struct extent_buffer *leaf;
 990         struct btrfs_key key;
 991         struct btrfs_key found_key;
 992         u32 new_size = sizeof(*item);
 993         u64 refs;
 994         int ret;
 995
 996         leaf = path->nodes[0];
 997         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 998
 999         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1000         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1001                              struct btrfs_extent_item_v0);
1002         refs = btrfs_extent_refs_v0(leaf, ei0);
1003
1004         if (owner == (u64)-1) {
1005                 while (1) {
1006                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1007                                 ret = btrfs_next_leaf(root, path);
1008                                 if (ret < 0)
1009                                         return ret;
1010                                 BUG_ON(ret > 0); /* Corruption */
1011                                 leaf = path->nodes[0];
1012                         }
1013                         btrfs_item_key_to_cpu(leaf, &found_key,
1014                                               path->slots[0]);
1015                         BUG_ON(key.objectid != found_key.objectid);
1016                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1017                                 path->slots[0]++;
1018                                 continue;
1019                         }
1020                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1021                                               struct btrfs_extent_ref_v0);
1022                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1023                         break;
1024                 }
1025         }
1026         btrfs_release_path(path);
1027
1028         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1029                 new_size += sizeof(*bi);
1030
1031         new_size -= sizeof(*ei0);
1032         ret = btrfs_search_slot(trans, root, &key, path,
1033                                 new_size + extra_size, 1);
1034         if (ret < 0)
1035                 return ret;
1036         BUG_ON(ret); /* Corruption */
1037
1038         btrfs_extend_item(root, path, new_size);
1039
1040         leaf = path->nodes[0];
1041         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1042         btrfs_set_extent_refs(leaf, item, refs);
1043         /* FIXME: get real generation */
1044         btrfs_set_extent_generation(leaf, item, 0);
1045         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1046                 btrfs_set_extent_flags(leaf, item,
1047                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1048                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1049                 bi = (struct btrfs_tree_block_info *)(item + 1);
1050                 /* FIXME: get first key of the block */
1051                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1052                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1053         } else {
1054                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1055         }
1056         btrfs_mark_buffer_dirty(leaf);
1057         return 0;
1058 }
1059 #endif
1060
1061 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1062 {
1063         u32 high_crc = ~(u32)0;
1064         u32 low_crc = ~(u32)0;
1065         __le64 lenum;
1066
1067         lenum = cpu_to_le64(root_objectid);
1068         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1069         lenum = cpu_to_le64(owner);
1070         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1071         lenum = cpu_to_le64(offset);
1072         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1073
1074         return ((u64)high_crc << 31) ^ (u64)low_crc;
1075 }
1076
1077 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1078                                      struct btrfs_extent_data_ref *ref)
1079 {
1080         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1081                                     btrfs_extent_data_ref_objectid(leaf, ref),
1082                                     btrfs_extent_data_ref_offset(leaf, ref));
1083 }
1084
1085 static int match_extent_data_ref(struct extent_buffer *leaf,
1086                                  struct btrfs_extent_data_ref *ref,
1087                                  u64 root_objectid, u64 owner, u64 offset)
1088 {
1089         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1090             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1091             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1092                 return 0;
1093         return 1;
1094 }
1095
1096 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1097                                            struct btrfs_root *root,
1098                                            struct btrfs_path *path,
1099                                            u64 bytenr, u64 parent,
1100                                            u64 root_objectid,
1101                                            u64 owner, u64 offset)
1102 {
1103         struct btrfs_key key;
1104         struct btrfs_extent_data_ref *ref;
1105         struct extent_buffer *leaf;
1106         u32 nritems;
1107         int ret;
1108         int recow;
1109         int err = -ENOENT;
1110
1111         key.objectid = bytenr;
1112         if (parent) {
1113                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1114                 key.offset = parent;
1115         } else {
1116                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1117                 key.offset = hash_extent_data_ref(root_objectid,
1118                                                   owner, offset);
1119         }
1120 again:
1121         recow = 0;
1122         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1123         if (ret < 0) {
1124                 err = ret;
1125                 goto fail;
1126         }
1127
1128         if (parent) {
1129                 if (!ret)
1130                         return 0;
1131 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1132                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1133                 btrfs_release_path(path);
1134                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1135                 if (ret < 0) {
1136                         err = ret;
1137                         goto fail;
1138                 }
1139                 if (!ret)
1140                         return 0;
1141 #endif
1142                 goto fail;
1143         }
1144
1145         leaf = path->nodes[0];
1146         nritems = btrfs_header_nritems(leaf);
1147         while (1) {
1148                 if (path->slots[0] >= nritems) {
1149                         ret = btrfs_next_leaf(root, path);
1150                         if (ret < 0)
1151                                 err = ret;
1152                         if (ret)
1153                                 goto fail;
1154
1155                         leaf = path->nodes[0];
1156                         nritems = btrfs_header_nritems(leaf);
1157                         recow = 1;
1158                 }
1159
1160                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1161                 if (key.objectid != bytenr ||
1162                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1163                         goto fail;
1164
1165                 ref = btrfs_item_ptr(leaf, path->slots[0],
1166                                      struct btrfs_extent_data_ref);
1167
1168                 if (match_extent_data_ref(leaf, ref, root_objectid,
1169                                           owner, offset)) {
1170                         if (recow) {
1171                                 btrfs_release_path(path);
1172                                 goto again;
1173                         }
1174                         err = 0;
1175                         break;
1176                 }
1177                 path->slots[0]++;
1178         }
1179 fail:
1180         return err;
1181 }
1182
1183 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1184                                            struct btrfs_root *root,
1185                                            struct btrfs_path *path,
1186                                            u64 bytenr, u64 parent,
1187                                            u64 root_objectid, u64 owner,
1188                                            u64 offset, int refs_to_add)
1189 {
1190         struct btrfs_key key;
1191         struct extent_buffer *leaf;
1192         u32 size;
1193         u32 num_refs;
1194         int ret;
1195
1196         key.objectid = bytenr;
1197         if (parent) {
1198                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1199                 key.offset = parent;
1200                 size = sizeof(struct btrfs_shared_data_ref);
1201         } else {
1202                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1203                 key.offset = hash_extent_data_ref(root_objectid,
1204                                                   owner, offset);
1205                 size = sizeof(struct btrfs_extent_data_ref);
1206         }
1207
1208         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1209         if (ret && ret != -EEXIST)
1210                 goto fail;
1211
1212         leaf = path->nodes[0];
1213         if (parent) {
1214                 struct btrfs_shared_data_ref *ref;
1215                 ref = btrfs_item_ptr(leaf, path->slots[0],
1216                                      struct btrfs_shared_data_ref);
1217                 if (ret == 0) {
1218                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1219                 } else {
1220                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1221                         num_refs += refs_to_add;
1222                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1223                 }
1224         } else {
1225                 struct btrfs_extent_data_ref *ref;
1226                 while (ret == -EEXIST) {
1227                         ref = btrfs_item_ptr(leaf, path->slots[0],
1228                                              struct btrfs_extent_data_ref);
1229                         if (match_extent_data_ref(leaf, ref, root_objectid,
1230                                                   owner, offset))
1231                                 break;
1232                         btrfs_release_path(path);
1233                         key.offset++;
1234                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1235                                                       size);
1236                         if (ret && ret != -EEXIST)
1237                                 goto fail;
1238
1239                         leaf = path->nodes[0];
1240                 }
1241                 ref = btrfs_item_ptr(leaf, path->slots[0],
1242                                      struct btrfs_extent_data_ref);
1243                 if (ret == 0) {
1244                         btrfs_set_extent_data_ref_root(leaf, ref,
1245                                                        root_objectid);
1246                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1247                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1248                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1249                 } else {
1250                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1251                         num_refs += refs_to_add;
1252                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1253                 }
1254         }
1255         btrfs_mark_buffer_dirty(leaf);
1256         ret = 0;
1257 fail:
1258         btrfs_release_path(path);
1259         return ret;
1260 }
1261
1262 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1263                                            struct btrfs_root *root,
1264                                            struct btrfs_path *path,
1265                                            int refs_to_drop)
1266 {
1267         struct btrfs_key key;
1268         struct btrfs_extent_data_ref *ref1 = NULL;
1269         struct btrfs_shared_data_ref *ref2 = NULL;
1270         struct extent_buffer *leaf;
1271         u32 num_refs = 0;
1272         int ret = 0;
1273
1274         leaf = path->nodes[0];
1275         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1276
1277         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1278                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1279                                       struct btrfs_extent_data_ref);
1280                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1281         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1282                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1283                                       struct btrfs_shared_data_ref);
1284                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1285 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1286         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1287                 struct btrfs_extent_ref_v0 *ref0;
1288                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1289                                       struct btrfs_extent_ref_v0);
1290                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1291 #endif
1292         } else {
1293                 BUG();
1294         }
1295
1296         BUG_ON(num_refs < refs_to_drop);
1297         num_refs -= refs_to_drop;
1298
1299         if (num_refs == 0) {
1300                 ret = btrfs_del_item(trans, root, path);
1301         } else {
1302                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1303                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1304                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1305                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1306 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1307                 else {
1308                         struct btrfs_extent_ref_v0 *ref0;
1309                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1310                                         struct btrfs_extent_ref_v0);
1311                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1312                 }
1313 #endif
1314                 btrfs_mark_buffer_dirty(leaf);
1315         }
1316         return ret;
1317 }
1318
1319 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1320                                           struct btrfs_path *path,
1321                                           struct btrfs_extent_inline_ref *iref)
1322 {
1323         struct btrfs_key key;
1324         struct extent_buffer *leaf;
1325         struct btrfs_extent_data_ref *ref1;
1326         struct btrfs_shared_data_ref *ref2;
1327         u32 num_refs = 0;
1328
1329         leaf = path->nodes[0];
1330         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1331         if (iref) {
1332                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1333                     BTRFS_EXTENT_DATA_REF_KEY) {
1334                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1335                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1336                 } else {
1337                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1338                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1339                 }
1340         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1341                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1342                                       struct btrfs_extent_data_ref);
1343                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1344         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1345                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1346                                       struct btrfs_shared_data_ref);
1347                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1348 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1349         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1350                 struct btrfs_extent_ref_v0 *ref0;
1351                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1352                                       struct btrfs_extent_ref_v0);
1353                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1354 #endif
1355         } else {
1356                 WARN_ON(1);
1357         }
1358         return num_refs;
1359 }
1360
1361 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1362                                           struct btrfs_root *root,
1363                                           struct btrfs_path *path,
1364                                           u64 bytenr, u64 parent,
1365                                           u64 root_objectid)
1366 {
1367         struct btrfs_key key;
1368         int ret;
1369
1370         key.objectid = bytenr;
1371         if (parent) {
1372                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1373                 key.offset = parent;
1374         } else {
1375                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1376                 key.offset = root_objectid;
1377         }
1378
1379         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1380         if (ret > 0)
1381                 ret = -ENOENT;
1382 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1383         if (ret == -ENOENT && parent) {
1384                 btrfs_release_path(path);
1385                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1386                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1387                 if (ret > 0)
1388                         ret = -ENOENT;
1389         }
1390 #endif
1391         return ret;
1392 }
1393
1394 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1395                                           struct btrfs_root *root,
1396                                           struct btrfs_path *path,
1397                                           u64 bytenr, u64 parent,
1398                                           u64 root_objectid)
1399 {
1400         struct btrfs_key key;
1401         int ret;
1402
1403         key.objectid = bytenr;
1404         if (parent) {
1405                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1406                 key.offset = parent;
1407         } else {
1408                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1409                 key.offset = root_objectid;
1410         }
1411
1412         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1413         btrfs_release_path(path);
1414         return ret;
1415 }
1416
1417 static inline int extent_ref_type(u64 parent, u64 owner)
1418 {
1419         int type;
1420         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1421                 if (parent > 0)
1422                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1423                 else
1424                         type = BTRFS_TREE_BLOCK_REF_KEY;
1425         } else {
1426                 if (parent > 0)
1427                         type = BTRFS_SHARED_DATA_REF_KEY;
1428                 else
1429                         type = BTRFS_EXTENT_DATA_REF_KEY;
1430         }
1431         return type;
1432 }
1433
1434 static int find_next_key(struct btrfs_path *path, int level,
1435                          struct btrfs_key *key)
1436
1437 {
1438         for (; level < BTRFS_MAX_LEVEL; level++) {
1439                 if (!path->nodes[level])
1440                         break;
1441                 if (path->slots[level] + 1 >=
1442                     btrfs_header_nritems(path->nodes[level]))
1443                         continue;
1444                 if (level == 0)
1445                         btrfs_item_key_to_cpu(path->nodes[level], key,
1446                                               path->slots[level] + 1);
1447                 else
1448                         btrfs_node_key_to_cpu(path->nodes[level], key,
1449                                               path->slots[level] + 1);
1450                 return 0;
1451         }
1452         return 1;
1453 }
1454
1455 /*
1456  * look for inline back ref. if back ref is found, *ref_ret is set
1457  * to the address of inline back ref, and 0 is returned.
1458  *
1459  * if back ref isn't found, *ref_ret is set to the address where it
1460  * should be inserted, and -ENOENT is returned.
1461  *
1462  * if insert is true and there are too many inline back refs, the path
1463  * points to the extent item, and -EAGAIN is returned.
1464  *
1465  * NOTE: inline back refs are ordered in the same way that back ref
1466  *       items in the tree are ordered.
1467  */
1468 static noinline_for_stack
1469 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1470                                  struct btrfs_root *root,
1471                                  struct btrfs_path *path,
1472                                  struct btrfs_extent_inline_ref **ref_ret,
1473                                  u64 bytenr, u64 num_bytes,
1474                                  u64 parent, u64 root_objectid,
1475                                  u64 owner, u64 offset, int insert)
1476 {
1477         struct btrfs_key key;
1478         struct extent_buffer *leaf;
1479         struct btrfs_extent_item *ei;
1480         struct btrfs_extent_inline_ref *iref;
1481         u64 flags;
1482         u64 item_size;
1483         unsigned long ptr;
1484         unsigned long end;
1485         int extra_size;
1486         int type;
1487         int want;
1488         int ret;
1489         int err = 0;
1490         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1491                                                  SKINNY_METADATA);
1492
1493         key.objectid = bytenr;
1494         key.type = BTRFS_EXTENT_ITEM_KEY;
1495         key.offset = num_bytes;
1496
1497         want = extent_ref_type(parent, owner);
1498         if (insert) {
1499                 extra_size = btrfs_extent_inline_ref_size(want);
1500                 path->keep_locks = 1;
1501         } else
1502                 extra_size = -1;
1503
1504         /*
1505          * Owner is our parent level, so we can just add one to get the level
1506          * for the block we are interested in.
1507          */
1508         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1509                 key.type = BTRFS_METADATA_ITEM_KEY;
1510                 key.offset = owner;
1511         }
1512
1513 again:
1514         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1515         if (ret < 0) {
1516                 err = ret;
1517                 goto out;
1518         }
1519
1520         /*
1521          * We may be a newly converted file system which still has the old fat
1522          * extent entries for metadata, so try and see if we have one of those.
1523          */
1524         if (ret > 0 && skinny_metadata) {
1525                 skinny_metadata = false;
1526                 if (path->slots[0]) {
1527                         path->slots[0]--;
1528                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1529                                               path->slots[0]);
1530                         if (key.objectid == bytenr &&
1531                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1532                             key.offset == num_bytes)
1533                                 ret = 0;
1534                 }
1535                 if (ret) {
1536                         key.type = BTRFS_EXTENT_ITEM_KEY;
1537                         key.offset = num_bytes;
1538                         btrfs_release_path(path);
1539                         goto again;
1540                 }
1541         }
1542
1543         if (ret && !insert) {
1544                 err = -ENOENT;
1545                 goto out;
1546         } else if (ret) {
1547                 err = -EIO;
1548                 WARN_ON(1);
1549                 goto out;
1550         }
1551
1552         leaf = path->nodes[0];
1553         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1554 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1555         if (item_size < sizeof(*ei)) {
1556                 if (!insert) {
1557                         err = -ENOENT;
1558                         goto out;
1559                 }
1560                 ret = convert_extent_item_v0(trans, root, path, owner,
1561                                              extra_size);
1562                 if (ret < 0) {
1563                         err = ret;
1564                         goto out;
1565                 }
1566                 leaf = path->nodes[0];
1567                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1568         }
1569 #endif
1570         BUG_ON(item_size < sizeof(*ei));
1571
1572         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1573         flags = btrfs_extent_flags(leaf, ei);
1574
1575         ptr = (unsigned long)(ei + 1);
1576         end = (unsigned long)ei + item_size;
1577
1578         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1579                 ptr += sizeof(struct btrfs_tree_block_info);
1580                 BUG_ON(ptr > end);
1581         }
1582
1583         err = -ENOENT;
1584         while (1) {
1585                 if (ptr >= end) {
1586                         WARN_ON(ptr > end);
1587                         break;
1588                 }
1589                 iref = (struct btrfs_extent_inline_ref *)ptr;
1590                 type = btrfs_extent_inline_ref_type(leaf, iref);
1591                 if (want < type)
1592                         break;
1593                 if (want > type) {
1594                         ptr += btrfs_extent_inline_ref_size(type);
1595                         continue;
1596                 }
1597
1598                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1599                         struct btrfs_extent_data_ref *dref;
1600                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1601                         if (match_extent_data_ref(leaf, dref, root_objectid,
1602                                                   owner, offset)) {
1603                                 err = 0;
1604                                 break;
1605                         }
1606                         if (hash_extent_data_ref_item(leaf, dref) <
1607                             hash_extent_data_ref(root_objectid, owner, offset))
1608                                 break;
1609                 } else {
1610                         u64 ref_offset;
1611                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1612                         if (parent > 0) {
1613                                 if (parent == ref_offset) {
1614                                         err = 0;
1615                                         break;
1616                                 }
1617                                 if (ref_offset < parent)
1618                                         break;
1619                         } else {
1620                                 if (root_objectid == ref_offset) {
1621                                         err = 0;
1622                                         break;
1623                                 }
1624                                 if (ref_offset < root_objectid)
1625                                         break;
1626                         }
1627                 }
1628                 ptr += btrfs_extent_inline_ref_size(type);
1629         }
1630         if (err == -ENOENT && insert) {
1631                 if (item_size + extra_size >=
1632                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1633                         err = -EAGAIN;
1634                         goto out;
1635                 }
1636                 /*
1637                  * To add new inline back ref, we have to make sure
1638                  * there is no corresponding back ref item.
1639                  * For simplicity, we just do not add new inline back
1640                  * ref if there is any kind of item for this block
1641                  */
1642                 if (find_next_key(path, 0, &key) == 0 &&
1643                     key.objectid == bytenr &&
1644                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1645                         err = -EAGAIN;
1646                         goto out;
1647                 }
1648         }
1649         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1650 out:
1651         if (insert) {
1652                 path->keep_locks = 0;
1653                 btrfs_unlock_up_safe(path, 1);
1654         }
1655         return err;
1656 }
1657
1658 /*
1659  * helper to add new inline back ref
1660  */
1661 static noinline_for_stack
1662 void setup_inline_extent_backref(struct btrfs_root *root,
1663                                  struct btrfs_path *path,
1664                                  struct btrfs_extent_inline_ref *iref,
1665                                  u64 parent, u64 root_objectid,
1666                                  u64 owner, u64 offset, int refs_to_add,
1667                                  struct btrfs_delayed_extent_op *extent_op)
1668 {
1669         struct extent_buffer *leaf;
1670         struct btrfs_extent_item *ei;
1671         unsigned long ptr;
1672         unsigned long end;
1673         unsigned long item_offset;
1674         u64 refs;
1675         int size;
1676         int type;
1677
1678         leaf = path->nodes[0];
1679         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1680         item_offset = (unsigned long)iref - (unsigned long)ei;
1681
1682         type = extent_ref_type(parent, owner);
1683         size = btrfs_extent_inline_ref_size(type);
1684
1685         btrfs_extend_item(root, path, size);
1686
1687         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1688         refs = btrfs_extent_refs(leaf, ei);
1689         refs += refs_to_add;
1690         btrfs_set_extent_refs(leaf, ei, refs);
1691         if (extent_op)
1692                 __run_delayed_extent_op(extent_op, leaf, ei);
1693
1694         ptr = (unsigned long)ei + item_offset;
1695         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1696         if (ptr < end - size)
1697                 memmove_extent_buffer(leaf, ptr + size, ptr,
1698                                       end - size - ptr);
1699
1700         iref = (struct btrfs_extent_inline_ref *)ptr;
1701         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1702         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1703                 struct btrfs_extent_data_ref *dref;
1704                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1705                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1706                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1707                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1708                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1709         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1710                 struct btrfs_shared_data_ref *sref;
1711                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1712                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1713                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1714         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1715                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1716         } else {
1717                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1718         }
1719         btrfs_mark_buffer_dirty(leaf);
1720 }
1721
1722 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1723                                  struct btrfs_root *root,
1724                                  struct btrfs_path *path,
1725                                  struct btrfs_extent_inline_ref **ref_ret,
1726                                  u64 bytenr, u64 num_bytes, u64 parent,
1727                                  u64 root_objectid, u64 owner, u64 offset)
1728 {
1729         int ret;
1730
1731         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1732                                            bytenr, num_bytes, parent,
1733                                            root_objectid, owner, offset, 0);
1734         if (ret != -ENOENT)
1735                 return ret;
1736
1737         btrfs_release_path(path);
1738         *ref_ret = NULL;
1739
1740         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1741                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1742                                             root_objectid);
1743         } else {
1744                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1745                                              root_objectid, owner, offset);
1746         }
1747         return ret;
1748 }
1749
1750 /*
1751  * helper to update/remove inline back ref
1752  */
1753 static noinline_for_stack
1754 void update_inline_extent_backref(struct btrfs_root *root,
1755                                   struct btrfs_path *path,
1756                                   struct btrfs_extent_inline_ref *iref,
1757                                   int refs_to_mod,
1758                                   struct btrfs_delayed_extent_op *extent_op)
1759 {
1760         struct extent_buffer *leaf;
1761         struct btrfs_extent_item *ei;
1762         struct btrfs_extent_data_ref *dref = NULL;
1763         struct btrfs_shared_data_ref *sref = NULL;
1764         unsigned long ptr;
1765         unsigned long end;
1766         u32 item_size;
1767         int size;
1768         int type;
1769         u64 refs;
1770
1771         leaf = path->nodes[0];
1772         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1773         refs = btrfs_extent_refs(leaf, ei);
1774         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1775         refs += refs_to_mod;
1776         btrfs_set_extent_refs(leaf, ei, refs);
1777         if (extent_op)
1778                 __run_delayed_extent_op(extent_op, leaf, ei);
1779
1780         type = btrfs_extent_inline_ref_type(leaf, iref);
1781
1782         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1783                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1784                 refs = btrfs_extent_data_ref_count(leaf, dref);
1785         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1786                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1787                 refs = btrfs_shared_data_ref_count(leaf, sref);
1788         } else {
1789                 refs = 1;
1790                 BUG_ON(refs_to_mod != -1);
1791         }
1792
1793         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1794         refs += refs_to_mod;
1795
1796         if (refs > 0) {
1797                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1798                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1799                 else
1800                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1801         } else {
1802                 size =  btrfs_extent_inline_ref_size(type);
1803                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1804                 ptr = (unsigned long)iref;
1805                 end = (unsigned long)ei + item_size;
1806                 if (ptr + size < end)
1807                         memmove_extent_buffer(leaf, ptr, ptr + size,
1808                                               end - ptr - size);
1809                 item_size -= size;
1810                 btrfs_truncate_item(root, path, item_size, 1);
1811         }
1812         btrfs_mark_buffer_dirty(leaf);
1813 }
1814
1815 static noinline_for_stack
1816 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1817                                  struct btrfs_root *root,
1818                                  struct btrfs_path *path,
1819                                  u64 bytenr, u64 num_bytes, u64 parent,
1820                                  u64 root_objectid, u64 owner,
1821                                  u64 offset, int refs_to_add,
1822                                  struct btrfs_delayed_extent_op *extent_op)
1823 {
1824         struct btrfs_extent_inline_ref *iref;
1825         int ret;
1826
1827         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1828                                            bytenr, num_bytes, parent,
1829                                            root_objectid, owner, offset, 1);
1830         if (ret == 0) {
1831                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1832                 update_inline_extent_backref(root, path, iref,
1833                                              refs_to_add, extent_op);
1834         } else if (ret == -ENOENT) {
1835                 setup_inline_extent_backref(root, path, iref, parent,
1836                                             root_objectid, owner, offset,
1837                                             refs_to_add, extent_op);
1838                 ret = 0;
1839         }
1840         return ret;
1841 }
1842
1843 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1844                                  struct btrfs_root *root,
1845                                  struct btrfs_path *path,
1846                                  u64 bytenr, u64 parent, u64 root_objectid,
1847                                  u64 owner, u64 offset, int refs_to_add)
1848 {
1849         int ret;
1850         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1851                 BUG_ON(refs_to_add != 1);
1852                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1853                                             parent, root_objectid);
1854         } else {
1855                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1856                                              parent, root_objectid,
1857                                              owner, offset, refs_to_add);
1858         }
1859         return ret;
1860 }
1861
1862 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1863                                  struct btrfs_root *root,
1864                                  struct btrfs_path *path,
1865                                  struct btrfs_extent_inline_ref *iref,
1866                                  int refs_to_drop, int is_data)
1867 {
1868         int ret = 0;
1869
1870         BUG_ON(!is_data && refs_to_drop != 1);
1871         if (iref) {
1872                 update_inline_extent_backref(root, path, iref,
1873                                              -refs_to_drop, NULL);
1874         } else if (is_data) {
1875                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1876         } else {
1877                 ret = btrfs_del_item(trans, root, path);
1878         }
1879         return ret;
1880 }
1881
1882 static int btrfs_issue_discard(struct block_device *bdev,
1883                                 u64 start, u64 len)
1884 {
1885         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1886 }
1887
1888 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1889                                 u64 num_bytes, u64 *actual_bytes)
1890 {
1891         int ret;
1892         u64 discarded_bytes = 0;
1893         struct btrfs_bio *bbio = NULL;
1894
1895
1896         /* Tell the block device(s) that the sectors can be discarded */
1897         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1898                               bytenr, &num_bytes, &bbio, 0);
1899         /* Error condition is -ENOMEM */
1900         if (!ret) {
1901                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1902                 int i;
1903
1904
1905                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1906                         if (!stripe->dev->can_discard)
1907                                 continue;
1908
1909                         ret = btrfs_issue_discard(stripe->dev->bdev,
1910                                                   stripe->physical,
1911                                                   stripe->length);
1912                         if (!ret)
1913                                 discarded_bytes += stripe->length;
1914                         else if (ret != -EOPNOTSUPP)
1915                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1916
1917                         /*
1918                          * Just in case we get back EOPNOTSUPP for some reason,
1919                          * just ignore the return value so we don't screw up
1920                          * people calling discard_extent.
1921                          */
1922                         ret = 0;
1923                 }
1924                 kfree(bbio);
1925         }
1926
1927         if (actual_bytes)
1928                 *actual_bytes = discarded_bytes;
1929
1930
1931         if (ret == -EOPNOTSUPP)
1932                 ret = 0;
1933         return ret;
1934 }
1935
1936 /* Can return -ENOMEM */
1937 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1938                          struct btrfs_root *root,
1939                          u64 bytenr, u64 num_bytes, u64 parent,
1940                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1941 {
1942         int ret;
1943         struct btrfs_fs_info *fs_info = root->fs_info;
1944
1945         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1946                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1947
1948         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1949                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1950                                         num_bytes,
1951                                         parent, root_objectid, (int)owner,
1952                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1953         } else {
1954                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1955                                         num_bytes,
1956                                         parent, root_objectid, owner, offset,
1957                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1958         }
1959         return ret;
1960 }
1961
1962 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1963                                   struct btrfs_root *root,
1964                                   u64 bytenr, u64 num_bytes,
1965                                   u64 parent, u64 root_objectid,
1966                                   u64 owner, u64 offset, int refs_to_add,
1967                                   struct btrfs_delayed_extent_op *extent_op)
1968 {
1969         struct btrfs_path *path;
1970         struct extent_buffer *leaf;
1971         struct btrfs_extent_item *item;
1972         u64 refs;
1973         int ret;
1974         int err = 0;
1975
1976         path = btrfs_alloc_path();
1977         if (!path)
1978                 return -ENOMEM;
1979
1980         path->reada = 1;
1981         path->leave_spinning = 1;
1982         /* this will setup the path even if it fails to insert the back ref */
1983         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1984                                            path, bytenr, num_bytes, parent,
1985                                            root_objectid, owner, offset,
1986                                            refs_to_add, extent_op);
1987         if (ret == 0)
1988                 goto out;
1989
1990         if (ret != -EAGAIN) {
1991                 err = ret;
1992                 goto out;
1993         }
1994
1995         leaf = path->nodes[0];
1996         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1997         refs = btrfs_extent_refs(leaf, item);
1998         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1999         if (extent_op)
2000                 __run_delayed_extent_op(extent_op, leaf, item);
2001
2002         btrfs_mark_buffer_dirty(leaf);
2003         btrfs_release_path(path);
2004
2005         path->reada = 1;
2006         path->leave_spinning = 1;
2007
2008         /* now insert the actual backref */
2009         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2010                                     path, bytenr, parent, root_objectid,
2011                                     owner, offset, refs_to_add);
2012         if (ret)
2013                 btrfs_abort_transaction(trans, root, ret);
2014 out:
2015         btrfs_free_path(path);
2016         return err;
2017 }
2018
2019 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2020                                 struct btrfs_root *root,
2021                                 struct btrfs_delayed_ref_node *node,
2022                                 struct btrfs_delayed_extent_op *extent_op,
2023                                 int insert_reserved)
2024 {
2025         int ret = 0;
2026         struct btrfs_delayed_data_ref *ref;
2027         struct btrfs_key ins;
2028         u64 parent = 0;
2029         u64 ref_root = 0;
2030         u64 flags = 0;
2031
2032         ins.objectid = node->bytenr;
2033         ins.offset = node->num_bytes;
2034         ins.type = BTRFS_EXTENT_ITEM_KEY;
2035
2036         ref = btrfs_delayed_node_to_data_ref(node);
2037         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2038                 parent = ref->parent;
2039         else
2040                 ref_root = ref->root;
2041
2042         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2043                 if (extent_op)
2044                         flags |= extent_op->flags_to_set;
2045                 ret = alloc_reserved_file_extent(trans, root,
2046                                                  parent, ref_root, flags,
2047                                                  ref->objectid, ref->offset,
2048                                                  &ins, node->ref_mod);
2049         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2050                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2051                                              node->num_bytes, parent,
2052                                              ref_root, ref->objectid,
2053                                              ref->offset, node->ref_mod,
2054                                              extent_op);
2055         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2056                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2057                                           node->num_bytes, parent,
2058                                           ref_root, ref->objectid,
2059                                           ref->offset, node->ref_mod,
2060                                           extent_op);
2061         } else {
2062                 BUG();
2063         }
2064         return ret;
2065 }
2066
2067 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2068                                     struct extent_buffer *leaf,
2069                                     struct btrfs_extent_item *ei)
2070 {
2071         u64 flags = btrfs_extent_flags(leaf, ei);
2072         if (extent_op->update_flags) {
2073                 flags |= extent_op->flags_to_set;
2074                 btrfs_set_extent_flags(leaf, ei, flags);
2075         }
2076
2077         if (extent_op->update_key) {
2078                 struct btrfs_tree_block_info *bi;
2079                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2080                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2081                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2082         }
2083 }
2084
2085 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2086                                  struct btrfs_root *root,
2087                                  struct btrfs_delayed_ref_node *node,
2088                                  struct btrfs_delayed_extent_op *extent_op)
2089 {
2090         struct btrfs_key key;
2091         struct btrfs_path *path;
2092         struct btrfs_extent_item *ei;
2093         struct extent_buffer *leaf;
2094         u32 item_size;
2095         int ret;
2096         int err = 0;
2097         int metadata = (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2098                         node->type == BTRFS_SHARED_BLOCK_REF_KEY);
2099
2100         if (trans->aborted)
2101                 return 0;
2102
2103         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2104                 metadata = 0;
2105
2106         path = btrfs_alloc_path();
2107         if (!path)
2108                 return -ENOMEM;
2109
2110         key.objectid = node->bytenr;
2111
2112         if (metadata) {
2113                 struct btrfs_delayed_tree_ref *tree_ref;
2114
2115                 tree_ref = btrfs_delayed_node_to_tree_ref(node);
2116                 key.type = BTRFS_METADATA_ITEM_KEY;
2117                 key.offset = tree_ref->level;
2118         } else {
2119                 key.type = BTRFS_EXTENT_ITEM_KEY;
2120                 key.offset = node->num_bytes;
2121         }
2122
2123 again:
2124         path->reada = 1;
2125         path->leave_spinning = 1;
2126         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2127                                 path, 0, 1);
2128         if (ret < 0) {
2129                 err = ret;
2130                 goto out;
2131         }
2132         if (ret > 0) {
2133                 if (metadata) {
2134                         btrfs_release_path(path);
2135                         metadata = 0;
2136
2137                         key.offset = node->num_bytes;
2138                         key.type = BTRFS_EXTENT_ITEM_KEY;
2139                         goto again;
2140                 }
2141                 err = -EIO;
2142                 goto out;
2143         }
2144
2145         leaf = path->nodes[0];
2146         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2147 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2148         if (item_size < sizeof(*ei)) {
2149                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2150                                              path, (u64)-1, 0);
2151                 if (ret < 0) {
2152                         err = ret;
2153                         goto out;
2154                 }
2155                 leaf = path->nodes[0];
2156                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2157         }
2158 #endif
2159         BUG_ON(item_size < sizeof(*ei));
2160         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2161         __run_delayed_extent_op(extent_op, leaf, ei);
2162
2163         btrfs_mark_buffer_dirty(leaf);
2164 out:
2165         btrfs_free_path(path);
2166         return err;
2167 }
2168
2169 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2170                                 struct btrfs_root *root,
2171                                 struct btrfs_delayed_ref_node *node,
2172                                 struct btrfs_delayed_extent_op *extent_op,
2173                                 int insert_reserved)
2174 {
2175         int ret = 0;
2176         struct btrfs_delayed_tree_ref *ref;
2177         struct btrfs_key ins;
2178         u64 parent = 0;
2179         u64 ref_root = 0;
2180         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2181                                                  SKINNY_METADATA);
2182
2183         ref = btrfs_delayed_node_to_tree_ref(node);
2184         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2185                 parent = ref->parent;
2186         else
2187                 ref_root = ref->root;
2188
2189         ins.objectid = node->bytenr;
2190         if (skinny_metadata) {
2191                 ins.offset = ref->level;
2192                 ins.type = BTRFS_METADATA_ITEM_KEY;
2193         } else {
2194                 ins.offset = node->num_bytes;
2195                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2196         }
2197
2198         BUG_ON(node->ref_mod != 1);
2199         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2200                 BUG_ON(!extent_op || !extent_op->update_flags);
2201                 ret = alloc_reserved_tree_block(trans, root,
2202                                                 parent, ref_root,
2203                                                 extent_op->flags_to_set,
2204                                                 &extent_op->key,
2205                                                 ref->level, &ins);
2206         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2207                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2208                                              node->num_bytes, parent, ref_root,
2209                                              ref->level, 0, 1, extent_op);
2210         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2211                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2212                                           node->num_bytes, parent, ref_root,
2213                                           ref->level, 0, 1, extent_op);
2214         } else {
2215                 BUG();
2216         }
2217         return ret;
2218 }
2219
2220 /* helper function to actually process a single delayed ref entry */
2221 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2222                                struct btrfs_root *root,
2223                                struct btrfs_delayed_ref_node *node,
2224                                struct btrfs_delayed_extent_op *extent_op,
2225                                int insert_reserved)
2226 {
2227         int ret = 0;
2228
2229         if (trans->aborted)
2230                 return 0;
2231
2232         if (btrfs_delayed_ref_is_head(node)) {
2233                 struct btrfs_delayed_ref_head *head;
2234                 /*
2235                  * we've hit the end of the chain and we were supposed
2236                  * to insert this extent into the tree.  But, it got
2237                  * deleted before we ever needed to insert it, so all
2238                  * we have to do is clean up the accounting
2239                  */
2240                 BUG_ON(extent_op);
2241                 head = btrfs_delayed_node_to_head(node);
2242                 if (insert_reserved) {
2243                         btrfs_pin_extent(root, node->bytenr,
2244                                          node->num_bytes, 1);
2245                         if (head->is_data) {
2246                                 ret = btrfs_del_csums(trans, root,
2247                                                       node->bytenr,
2248                                                       node->num_bytes);
2249                         }
2250                 }
2251                 return ret;
2252         }
2253
2254         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2255             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2256                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2257                                            insert_reserved);
2258         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2259                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2260                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2261                                            insert_reserved);
2262         else
2263                 BUG();
2264         return ret;
2265 }
2266
2267 static noinline struct btrfs_delayed_ref_node *
2268 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2269 {
2270         struct rb_node *node;
2271         struct btrfs_delayed_ref_node *ref;
2272         int action = BTRFS_ADD_DELAYED_REF;
2273 again:
2274         /*
2275          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2276          * this prevents ref count from going down to zero when
2277          * there still are pending delayed ref.
2278          */
2279         node = rb_prev(&head->node.rb_node);
2280         while (1) {
2281                 if (!node)
2282                         break;
2283                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2284                                 rb_node);
2285                 if (ref->bytenr != head->node.bytenr)
2286                         break;
2287                 if (ref->action == action)
2288                         return ref;
2289                 node = rb_prev(node);
2290         }
2291         if (action == BTRFS_ADD_DELAYED_REF) {
2292                 action = BTRFS_DROP_DELAYED_REF;
2293                 goto again;
2294         }
2295         return NULL;
2296 }
2297
2298 /*
2299  * Returns 0 on success or if called with an already aborted transaction.
2300  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2301  */
2302 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2303                                        struct btrfs_root *root,
2304                                        struct list_head *cluster)
2305 {
2306         struct btrfs_delayed_ref_root *delayed_refs;
2307         struct btrfs_delayed_ref_node *ref;
2308         struct btrfs_delayed_ref_head *locked_ref = NULL;
2309         struct btrfs_delayed_extent_op *extent_op;
2310         struct btrfs_fs_info *fs_info = root->fs_info;
2311         int ret;
2312         int count = 0;
2313         int must_insert_reserved = 0;
2314
2315         delayed_refs = &trans->transaction->delayed_refs;
2316         while (1) {
2317                 if (!locked_ref) {
2318                         /* pick a new head ref from the cluster list */
2319                         if (list_empty(cluster))
2320                                 break;
2321
2322                         locked_ref = list_entry(cluster->next,
2323                                      struct btrfs_delayed_ref_head, cluster);
2324
2325                         /* grab the lock that says we are going to process
2326                          * all the refs for this head */
2327                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2328
2329                         /*
2330                          * we may have dropped the spin lock to get the head
2331                          * mutex lock, and that might have given someone else
2332                          * time to free the head.  If that's true, it has been
2333                          * removed from our list and we can move on.
2334                          */
2335                         if (ret == -EAGAIN) {
2336                                 locked_ref = NULL;
2337                                 count++;
2338                                 continue;
2339                         }
2340                 }
2341
2342                 /*
2343                  * We need to try and merge add/drops of the same ref since we
2344                  * can run into issues with relocate dropping the implicit ref
2345                  * and then it being added back again before the drop can
2346                  * finish.  If we merged anything we need to re-loop so we can
2347                  * get a good ref.
2348                  */
2349                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2350                                          locked_ref);
2351
2352                 /*
2353                  * locked_ref is the head node, so we have to go one
2354                  * node back for any delayed ref updates
2355                  */
2356                 ref = select_delayed_ref(locked_ref);
2357
2358                 if (ref && ref->seq &&
2359                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2360                         /*
2361                          * there are still refs with lower seq numbers in the
2362                          * process of being added. Don't run this ref yet.
2363                          */
2364                         list_del_init(&locked_ref->cluster);
2365                         btrfs_delayed_ref_unlock(locked_ref);
2366                         locked_ref = NULL;
2367                         delayed_refs->num_heads_ready++;
2368                         spin_unlock(&delayed_refs->lock);
2369                         cond_resched();
2370                         spin_lock(&delayed_refs->lock);
2371                         continue;
2372                 }
2373
2374                 /*
2375                  * record the must insert reserved flag before we
2376                  * drop the spin lock.
2377                  */
2378                 must_insert_reserved = locked_ref->must_insert_reserved;
2379                 locked_ref->must_insert_reserved = 0;
2380
2381                 extent_op = locked_ref->extent_op;
2382                 locked_ref->extent_op = NULL;
2383
2384                 if (!ref) {
2385                         /* All delayed refs have been processed, Go ahead
2386                          * and send the head node to run_one_delayed_ref,
2387                          * so that any accounting fixes can happen
2388                          */
2389                         ref = &locked_ref->node;
2390
2391                         if (extent_op && must_insert_reserved) {
2392                                 btrfs_free_delayed_extent_op(extent_op);
2393                                 extent_op = NULL;
2394                         }
2395
2396                         if (extent_op) {
2397                                 spin_unlock(&delayed_refs->lock);
2398
2399                                 ret = run_delayed_extent_op(trans, root,
2400                                                             ref, extent_op);
2401                                 btrfs_free_delayed_extent_op(extent_op);
2402
2403                                 if (ret) {
2404                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2405                                         spin_lock(&delayed_refs->lock);
2406                                         btrfs_delayed_ref_unlock(locked_ref);
2407                                         return ret;
2408                                 }
2409
2410                                 goto next;
2411                         }
2412                 }
2413
2414                 ref->in_tree = 0;
2415                 rb_erase(&ref->rb_node, &delayed_refs->root);
2416                 delayed_refs->num_entries--;
2417                 if (!btrfs_delayed_ref_is_head(ref)) {
2418                         /*
2419                          * when we play the delayed ref, also correct the
2420                          * ref_mod on head
2421                          */
2422                         switch (ref->action) {
2423                         case BTRFS_ADD_DELAYED_REF:
2424                         case BTRFS_ADD_DELAYED_EXTENT:
2425                                 locked_ref->node.ref_mod -= ref->ref_mod;
2426                                 break;
2427                         case BTRFS_DROP_DELAYED_REF:
2428                                 locked_ref->node.ref_mod += ref->ref_mod;
2429                                 break;
2430                         default:
2431                                 WARN_ON(1);
2432                         }
2433                 }
2434                 spin_unlock(&delayed_refs->lock);
2435
2436                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2437                                           must_insert_reserved);
2438
2439                 btrfs_free_delayed_extent_op(extent_op);
2440                 if (ret) {
2441                         btrfs_delayed_ref_unlock(locked_ref);
2442                         btrfs_put_delayed_ref(ref);
2443                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2444                         spin_lock(&delayed_refs->lock);
2445                         return ret;
2446                 }
2447
2448                 /*
2449                  * If this node is a head, that means all the refs in this head
2450                  * have been dealt with, and we will pick the next head to deal
2451                  * with, so we must unlock the head and drop it from the cluster
2452                  * list before we release it.
2453                  */
2454                 if (btrfs_delayed_ref_is_head(ref)) {
2455                         list_del_init(&locked_ref->cluster);
2456                         btrfs_delayed_ref_unlock(locked_ref);
2457                         locked_ref = NULL;
2458                 }
2459                 btrfs_put_delayed_ref(ref);
2460                 count++;
2461 next:
2462                 cond_resched();
2463                 spin_lock(&delayed_refs->lock);
2464         }
2465         return count;
2466 }
2467
2468 #ifdef SCRAMBLE_DELAYED_REFS
2469 /*
2470  * Normally delayed refs get processed in ascending bytenr order. This
2471  * correlates in most cases to the order added. To expose dependencies on this
2472  * order, we start to process the tree in the middle instead of the beginning
2473  */
2474 static u64 find_middle(struct rb_root *root)
2475 {
2476         struct rb_node *n = root->rb_node;
2477         struct btrfs_delayed_ref_node *entry;
2478         int alt = 1;
2479         u64 middle;
2480         u64 first = 0, last = 0;
2481
2482         n = rb_first(root);
2483         if (n) {
2484                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2485                 first = entry->bytenr;
2486         }
2487         n = rb_last(root);
2488         if (n) {
2489                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2490                 last = entry->bytenr;
2491         }
2492         n = root->rb_node;
2493
2494         while (n) {
2495                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2496                 WARN_ON(!entry->in_tree);
2497
2498                 middle = entry->bytenr;
2499
2500                 if (alt)
2501                         n = n->rb_left;
2502                 else
2503                         n = n->rb_right;
2504
2505                 alt = 1 - alt;
2506         }
2507         return middle;
2508 }
2509 #endif
2510
2511 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2512                                          struct btrfs_fs_info *fs_info)
2513 {
2514         struct qgroup_update *qgroup_update;
2515         int ret = 0;
2516
2517         if (list_empty(&trans->qgroup_ref_list) !=
2518             !trans->delayed_ref_elem.seq) {
2519                 /* list without seq or seq without list */
2520                 btrfs_err(fs_info,
2521                         "qgroup accounting update error, list is%s empty, seq is %llu",
2522                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2523                         trans->delayed_ref_elem.seq);
2524                 BUG();
2525         }
2526
2527         if (!trans->delayed_ref_elem.seq)
2528                 return 0;
2529
2530         while (!list_empty(&trans->qgroup_ref_list)) {
2531                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2532                                                  struct qgroup_update, list);
2533                 list_del(&qgroup_update->list);
2534                 if (!ret)
2535                         ret = btrfs_qgroup_account_ref(
2536                                         trans, fs_info, qgroup_update->node,
2537                                         qgroup_update->extent_op);
2538                 kfree(qgroup_update);
2539         }
2540
2541         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2542
2543         return ret;
2544 }
2545
2546 static int refs_newer(struct btrfs_delayed_ref_root *delayed_refs, int seq,
2547                       int count)
2548 {
2549         int val = atomic_read(&delayed_refs->ref_seq);
2550
2551         if (val < seq || val >= seq + count)
2552                 return 1;
2553         return 0;
2554 }
2555
2556 /*
2557  * this starts processing the delayed reference count updates and
2558  * extent insertions we have queued up so far.  count can be
2559  * 0, which means to process everything in the tree at the start
2560  * of the run (but not newly added entries), or it can be some target
2561  * number you'd like to process.
2562  *
2563  * Returns 0 on success or if called with an aborted transaction
2564  * Returns <0 on error and aborts the transaction
2565  */
2566 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2567                            struct btrfs_root *root, unsigned long count)
2568 {
2569         struct rb_node *node;
2570         struct btrfs_delayed_ref_root *delayed_refs;
2571         struct btrfs_delayed_ref_node *ref;
2572         struct list_head cluster;
2573         int ret;
2574         u64 delayed_start;
2575         int run_all = count == (unsigned long)-1;
2576         int run_most = 0;
2577         int loops;
2578
2579         /* We'll clean this up in btrfs_cleanup_transaction */
2580         if (trans->aborted)
2581                 return 0;
2582
2583         if (root == root->fs_info->extent_root)
2584                 root = root->fs_info->tree_root;
2585
2586         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2587
2588         delayed_refs = &trans->transaction->delayed_refs;
2589         INIT_LIST_HEAD(&cluster);
2590         if (count == 0) {
2591                 count = delayed_refs->num_entries * 2;
2592                 run_most = 1;
2593         }
2594
2595         if (!run_all && !run_most) {
2596                 int old;
2597                 int seq = atomic_read(&delayed_refs->ref_seq);
2598
2599 progress:
2600                 old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2601                 if (old) {
2602                         DEFINE_WAIT(__wait);
2603                         if (delayed_refs->num_entries < 16348)
2604                                 return 0;
2605
2606                         prepare_to_wait(&delayed_refs->wait, &__wait,
2607                                         TASK_UNINTERRUPTIBLE);
2608
2609                         old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2610                         if (old) {
2611                                 schedule();
2612                                 finish_wait(&delayed_refs->wait, &__wait);
2613
2614                                 if (!refs_newer(delayed_refs, seq, 256))
2615                                         goto progress;
2616                                 else
2617                                         return 0;
2618                         } else {
2619                                 finish_wait(&delayed_refs->wait, &__wait);
2620                                 goto again;
2621                         }
2622                 }
2623
2624         } else {
2625                 atomic_inc(&delayed_refs->procs_running_refs);
2626         }
2627
2628 again:
2629         loops = 0;
2630         spin_lock(&delayed_refs->lock);
2631
2632 #ifdef SCRAMBLE_DELAYED_REFS
2633         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2634 #endif
2635
2636         while (1) {
2637                 if (!(run_all || run_most) &&
2638                     delayed_refs->num_heads_ready < 64)
2639                         break;
2640
2641                 /*
2642                  * go find something we can process in the rbtree.  We start at
2643                  * the beginning of the tree, and then build a cluster
2644                  * of refs to process starting at the first one we are able to
2645                  * lock
2646                  */
2647                 delayed_start = delayed_refs->run_delayed_start;
2648                 ret = btrfs_find_ref_cluster(trans, &cluster,
2649                                              delayed_refs->run_delayed_start);
2650                 if (ret)
2651                         break;
2652
2653                 ret = run_clustered_refs(trans, root, &cluster);
2654                 if (ret < 0) {
2655                         btrfs_release_ref_cluster(&cluster);
2656                         spin_unlock(&delayed_refs->lock);
2657                         btrfs_abort_transaction(trans, root, ret);
2658                         atomic_dec(&delayed_refs->procs_running_refs);
2659                         return ret;
2660                 }
2661
2662                 atomic_add(ret, &delayed_refs->ref_seq);
2663
2664                 count -= min_t(unsigned long, ret, count);
2665
2666                 if (count == 0)
2667                         break;
2668
2669                 if (delayed_start >= delayed_refs->run_delayed_start) {
2670                         if (loops == 0) {
2671                                 /*
2672                                  * btrfs_find_ref_cluster looped. let's do one
2673                                  * more cycle. if we don't run any delayed ref
2674                                  * during that cycle (because we can't because
2675                                  * all of them are blocked), bail out.
2676                                  */
2677                                 loops = 1;
2678                         } else {
2679                                 /*
2680                                  * no runnable refs left, stop trying
2681                                  */
2682                                 BUG_ON(run_all);
2683                                 break;
2684                         }
2685                 }
2686                 if (ret) {
2687                         /* refs were run, let's reset staleness detection */
2688                         loops = 0;
2689                 }
2690         }
2691
2692         if (run_all) {
2693                 if (!list_empty(&trans->new_bgs)) {
2694                         spin_unlock(&delayed_refs->lock);
2695                         btrfs_create_pending_block_groups(trans, root);
2696                         spin_lock(&delayed_refs->lock);
2697                 }
2698
2699                 node = rb_first(&delayed_refs->root);
2700                 if (!node)
2701                         goto out;
2702                 count = (unsigned long)-1;
2703
2704                 while (node) {
2705                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2706                                        rb_node);
2707                         if (btrfs_delayed_ref_is_head(ref)) {
2708                                 struct btrfs_delayed_ref_head *head;
2709
2710                                 head = btrfs_delayed_node_to_head(ref);
2711                                 atomic_inc(&ref->refs);
2712
2713                                 spin_unlock(&delayed_refs->lock);
2714                                 /*
2715                                  * Mutex was contended, block until it's
2716                                  * released and try again
2717                                  */
2718                                 mutex_lock(&head->mutex);
2719                                 mutex_unlock(&head->mutex);
2720
2721                                 btrfs_put_delayed_ref(ref);
2722                                 cond_resched();
2723                                 goto again;
2724                         }
2725                         node = rb_next(node);
2726                 }
2727                 spin_unlock(&delayed_refs->lock);
2728                 schedule_timeout(1);
2729                 goto again;
2730         }
2731 out:
2732         atomic_dec(&delayed_refs->procs_running_refs);
2733         smp_mb();
2734         if (waitqueue_active(&delayed_refs->wait))
2735                 wake_up(&delayed_refs->wait);
2736
2737         spin_unlock(&delayed_refs->lock);
2738         assert_qgroups_uptodate(trans);
2739         return 0;
2740 }
2741
2742 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2743                                 struct btrfs_root *root,
2744                                 u64 bytenr, u64 num_bytes, u64 flags,
2745                                 int is_data)
2746 {
2747         struct btrfs_delayed_extent_op *extent_op;
2748         int ret;
2749
2750         extent_op = btrfs_alloc_delayed_extent_op();
2751         if (!extent_op)
2752                 return -ENOMEM;
2753
2754         extent_op->flags_to_set = flags;
2755         extent_op->update_flags = 1;
2756         extent_op->update_key = 0;
2757         extent_op->is_data = is_data ? 1 : 0;
2758
2759         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2760                                           num_bytes, extent_op);
2761         if (ret)
2762                 btrfs_free_delayed_extent_op(extent_op);
2763         return ret;
2764 }
2765
2766 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2767                                       struct btrfs_root *root,
2768                                       struct btrfs_path *path,
2769                                       u64 objectid, u64 offset, u64 bytenr)
2770 {
2771         struct btrfs_delayed_ref_head *head;
2772         struct btrfs_delayed_ref_node *ref;
2773         struct btrfs_delayed_data_ref *data_ref;
2774         struct btrfs_delayed_ref_root *delayed_refs;
2775         struct rb_node *node;
2776         int ret = 0;
2777
2778         ret = -ENOENT;
2779         delayed_refs = &trans->transaction->delayed_refs;
2780         spin_lock(&delayed_refs->lock);
2781         head = btrfs_find_delayed_ref_head(trans, bytenr);
2782         if (!head)
2783                 goto out;
2784
2785         if (!mutex_trylock(&head->mutex)) {
2786                 atomic_inc(&head->node.refs);
2787                 spin_unlock(&delayed_refs->lock);
2788
2789                 btrfs_release_path(path);
2790
2791                 /*
2792                  * Mutex was contended, block until it's released and let
2793                  * caller try again
2794                  */
2795                 mutex_lock(&head->mutex);
2796                 mutex_unlock(&head->mutex);
2797                 btrfs_put_delayed_ref(&head->node);
2798                 return -EAGAIN;
2799         }
2800
2801         node = rb_prev(&head->node.rb_node);
2802         if (!node)
2803                 goto out_unlock;
2804
2805         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2806
2807         if (ref->bytenr != bytenr)
2808                 goto out_unlock;
2809
2810         ret = 1;
2811         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2812                 goto out_unlock;
2813
2814         data_ref = btrfs_delayed_node_to_data_ref(ref);
2815
2816         node = rb_prev(node);
2817         if (node) {
2818                 int seq = ref->seq;
2819
2820                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2821                 if (ref->bytenr == bytenr && ref->seq == seq)
2822                         goto out_unlock;
2823         }
2824
2825         if (data_ref->root != root->root_key.objectid ||
2826             data_ref->objectid != objectid || data_ref->offset != offset)
2827                 goto out_unlock;
2828
2829         ret = 0;
2830 out_unlock:
2831         mutex_unlock(&head->mutex);
2832 out:
2833         spin_unlock(&delayed_refs->lock);
2834         return ret;
2835 }
2836
2837 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2838                                         struct btrfs_root *root,
2839                                         struct btrfs_path *path,
2840                                         u64 objectid, u64 offset, u64 bytenr)
2841 {
2842         struct btrfs_root *extent_root = root->fs_info->extent_root;
2843         struct extent_buffer *leaf;
2844         struct btrfs_extent_data_ref *ref;
2845         struct btrfs_extent_inline_ref *iref;
2846         struct btrfs_extent_item *ei;
2847         struct btrfs_key key;
2848         u32 item_size;
2849         int ret;
2850
2851         key.objectid = bytenr;
2852         key.offset = (u64)-1;
2853         key.type = BTRFS_EXTENT_ITEM_KEY;
2854
2855         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2856         if (ret < 0)
2857                 goto out;
2858         BUG_ON(ret == 0); /* Corruption */
2859
2860         ret = -ENOENT;
2861         if (path->slots[0] == 0)
2862                 goto out;
2863
2864         path->slots[0]--;
2865         leaf = path->nodes[0];
2866         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2867
2868         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2869                 goto out;
2870
2871         ret = 1;
2872         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2873 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2874         if (item_size < sizeof(*ei)) {
2875                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2876                 goto out;
2877         }
2878 #endif
2879         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2880
2881         if (item_size != sizeof(*ei) +
2882             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2883                 goto out;
2884
2885         if (btrfs_extent_generation(leaf, ei) <=
2886             btrfs_root_last_snapshot(&root->root_item))
2887                 goto out;
2888
2889         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2890         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2891             BTRFS_EXTENT_DATA_REF_KEY)
2892                 goto out;
2893
2894         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2895         if (btrfs_extent_refs(leaf, ei) !=
2896             btrfs_extent_data_ref_count(leaf, ref) ||
2897             btrfs_extent_data_ref_root(leaf, ref) !=
2898             root->root_key.objectid ||
2899             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2900             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2901                 goto out;
2902
2903         ret = 0;
2904 out:
2905         return ret;
2906 }
2907
2908 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2909                           struct btrfs_root *root,
2910                           u64 objectid, u64 offset, u64 bytenr)
2911 {
2912         struct btrfs_path *path;
2913         int ret;
2914         int ret2;
2915
2916         path = btrfs_alloc_path();
2917         if (!path)
2918                 return -ENOENT;
2919
2920         do {
2921                 ret = check_committed_ref(trans, root, path, objectid,
2922                                           offset, bytenr);
2923                 if (ret && ret != -ENOENT)
2924                         goto out;
2925
2926                 ret2 = check_delayed_ref(trans, root, path, objectid,
2927                                          offset, bytenr);
2928         } while (ret2 == -EAGAIN);
2929
2930         if (ret2 && ret2 != -ENOENT) {
2931                 ret = ret2;
2932                 goto out;
2933         }
2934
2935         if (ret != -ENOENT || ret2 != -ENOENT)
2936                 ret = 0;
2937 out:
2938         btrfs_free_path(path);
2939         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2940                 WARN_ON(ret > 0);
2941         return ret;
2942 }
2943
2944 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2945                            struct btrfs_root *root,
2946                            struct extent_buffer *buf,
2947                            int full_backref, int inc, int for_cow)
2948 {
2949         u64 bytenr;
2950         u64 num_bytes;
2951         u64 parent;
2952         u64 ref_root;
2953         u32 nritems;
2954         struct btrfs_key key;
2955         struct btrfs_file_extent_item *fi;
2956         int i;
2957         int level;
2958         int ret = 0;
2959         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2960                             u64, u64, u64, u64, u64, u64, int);
2961
2962         ref_root = btrfs_header_owner(buf);
2963         nritems = btrfs_header_nritems(buf);
2964         level = btrfs_header_level(buf);
2965
2966         if (!root->ref_cows && level == 0)
2967                 return 0;
2968
2969         if (inc)
2970                 process_func = btrfs_inc_extent_ref;
2971         else
2972                 process_func = btrfs_free_extent;
2973
2974         if (full_backref)
2975                 parent = buf->start;
2976         else
2977                 parent = 0;
2978
2979         for (i = 0; i < nritems; i++) {
2980                 if (level == 0) {
2981                         btrfs_item_key_to_cpu(buf, &key, i);
2982                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2983                                 continue;
2984                         fi = btrfs_item_ptr(buf, i,
2985                                             struct btrfs_file_extent_item);
2986                         if (btrfs_file_extent_type(buf, fi) ==
2987                             BTRFS_FILE_EXTENT_INLINE)
2988                                 continue;
2989                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2990                         if (bytenr == 0)
2991                                 continue;
2992
2993                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2994                         key.offset -= btrfs_file_extent_offset(buf, fi);
2995                         ret = process_func(trans, root, bytenr, num_bytes,
2996                                            parent, ref_root, key.objectid,
2997                                            key.offset, for_cow);
2998                         if (ret)
2999                                 goto fail;
3000                 } else {
3001                         bytenr = btrfs_node_blockptr(buf, i);
3002                         num_bytes = btrfs_level_size(root, level - 1);
3003                         ret = process_func(trans, root, bytenr, num_bytes,
3004                                            parent, ref_root, level - 1, 0,
3005                                            for_cow);
3006                         if (ret)
3007                                 goto fail;
3008                 }
3009         }
3010         return 0;
3011 fail:
3012         return ret;
3013 }
3014
3015 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3016                   struct extent_buffer *buf, int full_backref, int for_cow)
3017 {
3018         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
3019 }
3020
3021 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3022                   struct extent_buffer *buf, int full_backref, int for_cow)
3023 {
3024         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
3025 }
3026
3027 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3028                                  struct btrfs_root *root,
3029                                  struct btrfs_path *path,
3030                                  struct btrfs_block_group_cache *cache)
3031 {
3032         int ret;
3033         struct btrfs_root *extent_root = root->fs_info->extent_root;
3034         unsigned long bi;
3035         struct extent_buffer *leaf;
3036
3037         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3038         if (ret < 0)
3039                 goto fail;
3040         BUG_ON(ret); /* Corruption */
3041
3042         leaf = path->nodes[0];
3043         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3044         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3045         btrfs_mark_buffer_dirty(leaf);
3046         btrfs_release_path(path);
3047 fail:
3048         if (ret) {
3049                 btrfs_abort_transaction(trans, root, ret);
3050                 return ret;
3051         }
3052         return 0;
3053
3054 }
3055
3056 static struct btrfs_block_group_cache *
3057 next_block_group(struct btrfs_root *root,
3058                  struct btrfs_block_group_cache *cache)
3059 {
3060         struct rb_node *node;
3061         spin_lock(&root->fs_info->block_group_cache_lock);
3062         node = rb_next(&cache->cache_node);
3063         btrfs_put_block_group(cache);
3064         if (node) {
3065                 cache = rb_entry(node, struct btrfs_block_group_cache,
3066                                  cache_node);
3067                 btrfs_get_block_group(cache);
3068         } else
3069                 cache = NULL;
3070         spin_unlock(&root->fs_info->block_group_cache_lock);
3071         return cache;
3072 }
3073
3074 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3075                             struct btrfs_trans_handle *trans,
3076                             struct btrfs_path *path)
3077 {
3078         struct btrfs_root *root = block_group->fs_info->tree_root;
3079         struct inode *inode = NULL;
3080         u64 alloc_hint = 0;
3081         int dcs = BTRFS_DC_ERROR;
3082         int num_pages = 0;
3083         int retries = 0;
3084         int ret = 0;
3085
3086         /*
3087          * If this block group is smaller than 100 megs don't bother caching the
3088          * block group.
3089          */
3090         if (block_group->key.offset < (100 * 1024 * 1024)) {
3091                 spin_lock(&block_group->lock);
3092                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3093                 spin_unlock(&block_group->lock);
3094                 return 0;
3095         }
3096
3097 again:
3098         inode = lookup_free_space_inode(root, block_group, path);
3099         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3100                 ret = PTR_ERR(inode);
3101                 btrfs_release_path(path);
3102                 goto out;
3103         }
3104
3105         if (IS_ERR(inode)) {
3106                 BUG_ON(retries);
3107                 retries++;
3108
3109                 if (block_group->ro)
3110                         goto out_free;
3111
3112                 ret = create_free_space_inode(root, trans, block_group, path);
3113                 if (ret)
3114                         goto out_free;
3115                 goto again;
3116         }
3117
3118         /* We've already setup this transaction, go ahead and exit */
3119         if (block_group->cache_generation == trans->transid &&
3120             i_size_read(inode)) {
3121                 dcs = BTRFS_DC_SETUP;
3122                 goto out_put;
3123         }
3124
3125         /*
3126          * We want to set the generation to 0, that way if anything goes wrong
3127          * from here on out we know not to trust this cache when we load up next
3128          * time.
3129          */
3130         BTRFS_I(inode)->generation = 0;
3131         ret = btrfs_update_inode(trans, root, inode);
3132         WARN_ON(ret);
3133
3134         if (i_size_read(inode) > 0) {
3135                 ret = btrfs_truncate_free_space_cache(root, trans, path,
3136                                                       inode);
3137                 if (ret)
3138                         goto out_put;
3139         }
3140
3141         spin_lock(&block_group->lock);
3142         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3143             !btrfs_test_opt(root, SPACE_CACHE)) {
3144                 /*
3145                  * don't bother trying to write stuff out _if_
3146                  * a) we're not cached,
3147                  * b) we're with nospace_cache mount option.
3148                  */
3149                 dcs = BTRFS_DC_WRITTEN;
3150                 spin_unlock(&block_group->lock);
3151                 goto out_put;
3152         }
3153         spin_unlock(&block_group->lock);
3154
3155         /*
3156          * Try to preallocate enough space based on how big the block group is.
3157          * Keep in mind this has to include any pinned space which could end up
3158          * taking up quite a bit since it's not folded into the other space
3159          * cache.
3160          */
3161         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3162         if (!num_pages)
3163                 num_pages = 1;
3164
3165         num_pages *= 16;
3166         num_pages *= PAGE_CACHE_SIZE;
3167
3168         ret = btrfs_check_data_free_space(inode, num_pages);
3169         if (ret)
3170                 goto out_put;
3171
3172         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3173                                               num_pages, num_pages,
3174                                               &alloc_hint);
3175         if (!ret)
3176                 dcs = BTRFS_DC_SETUP;
3177         btrfs_free_reserved_data_space(inode, num_pages);
3178
3179 out_put:
3180         iput(inode);
3181 out_free:
3182         btrfs_release_path(path);
3183 out:
3184         spin_lock(&block_group->lock);
3185         if (!ret && dcs == BTRFS_DC_SETUP)
3186                 block_group->cache_generation = trans->transid;
3187         block_group->disk_cache_state = dcs;
3188         spin_unlock(&block_group->lock);
3189
3190         return ret;
3191 }
3192
3193 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3194                                    struct btrfs_root *root)
3195 {
3196         struct btrfs_block_group_cache *cache;
3197         int err = 0;
3198         struct btrfs_path *path;
3199         u64 last = 0;
3200
3201         path = btrfs_alloc_path();
3202         if (!path)
3203                 return -ENOMEM;
3204
3205 again:
3206         while (1) {
3207                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3208                 while (cache) {
3209                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3210                                 break;
3211                         cache = next_block_group(root, cache);
3212                 }
3213                 if (!cache) {
3214                         if (last == 0)
3215                                 break;
3216                         last = 0;
3217                         continue;
3218                 }
3219                 err = cache_save_setup(cache, trans, path);
3220                 last = cache->key.objectid + cache->key.offset;
3221                 btrfs_put_block_group(cache);
3222         }
3223
3224         while (1) {
3225                 if (last == 0) {
3226                         err = btrfs_run_delayed_refs(trans, root,
3227                                                      (unsigned long)-1);
3228                         if (err) /* File system offline */
3229                                 goto out;
3230                 }
3231
3232                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3233                 while (cache) {
3234                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3235                                 btrfs_put_block_group(cache);
3236                                 goto again;
3237                         }
3238
3239                         if (cache->dirty)
3240                                 break;
3241                         cache = next_block_group(root, cache);
3242                 }
3243                 if (!cache) {
3244                         if (last == 0)
3245                                 break;
3246                         last = 0;
3247                         continue;
3248                 }
3249
3250                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3251                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3252                 cache->dirty = 0;
3253                 last = cache->key.objectid + cache->key.offset;
3254
3255                 err = write_one_cache_group(trans, root, path, cache);
3256                 if (err) /* File system offline */
3257                         goto out;
3258
3259                 btrfs_put_block_group(cache);
3260         }
3261
3262         while (1) {
3263                 /*
3264                  * I don't think this is needed since we're just marking our
3265                  * preallocated extent as written, but just in case it can't
3266                  * hurt.
3267                  */
3268                 if (last == 0) {
3269                         err = btrfs_run_delayed_refs(trans, root,
3270                                                      (unsigned long)-1);
3271                         if (err) /* File system offline */
3272                                 goto out;
3273                 }
3274
3275                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3276                 while (cache) {
3277                         /*
3278                          * Really this shouldn't happen, but it could if we
3279                          * couldn't write the entire preallocated extent and
3280                          * splitting the extent resulted in a new block.
3281                          */
3282                         if (cache->dirty) {
3283                                 btrfs_put_block_group(cache);
3284                                 goto again;
3285                         }
3286                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3287                                 break;
3288                         cache = next_block_group(root, cache);
3289                 }
3290                 if (!cache) {
3291                         if (last == 0)
3292                                 break;
3293                         last = 0;
3294                         continue;
3295                 }
3296
3297                 err = btrfs_write_out_cache(root, trans, cache, path);
3298
3299                 /*
3300                  * If we didn't have an error then the cache state is still
3301                  * NEED_WRITE, so we can set it to WRITTEN.
3302                  */
3303                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3304                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3305                 last = cache->key.objectid + cache->key.offset;
3306                 btrfs_put_block_group(cache);
3307         }
3308 out:
3309
3310         btrfs_free_path(path);
3311         return err;
3312 }
3313
3314 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3315 {
3316         struct btrfs_block_group_cache *block_group;
3317         int readonly = 0;
3318
3319         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3320         if (!block_group || block_group->ro)
3321                 readonly = 1;
3322         if (block_group)
3323                 btrfs_put_block_group(block_group);
3324         return readonly;
3325 }
3326
3327 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3328                              u64 total_bytes, u64 bytes_used,
3329                              struct btrfs_space_info **space_info)
3330 {
3331         struct btrfs_space_info *found;
3332         int i;
3333         int factor;
3334
3335         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3336                      BTRFS_BLOCK_GROUP_RAID10))
3337                 factor = 2;
3338         else
3339                 factor = 1;
3340
3341         found = __find_space_info(info, flags);
3342         if (found) {
3343                 spin_lock(&found->lock);
3344                 found->total_bytes += total_bytes;
3345                 found->disk_total += total_bytes * factor;
3346                 found->bytes_used += bytes_used;
3347                 found->disk_used += bytes_used * factor;
3348                 found->full = 0;
3349                 spin_unlock(&found->lock);
3350                 *space_info = found;
3351                 return 0;
3352         }
3353         found = kzalloc(sizeof(*found), GFP_NOFS);
3354         if (!found)
3355                 return -ENOMEM;
3356
3357         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3358                 INIT_LIST_HEAD(&found->block_groups[i]);
3359         init_rwsem(&found->groups_sem);
3360         spin_lock_init(&found->lock);
3361         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3362         found->total_bytes = total_bytes;
3363         found->disk_total = total_bytes * factor;
3364         found->bytes_used = bytes_used;
3365         found->disk_used = bytes_used * factor;
3366         found->bytes_pinned = 0;
3367         found->bytes_reserved = 0;
3368         found->bytes_readonly = 0;
3369         found->bytes_may_use = 0;
3370         found->full = 0;
3371         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3372         found->chunk_alloc = 0;
3373         found->flush = 0;
3374         init_waitqueue_head(&found->wait);
3375         *space_info = found;
3376         list_add_rcu(&found->list, &info->space_info);
3377         if (flags & BTRFS_BLOCK_GROUP_DATA)
3378                 info->data_sinfo = found;
3379         return 0;
3380 }
3381
3382 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3383 {
3384         u64 extra_flags = chunk_to_extended(flags) &
3385                                 BTRFS_EXTENDED_PROFILE_MASK;
3386
3387         write_seqlock(&fs_info->profiles_lock);
3388         if (flags & BTRFS_BLOCK_GROUP_DATA)
3389                 fs_info->avail_data_alloc_bits |= extra_flags;
3390         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3391                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3392         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3393                 fs_info->avail_system_alloc_bits |= extra_flags;
3394         write_sequnlock(&fs_info->profiles_lock);
3395 }
3396
3397 /*
3398  * returns target flags in extended format or 0 if restripe for this
3399  * chunk_type is not in progress
3400  *
3401  * should be called with either volume_mutex or balance_lock held
3402  */
3403 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3404 {
3405         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3406         u64 target = 0;
3407
3408         if (!bctl)
3409                 return 0;
3410
3411         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3412             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3413                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3414         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3415                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3416                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3417         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3418                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3419                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3420         }
3421
3422         return target;
3423 }
3424
3425 /*
3426  * @flags: available profiles in extended format (see ctree.h)
3427  *
3428  * Returns reduced profile in chunk format.  If profile changing is in
3429  * progress (either running or paused) picks the target profile (if it's
3430  * already available), otherwise falls back to plain reducing.
3431  */
3432 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3433 {
3434         /*
3435          * we add in the count of missing devices because we want
3436          * to make sure that any RAID levels on a degraded FS
3437          * continue to be honored.
3438          */
3439         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3440                 root->fs_info->fs_devices->missing_devices;
3441         u64 target;
3442         u64 tmp;
3443
3444         /*
3445          * see if restripe for this chunk_type is in progress, if so
3446          * try to reduce to the target profile
3447          */
3448         spin_lock(&root->fs_info->balance_lock);
3449         target = get_restripe_target(root->fs_info, flags);
3450         if (target) {
3451                 /* pick target profile only if it's already available */
3452                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3453                         spin_unlock(&root->fs_info->balance_lock);
3454                         return extended_to_chunk(target);
3455                 }
3456         }
3457         spin_unlock(&root->fs_info->balance_lock);
3458
3459         /* First, mask out the RAID levels which aren't possible */
3460         if (num_devices == 1)
3461                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3462                            BTRFS_BLOCK_GROUP_RAID5);
3463         if (num_devices < 3)
3464                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3465         if (num_devices < 4)
3466                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3467
3468         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3469                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3470                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3471         flags &= ~tmp;
3472
3473         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3474                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3475         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3476                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3477         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3478                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3479         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3480                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3481         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3482                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3483
3484         return extended_to_chunk(flags | tmp);
3485 }
3486
3487 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3488 {
3489         unsigned seq;
3490
3491         do {
3492                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3493
3494                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3495                         flags |= root->fs_info->avail_data_alloc_bits;
3496                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3497                         flags |= root->fs_info->avail_system_alloc_bits;
3498                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3499                         flags |= root->fs_info->avail_metadata_alloc_bits;
3500         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3501
3502         return btrfs_reduce_alloc_profile(root, flags);
3503 }
3504
3505 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3506 {
3507         u64 flags;
3508         u64 ret;
3509
3510         if (data)
3511                 flags = BTRFS_BLOCK_GROUP_DATA;
3512         else if (root == root->fs_info->chunk_root)
3513                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3514         else
3515                 flags = BTRFS_BLOCK_GROUP_METADATA;
3516
3517         ret = get_alloc_profile(root, flags);
3518         return ret;
3519 }
3520
3521 /*
3522  * This will check the space that the inode allocates from to make sure we have
3523  * enough space for bytes.
3524  */
3525 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3526 {
3527         struct btrfs_space_info *data_sinfo;
3528         struct btrfs_root *root = BTRFS_I(inode)->root;
3529         struct btrfs_fs_info *fs_info = root->fs_info;
3530         u64 used;
3531         int ret = 0, committed = 0, alloc_chunk = 1;
3532
3533         /* make sure bytes are sectorsize aligned */
3534         bytes = ALIGN(bytes, root->sectorsize);
3535
3536         if (root == root->fs_info->tree_root ||
3537             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3538                 alloc_chunk = 0;
3539                 committed = 1;
3540         }
3541
3542         data_sinfo = fs_info->data_sinfo;
3543         if (!data_sinfo)
3544                 goto alloc;
3545
3546 again:
3547         /* make sure we have enough space to handle the data first */
3548         spin_lock(&data_sinfo->lock);
3549         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3550                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3551                 data_sinfo->bytes_may_use;
3552
3553         if (used + bytes > data_sinfo->total_bytes) {
3554                 struct btrfs_trans_handle *trans;
3555
3556                 /*
3557                  * if we don't have enough free bytes in this space then we need
3558                  * to alloc a new chunk.
3559                  */
3560                 if (!data_sinfo->full && alloc_chunk) {
3561                         u64 alloc_target;
3562
3563                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3564                         spin_unlock(&data_sinfo->lock);
3565 alloc:
3566                         alloc_target = btrfs_get_alloc_profile(root, 1);
3567                         trans = btrfs_join_transaction(root);
3568                         if (IS_ERR(trans))
3569                                 return PTR_ERR(trans);
3570
3571                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3572                                              alloc_target,
3573                                              CHUNK_ALLOC_NO_FORCE);
3574                         btrfs_end_transaction(trans, root);
3575                         if (ret < 0) {
3576                                 if (ret != -ENOSPC)
3577                                         return ret;
3578                                 else
3579                                         goto commit_trans;
3580                         }
3581
3582                         if (!data_sinfo)
3583                                 data_sinfo = fs_info->data_sinfo;
3584
3585                         goto again;
3586                 }
3587
3588                 /*
3589                  * If we have less pinned bytes than we want to allocate then
3590                  * don't bother committing the transaction, it won't help us.
3591                  */
3592                 if (data_sinfo->bytes_pinned < bytes)
3593                         committed = 1;
3594                 spin_unlock(&data_sinfo->lock);
3595
3596                 /* commit the current transaction and try again */
3597 commit_trans:
3598                 if (!committed &&
3599                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3600                         committed = 1;
3601                         trans = btrfs_join_transaction(root);
3602                         if (IS_ERR(trans))
3603                                 return PTR_ERR(trans);
3604                         ret = btrfs_commit_transaction(trans, root);
3605                         if (ret)
3606                                 return ret;
3607                         goto again;
3608                 }
3609
3610                 return -ENOSPC;
3611         }
3612         data_sinfo->bytes_may_use += bytes;
3613         trace_btrfs_space_reservation(root->fs_info, "space_info",
3614                                       data_sinfo->flags, bytes, 1);
3615         spin_unlock(&data_sinfo->lock);
3616
3617         return 0;
3618 }
3619
3620 /*
3621  * Called if we need to clear a data reservation for this inode.
3622  */
3623 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3624 {
3625         struct btrfs_root *root = BTRFS_I(inode)->root;
3626         struct btrfs_space_info *data_sinfo;
3627
3628         /* make sure bytes are sectorsize aligned */
3629         bytes = ALIGN(bytes, root->sectorsize);
3630
3631         data_sinfo = root->fs_info->data_sinfo;
3632         spin_lock(&data_sinfo->lock);
3633         data_sinfo->bytes_may_use -= bytes;
3634         trace_btrfs_space_reservation(root->fs_info, "space_info",
3635                                       data_sinfo->flags, bytes, 0);
3636         spin_unlock(&data_sinfo->lock);
3637 }
3638
3639 static void force_metadata_allocation(struct btrfs_fs_info *info)
3640 {
3641         struct list_head *head = &info->space_info;
3642         struct btrfs_space_info *found;
3643
3644         rcu_read_lock();
3645         list_for_each_entry_rcu(found, head, list) {
3646                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3647                         found->force_alloc = CHUNK_ALLOC_FORCE;
3648         }
3649         rcu_read_unlock();
3650 }
3651
3652 static int should_alloc_chunk(struct btrfs_root *root,
3653                               struct btrfs_space_info *sinfo, int force)
3654 {
3655         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3656         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3657         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3658         u64 thresh;
3659
3660         if (force == CHUNK_ALLOC_FORCE)
3661                 return 1;
3662
3663         /*
3664          * We need to take into account the global rsv because for all intents
3665          * and purposes it's used space.  Don't worry about locking the
3666          * global_rsv, it doesn't change except when the transaction commits.
3667          */
3668         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3669                 num_allocated += global_rsv->size;
3670
3671         /*
3672          * in limited mode, we want to have some free space up to
3673          * about 1% of the FS size.
3674          */
3675         if (force == CHUNK_ALLOC_LIMITED) {
3676                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3677                 thresh = max_t(u64, 64 * 1024 * 1024,
3678                                div_factor_fine(thresh, 1));
3679
3680                 if (num_bytes - num_allocated < thresh)
3681                         return 1;
3682         }
3683
3684         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3685                 return 0;
3686         return 1;
3687 }
3688
3689 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3690 {
3691         u64 num_dev;
3692
3693         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3694                     BTRFS_BLOCK_GROUP_RAID0 |
3695                     BTRFS_BLOCK_GROUP_RAID5 |
3696                     BTRFS_BLOCK_GROUP_RAID6))
3697                 num_dev = root->fs_info->fs_devices->rw_devices;
3698         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3699                 num_dev = 2;
3700         else
3701                 num_dev = 1;    /* DUP or single */
3702
3703         /* metadata for updaing devices and chunk tree */
3704         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3705 }
3706
3707 static void check_system_chunk(struct btrfs_trans_handle *trans,
3708                                struct btrfs_root *root, u64 type)
3709 {
3710         struct btrfs_space_info *info;
3711         u64 left;
3712         u64 thresh;
3713
3714         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3715         spin_lock(&info->lock);
3716         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3717                 info->bytes_reserved - info->bytes_readonly;
3718         spin_unlock(&info->lock);
3719
3720         thresh = get_system_chunk_thresh(root, type);
3721         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3722                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3723                         left, thresh, type);
3724                 dump_space_info(info, 0, 0);
3725         }
3726
3727         if (left < thresh) {
3728                 u64 flags;
3729
3730                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3731                 btrfs_alloc_chunk(trans, root, flags);
3732         }
3733 }
3734
3735 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3736                           struct btrfs_root *extent_root, u64 flags, int force)
3737 {
3738         struct btrfs_space_info *space_info;
3739         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3740         int wait_for_alloc = 0;
3741         int ret = 0;
3742
3743         /* Don't re-enter if we're already allocating a chunk */
3744         if (trans->allocating_chunk)
3745                 return -ENOSPC;
3746
3747         space_info = __find_space_info(extent_root->fs_info, flags);
3748         if (!space_info) {
3749                 ret = update_space_info(extent_root->fs_info, flags,
3750                                         0, 0, &space_info);
3751                 BUG_ON(ret); /* -ENOMEM */
3752         }
3753         BUG_ON(!space_info); /* Logic error */
3754
3755 again:
3756         spin_lock(&space_info->lock);
3757         if (force < space_info->force_alloc)
3758                 force = space_info->force_alloc;
3759         if (space_info->full) {
3760                 spin_unlock(&space_info->lock);
3761                 return 0;
3762         }
3763
3764         if (!should_alloc_chunk(extent_root, space_info, force)) {
3765                 spin_unlock(&space_info->lock);
3766                 return 0;
3767         } else if (space_info->chunk_alloc) {
3768                 wait_for_alloc = 1;
3769         } else {
3770                 space_info->chunk_alloc = 1;
3771         }
3772
3773         spin_unlock(&space_info->lock);
3774
3775         mutex_lock(&fs_info->chunk_mutex);
3776
3777         /*
3778          * The chunk_mutex is held throughout the entirety of a chunk
3779          * allocation, so once we've acquired the chunk_mutex we know that the
3780          * other guy is done and we need to recheck and see if we should
3781          * allocate.
3782          */
3783         if (wait_for_alloc) {
3784                 mutex_unlock(&fs_info->chunk_mutex);
3785                 wait_for_alloc = 0;
3786                 goto again;
3787         }
3788
3789         trans->allocating_chunk = true;
3790
3791         /*
3792          * If we have mixed data/metadata chunks we want to make sure we keep
3793          * allocating mixed chunks instead of individual chunks.
3794          */
3795         if (btrfs_mixed_space_info(space_info))
3796                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3797
3798         /*
3799          * if we're doing a data chunk, go ahead and make sure that
3800          * we keep a reasonable number of metadata chunks allocated in the
3801          * FS as well.
3802          */
3803         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3804                 fs_info->data_chunk_allocations++;
3805                 if (!(fs_info->data_chunk_allocations %
3806                       fs_info->metadata_ratio))
3807                         force_metadata_allocation(fs_info);
3808         }
3809
3810         /*
3811          * Check if we have enough space in SYSTEM chunk because we may need
3812          * to update devices.
3813          */
3814         check_system_chunk(trans, extent_root, flags);
3815
3816         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3817         trans->allocating_chunk = false;
3818
3819         spin_lock(&space_info->lock);
3820         if (ret < 0 && ret != -ENOSPC)
3821                 goto out;
3822         if (ret)
3823                 space_info->full = 1;
3824         else
3825                 ret = 1;
3826
3827         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3828 out:
3829         space_info->chunk_alloc = 0;
3830         spin_unlock(&space_info->lock);
3831         mutex_unlock(&fs_info->chunk_mutex);
3832         return ret;
3833 }
3834
3835 static int can_overcommit(struct btrfs_root *root,
3836                           struct btrfs_space_info *space_info, u64 bytes,
3837                           enum btrfs_reserve_flush_enum flush)
3838 {
3839         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3840         u64 profile = btrfs_get_alloc_profile(root, 0);
3841         u64 rsv_size = 0;
3842         u64 avail;
3843         u64 used;
3844         u64 to_add;
3845
3846         used = space_info->bytes_used + space_info->bytes_reserved +
3847                 space_info->bytes_pinned + space_info->bytes_readonly;
3848
3849         spin_lock(&global_rsv->lock);
3850         rsv_size = global_rsv->size;
3851         spin_unlock(&global_rsv->lock);
3852
3853         /*
3854          * We only want to allow over committing if we have lots of actual space
3855          * free, but if we don't have enough space to handle the global reserve
3856          * space then we could end up having a real enospc problem when trying
3857          * to allocate a chunk or some other such important allocation.
3858          */
3859         rsv_size <<= 1;
3860         if (used + rsv_size >= space_info->total_bytes)
3861                 return 0;
3862
3863         used += space_info->bytes_may_use;
3864
3865         spin_lock(&root->fs_info->free_chunk_lock);
3866         avail = root->fs_info->free_chunk_space;
3867         spin_unlock(&root->fs_info->free_chunk_lock);
3868
3869         /*
3870          * If we have dup, raid1 or raid10 then only half of the free
3871          * space is actually useable.  For raid56, the space info used
3872          * doesn't include the parity drive, so we don't have to
3873          * change the math
3874          */
3875         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3876                        BTRFS_BLOCK_GROUP_RAID1 |
3877                        BTRFS_BLOCK_GROUP_RAID10))
3878                 avail >>= 1;
3879
3880         to_add = space_info->total_bytes;
3881
3882         /*
3883          * If we aren't flushing all things, let us overcommit up to
3884          * 1/2th of the space. If we can flush, don't let us overcommit
3885          * too much, let it overcommit up to 1/8 of the space.
3886          */
3887         if (flush == BTRFS_RESERVE_FLUSH_ALL)
3888                 to_add >>= 3;
3889         else
3890                 to_add >>= 1;
3891
3892         /*
3893          * Limit the overcommit to the amount of free space we could possibly
3894          * allocate for chunks.
3895          */
3896         to_add = min(avail, to_add);
3897
3898         if (used + bytes < space_info->total_bytes + to_add)
3899                 return 1;
3900         return 0;
3901 }
3902
3903 void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3904                                   unsigned long nr_pages)
3905 {
3906         struct super_block *sb = root->fs_info->sb;
3907         int started;
3908
3909         /* If we can not start writeback, just sync all the delalloc file. */
3910         started = try_to_writeback_inodes_sb_nr(sb, nr_pages,
3911                                                       WB_REASON_FS_FREE_SPACE);
3912         if (!started) {
3913                 /*
3914                  * We needn't worry the filesystem going from r/w to r/o though
3915                  * we don't acquire ->s_umount mutex, because the filesystem
3916                  * should guarantee the delalloc inodes list be empty after
3917                  * the filesystem is readonly(all dirty pages are written to
3918                  * the disk).
3919                  */
3920                 btrfs_start_delalloc_inodes(root, 0);
3921                 if (!current->journal_info)
3922                         btrfs_wait_ordered_extents(root, 0);
3923         }
3924 }
3925
3926 /*
3927  * shrink metadata reservation for delalloc
3928  */
3929 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3930                             bool wait_ordered)
3931 {
3932         struct btrfs_block_rsv *block_rsv;
3933         struct btrfs_space_info *space_info;
3934         struct btrfs_trans_handle *trans;
3935         u64 delalloc_bytes;
3936         u64 max_reclaim;
3937         long time_left;
3938         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3939         int loops = 0;
3940         enum btrfs_reserve_flush_enum flush;
3941
3942         trans = (struct btrfs_trans_handle *)current->journal_info;
3943         block_rsv = &root->fs_info->delalloc_block_rsv;
3944         space_info = block_rsv->space_info;
3945
3946         smp_mb();
3947         delalloc_bytes = percpu_counter_sum_positive(
3948                                                 &root->fs_info->delalloc_bytes);
3949         if (delalloc_bytes == 0) {
3950                 if (trans)
3951                         return;
3952                 btrfs_wait_ordered_extents(root, 0);
3953                 return;
3954         }
3955
3956         while (delalloc_bytes && loops < 3) {
3957                 max_reclaim = min(delalloc_bytes, to_reclaim);
3958                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3959                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
3960                 /*
3961                  * We need to wait for the async pages to actually start before
3962                  * we do anything.
3963                  */
3964                 wait_event(root->fs_info->async_submit_wait,
3965                            !atomic_read(&root->fs_info->async_delalloc_pages));
3966
3967                 if (!trans)
3968                         flush = BTRFS_RESERVE_FLUSH_ALL;
3969                 else
3970                         flush = BTRFS_RESERVE_NO_FLUSH;
3971                 spin_lock(&space_info->lock);
3972                 if (can_overcommit(root, space_info, orig, flush)) {
3973                         spin_unlock(&space_info->lock);
3974                         break;
3975                 }
3976                 spin_unlock(&space_info->lock);
3977
3978                 loops++;
3979                 if (wait_ordered && !trans) {
3980                         btrfs_wait_ordered_extents(root, 0);
3981                 } else {
3982                         time_left = schedule_timeout_killable(1);
3983                         if (time_left)
3984                                 break;
3985                 }
3986                 smp_mb();
3987                 delalloc_bytes = percpu_counter_sum_positive(
3988                                                 &root->fs_info->delalloc_bytes);
3989         }
3990 }
3991
3992 /**
3993  * maybe_commit_transaction - possibly commit the transaction if its ok to
3994  * @root - the root we're allocating for
3995  * @bytes - the number of bytes we want to reserve
3996  * @force - force the commit
3997  *
3998  * This will check to make sure that committing the transaction will actually
3999  * get us somewhere and then commit the transaction if it does.  Otherwise it
4000  * will return -ENOSPC.
4001  */
4002 static int may_commit_transaction(struct btrfs_root *root,
4003                                   struct btrfs_space_info *space_info,
4004                                   u64 bytes, int force)
4005 {
4006         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4007         struct btrfs_trans_handle *trans;
4008
4009         trans = (struct btrfs_trans_handle *)current->journal_info;
4010         if (trans)
4011                 return -EAGAIN;
4012
4013         if (force)
4014                 goto commit;
4015
4016         /* See if there is enough pinned space to make this reservation */
4017         spin_lock(&space_info->lock);
4018         if (space_info->bytes_pinned >= bytes) {
4019                 spin_unlock(&space_info->lock);
4020                 goto commit;
4021         }
4022         spin_unlock(&space_info->lock);
4023
4024         /*
4025          * See if there is some space in the delayed insertion reservation for
4026          * this reservation.
4027          */
4028         if (space_info != delayed_rsv->space_info)
4029                 return -ENOSPC;
4030
4031         spin_lock(&space_info->lock);
4032         spin_lock(&delayed_rsv->lock);
4033         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
4034                 spin_unlock(&delayed_rsv->lock);
4035                 spin_unlock(&space_info->lock);
4036                 return -ENOSPC;
4037         }
4038         spin_unlock(&delayed_rsv->lock);
4039         spin_unlock(&space_info->lock);
4040
4041 commit:
4042         trans = btrfs_join_transaction(root);
4043         if (IS_ERR(trans))
4044                 return -ENOSPC;
4045
4046         return btrfs_commit_transaction(trans, root);
4047 }
4048
4049 enum flush_state {
4050         FLUSH_DELAYED_ITEMS_NR  =       1,
4051         FLUSH_DELAYED_ITEMS     =       2,
4052         FLUSH_DELALLOC          =       3,
4053         FLUSH_DELALLOC_WAIT     =       4,
4054         ALLOC_CHUNK             =       5,
4055         COMMIT_TRANS            =       6,
4056 };
4057
4058 static int flush_space(struct btrfs_root *root,
4059                        struct btrfs_space_info *space_info, u64 num_bytes,
4060                        u64 orig_bytes, int state)
4061 {
4062         struct btrfs_trans_handle *trans;
4063         int nr;
4064         int ret = 0;
4065
4066         switch (state) {
4067         case FLUSH_DELAYED_ITEMS_NR:
4068         case FLUSH_DELAYED_ITEMS:
4069                 if (state == FLUSH_DELAYED_ITEMS_NR) {
4070                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
4071
4072                         nr = (int)div64_u64(num_bytes, bytes);
4073                         if (!nr)
4074                                 nr = 1;
4075                         nr *= 2;
4076                 } else {
4077                         nr = -1;
4078                 }
4079                 trans = btrfs_join_transaction(root);
4080                 if (IS_ERR(trans)) {
4081                         ret = PTR_ERR(trans);
4082                         break;
4083                 }
4084                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4085                 btrfs_end_transaction(trans, root);
4086                 break;
4087         case FLUSH_DELALLOC:
4088         case FLUSH_DELALLOC_WAIT:
4089                 shrink_delalloc(root, num_bytes, orig_bytes,
4090                                 state == FLUSH_DELALLOC_WAIT);
4091                 break;
4092         case ALLOC_CHUNK:
4093                 trans = btrfs_join_transaction(root);
4094                 if (IS_ERR(trans)) {
4095                         ret = PTR_ERR(trans);
4096                         break;
4097                 }
4098                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4099                                      btrfs_get_alloc_profile(root, 0),
4100                                      CHUNK_ALLOC_NO_FORCE);
4101                 btrfs_end_transaction(trans, root);
4102                 if (ret == -ENOSPC)
4103                         ret = 0;
4104                 break;
4105         case COMMIT_TRANS:
4106                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4107                 break;
4108         default:
4109                 ret = -ENOSPC;
4110                 break;
4111         }
4112
4113         return ret;
4114 }
4115 /**
4116  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4117  * @root - the root we're allocating for
4118  * @block_rsv - the block_rsv we're allocating for
4119  * @orig_bytes - the number of bytes we want
4120  * @flush - whether or not we can flush to make our reservation
4121  *
4122  * This will reserve orgi_bytes number of bytes from the space info associated
4123  * with the block_rsv.  If there is not enough space it will make an attempt to
4124  * flush out space to make room.  It will do this by flushing delalloc if
4125  * possible or committing the transaction.  If flush is 0 then no attempts to
4126  * regain reservations will be made and this will fail if there is not enough
4127  * space already.
4128  */
4129 static int reserve_metadata_bytes(struct btrfs_root *root,
4130                                   struct btrfs_block_rsv *block_rsv,
4131                                   u64 orig_bytes,
4132                                   enum btrfs_reserve_flush_enum flush)
4133 {
4134         struct btrfs_space_info *space_info = block_rsv->space_info;
4135         u64 used;
4136         u64 num_bytes = orig_bytes;
4137         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4138         int ret = 0;
4139         bool flushing = false;
4140
4141 again:
4142         ret = 0;
4143         spin_lock(&space_info->lock);
4144         /*
4145          * We only want to wait if somebody other than us is flushing and we
4146          * are actually allowed to flush all things.
4147          */
4148         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4149                space_info->flush) {
4150                 spin_unlock(&space_info->lock);
4151                 /*
4152                  * If we have a trans handle we can't wait because the flusher
4153                  * may have to commit the transaction, which would mean we would
4154                  * deadlock since we are waiting for the flusher to finish, but
4155                  * hold the current transaction open.
4156                  */
4157                 if (current->journal_info)
4158                         return -EAGAIN;
4159                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4160                 /* Must have been killed, return */
4161                 if (ret)
4162                         return -EINTR;
4163
4164                 spin_lock(&space_info->lock);
4165         }
4166
4167         ret = -ENOSPC;
4168         used = space_info->bytes_used + space_info->bytes_reserved +
4169                 space_info->bytes_pinned + space_info->bytes_readonly +
4170                 space_info->bytes_may_use;
4171
4172         /*
4173          * The idea here is that we've not already over-reserved the block group
4174          * then we can go ahead and save our reservation first and then start
4175          * flushing if we need to.  Otherwise if we've already overcommitted
4176          * lets start flushing stuff first and then come back and try to make
4177          * our reservation.
4178          */
4179         if (used <= space_info->total_bytes) {
4180                 if (used + orig_bytes <= space_info->total_bytes) {
4181                         space_info->bytes_may_use += orig_bytes;
4182                         trace_btrfs_space_reservation(root->fs_info,
4183                                 "space_info", space_info->flags, orig_bytes, 1);
4184                         ret = 0;
4185                 } else {
4186                         /*
4187                          * Ok set num_bytes to orig_bytes since we aren't
4188                          * overocmmitted, this way we only try and reclaim what
4189                          * we need.
4190                          */
4191                         num_bytes = orig_bytes;
4192                 }
4193         } else {
4194                 /*
4195                  * Ok we're over committed, set num_bytes to the overcommitted
4196                  * amount plus the amount of bytes that we need for this
4197                  * reservation.
4198                  */
4199                 num_bytes = used - space_info->total_bytes +
4200                         (orig_bytes * 2);
4201         }
4202
4203         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4204                 space_info->bytes_may_use += orig_bytes;
4205                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4206                                               space_info->flags, orig_bytes,
4207                                               1);
4208                 ret = 0;
4209         }
4210
4211         /*
4212          * Couldn't make our reservation, save our place so while we're trying
4213          * to reclaim space we can actually use it instead of somebody else
4214          * stealing it from us.
4215          *
4216          * We make the other tasks wait for the flush only when we can flush
4217          * all things.
4218          */
4219         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4220                 flushing = true;
4221                 space_info->flush = 1;
4222         }
4223
4224         spin_unlock(&space_info->lock);
4225
4226         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4227                 goto out;
4228
4229         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4230                           flush_state);
4231         flush_state++;
4232
4233         /*
4234          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4235          * would happen. So skip delalloc flush.
4236          */
4237         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4238             (flush_state == FLUSH_DELALLOC ||
4239              flush_state == FLUSH_DELALLOC_WAIT))
4240                 flush_state = ALLOC_CHUNK;
4241
4242         if (!ret)
4243                 goto again;
4244         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4245                  flush_state < COMMIT_TRANS)
4246                 goto again;
4247         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4248                  flush_state <= COMMIT_TRANS)
4249                 goto again;
4250
4251 out:
4252         if (ret == -ENOSPC &&
4253             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4254                 struct btrfs_block_rsv *global_rsv =
4255                         &root->fs_info->global_block_rsv;
4256
4257                 if (block_rsv != global_rsv &&
4258                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4259                         ret = 0;
4260         }
4261         if (flushing) {
4262                 spin_lock(&space_info->lock);
4263                 space_info->flush = 0;
4264                 wake_up_all(&space_info->wait);
4265                 spin_unlock(&space_info->lock);
4266         }
4267         return ret;
4268 }
4269
4270 static struct btrfs_block_rsv *get_block_rsv(
4271                                         const struct btrfs_trans_handle *trans,
4272                                         const struct btrfs_root *root)
4273 {
4274         struct btrfs_block_rsv *block_rsv = NULL;
4275
4276         if (root->ref_cows)
4277                 block_rsv = trans->block_rsv;
4278
4279         if (root == root->fs_info->csum_root && trans->adding_csums)
4280                 block_rsv = trans->block_rsv;
4281
4282         if (!block_rsv)
4283                 block_rsv = root->block_rsv;
4284
4285         if (!block_rsv)
4286                 block_rsv = &root->fs_info->empty_block_rsv;
4287
4288         return block_rsv;
4289 }
4290
4291 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4292                                u64 num_bytes)
4293 {
4294         int ret = -ENOSPC;
4295         spin_lock(&block_rsv->lock);
4296         if (block_rsv->reserved >= num_bytes) {
4297                 block_rsv->reserved -= num_bytes;
4298                 if (block_rsv->reserved < block_rsv->size)
4299                         block_rsv->full = 0;
4300                 ret = 0;
4301         }
4302         spin_unlock(&block_rsv->lock);
4303         return ret;
4304 }
4305
4306 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4307                                 u64 num_bytes, int update_size)
4308 {
4309         spin_lock(&block_rsv->lock);
4310         block_rsv->reserved += num_bytes;
4311         if (update_size)
4312                 block_rsv->size += num_bytes;
4313         else if (block_rsv->reserved >= block_rsv->size)
4314                 block_rsv->full = 1;
4315         spin_unlock(&block_rsv->lock);
4316 }
4317
4318 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4319                                     struct btrfs_block_rsv *block_rsv,
4320                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4321 {
4322         struct btrfs_space_info *space_info = block_rsv->space_info;
4323
4324         spin_lock(&block_rsv->lock);
4325         if (num_bytes == (u64)-1)
4326                 num_bytes = block_rsv->size;
4327         block_rsv->size -= num_bytes;
4328         if (block_rsv->reserved >= block_rsv->size) {
4329                 num_bytes = block_rsv->reserved - block_rsv->size;
4330                 block_rsv->reserved = block_rsv->size;
4331                 block_rsv->full = 1;
4332         } else {
4333                 num_bytes = 0;
4334         }
4335         spin_unlock(&block_rsv->lock);
4336
4337         if (num_bytes > 0) {
4338                 if (dest) {
4339                         spin_lock(&dest->lock);
4340                         if (!dest->full) {
4341                                 u64 bytes_to_add;
4342
4343                                 bytes_to_add = dest->size - dest->reserved;
4344                                 bytes_to_add = min(num_bytes, bytes_to_add);
4345                                 dest->reserved += bytes_to_add;
4346                                 if (dest->reserved >= dest->size)
4347                                         dest->full = 1;
4348                                 num_bytes -= bytes_to_add;
4349                         }
4350                         spin_unlock(&dest->lock);
4351                 }
4352                 if (num_bytes) {
4353                         spin_lock(&space_info->lock);
4354                         space_info->bytes_may_use -= num_bytes;
4355                         trace_btrfs_space_reservation(fs_info, "space_info",
4356                                         space_info->flags, num_bytes, 0);
4357                         space_info->reservation_progress++;
4358                         spin_unlock(&space_info->lock);
4359                 }
4360         }
4361 }
4362
4363 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4364                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4365 {
4366         int ret;
4367
4368         ret = block_rsv_use_bytes(src, num_bytes);
4369         if (ret)
4370                 return ret;
4371
4372         block_rsv_add_bytes(dst, num_bytes, 1);
4373         return 0;
4374 }
4375
4376 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4377 {
4378         memset(rsv, 0, sizeof(*rsv));
4379         spin_lock_init(&rsv->lock);
4380         rsv->type = type;
4381 }
4382
4383 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4384                                               unsigned short type)
4385 {
4386         struct btrfs_block_rsv *block_rsv;
4387         struct btrfs_fs_info *fs_info = root->fs_info;
4388
4389         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4390         if (!block_rsv)
4391                 return NULL;
4392
4393         btrfs_init_block_rsv(block_rsv, type);
4394         block_rsv->space_info = __find_space_info(fs_info,
4395                                                   BTRFS_BLOCK_GROUP_METADATA);
4396         return block_rsv;
4397 }
4398
4399 void btrfs_free_block_rsv(struct btrfs_root *root,
4400                           struct btrfs_block_rsv *rsv)
4401 {
4402         if (!rsv)
4403                 return;
4404         btrfs_block_rsv_release(root, rsv, (u64)-1);
4405         kfree(rsv);
4406 }
4407
4408 int btrfs_block_rsv_add(struct btrfs_root *root,
4409                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4410                         enum btrfs_reserve_flush_enum flush)
4411 {
4412         int ret;
4413
4414         if (num_bytes == 0)
4415                 return 0;
4416
4417         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4418         if (!ret) {
4419                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4420                 return 0;
4421         }
4422
4423         return ret;
4424 }
4425
4426 int btrfs_block_rsv_check(struct btrfs_root *root,
4427                           struct btrfs_block_rsv *block_rsv, int min_factor)
4428 {
4429         u64 num_bytes = 0;
4430         int ret = -ENOSPC;
4431
4432         if (!block_rsv)
4433                 return 0;
4434
4435         spin_lock(&block_rsv->lock);
4436         num_bytes = div_factor(block_rsv->size, min_factor);
4437         if (block_rsv->reserved >= num_bytes)
4438                 ret = 0;
4439         spin_unlock(&block_rsv->lock);
4440
4441         return ret;
4442 }
4443
4444 int btrfs_block_rsv_refill(struct btrfs_root *root,
4445                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4446                            enum btrfs_reserve_flush_enum flush)
4447 {
4448         u64 num_bytes = 0;
4449         int ret = -ENOSPC;
4450
4451         if (!block_rsv)
4452                 return 0;
4453
4454         spin_lock(&block_rsv->lock);
4455         num_bytes = min_reserved;
4456         if (block_rsv->reserved >= num_bytes)
4457                 ret = 0;
4458         else
4459                 num_bytes -= block_rsv->reserved;
4460         spin_unlock(&block_rsv->lock);
4461
4462         if (!ret)
4463                 return 0;
4464
4465         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4466         if (!ret) {
4467                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4468                 return 0;
4469         }
4470
4471         return ret;
4472 }
4473
4474 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4475                             struct btrfs_block_rsv *dst_rsv,
4476                             u64 num_bytes)
4477 {
4478         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4479 }
4480
4481 void btrfs_block_rsv_release(struct btrfs_root *root,
4482                              struct btrfs_block_rsv *block_rsv,
4483                              u64 num_bytes)
4484 {
4485         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4486         if (global_rsv->full || global_rsv == block_rsv ||
4487             block_rsv->space_info != global_rsv->space_info)
4488                 global_rsv = NULL;
4489         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4490                                 num_bytes);
4491 }
4492
4493 /*
4494  * helper to calculate size of global block reservation.
4495  * the desired value is sum of space used by extent tree,
4496  * checksum tree and root tree
4497  */
4498 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4499 {
4500         struct btrfs_space_info *sinfo;
4501         u64 num_bytes;
4502         u64 meta_used;
4503         u64 data_used;
4504         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4505
4506         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4507         spin_lock(&sinfo->lock);
4508         data_used = sinfo->bytes_used;
4509         spin_unlock(&sinfo->lock);
4510
4511         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4512         spin_lock(&sinfo->lock);
4513         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4514                 data_used = 0;
4515         meta_used = sinfo->bytes_used;
4516         spin_unlock(&sinfo->lock);
4517
4518         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4519                     csum_size * 2;
4520         num_bytes += div64_u64(data_used + meta_used, 50);
4521
4522         if (num_bytes * 3 > meta_used)
4523                 num_bytes = div64_u64(meta_used, 3);
4524
4525         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4526 }
4527
4528 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4529 {
4530         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4531         struct btrfs_space_info *sinfo = block_rsv->space_info;
4532         u64 num_bytes;
4533
4534         num_bytes = calc_global_metadata_size(fs_info);
4535
4536         spin_lock(&sinfo->lock);
4537         spin_lock(&block_rsv->lock);
4538
4539         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4540
4541         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4542                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4543                     sinfo->bytes_may_use;
4544
4545         if (sinfo->total_bytes > num_bytes) {
4546                 num_bytes = sinfo->total_bytes - num_bytes;
4547                 block_rsv->reserved += num_bytes;
4548                 sinfo->bytes_may_use += num_bytes;
4549                 trace_btrfs_space_reservation(fs_info, "space_info",
4550                                       sinfo->flags, num_bytes, 1);
4551         }
4552
4553         if (block_rsv->reserved >= block_rsv->size) {
4554                 num_bytes = block_rsv->reserved - block_rsv->size;
4555                 sinfo->bytes_may_use -= num_bytes;
4556                 trace_btrfs_space_reservation(fs_info, "space_info",
4557                                       sinfo->flags, num_bytes, 0);
4558                 sinfo->reservation_progress++;
4559                 block_rsv->reserved = block_rsv->size;
4560                 block_rsv->full = 1;
4561         }
4562
4563         spin_unlock(&block_rsv->lock);
4564         spin_unlock(&sinfo->lock);
4565 }
4566
4567 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4568 {
4569         struct btrfs_space_info *space_info;
4570
4571         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4572         fs_info->chunk_block_rsv.space_info = space_info;
4573
4574         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4575         fs_info->global_block_rsv.space_info = space_info;
4576         fs_info->delalloc_block_rsv.space_info = space_info;
4577         fs_info->trans_block_rsv.space_info = space_info;
4578         fs_info->empty_block_rsv.space_info = space_info;
4579         fs_info->delayed_block_rsv.space_info = space_info;
4580
4581         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4582         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4583         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4584         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4585         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4586
4587         update_global_block_rsv(fs_info);
4588 }
4589
4590 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4591 {
4592         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4593                                 (u64)-1);
4594         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4595         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4596         WARN_ON(fs_info->trans_block_rsv.size > 0);
4597         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4598         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4599         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4600         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4601         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4602 }
4603
4604 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4605                                   struct btrfs_root *root)
4606 {
4607         if (!trans->block_rsv)
4608                 return;
4609
4610         if (!trans->bytes_reserved)
4611                 return;
4612
4613         trace_btrfs_space_reservation(root->fs_info, "transaction",
4614                                       trans->transid, trans->bytes_reserved, 0);
4615         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4616         trans->bytes_reserved = 0;
4617 }
4618
4619 /* Can only return 0 or -ENOSPC */
4620 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4621                                   struct inode *inode)
4622 {
4623         struct btrfs_root *root = BTRFS_I(inode)->root;
4624         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4625         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4626
4627         /*
4628          * We need to hold space in order to delete our orphan item once we've
4629          * added it, so this takes the reservation so we can release it later
4630          * when we are truly done with the orphan item.
4631          */
4632         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4633         trace_btrfs_space_reservation(root->fs_info, "orphan",
4634                                       btrfs_ino(inode), num_bytes, 1);
4635         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4636 }
4637
4638 void btrfs_orphan_release_metadata(struct inode *inode)
4639 {
4640         struct btrfs_root *root = BTRFS_I(inode)->root;
4641         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4642         trace_btrfs_space_reservation(root->fs_info, "orphan",
4643                                       btrfs_ino(inode), num_bytes, 0);
4644         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4645 }
4646
4647 /*
4648  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4649  * root: the root of the parent directory
4650  * rsv: block reservation
4651  * items: the number of items that we need do reservation
4652  * qgroup_reserved: used to return the reserved size in qgroup
4653  *
4654  * This function is used to reserve the space for snapshot/subvolume
4655  * creation and deletion. Those operations are different with the
4656  * common file/directory operations, they change two fs/file trees
4657  * and root tree, the number of items that the qgroup reserves is
4658  * different with the free space reservation. So we can not use
4659  * the space reseravtion mechanism in start_transaction().
4660  */
4661 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4662                                      struct btrfs_block_rsv *rsv,
4663                                      int items,
4664                                      u64 *qgroup_reserved)
4665 {
4666         u64 num_bytes;
4667         int ret;
4668
4669         if (root->fs_info->quota_enabled) {
4670                 /* One for parent inode, two for dir entries */
4671                 num_bytes = 3 * root->leafsize;
4672                 ret = btrfs_qgroup_reserve(root, num_bytes);
4673                 if (ret)
4674                         return ret;
4675         } else {
4676                 num_bytes = 0;
4677         }
4678
4679         *qgroup_reserved = num_bytes;
4680
4681         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4682         rsv->space_info = __find_space_info(root->fs_info,
4683                                             BTRFS_BLOCK_GROUP_METADATA);
4684         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4685                                   BTRFS_RESERVE_FLUSH_ALL);
4686         if (ret) {
4687                 if (*qgroup_reserved)
4688                         btrfs_qgroup_free(root, *qgroup_reserved);
4689         }
4690
4691         return ret;
4692 }
4693
4694 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4695                                       struct btrfs_block_rsv *rsv,
4696                                       u64 qgroup_reserved)
4697 {
4698         btrfs_block_rsv_release(root, rsv, (u64)-1);
4699         if (qgroup_reserved)
4700                 btrfs_qgroup_free(root, qgroup_reserved);
4701 }
4702
4703 /**
4704  * drop_outstanding_extent - drop an outstanding extent
4705  * @inode: the inode we're dropping the extent for
4706  *
4707  * This is called when we are freeing up an outstanding extent, either called
4708  * after an error or after an extent is written.  This will return the number of
4709  * reserved extents that need to be freed.  This must be called with
4710  * BTRFS_I(inode)->lock held.
4711  */
4712 static unsigned drop_outstanding_extent(struct inode *inode)
4713 {
4714         unsigned drop_inode_space = 0;
4715         unsigned dropped_extents = 0;
4716
4717         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4718         BTRFS_I(inode)->outstanding_extents--;
4719
4720         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4721             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4722                                &BTRFS_I(inode)->runtime_flags))
4723                 drop_inode_space = 1;
4724
4725         /*
4726          * If we have more or the same amount of outsanding extents than we have
4727          * reserved then we need to leave the reserved extents count alone.
4728          */
4729         if (BTRFS_I(inode)->outstanding_extents >=
4730             BTRFS_I(inode)->reserved_extents)
4731                 return drop_inode_space;
4732
4733         dropped_extents = BTRFS_I(inode)->reserved_extents -
4734                 BTRFS_I(inode)->outstanding_extents;
4735         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4736         return dropped_extents + drop_inode_space;
4737 }
4738
4739 /**
4740  * calc_csum_metadata_size - return the amount of metada space that must be
4741  *      reserved/free'd for the given bytes.
4742  * @inode: the inode we're manipulating
4743  * @num_bytes: the number of bytes in question
4744  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4745  *
4746  * This adjusts the number of csum_bytes in the inode and then returns the
4747  * correct amount of metadata that must either be reserved or freed.  We
4748  * calculate how many checksums we can fit into one leaf and then divide the
4749  * number of bytes that will need to be checksumed by this value to figure out
4750  * how many checksums will be required.  If we are adding bytes then the number
4751  * may go up and we will return the number of additional bytes that must be
4752  * reserved.  If it is going down we will return the number of bytes that must
4753  * be freed.
4754  *
4755  * This must be called with BTRFS_I(inode)->lock held.
4756  */
4757 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4758                                    int reserve)
4759 {
4760         struct btrfs_root *root = BTRFS_I(inode)->root;
4761         u64 csum_size;
4762         int num_csums_per_leaf;
4763         int num_csums;
4764         int old_csums;
4765
4766         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4767             BTRFS_I(inode)->csum_bytes == 0)
4768                 return 0;
4769
4770         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4771         if (reserve)
4772                 BTRFS_I(inode)->csum_bytes += num_bytes;
4773         else
4774                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4775         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4776         num_csums_per_leaf = (int)div64_u64(csum_size,
4777                                             sizeof(struct btrfs_csum_item) +
4778                                             sizeof(struct btrfs_disk_key));
4779         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4780         num_csums = num_csums + num_csums_per_leaf - 1;
4781         num_csums = num_csums / num_csums_per_leaf;
4782
4783         old_csums = old_csums + num_csums_per_leaf - 1;
4784         old_csums = old_csums / num_csums_per_leaf;
4785
4786         /* No change, no need to reserve more */
4787         if (old_csums == num_csums)
4788                 return 0;
4789
4790         if (reserve)
4791                 return btrfs_calc_trans_metadata_size(root,
4792                                                       num_csums - old_csums);
4793
4794         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4795 }
4796
4797 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4798 {
4799         struct btrfs_root *root = BTRFS_I(inode)->root;
4800         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4801         u64 to_reserve = 0;
4802         u64 csum_bytes;
4803         unsigned nr_extents = 0;
4804         int extra_reserve = 0;
4805         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
4806         int ret = 0;
4807         bool delalloc_lock = true;
4808         u64 to_free = 0;
4809         unsigned dropped;
4810
4811         /* If we are a free space inode we need to not flush since we will be in
4812          * the middle of a transaction commit.  We also don't need the delalloc
4813          * mutex since we won't race with anybody.  We need this mostly to make
4814          * lockdep shut its filthy mouth.
4815          */
4816         if (btrfs_is_free_space_inode(inode)) {
4817                 flush = BTRFS_RESERVE_NO_FLUSH;
4818                 delalloc_lock = false;
4819         }
4820
4821         if (flush != BTRFS_RESERVE_NO_FLUSH &&
4822             btrfs_transaction_in_commit(root->fs_info))
4823                 schedule_timeout(1);
4824
4825         if (delalloc_lock)
4826                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4827
4828         num_bytes = ALIGN(num_bytes, root->sectorsize);
4829
4830         spin_lock(&BTRFS_I(inode)->lock);
4831         BTRFS_I(inode)->outstanding_extents++;
4832
4833         if (BTRFS_I(inode)->outstanding_extents >
4834             BTRFS_I(inode)->reserved_extents)
4835                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4836                         BTRFS_I(inode)->reserved_extents;
4837
4838         /*
4839          * Add an item to reserve for updating the inode when we complete the
4840          * delalloc io.
4841          */
4842         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4843                       &BTRFS_I(inode)->runtime_flags)) {
4844                 nr_extents++;
4845                 extra_reserve = 1;
4846         }
4847
4848         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4849         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4850         csum_bytes = BTRFS_I(inode)->csum_bytes;
4851         spin_unlock(&BTRFS_I(inode)->lock);
4852
4853         if (root->fs_info->quota_enabled) {
4854                 ret = btrfs_qgroup_reserve(root, num_bytes +
4855                                            nr_extents * root->leafsize);
4856                 if (ret)
4857                         goto out_fail;
4858         }
4859
4860         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4861         if (unlikely(ret)) {
4862                 if (root->fs_info->quota_enabled)
4863                         btrfs_qgroup_free(root, num_bytes +
4864                                                 nr_extents * root->leafsize);
4865                 goto out_fail;
4866         }
4867
4868         spin_lock(&BTRFS_I(inode)->lock);
4869         if (extra_reserve) {
4870                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4871                         &BTRFS_I(inode)->runtime_flags);
4872                 nr_extents--;
4873         }
4874         BTRFS_I(inode)->reserved_extents += nr_extents;
4875         spin_unlock(&BTRFS_I(inode)->lock);
4876
4877         if (delalloc_lock)
4878                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4879
4880         if (to_reserve)
4881                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4882                                               btrfs_ino(inode), to_reserve, 1);
4883         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4884
4885         return 0;
4886
4887 out_fail:
4888         spin_lock(&BTRFS_I(inode)->lock);
4889         dropped = drop_outstanding_extent(inode);
4890         /*
4891          * If the inodes csum_bytes is the same as the original
4892          * csum_bytes then we know we haven't raced with any free()ers
4893          * so we can just reduce our inodes csum bytes and carry on.
4894          */
4895         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
4896                 calc_csum_metadata_size(inode, num_bytes, 0);
4897         } else {
4898                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
4899                 u64 bytes;
4900
4901                 /*
4902                  * This is tricky, but first we need to figure out how much we
4903                  * free'd from any free-ers that occured during this
4904                  * reservation, so we reset ->csum_bytes to the csum_bytes
4905                  * before we dropped our lock, and then call the free for the
4906                  * number of bytes that were freed while we were trying our
4907                  * reservation.
4908                  */
4909                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
4910                 BTRFS_I(inode)->csum_bytes = csum_bytes;
4911                 to_free = calc_csum_metadata_size(inode, bytes, 0);
4912
4913
4914                 /*
4915                  * Now we need to see how much we would have freed had we not
4916                  * been making this reservation and our ->csum_bytes were not
4917                  * artificially inflated.
4918                  */
4919                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
4920                 bytes = csum_bytes - orig_csum_bytes;
4921                 bytes = calc_csum_metadata_size(inode, bytes, 0);
4922
4923                 /*
4924                  * Now reset ->csum_bytes to what it should be.  If bytes is
4925                  * more than to_free then we would have free'd more space had we
4926                  * not had an artificially high ->csum_bytes, so we need to free
4927                  * the remainder.  If bytes is the same or less then we don't
4928                  * need to do anything, the other free-ers did the correct
4929                  * thing.
4930                  */
4931                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
4932                 if (bytes > to_free)
4933                         to_free = bytes - to_free;
4934                 else
4935                         to_free = 0;
4936         }
4937         spin_unlock(&BTRFS_I(inode)->lock);
4938         if (dropped)
4939                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4940
4941         if (to_free) {
4942                 btrfs_block_rsv_release(root, block_rsv, to_free);
4943                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
4944                                               btrfs_ino(inode), to_free, 0);
4945         }
4946         if (delalloc_lock)
4947                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4948         return ret;
4949 }
4950
4951 /**
4952  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4953  * @inode: the inode to release the reservation for
4954  * @num_bytes: the number of bytes we're releasing
4955  *
4956  * This will release the metadata reservation for an inode.  This can be called
4957  * once we complete IO for a given set of bytes to release their metadata
4958  * reservations.
4959  */
4960 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4961 {
4962         struct btrfs_root *root = BTRFS_I(inode)->root;
4963         u64 to_free = 0;
4964         unsigned dropped;
4965
4966         num_bytes = ALIGN(num_bytes, root->sectorsize);
4967         spin_lock(&BTRFS_I(inode)->lock);
4968         dropped = drop_outstanding_extent(inode);
4969
4970         if (num_bytes)
4971                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4972         spin_unlock(&BTRFS_I(inode)->lock);
4973         if (dropped > 0)
4974                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4975
4976         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4977                                       btrfs_ino(inode), to_free, 0);
4978         if (root->fs_info->quota_enabled) {
4979                 btrfs_qgroup_free(root, num_bytes +
4980                                         dropped * root->leafsize);
4981         }
4982
4983         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4984                                 to_free);
4985 }
4986
4987 /**
4988  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4989  * @inode: inode we're writing to
4990  * @num_bytes: the number of bytes we want to allocate
4991  *
4992  * This will do the following things
4993  *
4994  * o reserve space in the data space info for num_bytes
4995  * o reserve space in the metadata space info based on number of outstanding
4996  *   extents and how much csums will be needed
4997  * o add to the inodes ->delalloc_bytes
4998  * o add it to the fs_info's delalloc inodes list.
4999  *
5000  * This will return 0 for success and -ENOSPC if there is no space left.
5001  */
5002 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5003 {
5004         int ret;
5005
5006         ret = btrfs_check_data_free_space(inode, num_bytes);
5007         if (ret)
5008                 return ret;
5009
5010         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5011         if (ret) {
5012                 btrfs_free_reserved_data_space(inode, num_bytes);
5013                 return ret;
5014         }
5015
5016         return 0;
5017 }
5018
5019 /**
5020  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5021  * @inode: inode we're releasing space for
5022  * @num_bytes: the number of bytes we want to free up
5023  *
5024  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5025  * called in the case that we don't need the metadata AND data reservations
5026  * anymore.  So if there is an error or we insert an inline extent.
5027  *
5028  * This function will release the metadata space that was not used and will
5029  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5030  * list if there are no delalloc bytes left.
5031  */
5032 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5033 {
5034         btrfs_delalloc_release_metadata(inode, num_bytes);
5035         btrfs_free_reserved_data_space(inode, num_bytes);
5036 }
5037
5038 static int update_block_group(struct btrfs_root *root,
5039                               u64 bytenr, u64 num_bytes, int alloc)
5040 {
5041         struct btrfs_block_group_cache *cache = NULL;
5042         struct btrfs_fs_info *info = root->fs_info;
5043         u64 total = num_bytes;
5044         u64 old_val;
5045         u64 byte_in_group;
5046         int factor;
5047
5048         /* block accounting for super block */
5049         spin_lock(&info->delalloc_lock);
5050         old_val = btrfs_super_bytes_used(info->super_copy);
5051         if (alloc)
5052                 old_val += num_bytes;
5053         else
5054                 old_val -= num_bytes;
5055         btrfs_set_super_bytes_used(info->super_copy, old_val);
5056         spin_unlock(&info->delalloc_lock);
5057
5058         while (total) {
5059                 cache = btrfs_lookup_block_group(info, bytenr);
5060                 if (!cache)
5061                         return -ENOENT;
5062                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5063                                     BTRFS_BLOCK_GROUP_RAID1 |
5064                                     BTRFS_BLOCK_GROUP_RAID10))
5065                         factor = 2;
5066                 else
5067                         factor = 1;
5068                 /*
5069                  * If this block group has free space cache written out, we
5070                  * need to make sure to load it if we are removing space.  This
5071                  * is because we need the unpinning stage to actually add the
5072                  * space back to the block group, otherwise we will leak space.
5073                  */
5074                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5075                         cache_block_group(cache, 1);
5076
5077                 byte_in_group = bytenr - cache->key.objectid;
5078                 WARN_ON(byte_in_group > cache->key.offset);
5079
5080                 spin_lock(&cache->space_info->lock);
5081                 spin_lock(&cache->lock);
5082
5083                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5084                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5085                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5086
5087                 cache->dirty = 1;
5088                 old_val = btrfs_block_group_used(&cache->item);
5089                 num_bytes = min(total, cache->key.offset - byte_in_group);
5090                 if (alloc) {
5091                         old_val += num_bytes;
5092                         btrfs_set_block_group_used(&cache->item, old_val);
5093                         cache->reserved -= num_bytes;
5094                         cache->space_info->bytes_reserved -= num_bytes;
5095                         cache->space_info->bytes_used += num_bytes;
5096                         cache->space_info->disk_used += num_bytes * factor;
5097                         spin_unlock(&cache->lock);
5098                         spin_unlock(&cache->space_info->lock);
5099                 } else {
5100                         old_val -= num_bytes;
5101                         btrfs_set_block_group_used(&cache->item, old_val);
5102                         cache->pinned += num_bytes;
5103                         cache->space_info->bytes_pinned += num_bytes;
5104                         cache->space_info->bytes_used -= num_bytes;
5105                         cache->space_info->disk_used -= num_bytes * factor;
5106                         spin_unlock(&cache->lock);
5107                         spin_unlock(&cache->space_info->lock);
5108
5109                         set_extent_dirty(info->pinned_extents,
5110                                          bytenr, bytenr + num_bytes - 1,
5111                                          GFP_NOFS | __GFP_NOFAIL);
5112                 }
5113                 btrfs_put_block_group(cache);
5114                 total -= num_bytes;
5115                 bytenr += num_bytes;
5116         }
5117         return 0;
5118 }
5119
5120 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5121 {
5122         struct btrfs_block_group_cache *cache;
5123         u64 bytenr;
5124
5125         spin_lock(&root->fs_info->block_group_cache_lock);
5126         bytenr = root->fs_info->first_logical_byte;
5127         spin_unlock(&root->fs_info->block_group_cache_lock);
5128
5129         if (bytenr < (u64)-1)
5130                 return bytenr;
5131
5132         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5133         if (!cache)
5134                 return 0;
5135
5136         bytenr = cache->key.objectid;
5137         btrfs_put_block_group(cache);
5138
5139         return bytenr;
5140 }
5141
5142 static int pin_down_extent(struct btrfs_root *root,
5143                            struct btrfs_block_group_cache *cache,
5144                            u64 bytenr, u64 num_bytes, int reserved)
5145 {
5146         spin_lock(&cache->space_info->lock);
5147         spin_lock(&cache->lock);
5148         cache->pinned += num_bytes;
5149         cache->space_info->bytes_pinned += num_bytes;
5150         if (reserved) {
5151                 cache->reserved -= num_bytes;
5152                 cache->space_info->bytes_reserved -= num_bytes;
5153         }
5154         spin_unlock(&cache->lock);
5155         spin_unlock(&cache->space_info->lock);
5156
5157         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5158                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5159         return 0;
5160 }
5161
5162 /*
5163  * this function must be called within transaction
5164  */
5165 int btrfs_pin_extent(struct btrfs_root *root,
5166                      u64 bytenr, u64 num_bytes, int reserved)
5167 {
5168         struct btrfs_block_group_cache *cache;
5169
5170         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5171         BUG_ON(!cache); /* Logic error */
5172
5173         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5174
5175         btrfs_put_block_group(cache);
5176         return 0;
5177 }
5178
5179 /*
5180  * this function must be called within transaction
5181  */
5182 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5183                                     u64 bytenr, u64 num_bytes)
5184 {
5185         struct btrfs_block_group_cache *cache;
5186
5187         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5188         BUG_ON(!cache); /* Logic error */
5189
5190         /*
5191          * pull in the free space cache (if any) so that our pin
5192          * removes the free space from the cache.  We have load_only set
5193          * to one because the slow code to read in the free extents does check
5194          * the pinned extents.
5195          */
5196         cache_block_group(cache, 1);
5197
5198         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5199
5200         /* remove us from the free space cache (if we're there at all) */
5201         btrfs_remove_free_space(cache, bytenr, num_bytes);
5202         btrfs_put_block_group(cache);
5203         return 0;
5204 }
5205
5206 /**
5207  * btrfs_update_reserved_bytes - update the block_group and space info counters
5208  * @cache:      The cache we are manipulating
5209  * @num_bytes:  The number of bytes in question
5210  * @reserve:    One of the reservation enums
5211  *
5212  * This is called by the allocator when it reserves space, or by somebody who is
5213  * freeing space that was never actually used on disk.  For example if you
5214  * reserve some space for a new leaf in transaction A and before transaction A
5215  * commits you free that leaf, you call this with reserve set to 0 in order to
5216  * clear the reservation.
5217  *
5218  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5219  * ENOSPC accounting.  For data we handle the reservation through clearing the
5220  * delalloc bits in the io_tree.  We have to do this since we could end up
5221  * allocating less disk space for the amount of data we have reserved in the
5222  * case of compression.
5223  *
5224  * If this is a reservation and the block group has become read only we cannot
5225  * make the reservation and return -EAGAIN, otherwise this function always
5226  * succeeds.
5227  */
5228 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5229                                        u64 num_bytes, int reserve)
5230 {
5231         struct btrfs_space_info *space_info = cache->space_info;
5232         int ret = 0;
5233
5234         spin_lock(&space_info->lock);
5235         spin_lock(&cache->lock);
5236         if (reserve != RESERVE_FREE) {
5237                 if (cache->ro) {
5238                         ret = -EAGAIN;
5239                 } else {
5240                         cache->reserved += num_bytes;
5241                         space_info->bytes_reserved += num_bytes;
5242                         if (reserve == RESERVE_ALLOC) {
5243                                 trace_btrfs_space_reservation(cache->fs_info,
5244                                                 "space_info", space_info->flags,
5245                                                 num_bytes, 0);
5246                                 space_info->bytes_may_use -= num_bytes;
5247                         }
5248                 }
5249         } else {
5250                 if (cache->ro)
5251                         space_info->bytes_readonly += num_bytes;
5252                 cache->reserved -= num_bytes;
5253                 space_info->bytes_reserved -= num_bytes;
5254                 space_info->reservation_progress++;
5255         }
5256         spin_unlock(&cache->lock);
5257         spin_unlock(&space_info->lock);
5258         return ret;
5259 }
5260
5261 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5262                                 struct btrfs_root *root)
5263 {
5264         struct btrfs_fs_info *fs_info = root->fs_info;
5265         struct btrfs_caching_control *next;
5266         struct btrfs_caching_control *caching_ctl;
5267         struct btrfs_block_group_cache *cache;
5268
5269         down_write(&fs_info->extent_commit_sem);
5270
5271         list_for_each_entry_safe(caching_ctl, next,
5272                                  &fs_info->caching_block_groups, list) {
5273                 cache = caching_ctl->block_group;
5274                 if (block_group_cache_done(cache)) {
5275                         cache->last_byte_to_unpin = (u64)-1;
5276                         list_del_init(&caching_ctl->list);
5277                         put_caching_control(caching_ctl);
5278                 } else {
5279                         cache->last_byte_to_unpin = caching_ctl->progress;
5280                 }
5281         }
5282
5283         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5284                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5285         else
5286                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5287
5288         up_write(&fs_info->extent_commit_sem);
5289
5290         update_global_block_rsv(fs_info);
5291 }
5292
5293 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5294 {
5295         struct btrfs_fs_info *fs_info = root->fs_info;
5296         struct btrfs_block_group_cache *cache = NULL;
5297         struct btrfs_space_info *space_info;
5298         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5299         u64 len;
5300         bool readonly;
5301
5302         while (start <= end) {
5303                 readonly = false;
5304                 if (!cache ||
5305                     start >= cache->key.objectid + cache->key.offset) {
5306                         if (cache)
5307                                 btrfs_put_block_group(cache);
5308                         cache = btrfs_lookup_block_group(fs_info, start);
5309                         BUG_ON(!cache); /* Logic error */
5310                 }
5311
5312                 len = cache->key.objectid + cache->key.offset - start;
5313                 len = min(len, end + 1 - start);
5314
5315                 if (start < cache->last_byte_to_unpin) {
5316                         len = min(len, cache->last_byte_to_unpin - start);
5317                         btrfs_add_free_space(cache, start, len);
5318                 }
5319
5320                 start += len;
5321                 space_info = cache->space_info;
5322
5323                 spin_lock(&space_info->lock);
5324                 spin_lock(&cache->lock);
5325                 cache->pinned -= len;
5326                 space_info->bytes_pinned -= len;
5327                 if (cache->ro) {
5328                         space_info->bytes_readonly += len;
5329                         readonly = true;
5330                 }
5331                 spin_unlock(&cache->lock);
5332                 if (!readonly && global_rsv->space_info == space_info) {
5333                         spin_lock(&global_rsv->lock);
5334                         if (!global_rsv->full) {
5335                                 len = min(len, global_rsv->size -
5336                                           global_rsv->reserved);
5337                                 global_rsv->reserved += len;
5338                                 space_info->bytes_may_use += len;
5339                                 if (global_rsv->reserved >= global_rsv->size)
5340                                         global_rsv->full = 1;
5341                         }
5342                         spin_unlock(&global_rsv->lock);
5343                 }
5344                 spin_unlock(&space_info->lock);
5345         }
5346
5347         if (cache)
5348                 btrfs_put_block_group(cache);
5349         return 0;
5350 }
5351
5352 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5353                                struct btrfs_root *root)
5354 {
5355         struct btrfs_fs_info *fs_info = root->fs_info;
5356         struct extent_io_tree *unpin;
5357         u64 start;
5358         u64 end;
5359         int ret;
5360
5361         if (trans->aborted)
5362                 return 0;
5363
5364         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5365                 unpin = &fs_info->freed_extents[1];
5366         else
5367                 unpin = &fs_info->freed_extents[0];
5368
5369         while (1) {
5370                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5371                                             EXTENT_DIRTY, NULL);
5372                 if (ret)
5373                         break;
5374
5375                 if (btrfs_test_opt(root, DISCARD))
5376                         ret = btrfs_discard_extent(root, start,
5377                                                    end + 1 - start, NULL);
5378
5379                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5380                 unpin_extent_range(root, start, end);
5381                 cond_resched();
5382         }
5383
5384         return 0;
5385 }
5386
5387 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5388                                 struct btrfs_root *root,
5389                                 u64 bytenr, u64 num_bytes, u64 parent,
5390                                 u64 root_objectid, u64 owner_objectid,
5391                                 u64 owner_offset, int refs_to_drop,
5392                                 struct btrfs_delayed_extent_op *extent_op)
5393 {
5394         struct btrfs_key key;
5395         struct btrfs_path *path;
5396         struct btrfs_fs_info *info = root->fs_info;
5397         struct btrfs_root *extent_root = info->extent_root;
5398         struct extent_buffer *leaf;
5399         struct btrfs_extent_item *ei;
5400         struct btrfs_extent_inline_ref *iref;
5401         int ret;
5402         int is_data;
5403         int extent_slot = 0;
5404         int found_extent = 0;
5405         int num_to_del = 1;
5406         u32 item_size;
5407         u64 refs;
5408         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5409                                                  SKINNY_METADATA);
5410
5411         path = btrfs_alloc_path();
5412         if (!path)
5413                 return -ENOMEM;
5414
5415         path->reada = 1;
5416         path->leave_spinning = 1;
5417
5418         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5419         BUG_ON(!is_data && refs_to_drop != 1);
5420
5421         if (is_data)
5422                 skinny_metadata = 0;
5423
5424         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5425                                     bytenr, num_bytes, parent,
5426                                     root_objectid, owner_objectid,
5427                                     owner_offset);
5428         if (ret == 0) {
5429                 extent_slot = path->slots[0];
5430                 while (extent_slot >= 0) {
5431                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5432                                               extent_slot);
5433                         if (key.objectid != bytenr)
5434                                 break;
5435                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5436                             key.offset == num_bytes) {
5437                                 found_extent = 1;
5438                                 break;
5439                         }
5440                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5441                             key.offset == owner_objectid) {
5442                                 found_extent = 1;
5443                                 break;
5444                         }
5445                         if (path->slots[0] - extent_slot > 5)
5446                                 break;
5447                         extent_slot--;
5448                 }
5449 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5450                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5451                 if (found_extent && item_size < sizeof(*ei))
5452                         found_extent = 0;
5453 #endif
5454                 if (!found_extent) {
5455                         BUG_ON(iref);
5456                         ret = remove_extent_backref(trans, extent_root, path,
5457                                                     NULL, refs_to_drop,
5458                                                     is_data);
5459                         if (ret) {
5460                                 btrfs_abort_transaction(trans, extent_root, ret);
5461                                 goto out;
5462                         }
5463                         btrfs_release_path(path);
5464                         path->leave_spinning = 1;
5465
5466                         key.objectid = bytenr;
5467                         key.type = BTRFS_EXTENT_ITEM_KEY;
5468                         key.offset = num_bytes;
5469
5470                         if (!is_data && skinny_metadata) {
5471                                 key.type = BTRFS_METADATA_ITEM_KEY;
5472                                 key.offset = owner_objectid;
5473                         }
5474
5475                         ret = btrfs_search_slot(trans, extent_root,
5476                                                 &key, path, -1, 1);
5477                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5478                                 /*
5479                                  * Couldn't find our skinny metadata item,
5480                                  * see if we have ye olde extent item.
5481                                  */
5482                                 path->slots[0]--;
5483                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5484                                                       path->slots[0]);
5485                                 if (key.objectid == bytenr &&
5486                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5487                                     key.offset == num_bytes)
5488                                         ret = 0;
5489                         }
5490
5491                         if (ret > 0 && skinny_metadata) {
5492                                 skinny_metadata = false;
5493                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5494                                 key.offset = num_bytes;
5495                                 btrfs_release_path(path);
5496                                 ret = btrfs_search_slot(trans, extent_root,
5497                                                         &key, path, -1, 1);
5498                         }
5499
5500                         if (ret) {
5501                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5502                                         ret, (unsigned long long)bytenr);
5503                                 if (ret > 0)
5504                                         btrfs_print_leaf(extent_root,
5505                                                          path->nodes[0]);
5506                         }
5507                         if (ret < 0) {
5508                                 btrfs_abort_transaction(trans, extent_root, ret);
5509                                 goto out;
5510                         }
5511                         extent_slot = path->slots[0];
5512                 }
5513         } else if (ret == -ENOENT) {
5514                 btrfs_print_leaf(extent_root, path->nodes[0]);
5515                 WARN_ON(1);
5516                 btrfs_err(info,
5517                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5518                         (unsigned long long)bytenr,
5519                         (unsigned long long)parent,
5520                         (unsigned long long)root_objectid,
5521                         (unsigned long long)owner_objectid,
5522                         (unsigned long long)owner_offset);
5523         } else {
5524                 btrfs_abort_transaction(trans, extent_root, ret);
5525                 goto out;
5526         }
5527
5528         leaf = path->nodes[0];
5529         item_size = btrfs_item_size_nr(leaf, extent_slot);
5530 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5531         if (item_size < sizeof(*ei)) {
5532                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5533                 ret = convert_extent_item_v0(trans, extent_root, path,
5534                                              owner_objectid, 0);
5535                 if (ret < 0) {
5536                         btrfs_abort_transaction(trans, extent_root, ret);
5537                         goto out;
5538                 }
5539
5540                 btrfs_release_path(path);
5541                 path->leave_spinning = 1;
5542
5543                 key.objectid = bytenr;
5544                 key.type = BTRFS_EXTENT_ITEM_KEY;
5545                 key.offset = num_bytes;
5546
5547                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5548                                         -1, 1);
5549                 if (ret) {
5550                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5551                                 ret, (unsigned long long)bytenr);
5552                         btrfs_print_leaf(extent_root, path->nodes[0]);
5553                 }
5554                 if (ret < 0) {
5555                         btrfs_abort_transaction(trans, extent_root, ret);
5556                         goto out;
5557                 }
5558
5559                 extent_slot = path->slots[0];
5560                 leaf = path->nodes[0];
5561                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5562         }
5563 #endif
5564         BUG_ON(item_size < sizeof(*ei));
5565         ei = btrfs_item_ptr(leaf, extent_slot,
5566                             struct btrfs_extent_item);
5567         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
5568             key.type == BTRFS_EXTENT_ITEM_KEY) {
5569                 struct btrfs_tree_block_info *bi;
5570                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5571                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5572                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5573         }
5574
5575         refs = btrfs_extent_refs(leaf, ei);
5576         BUG_ON(refs < refs_to_drop);
5577         refs -= refs_to_drop;
5578
5579         if (refs > 0) {
5580                 if (extent_op)
5581                         __run_delayed_extent_op(extent_op, leaf, ei);
5582                 /*
5583                  * In the case of inline back ref, reference count will
5584                  * be updated by remove_extent_backref
5585                  */
5586                 if (iref) {
5587                         BUG_ON(!found_extent);
5588                 } else {
5589                         btrfs_set_extent_refs(leaf, ei, refs);
5590                         btrfs_mark_buffer_dirty(leaf);
5591                 }
5592                 if (found_extent) {
5593                         ret = remove_extent_backref(trans, extent_root, path,
5594                                                     iref, refs_to_drop,
5595                                                     is_data);
5596                         if (ret) {
5597                                 btrfs_abort_transaction(trans, extent_root, ret);
5598                                 goto out;
5599                         }
5600                 }
5601         } else {
5602                 if (found_extent) {
5603                         BUG_ON(is_data && refs_to_drop !=
5604                                extent_data_ref_count(root, path, iref));
5605                         if (iref) {
5606                                 BUG_ON(path->slots[0] != extent_slot);
5607                         } else {
5608                                 BUG_ON(path->slots[0] != extent_slot + 1);
5609                                 path->slots[0] = extent_slot;
5610                                 num_to_del = 2;
5611                         }
5612                 }
5613
5614                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5615                                       num_to_del);
5616                 if (ret) {
5617                         btrfs_abort_transaction(trans, extent_root, ret);
5618                         goto out;
5619                 }
5620                 btrfs_release_path(path);
5621
5622                 if (is_data) {
5623                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5624                         if (ret) {
5625                                 btrfs_abort_transaction(trans, extent_root, ret);
5626                                 goto out;
5627                         }
5628                 }
5629
5630                 ret = update_block_group(root, bytenr, num_bytes, 0);
5631                 if (ret) {
5632                         btrfs_abort_transaction(trans, extent_root, ret);
5633                         goto out;
5634                 }
5635         }
5636 out:
5637         btrfs_free_path(path);
5638         return ret;
5639 }
5640
5641 /*
5642  * when we free an block, it is possible (and likely) that we free the last
5643  * delayed ref for that extent as well.  This searches the delayed ref tree for
5644  * a given extent, and if there are no other delayed refs to be processed, it
5645  * removes it from the tree.
5646  */
5647 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5648                                       struct btrfs_root *root, u64 bytenr)
5649 {
5650         struct btrfs_delayed_ref_head *head;
5651         struct btrfs_delayed_ref_root *delayed_refs;
5652         struct btrfs_delayed_ref_node *ref;
5653         struct rb_node *node;
5654         int ret = 0;
5655
5656         delayed_refs = &trans->transaction->delayed_refs;
5657         spin_lock(&delayed_refs->lock);
5658         head = btrfs_find_delayed_ref_head(trans, bytenr);
5659         if (!head)
5660                 goto out;
5661
5662         node = rb_prev(&head->node.rb_node);
5663         if (!node)
5664                 goto out;
5665
5666         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5667
5668         /* there are still entries for this ref, we can't drop it */
5669         if (ref->bytenr == bytenr)
5670                 goto out;
5671
5672         if (head->extent_op) {
5673                 if (!head->must_insert_reserved)
5674                         goto out;
5675                 btrfs_free_delayed_extent_op(head->extent_op);
5676                 head->extent_op = NULL;
5677         }
5678
5679         /*
5680          * waiting for the lock here would deadlock.  If someone else has it
5681          * locked they are already in the process of dropping it anyway
5682          */
5683         if (!mutex_trylock(&head->mutex))
5684                 goto out;
5685
5686         /*
5687          * at this point we have a head with no other entries.  Go
5688          * ahead and process it.
5689          */
5690         head->node.in_tree = 0;
5691         rb_erase(&head->node.rb_node, &delayed_refs->root);
5692
5693         delayed_refs->num_entries--;
5694
5695         /*
5696          * we don't take a ref on the node because we're removing it from the
5697          * tree, so we just steal the ref the tree was holding.
5698          */
5699         delayed_refs->num_heads--;
5700         if (list_empty(&head->cluster))
5701                 delayed_refs->num_heads_ready--;
5702
5703         list_del_init(&head->cluster);
5704         spin_unlock(&delayed_refs->lock);
5705
5706         BUG_ON(head->extent_op);
5707         if (head->must_insert_reserved)
5708                 ret = 1;
5709
5710         mutex_unlock(&head->mutex);
5711         btrfs_put_delayed_ref(&head->node);
5712         return ret;
5713 out:
5714         spin_unlock(&delayed_refs->lock);
5715         return 0;
5716 }
5717
5718 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5719                            struct btrfs_root *root,
5720                            struct extent_buffer *buf,
5721                            u64 parent, int last_ref)
5722 {
5723         struct btrfs_block_group_cache *cache = NULL;
5724         int ret;
5725
5726         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5727                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5728                                         buf->start, buf->len,
5729                                         parent, root->root_key.objectid,
5730                                         btrfs_header_level(buf),
5731                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5732                 BUG_ON(ret); /* -ENOMEM */
5733         }
5734
5735         if (!last_ref)
5736                 return;
5737
5738         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5739
5740         if (btrfs_header_generation(buf) == trans->transid) {
5741                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5742                         ret = check_ref_cleanup(trans, root, buf->start);
5743                         if (!ret)
5744                                 goto out;
5745                 }
5746
5747                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5748                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5749                         goto out;
5750                 }
5751
5752                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5753
5754                 btrfs_add_free_space(cache, buf->start, buf->len);
5755                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5756         }
5757 out:
5758         /*
5759          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5760          * anymore.
5761          */
5762         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5763         btrfs_put_block_group(cache);
5764 }
5765
5766 /* Can return -ENOMEM */
5767 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5768                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5769                       u64 owner, u64 offset, int for_cow)
5770 {
5771         int ret;
5772         struct btrfs_fs_info *fs_info = root->fs_info;
5773
5774         /*
5775          * tree log blocks never actually go into the extent allocation
5776          * tree, just update pinning info and exit early.
5777          */
5778         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5779                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5780                 /* unlocks the pinned mutex */
5781                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5782                 ret = 0;
5783         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5784                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5785                                         num_bytes,
5786                                         parent, root_objectid, (int)owner,
5787                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5788         } else {
5789                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5790                                                 num_bytes,
5791                                                 parent, root_objectid, owner,
5792                                                 offset, BTRFS_DROP_DELAYED_REF,
5793                                                 NULL, for_cow);
5794         }
5795         return ret;
5796 }
5797
5798 static u64 stripe_align(struct btrfs_root *root,
5799                         struct btrfs_block_group_cache *cache,
5800                         u64 val, u64 num_bytes)
5801 {
5802         u64 ret = ALIGN(val, root->stripesize);
5803         return ret;
5804 }
5805
5806 /*
5807  * when we wait for progress in the block group caching, its because
5808  * our allocation attempt failed at least once.  So, we must sleep
5809  * and let some progress happen before we try again.
5810  *
5811  * This function will sleep at least once waiting for new free space to
5812  * show up, and then it will check the block group free space numbers
5813  * for our min num_bytes.  Another option is to have it go ahead
5814  * and look in the rbtree for a free extent of a given size, but this
5815  * is a good start.
5816  */
5817 static noinline int
5818 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5819                                 u64 num_bytes)
5820 {
5821         struct btrfs_caching_control *caching_ctl;
5822
5823         caching_ctl = get_caching_control(cache);
5824         if (!caching_ctl)
5825                 return 0;
5826
5827         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5828                    (cache->free_space_ctl->free_space >= num_bytes));
5829
5830         put_caching_control(caching_ctl);
5831         return 0;
5832 }
5833
5834 static noinline int
5835 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5836 {
5837         struct btrfs_caching_control *caching_ctl;
5838
5839         caching_ctl = get_caching_control(cache);
5840         if (!caching_ctl)
5841                 return 0;
5842
5843         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5844
5845         put_caching_control(caching_ctl);
5846         return 0;
5847 }
5848
5849 int __get_raid_index(u64 flags)
5850 {
5851         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5852                 return BTRFS_RAID_RAID10;
5853         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5854                 return BTRFS_RAID_RAID1;
5855         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5856                 return BTRFS_RAID_DUP;
5857         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5858                 return BTRFS_RAID_RAID0;
5859         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
5860                 return BTRFS_RAID_RAID5;
5861         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
5862                 return BTRFS_RAID_RAID6;
5863
5864         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
5865 }
5866
5867 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5868 {
5869         return __get_raid_index(cache->flags);
5870 }
5871
5872 enum btrfs_loop_type {
5873         LOOP_CACHING_NOWAIT = 0,
5874         LOOP_CACHING_WAIT = 1,
5875         LOOP_ALLOC_CHUNK = 2,
5876         LOOP_NO_EMPTY_SIZE = 3,
5877 };
5878
5879 /*
5880  * walks the btree of allocated extents and find a hole of a given size.
5881  * The key ins is changed to record the hole:
5882  * ins->objectid == block start
5883  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5884  * ins->offset == number of blocks
5885  * Any available blocks before search_start are skipped.
5886  */
5887 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5888                                      struct btrfs_root *orig_root,
5889                                      u64 num_bytes, u64 empty_size,
5890                                      u64 hint_byte, struct btrfs_key *ins,
5891                                      u64 data)
5892 {
5893         int ret = 0;
5894         struct btrfs_root *root = orig_root->fs_info->extent_root;
5895         struct btrfs_free_cluster *last_ptr = NULL;
5896         struct btrfs_block_group_cache *block_group = NULL;
5897         struct btrfs_block_group_cache *used_block_group;
5898         u64 search_start = 0;
5899         int empty_cluster = 2 * 1024 * 1024;
5900         struct btrfs_space_info *space_info;
5901         int loop = 0;
5902         int index = __get_raid_index(data);
5903         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5904                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5905         bool found_uncached_bg = false;
5906         bool failed_cluster_refill = false;
5907         bool failed_alloc = false;
5908         bool use_cluster = true;
5909         bool have_caching_bg = false;
5910
5911         WARN_ON(num_bytes < root->sectorsize);
5912         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5913         ins->objectid = 0;
5914         ins->offset = 0;
5915
5916         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5917
5918         space_info = __find_space_info(root->fs_info, data);
5919         if (!space_info) {
5920                 btrfs_err(root->fs_info, "No space info for %llu", data);
5921                 return -ENOSPC;
5922         }
5923
5924         /*
5925          * If the space info is for both data and metadata it means we have a
5926          * small filesystem and we can't use the clustering stuff.
5927          */
5928         if (btrfs_mixed_space_info(space_info))
5929                 use_cluster = false;
5930
5931         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5932                 last_ptr = &root->fs_info->meta_alloc_cluster;
5933                 if (!btrfs_test_opt(root, SSD))
5934                         empty_cluster = 64 * 1024;
5935         }
5936
5937         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5938             btrfs_test_opt(root, SSD)) {
5939                 last_ptr = &root->fs_info->data_alloc_cluster;
5940         }
5941
5942         if (last_ptr) {
5943                 spin_lock(&last_ptr->lock);
5944                 if (last_ptr->block_group)
5945                         hint_byte = last_ptr->window_start;
5946                 spin_unlock(&last_ptr->lock);
5947         }
5948
5949         search_start = max(search_start, first_logical_byte(root, 0));
5950         search_start = max(search_start, hint_byte);
5951
5952         if (!last_ptr)
5953                 empty_cluster = 0;
5954
5955         if (search_start == hint_byte) {
5956                 block_group = btrfs_lookup_block_group(root->fs_info,
5957                                                        search_start);
5958                 used_block_group = block_group;
5959                 /*
5960                  * we don't want to use the block group if it doesn't match our
5961                  * allocation bits, or if its not cached.
5962                  *
5963                  * However if we are re-searching with an ideal block group
5964                  * picked out then we don't care that the block group is cached.
5965                  */
5966                 if (block_group && block_group_bits(block_group, data) &&
5967                     block_group->cached != BTRFS_CACHE_NO) {
5968                         down_read(&space_info->groups_sem);
5969                         if (list_empty(&block_group->list) ||
5970                             block_group->ro) {
5971                                 /*
5972                                  * someone is removing this block group,
5973                                  * we can't jump into the have_block_group
5974                                  * target because our list pointers are not
5975                                  * valid
5976                                  */
5977                                 btrfs_put_block_group(block_group);
5978                                 up_read(&space_info->groups_sem);
5979                         } else {
5980                                 index = get_block_group_index(block_group);
5981                                 goto have_block_group;
5982                         }
5983                 } else if (block_group) {
5984                         btrfs_put_block_group(block_group);
5985                 }
5986         }
5987 search:
5988         have_caching_bg = false;
5989         down_read(&space_info->groups_sem);
5990         list_for_each_entry(block_group, &space_info->block_groups[index],
5991                             list) {
5992                 u64 offset;
5993                 int cached;
5994
5995                 used_block_group = block_group;
5996                 btrfs_get_block_group(block_group);
5997                 search_start = block_group->key.objectid;
5998
5999                 /*
6000                  * this can happen if we end up cycling through all the
6001                  * raid types, but we want to make sure we only allocate
6002                  * for the proper type.
6003                  */
6004                 if (!block_group_bits(block_group, data)) {
6005                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6006                                 BTRFS_BLOCK_GROUP_RAID1 |
6007                                 BTRFS_BLOCK_GROUP_RAID5 |
6008                                 BTRFS_BLOCK_GROUP_RAID6 |
6009                                 BTRFS_BLOCK_GROUP_RAID10;
6010
6011                         /*
6012                          * if they asked for extra copies and this block group
6013                          * doesn't provide them, bail.  This does allow us to
6014                          * fill raid0 from raid1.
6015                          */
6016                         if ((data & extra) && !(block_group->flags & extra))
6017                                 goto loop;
6018                 }
6019
6020 have_block_group:
6021                 cached = block_group_cache_done(block_group);
6022                 if (unlikely(!cached)) {
6023                         found_uncached_bg = true;
6024                         ret = cache_block_group(block_group, 0);
6025                         BUG_ON(ret < 0);
6026                         ret = 0;
6027                 }
6028
6029                 if (unlikely(block_group->ro))
6030                         goto loop;
6031
6032                 /*
6033                  * Ok we want to try and use the cluster allocator, so
6034                  * lets look there
6035                  */
6036                 if (last_ptr) {
6037                         unsigned long aligned_cluster;
6038                         /*
6039                          * the refill lock keeps out other
6040                          * people trying to start a new cluster
6041                          */
6042                         spin_lock(&last_ptr->refill_lock);
6043                         used_block_group = last_ptr->block_group;
6044                         if (used_block_group != block_group &&
6045                             (!used_block_group ||
6046                              used_block_group->ro ||
6047                              !block_group_bits(used_block_group, data))) {
6048                                 used_block_group = block_group;
6049                                 goto refill_cluster;
6050                         }
6051
6052                         if (used_block_group != block_group)
6053                                 btrfs_get_block_group(used_block_group);
6054
6055                         offset = btrfs_alloc_from_cluster(used_block_group,
6056                           last_ptr, num_bytes, used_block_group->key.objectid);
6057                         if (offset) {
6058                                 /* we have a block, we're done */
6059                                 spin_unlock(&last_ptr->refill_lock);
6060                                 trace_btrfs_reserve_extent_cluster(root,
6061                                         block_group, search_start, num_bytes);
6062                                 goto checks;
6063                         }
6064
6065                         WARN_ON(last_ptr->block_group != used_block_group);
6066                         if (used_block_group != block_group) {
6067                                 btrfs_put_block_group(used_block_group);
6068                                 used_block_group = block_group;
6069                         }
6070 refill_cluster:
6071                         BUG_ON(used_block_group != block_group);
6072                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6073                          * set up a new clusters, so lets just skip it
6074                          * and let the allocator find whatever block
6075                          * it can find.  If we reach this point, we
6076                          * will have tried the cluster allocator
6077                          * plenty of times and not have found
6078                          * anything, so we are likely way too
6079                          * fragmented for the clustering stuff to find
6080                          * anything.
6081                          *
6082                          * However, if the cluster is taken from the
6083                          * current block group, release the cluster
6084                          * first, so that we stand a better chance of
6085                          * succeeding in the unclustered
6086                          * allocation.  */
6087                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6088                             last_ptr->block_group != block_group) {
6089                                 spin_unlock(&last_ptr->refill_lock);
6090                                 goto unclustered_alloc;
6091                         }
6092
6093                         /*
6094                          * this cluster didn't work out, free it and
6095                          * start over
6096                          */
6097                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6098
6099                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6100                                 spin_unlock(&last_ptr->refill_lock);
6101                                 goto unclustered_alloc;
6102                         }
6103
6104                         aligned_cluster = max_t(unsigned long,
6105                                                 empty_cluster + empty_size,
6106                                               block_group->full_stripe_len);
6107
6108                         /* allocate a cluster in this block group */
6109                         ret = btrfs_find_space_cluster(trans, root,
6110                                                block_group, last_ptr,
6111                                                search_start, num_bytes,
6112                                                aligned_cluster);
6113                         if (ret == 0) {
6114                                 /*
6115                                  * now pull our allocation out of this
6116                                  * cluster
6117                                  */
6118                                 offset = btrfs_alloc_from_cluster(block_group,
6119                                                   last_ptr, num_bytes,
6120                                                   search_start);
6121                                 if (offset) {
6122                                         /* we found one, proceed */
6123                                         spin_unlock(&last_ptr->refill_lock);
6124                                         trace_btrfs_reserve_extent_cluster(root,
6125                                                 block_group, search_start,
6126                                                 num_bytes);
6127                                         goto checks;
6128                                 }
6129                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6130                                    && !failed_cluster_refill) {
6131                                 spin_unlock(&last_ptr->refill_lock);
6132
6133                                 failed_cluster_refill = true;
6134                                 wait_block_group_cache_progress(block_group,
6135                                        num_bytes + empty_cluster + empty_size);
6136                                 goto have_block_group;
6137                         }
6138
6139                         /*
6140                          * at this point we either didn't find a cluster
6141                          * or we weren't able to allocate a block from our
6142                          * cluster.  Free the cluster we've been trying
6143                          * to use, and go to the next block group
6144                          */
6145                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6146                         spin_unlock(&last_ptr->refill_lock);
6147                         goto loop;
6148                 }
6149
6150 unclustered_alloc:
6151                 spin_lock(&block_group->free_space_ctl->tree_lock);
6152                 if (cached &&
6153                     block_group->free_space_ctl->free_space <
6154                     num_bytes + empty_cluster + empty_size) {
6155                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6156                         goto loop;
6157                 }
6158                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6159
6160                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6161                                                     num_bytes, empty_size);
6162                 /*
6163                  * If we didn't find a chunk, and we haven't failed on this
6164                  * block group before, and this block group is in the middle of
6165                  * caching and we are ok with waiting, then go ahead and wait
6166                  * for progress to be made, and set failed_alloc to true.
6167                  *
6168                  * If failed_alloc is true then we've already waited on this
6169                  * block group once and should move on to the next block group.
6170                  */
6171                 if (!offset && !failed_alloc && !cached &&
6172                     loop > LOOP_CACHING_NOWAIT) {
6173                         wait_block_group_cache_progress(block_group,
6174                                                 num_bytes + empty_size);
6175                         failed_alloc = true;
6176                         goto have_block_group;
6177                 } else if (!offset) {
6178                         if (!cached)
6179                                 have_caching_bg = true;
6180                         goto loop;
6181                 }
6182 checks:
6183                 search_start = stripe_align(root, used_block_group,
6184                                             offset, num_bytes);
6185
6186                 /* move on to the next group */
6187                 if (search_start + num_bytes >
6188                     used_block_group->key.objectid + used_block_group->key.offset) {
6189                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6190                         goto loop;
6191                 }
6192
6193                 if (offset < search_start)
6194                         btrfs_add_free_space(used_block_group, offset,
6195                                              search_start - offset);
6196                 BUG_ON(offset > search_start);
6197
6198                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
6199                                                   alloc_type);
6200                 if (ret == -EAGAIN) {
6201                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6202                         goto loop;
6203                 }
6204
6205                 /* we are all good, lets return */
6206                 ins->objectid = search_start;
6207                 ins->offset = num_bytes;
6208
6209                 trace_btrfs_reserve_extent(orig_root, block_group,
6210                                            search_start, num_bytes);
6211                 if (used_block_group != block_group)
6212                         btrfs_put_block_group(used_block_group);
6213                 btrfs_put_block_group(block_group);
6214                 break;
6215 loop:
6216                 failed_cluster_refill = false;
6217                 failed_alloc = false;
6218                 BUG_ON(index != get_block_group_index(block_group));
6219                 if (used_block_group != block_group)
6220                         btrfs_put_block_group(used_block_group);
6221                 btrfs_put_block_group(block_group);
6222         }
6223         up_read(&space_info->groups_sem);
6224
6225         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6226                 goto search;
6227
6228         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6229                 goto search;
6230
6231         /*
6232          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6233          *                      caching kthreads as we move along
6234          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6235          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6236          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6237          *                      again
6238          */
6239         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6240                 index = 0;
6241                 loop++;
6242                 if (loop == LOOP_ALLOC_CHUNK) {
6243                         ret = do_chunk_alloc(trans, root, data,
6244                                              CHUNK_ALLOC_FORCE);
6245                         /*
6246                          * Do not bail out on ENOSPC since we
6247                          * can do more things.
6248                          */
6249                         if (ret < 0 && ret != -ENOSPC) {
6250                                 btrfs_abort_transaction(trans,
6251                                                         root, ret);
6252                                 goto out;
6253                         }
6254                 }
6255
6256                 if (loop == LOOP_NO_EMPTY_SIZE) {
6257                         empty_size = 0;
6258                         empty_cluster = 0;
6259                 }
6260
6261                 goto search;
6262         } else if (!ins->objectid) {
6263                 ret = -ENOSPC;
6264         } else if (ins->objectid) {
6265                 ret = 0;
6266         }
6267 out:
6268
6269         return ret;
6270 }
6271
6272 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6273                             int dump_block_groups)
6274 {
6275         struct btrfs_block_group_cache *cache;
6276         int index = 0;
6277
6278         spin_lock(&info->lock);
6279         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
6280                (unsigned long long)info->flags,
6281                (unsigned long long)(info->total_bytes - info->bytes_used -
6282                                     info->bytes_pinned - info->bytes_reserved -
6283                                     info->bytes_readonly),
6284                (info->full) ? "" : "not ");
6285         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
6286                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6287                (unsigned long long)info->total_bytes,
6288                (unsigned long long)info->bytes_used,
6289                (unsigned long long)info->bytes_pinned,
6290                (unsigned long long)info->bytes_reserved,
6291                (unsigned long long)info->bytes_may_use,
6292                (unsigned long long)info->bytes_readonly);
6293         spin_unlock(&info->lock);
6294
6295         if (!dump_block_groups)
6296                 return;
6297
6298         down_read(&info->groups_sem);
6299 again:
6300         list_for_each_entry(cache, &info->block_groups[index], list) {
6301                 spin_lock(&cache->lock);
6302                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
6303                        (unsigned long long)cache->key.objectid,
6304                        (unsigned long long)cache->key.offset,
6305                        (unsigned long long)btrfs_block_group_used(&cache->item),
6306                        (unsigned long long)cache->pinned,
6307                        (unsigned long long)cache->reserved,
6308                        cache->ro ? "[readonly]" : "");
6309                 btrfs_dump_free_space(cache, bytes);
6310                 spin_unlock(&cache->lock);
6311         }
6312         if (++index < BTRFS_NR_RAID_TYPES)
6313                 goto again;
6314         up_read(&info->groups_sem);
6315 }
6316
6317 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
6318                          struct btrfs_root *root,
6319                          u64 num_bytes, u64 min_alloc_size,
6320                          u64 empty_size, u64 hint_byte,
6321                          struct btrfs_key *ins, u64 data)
6322 {
6323         bool final_tried = false;
6324         int ret;
6325
6326         data = btrfs_get_alloc_profile(root, data);
6327 again:
6328         WARN_ON(num_bytes < root->sectorsize);
6329         ret = find_free_extent(trans, root, num_bytes, empty_size,
6330                                hint_byte, ins, data);
6331
6332         if (ret == -ENOSPC) {
6333                 if (!final_tried) {
6334                         num_bytes = num_bytes >> 1;
6335                         num_bytes = round_down(num_bytes, root->sectorsize);
6336                         num_bytes = max(num_bytes, min_alloc_size);
6337                         if (num_bytes == min_alloc_size)
6338                                 final_tried = true;
6339                         goto again;
6340                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6341                         struct btrfs_space_info *sinfo;
6342
6343                         sinfo = __find_space_info(root->fs_info, data);
6344                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6345                                 (unsigned long long)data,
6346                                 (unsigned long long)num_bytes);
6347                         if (sinfo)
6348                                 dump_space_info(sinfo, num_bytes, 1);
6349                 }
6350         }
6351
6352         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6353
6354         return ret;
6355 }
6356
6357 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6358                                         u64 start, u64 len, int pin)
6359 {
6360         struct btrfs_block_group_cache *cache;
6361         int ret = 0;
6362
6363         cache = btrfs_lookup_block_group(root->fs_info, start);
6364         if (!cache) {
6365                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6366                         (unsigned long long)start);
6367                 return -ENOSPC;
6368         }
6369
6370         if (btrfs_test_opt(root, DISCARD))
6371                 ret = btrfs_discard_extent(root, start, len, NULL);
6372
6373         if (pin)
6374                 pin_down_extent(root, cache, start, len, 1);
6375         else {
6376                 btrfs_add_free_space(cache, start, len);
6377                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6378         }
6379         btrfs_put_block_group(cache);
6380
6381         trace_btrfs_reserved_extent_free(root, start, len);
6382
6383         return ret;
6384 }
6385
6386 int btrfs_free_reserved_extent(struct btrfs_root *root,
6387                                         u64 start, u64 len)
6388 {
6389         return __btrfs_free_reserved_extent(root, start, len, 0);
6390 }
6391
6392 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6393                                        u64 start, u64 len)
6394 {
6395         return __btrfs_free_reserved_extent(root, start, len, 1);
6396 }
6397
6398 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6399                                       struct btrfs_root *root,
6400                                       u64 parent, u64 root_objectid,
6401                                       u64 flags, u64 owner, u64 offset,
6402                                       struct btrfs_key *ins, int ref_mod)
6403 {
6404         int ret;
6405         struct btrfs_fs_info *fs_info = root->fs_info;
6406         struct btrfs_extent_item *extent_item;
6407         struct btrfs_extent_inline_ref *iref;
6408         struct btrfs_path *path;
6409         struct extent_buffer *leaf;
6410         int type;
6411         u32 size;
6412
6413         if (parent > 0)
6414                 type = BTRFS_SHARED_DATA_REF_KEY;
6415         else
6416                 type = BTRFS_EXTENT_DATA_REF_KEY;
6417
6418         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6419
6420         path = btrfs_alloc_path();
6421         if (!path)
6422                 return -ENOMEM;
6423
6424         path->leave_spinning = 1;
6425         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6426                                       ins, size);
6427         if (ret) {
6428                 btrfs_free_path(path);
6429                 return ret;
6430         }
6431
6432         leaf = path->nodes[0];
6433         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6434                                      struct btrfs_extent_item);
6435         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6436         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6437         btrfs_set_extent_flags(leaf, extent_item,
6438                                flags | BTRFS_EXTENT_FLAG_DATA);
6439
6440         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6441         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6442         if (parent > 0) {
6443                 struct btrfs_shared_data_ref *ref;
6444                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6445                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6446                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6447         } else {
6448                 struct btrfs_extent_data_ref *ref;
6449                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6450                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6451                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6452                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6453                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6454         }
6455
6456         btrfs_mark_buffer_dirty(path->nodes[0]);
6457         btrfs_free_path(path);
6458
6459         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6460         if (ret) { /* -ENOENT, logic error */
6461                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6462                         (unsigned long long)ins->objectid,
6463                         (unsigned long long)ins->offset);
6464                 BUG();
6465         }
6466         return ret;
6467 }
6468
6469 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6470                                      struct btrfs_root *root,
6471                                      u64 parent, u64 root_objectid,
6472                                      u64 flags, struct btrfs_disk_key *key,
6473                                      int level, struct btrfs_key *ins)
6474 {
6475         int ret;
6476         struct btrfs_fs_info *fs_info = root->fs_info;
6477         struct btrfs_extent_item *extent_item;
6478         struct btrfs_tree_block_info *block_info;
6479         struct btrfs_extent_inline_ref *iref;
6480         struct btrfs_path *path;
6481         struct extent_buffer *leaf;
6482         u32 size = sizeof(*extent_item) + sizeof(*iref);
6483         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6484                                                  SKINNY_METADATA);
6485
6486         if (!skinny_metadata)
6487                 size += sizeof(*block_info);
6488
6489         path = btrfs_alloc_path();
6490         if (!path)
6491                 return -ENOMEM;
6492
6493         path->leave_spinning = 1;
6494         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6495                                       ins, size);
6496         if (ret) {
6497                 btrfs_free_path(path);
6498                 return ret;
6499         }
6500
6501         leaf = path->nodes[0];
6502         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6503                                      struct btrfs_extent_item);
6504         btrfs_set_extent_refs(leaf, extent_item, 1);
6505         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6506         btrfs_set_extent_flags(leaf, extent_item,
6507                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6508
6509         if (skinny_metadata) {
6510                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6511         } else {
6512                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6513                 btrfs_set_tree_block_key(leaf, block_info, key);
6514                 btrfs_set_tree_block_level(leaf, block_info, level);
6515                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6516         }
6517
6518         if (parent > 0) {
6519                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6520                 btrfs_set_extent_inline_ref_type(leaf, iref,
6521                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6522                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6523         } else {
6524                 btrfs_set_extent_inline_ref_type(leaf, iref,
6525                                                  BTRFS_TREE_BLOCK_REF_KEY);
6526                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6527         }
6528
6529         btrfs_mark_buffer_dirty(leaf);
6530         btrfs_free_path(path);
6531
6532         ret = update_block_group(root, ins->objectid, root->leafsize, 1);
6533         if (ret) { /* -ENOENT, logic error */
6534                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6535                         (unsigned long long)ins->objectid,
6536                         (unsigned long long)ins->offset);
6537                 BUG();
6538         }
6539         return ret;
6540 }
6541
6542 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6543                                      struct btrfs_root *root,
6544                                      u64 root_objectid, u64 owner,
6545                                      u64 offset, struct btrfs_key *ins)
6546 {
6547         int ret;
6548
6549         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6550
6551         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6552                                          ins->offset, 0,
6553                                          root_objectid, owner, offset,
6554                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6555         return ret;
6556 }
6557
6558 /*
6559  * this is used by the tree logging recovery code.  It records that
6560  * an extent has been allocated and makes sure to clear the free
6561  * space cache bits as well
6562  */
6563 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6564                                    struct btrfs_root *root,
6565                                    u64 root_objectid, u64 owner, u64 offset,
6566                                    struct btrfs_key *ins)
6567 {
6568         int ret;
6569         struct btrfs_block_group_cache *block_group;
6570         struct btrfs_caching_control *caching_ctl;
6571         u64 start = ins->objectid;
6572         u64 num_bytes = ins->offset;
6573
6574         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6575         cache_block_group(block_group, 0);
6576         caching_ctl = get_caching_control(block_group);
6577
6578         if (!caching_ctl) {
6579                 BUG_ON(!block_group_cache_done(block_group));
6580                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6581                 BUG_ON(ret); /* -ENOMEM */
6582         } else {
6583                 mutex_lock(&caching_ctl->mutex);
6584
6585                 if (start >= caching_ctl->progress) {
6586                         ret = add_excluded_extent(root, start, num_bytes);
6587                         BUG_ON(ret); /* -ENOMEM */
6588                 } else if (start + num_bytes <= caching_ctl->progress) {
6589                         ret = btrfs_remove_free_space(block_group,
6590                                                       start, num_bytes);
6591                         BUG_ON(ret); /* -ENOMEM */
6592                 } else {
6593                         num_bytes = caching_ctl->progress - start;
6594                         ret = btrfs_remove_free_space(block_group,
6595                                                       start, num_bytes);
6596                         BUG_ON(ret); /* -ENOMEM */
6597
6598                         start = caching_ctl->progress;
6599                         num_bytes = ins->objectid + ins->offset -
6600                                     caching_ctl->progress;
6601                         ret = add_excluded_extent(root, start, num_bytes);
6602                         BUG_ON(ret); /* -ENOMEM */
6603                 }
6604
6605                 mutex_unlock(&caching_ctl->mutex);
6606                 put_caching_control(caching_ctl);
6607         }
6608
6609         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6610                                           RESERVE_ALLOC_NO_ACCOUNT);
6611         BUG_ON(ret); /* logic error */
6612         btrfs_put_block_group(block_group);
6613         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6614                                          0, owner, offset, ins, 1);
6615         return ret;
6616 }
6617
6618 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6619                                             struct btrfs_root *root,
6620                                             u64 bytenr, u32 blocksize,
6621                                             int level)
6622 {
6623         struct extent_buffer *buf;
6624
6625         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6626         if (!buf)
6627                 return ERR_PTR(-ENOMEM);
6628         btrfs_set_header_generation(buf, trans->transid);
6629         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6630         btrfs_tree_lock(buf);
6631         clean_tree_block(trans, root, buf);
6632         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6633
6634         btrfs_set_lock_blocking(buf);
6635         btrfs_set_buffer_uptodate(buf);
6636
6637         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6638                 /*
6639                  * we allow two log transactions at a time, use different
6640                  * EXENT bit to differentiate dirty pages.
6641                  */
6642                 if (root->log_transid % 2 == 0)
6643                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6644                                         buf->start + buf->len - 1, GFP_NOFS);
6645                 else
6646                         set_extent_new(&root->dirty_log_pages, buf->start,
6647                                         buf->start + buf->len - 1, GFP_NOFS);
6648         } else {
6649                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6650                          buf->start + buf->len - 1, GFP_NOFS);
6651         }
6652         trans->blocks_used++;
6653         /* this returns a buffer locked for blocking */
6654         return buf;
6655 }
6656
6657 static struct btrfs_block_rsv *
6658 use_block_rsv(struct btrfs_trans_handle *trans,
6659               struct btrfs_root *root, u32 blocksize)
6660 {
6661         struct btrfs_block_rsv *block_rsv;
6662         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6663         int ret;
6664
6665         block_rsv = get_block_rsv(trans, root);
6666
6667         if (block_rsv->size == 0) {
6668                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6669                                              BTRFS_RESERVE_NO_FLUSH);
6670                 /*
6671                  * If we couldn't reserve metadata bytes try and use some from
6672                  * the global reserve.
6673                  */
6674                 if (ret && block_rsv != global_rsv) {
6675                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6676                         if (!ret)
6677                                 return global_rsv;
6678                         return ERR_PTR(ret);
6679                 } else if (ret) {
6680                         return ERR_PTR(ret);
6681                 }
6682                 return block_rsv;
6683         }
6684
6685         ret = block_rsv_use_bytes(block_rsv, blocksize);
6686         if (!ret)
6687                 return block_rsv;
6688         if (ret && !block_rsv->failfast) {
6689                 if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6690                         static DEFINE_RATELIMIT_STATE(_rs,
6691                                         DEFAULT_RATELIMIT_INTERVAL * 10,
6692                                         /*DEFAULT_RATELIMIT_BURST*/ 1);
6693                         if (__ratelimit(&_rs))
6694                                 WARN(1, KERN_DEBUG
6695                                         "btrfs: block rsv returned %d\n", ret);
6696                 }
6697                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6698                                              BTRFS_RESERVE_NO_FLUSH);
6699                 if (!ret) {
6700                         return block_rsv;
6701                 } else if (ret && block_rsv != global_rsv) {
6702                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6703                         if (!ret)
6704                                 return global_rsv;
6705                 }
6706         }
6707
6708         return ERR_PTR(-ENOSPC);
6709 }
6710
6711 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6712                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6713 {
6714         block_rsv_add_bytes(block_rsv, blocksize, 0);
6715         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6716 }
6717
6718 /*
6719  * finds a free extent and does all the dirty work required for allocation
6720  * returns the key for the extent through ins, and a tree buffer for
6721  * the first block of the extent through buf.
6722  *
6723  * returns the tree buffer or NULL.
6724  */
6725 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6726                                         struct btrfs_root *root, u32 blocksize,
6727                                         u64 parent, u64 root_objectid,
6728                                         struct btrfs_disk_key *key, int level,
6729                                         u64 hint, u64 empty_size)
6730 {
6731         struct btrfs_key ins;
6732         struct btrfs_block_rsv *block_rsv;
6733         struct extent_buffer *buf;
6734         u64 flags = 0;
6735         int ret;
6736         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6737                                                  SKINNY_METADATA);
6738
6739         block_rsv = use_block_rsv(trans, root, blocksize);
6740         if (IS_ERR(block_rsv))
6741                 return ERR_CAST(block_rsv);
6742
6743         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6744                                    empty_size, hint, &ins, 0);
6745         if (ret) {
6746                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6747                 return ERR_PTR(ret);
6748         }
6749
6750         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6751                                     blocksize, level);
6752         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6753
6754         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6755                 if (parent == 0)
6756                         parent = ins.objectid;
6757                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6758         } else
6759                 BUG_ON(parent > 0);
6760
6761         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6762                 struct btrfs_delayed_extent_op *extent_op;
6763                 extent_op = btrfs_alloc_delayed_extent_op();
6764                 BUG_ON(!extent_op); /* -ENOMEM */
6765                 if (key)
6766                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6767                 else
6768                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6769                 extent_op->flags_to_set = flags;
6770                 if (skinny_metadata)
6771                         extent_op->update_key = 0;
6772                 else
6773                         extent_op->update_key = 1;
6774                 extent_op->update_flags = 1;
6775                 extent_op->is_data = 0;
6776
6777                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6778                                         ins.objectid,
6779                                         ins.offset, parent, root_objectid,
6780                                         level, BTRFS_ADD_DELAYED_EXTENT,
6781                                         extent_op, 0);
6782                 BUG_ON(ret); /* -ENOMEM */
6783         }
6784         return buf;
6785 }
6786
6787 struct walk_control {
6788         u64 refs[BTRFS_MAX_LEVEL];
6789         u64 flags[BTRFS_MAX_LEVEL];
6790         struct btrfs_key update_progress;
6791         int stage;
6792         int level;
6793         int shared_level;
6794         int update_ref;
6795         int keep_locks;
6796         int reada_slot;
6797         int reada_count;
6798         int for_reloc;
6799 };
6800
6801 #define DROP_REFERENCE  1
6802 #define UPDATE_BACKREF  2
6803
6804 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6805                                      struct btrfs_root *root,
6806                                      struct walk_control *wc,
6807                                      struct btrfs_path *path)
6808 {
6809         u64 bytenr;
6810         u64 generation;
6811         u64 refs;
6812         u64 flags;
6813         u32 nritems;
6814         u32 blocksize;
6815         struct btrfs_key key;
6816         struct extent_buffer *eb;
6817         int ret;
6818         int slot;
6819         int nread = 0;
6820
6821         if (path->slots[wc->level] < wc->reada_slot) {
6822                 wc->reada_count = wc->reada_count * 2 / 3;
6823                 wc->reada_count = max(wc->reada_count, 2);
6824         } else {
6825                 wc->reada_count = wc->reada_count * 3 / 2;
6826                 wc->reada_count = min_t(int, wc->reada_count,
6827                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6828         }
6829
6830         eb = path->nodes[wc->level];
6831         nritems = btrfs_header_nritems(eb);
6832         blocksize = btrfs_level_size(root, wc->level - 1);
6833
6834         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6835                 if (nread >= wc->reada_count)
6836                         break;
6837
6838                 cond_resched();
6839                 bytenr = btrfs_node_blockptr(eb, slot);
6840                 generation = btrfs_node_ptr_generation(eb, slot);
6841
6842                 if (slot == path->slots[wc->level])
6843                         goto reada;
6844
6845                 if (wc->stage == UPDATE_BACKREF &&
6846                     generation <= root->root_key.offset)
6847                         continue;
6848
6849                 /* We don't lock the tree block, it's OK to be racy here */
6850                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
6851                                                wc->level - 1, 1, &refs,
6852                                                &flags);
6853                 /* We don't care about errors in readahead. */
6854                 if (ret < 0)
6855                         continue;
6856                 BUG_ON(refs == 0);
6857
6858                 if (wc->stage == DROP_REFERENCE) {
6859                         if (refs == 1)
6860                                 goto reada;
6861
6862                         if (wc->level == 1 &&
6863                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6864                                 continue;
6865                         if (!wc->update_ref ||
6866                             generation <= root->root_key.offset)
6867                                 continue;
6868                         btrfs_node_key_to_cpu(eb, &key, slot);
6869                         ret = btrfs_comp_cpu_keys(&key,
6870                                                   &wc->update_progress);
6871                         if (ret < 0)
6872                                 continue;
6873                 } else {
6874                         if (wc->level == 1 &&
6875                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6876                                 continue;
6877                 }
6878 reada:
6879                 ret = readahead_tree_block(root, bytenr, blocksize,
6880                                            generation);
6881                 if (ret)
6882                         break;
6883                 nread++;
6884         }
6885         wc->reada_slot = slot;
6886 }
6887
6888 /*
6889  * helper to process tree block while walking down the tree.
6890  *
6891  * when wc->stage == UPDATE_BACKREF, this function updates
6892  * back refs for pointers in the block.
6893  *
6894  * NOTE: return value 1 means we should stop walking down.
6895  */
6896 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6897                                    struct btrfs_root *root,
6898                                    struct btrfs_path *path,
6899                                    struct walk_control *wc, int lookup_info)
6900 {
6901         int level = wc->level;
6902         struct extent_buffer *eb = path->nodes[level];
6903         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6904         int ret;
6905
6906         if (wc->stage == UPDATE_BACKREF &&
6907             btrfs_header_owner(eb) != root->root_key.objectid)
6908                 return 1;
6909
6910         /*
6911          * when reference count of tree block is 1, it won't increase
6912          * again. once full backref flag is set, we never clear it.
6913          */
6914         if (lookup_info &&
6915             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6916              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6917                 BUG_ON(!path->locks[level]);
6918                 ret = btrfs_lookup_extent_info(trans, root,
6919                                                eb->start, level, 1,
6920                                                &wc->refs[level],
6921                                                &wc->flags[level]);
6922                 BUG_ON(ret == -ENOMEM);
6923                 if (ret)
6924                         return ret;
6925                 BUG_ON(wc->refs[level] == 0);
6926         }
6927
6928         if (wc->stage == DROP_REFERENCE) {
6929                 if (wc->refs[level] > 1)
6930                         return 1;
6931
6932                 if (path->locks[level] && !wc->keep_locks) {
6933                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6934                         path->locks[level] = 0;
6935                 }
6936                 return 0;
6937         }
6938
6939         /* wc->stage == UPDATE_BACKREF */
6940         if (!(wc->flags[level] & flag)) {
6941                 BUG_ON(!path->locks[level]);
6942                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6943                 BUG_ON(ret); /* -ENOMEM */
6944                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6945                 BUG_ON(ret); /* -ENOMEM */
6946                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6947                                                   eb->len, flag, 0);
6948                 BUG_ON(ret); /* -ENOMEM */
6949                 wc->flags[level] |= flag;
6950         }
6951
6952         /*
6953          * the block is shared by multiple trees, so it's not good to
6954          * keep the tree lock
6955          */
6956         if (path->locks[level] && level > 0) {
6957                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6958                 path->locks[level] = 0;
6959         }
6960         return 0;
6961 }
6962
6963 /*
6964  * helper to process tree block pointer.
6965  *
6966  * when wc->stage == DROP_REFERENCE, this function checks
6967  * reference count of the block pointed to. if the block
6968  * is shared and we need update back refs for the subtree
6969  * rooted at the block, this function changes wc->stage to
6970  * UPDATE_BACKREF. if the block is shared and there is no
6971  * need to update back, this function drops the reference
6972  * to the block.
6973  *
6974  * NOTE: return value 1 means we should stop walking down.
6975  */
6976 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6977                                  struct btrfs_root *root,
6978                                  struct btrfs_path *path,
6979                                  struct walk_control *wc, int *lookup_info)
6980 {
6981         u64 bytenr;
6982         u64 generation;
6983         u64 parent;
6984         u32 blocksize;
6985         struct btrfs_key key;
6986         struct extent_buffer *next;
6987         int level = wc->level;
6988         int reada = 0;
6989         int ret = 0;
6990
6991         generation = btrfs_node_ptr_generation(path->nodes[level],
6992                                                path->slots[level]);
6993         /*
6994          * if the lower level block was created before the snapshot
6995          * was created, we know there is no need to update back refs
6996          * for the subtree
6997          */
6998         if (wc->stage == UPDATE_BACKREF &&
6999             generation <= root->root_key.offset) {
7000                 *lookup_info = 1;
7001                 return 1;
7002         }
7003
7004         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7005         blocksize = btrfs_level_size(root, level - 1);
7006
7007         next = btrfs_find_tree_block(root, bytenr, blocksize);
7008         if (!next) {
7009                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7010                 if (!next)
7011                         return -ENOMEM;
7012                 reada = 1;
7013         }
7014         btrfs_tree_lock(next);
7015         btrfs_set_lock_blocking(next);
7016
7017         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7018                                        &wc->refs[level - 1],
7019                                        &wc->flags[level - 1]);
7020         if (ret < 0) {
7021                 btrfs_tree_unlock(next);
7022                 return ret;
7023         }
7024
7025         if (unlikely(wc->refs[level - 1] == 0)) {
7026                 btrfs_err(root->fs_info, "Missing references.");
7027                 BUG();
7028         }
7029         *lookup_info = 0;
7030
7031         if (wc->stage == DROP_REFERENCE) {
7032                 if (wc->refs[level - 1] > 1) {
7033                         if (level == 1 &&
7034                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7035                                 goto skip;
7036
7037                         if (!wc->update_ref ||
7038                             generation <= root->root_key.offset)
7039                                 goto skip;
7040
7041                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7042                                               path->slots[level]);
7043                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7044                         if (ret < 0)
7045                                 goto skip;
7046
7047                         wc->stage = UPDATE_BACKREF;
7048                         wc->shared_level = level - 1;
7049                 }
7050         } else {
7051                 if (level == 1 &&
7052                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7053                         goto skip;
7054         }
7055
7056         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7057                 btrfs_tree_unlock(next);
7058                 free_extent_buffer(next);
7059                 next = NULL;
7060                 *lookup_info = 1;
7061         }
7062
7063         if (!next) {
7064                 if (reada && level == 1)
7065                         reada_walk_down(trans, root, wc, path);
7066                 next = read_tree_block(root, bytenr, blocksize, generation);
7067                 if (!next)
7068                         return -EIO;
7069                 btrfs_tree_lock(next);
7070                 btrfs_set_lock_blocking(next);
7071         }
7072
7073         level--;
7074         BUG_ON(level != btrfs_header_level(next));
7075         path->nodes[level] = next;
7076         path->slots[level] = 0;
7077         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7078         wc->level = level;
7079         if (wc->level == 1)
7080                 wc->reada_slot = 0;
7081         return 0;
7082 skip:
7083         wc->refs[level - 1] = 0;
7084         wc->flags[level - 1] = 0;
7085         if (wc->stage == DROP_REFERENCE) {
7086                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7087                         parent = path->nodes[level]->start;
7088                 } else {
7089                         BUG_ON(root->root_key.objectid !=
7090                                btrfs_header_owner(path->nodes[level]));
7091                         parent = 0;
7092                 }
7093
7094                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7095                                 root->root_key.objectid, level - 1, 0, 0);
7096                 BUG_ON(ret); /* -ENOMEM */
7097         }
7098         btrfs_tree_unlock(next);
7099         free_extent_buffer(next);
7100         *lookup_info = 1;
7101         return 1;
7102 }
7103
7104 /*
7105  * helper to process tree block while walking up the tree.
7106  *
7107  * when wc->stage == DROP_REFERENCE, this function drops
7108  * reference count on the block.
7109  *
7110  * when wc->stage == UPDATE_BACKREF, this function changes
7111  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7112  * to UPDATE_BACKREF previously while processing the block.
7113  *
7114  * NOTE: return value 1 means we should stop walking up.
7115  */
7116 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7117                                  struct btrfs_root *root,
7118                                  struct btrfs_path *path,
7119                                  struct walk_control *wc)
7120 {
7121         int ret;
7122         int level = wc->level;
7123         struct extent_buffer *eb = path->nodes[level];
7124         u64 parent = 0;
7125
7126         if (wc->stage == UPDATE_BACKREF) {
7127                 BUG_ON(wc->shared_level < level);
7128                 if (level < wc->shared_level)
7129                         goto out;
7130
7131                 ret = find_next_key(path, level + 1, &wc->update_progress);
7132                 if (ret > 0)
7133                         wc->update_ref = 0;
7134
7135                 wc->stage = DROP_REFERENCE;
7136                 wc->shared_level = -1;
7137                 path->slots[level] = 0;
7138
7139                 /*
7140                  * check reference count again if the block isn't locked.
7141                  * we should start walking down the tree again if reference
7142                  * count is one.
7143                  */
7144                 if (!path->locks[level]) {
7145                         BUG_ON(level == 0);
7146                         btrfs_tree_lock(eb);
7147                         btrfs_set_lock_blocking(eb);
7148                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7149
7150                         ret = btrfs_lookup_extent_info(trans, root,
7151                                                        eb->start, level, 1,
7152                                                        &wc->refs[level],
7153                                                        &wc->flags[level]);
7154                         if (ret < 0) {
7155                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7156                                 path->locks[level] = 0;
7157                                 return ret;
7158                         }
7159                         BUG_ON(wc->refs[level] == 0);
7160                         if (wc->refs[level] == 1) {
7161                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7162                                 path->locks[level] = 0;
7163                                 return 1;
7164                         }
7165                 }
7166         }
7167
7168         /* wc->stage == DROP_REFERENCE */
7169         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7170
7171         if (wc->refs[level] == 1) {
7172                 if (level == 0) {
7173                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7174                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7175                                                     wc->for_reloc);
7176                         else
7177                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7178                                                     wc->for_reloc);
7179                         BUG_ON(ret); /* -ENOMEM */
7180                 }
7181                 /* make block locked assertion in clean_tree_block happy */
7182                 if (!path->locks[level] &&
7183                     btrfs_header_generation(eb) == trans->transid) {
7184                         btrfs_tree_lock(eb);
7185                         btrfs_set_lock_blocking(eb);
7186                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7187                 }
7188                 clean_tree_block(trans, root, eb);
7189         }
7190
7191         if (eb == root->node) {
7192                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7193                         parent = eb->start;
7194                 else
7195                         BUG_ON(root->root_key.objectid !=
7196                                btrfs_header_owner(eb));
7197         } else {
7198                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7199                         parent = path->nodes[level + 1]->start;
7200                 else
7201                         BUG_ON(root->root_key.objectid !=
7202                                btrfs_header_owner(path->nodes[level + 1]));
7203         }
7204
7205         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7206 out:
7207         wc->refs[level] = 0;
7208         wc->flags[level] = 0;
7209         return 0;
7210 }
7211
7212 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7213                                    struct btrfs_root *root,
7214                                    struct btrfs_path *path,
7215                                    struct walk_control *wc)
7216 {
7217         int level = wc->level;
7218         int lookup_info = 1;
7219         int ret;
7220
7221         while (level >= 0) {
7222                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7223                 if (ret > 0)
7224                         break;
7225
7226                 if (level == 0)
7227                         break;
7228
7229                 if (path->slots[level] >=
7230                     btrfs_header_nritems(path->nodes[level]))
7231                         break;
7232
7233                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7234                 if (ret > 0) {
7235                         path->slots[level]++;
7236                         continue;
7237                 } else if (ret < 0)
7238                         return ret;
7239                 level = wc->level;
7240         }
7241         return 0;
7242 }
7243
7244 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7245                                  struct btrfs_root *root,
7246                                  struct btrfs_path *path,
7247                                  struct walk_control *wc, int max_level)
7248 {
7249         int level = wc->level;
7250         int ret;
7251
7252         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7253         while (level < max_level && path->nodes[level]) {
7254                 wc->level = level;
7255                 if (path->slots[level] + 1 <
7256                     btrfs_header_nritems(path->nodes[level])) {
7257                         path->slots[level]++;
7258                         return 0;
7259                 } else {
7260                         ret = walk_up_proc(trans, root, path, wc);
7261                         if (ret > 0)
7262                                 return 0;
7263
7264                         if (path->locks[level]) {
7265                                 btrfs_tree_unlock_rw(path->nodes[level],
7266                                                      path->locks[level]);
7267                                 path->locks[level] = 0;
7268                         }
7269                         free_extent_buffer(path->nodes[level]);
7270                         path->nodes[level] = NULL;
7271                         level++;
7272                 }
7273         }
7274         return 1;
7275 }
7276
7277 /*
7278  * drop a subvolume tree.
7279  *
7280  * this function traverses the tree freeing any blocks that only
7281  * referenced by the tree.
7282  *
7283  * when a shared tree block is found. this function decreases its
7284  * reference count by one. if update_ref is true, this function
7285  * also make sure backrefs for the shared block and all lower level
7286  * blocks are properly updated.
7287  *
7288  * If called with for_reloc == 0, may exit early with -EAGAIN
7289  */
7290 int btrfs_drop_snapshot(struct btrfs_root *root,
7291                          struct btrfs_block_rsv *block_rsv, int update_ref,
7292                          int for_reloc)
7293 {
7294         struct btrfs_path *path;
7295         struct btrfs_trans_handle *trans;
7296         struct btrfs_root *tree_root = root->fs_info->tree_root;
7297         struct btrfs_root_item *root_item = &root->root_item;
7298         struct walk_control *wc;
7299         struct btrfs_key key;
7300         int err = 0;
7301         int ret;
7302         int level;
7303
7304         path = btrfs_alloc_path();
7305         if (!path) {
7306                 err = -ENOMEM;
7307                 goto out;
7308         }
7309
7310         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7311         if (!wc) {
7312                 btrfs_free_path(path);
7313                 err = -ENOMEM;
7314                 goto out;
7315         }
7316
7317         trans = btrfs_start_transaction(tree_root, 0);
7318         if (IS_ERR(trans)) {
7319                 err = PTR_ERR(trans);
7320                 goto out_free;
7321         }
7322
7323         if (block_rsv)
7324                 trans->block_rsv = block_rsv;
7325
7326         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7327                 level = btrfs_header_level(root->node);
7328                 path->nodes[level] = btrfs_lock_root_node(root);
7329                 btrfs_set_lock_blocking(path->nodes[level]);
7330                 path->slots[level] = 0;
7331                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7332                 memset(&wc->update_progress, 0,
7333                        sizeof(wc->update_progress));
7334         } else {
7335                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7336                 memcpy(&wc->update_progress, &key,
7337                        sizeof(wc->update_progress));
7338
7339                 level = root_item->drop_level;
7340                 BUG_ON(level == 0);
7341                 path->lowest_level = level;
7342                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7343                 path->lowest_level = 0;
7344                 if (ret < 0) {
7345                         err = ret;
7346                         goto out_end_trans;
7347                 }
7348                 WARN_ON(ret > 0);
7349
7350                 /*
7351                  * unlock our path, this is safe because only this
7352                  * function is allowed to delete this snapshot
7353                  */
7354                 btrfs_unlock_up_safe(path, 0);
7355
7356                 level = btrfs_header_level(root->node);
7357                 while (1) {
7358                         btrfs_tree_lock(path->nodes[level]);
7359                         btrfs_set_lock_blocking(path->nodes[level]);
7360
7361                         ret = btrfs_lookup_extent_info(trans, root,
7362                                                 path->nodes[level]->start,
7363                                                 level, 1, &wc->refs[level],
7364                                                 &wc->flags[level]);
7365                         if (ret < 0) {
7366                                 err = ret;
7367                                 goto out_end_trans;
7368                         }
7369                         BUG_ON(wc->refs[level] == 0);
7370
7371                         if (level == root_item->drop_level)
7372                                 break;
7373
7374                         btrfs_tree_unlock(path->nodes[level]);
7375                         WARN_ON(wc->refs[level] != 1);
7376                         level--;
7377                 }
7378         }
7379
7380         wc->level = level;
7381         wc->shared_level = -1;
7382         wc->stage = DROP_REFERENCE;
7383         wc->update_ref = update_ref;
7384         wc->keep_locks = 0;
7385         wc->for_reloc = for_reloc;
7386         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7387
7388         while (1) {
7389                 if (!for_reloc && btrfs_fs_closing(root->fs_info)) {
7390                         pr_debug("btrfs: drop snapshot early exit\n");
7391                         err = -EAGAIN;
7392                         goto out_end_trans;
7393                 }
7394
7395                 ret = walk_down_tree(trans, root, path, wc);
7396                 if (ret < 0) {
7397                         err = ret;
7398                         break;
7399                 }
7400
7401                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7402                 if (ret < 0) {
7403                         err = ret;
7404                         break;
7405                 }
7406
7407                 if (ret > 0) {
7408                         BUG_ON(wc->stage != DROP_REFERENCE);
7409                         break;
7410                 }
7411
7412                 if (wc->stage == DROP_REFERENCE) {
7413                         level = wc->level;
7414                         btrfs_node_key(path->nodes[level],
7415                                        &root_item->drop_progress,
7416                                        path->slots[level]);
7417                         root_item->drop_level = level;
7418                 }
7419
7420                 BUG_ON(wc->level == 0);
7421                 if (btrfs_should_end_transaction(trans, tree_root)) {
7422                         ret = btrfs_update_root(trans, tree_root,
7423                                                 &root->root_key,
7424                                                 root_item);
7425                         if (ret) {
7426                                 btrfs_abort_transaction(trans, tree_root, ret);
7427                                 err = ret;
7428                                 goto out_end_trans;
7429                         }
7430
7431                         btrfs_end_transaction_throttle(trans, tree_root);
7432                         trans = btrfs_start_transaction(tree_root, 0);
7433                         if (IS_ERR(trans)) {
7434                                 err = PTR_ERR(trans);
7435                                 goto out_free;
7436                         }
7437                         if (block_rsv)
7438                                 trans->block_rsv = block_rsv;
7439                 }
7440         }
7441         btrfs_release_path(path);
7442         if (err)
7443                 goto out_end_trans;
7444
7445         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7446         if (ret) {
7447                 btrfs_abort_transaction(trans, tree_root, ret);
7448                 goto out_end_trans;
7449         }
7450
7451         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7452                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
7453                                            NULL, NULL);
7454                 if (ret < 0) {
7455                         btrfs_abort_transaction(trans, tree_root, ret);
7456                         err = ret;
7457                         goto out_end_trans;
7458                 } else if (ret > 0) {
7459                         /* if we fail to delete the orphan item this time
7460                          * around, it'll get picked up the next time.
7461                          *
7462                          * The most common failure here is just -ENOENT.
7463                          */
7464                         btrfs_del_orphan_item(trans, tree_root,
7465                                               root->root_key.objectid);
7466                 }
7467         }
7468
7469         if (root->in_radix) {
7470                 btrfs_free_fs_root(tree_root->fs_info, root);
7471         } else {
7472                 free_extent_buffer(root->node);
7473                 free_extent_buffer(root->commit_root);
7474                 kfree(root);
7475         }
7476 out_end_trans:
7477         btrfs_end_transaction_throttle(trans, tree_root);
7478 out_free:
7479         kfree(wc);
7480         btrfs_free_path(path);
7481 out:
7482         if (err)
7483                 btrfs_std_error(root->fs_info, err);
7484         return err;
7485 }
7486
7487 /*
7488  * drop subtree rooted at tree block 'node'.
7489  *
7490  * NOTE: this function will unlock and release tree block 'node'
7491  * only used by relocation code
7492  */
7493 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7494                         struct btrfs_root *root,
7495                         struct extent_buffer *node,
7496                         struct extent_buffer *parent)
7497 {
7498         struct btrfs_path *path;
7499         struct walk_control *wc;
7500         int level;
7501         int parent_level;
7502         int ret = 0;
7503         int wret;
7504
7505         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7506
7507         path = btrfs_alloc_path();
7508         if (!path)
7509                 return -ENOMEM;
7510
7511         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7512         if (!wc) {
7513                 btrfs_free_path(path);
7514                 return -ENOMEM;
7515         }
7516
7517         btrfs_assert_tree_locked(parent);
7518         parent_level = btrfs_header_level(parent);
7519         extent_buffer_get(parent);
7520         path->nodes[parent_level] = parent;
7521         path->slots[parent_level] = btrfs_header_nritems(parent);
7522
7523         btrfs_assert_tree_locked(node);
7524         level = btrfs_header_level(node);
7525         path->nodes[level] = node;
7526         path->slots[level] = 0;
7527         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7528
7529         wc->refs[parent_level] = 1;
7530         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7531         wc->level = level;
7532         wc->shared_level = -1;
7533         wc->stage = DROP_REFERENCE;
7534         wc->update_ref = 0;
7535         wc->keep_locks = 1;
7536         wc->for_reloc = 1;
7537         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7538
7539         while (1) {
7540                 wret = walk_down_tree(trans, root, path, wc);
7541                 if (wret < 0) {
7542                         ret = wret;
7543                         break;
7544                 }
7545
7546                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7547                 if (wret < 0)
7548                         ret = wret;
7549                 if (wret != 0)
7550                         break;
7551         }
7552
7553         kfree(wc);
7554         btrfs_free_path(path);
7555         return ret;
7556 }
7557
7558 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7559 {
7560         u64 num_devices;
7561         u64 stripped;
7562
7563         /*
7564          * if restripe for this chunk_type is on pick target profile and
7565          * return, otherwise do the usual balance
7566          */
7567         stripped = get_restripe_target(root->fs_info, flags);
7568         if (stripped)
7569                 return extended_to_chunk(stripped);
7570
7571         /*
7572          * we add in the count of missing devices because we want
7573          * to make sure that any RAID levels on a degraded FS
7574          * continue to be honored.
7575          */
7576         num_devices = root->fs_info->fs_devices->rw_devices +
7577                 root->fs_info->fs_devices->missing_devices;
7578
7579         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7580                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7581                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7582
7583         if (num_devices == 1) {
7584                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7585                 stripped = flags & ~stripped;
7586
7587                 /* turn raid0 into single device chunks */
7588                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7589                         return stripped;
7590
7591                 /* turn mirroring into duplication */
7592                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7593                              BTRFS_BLOCK_GROUP_RAID10))
7594                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7595         } else {
7596                 /* they already had raid on here, just return */
7597                 if (flags & stripped)
7598                         return flags;
7599
7600                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7601                 stripped = flags & ~stripped;
7602
7603                 /* switch duplicated blocks with raid1 */
7604                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7605                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7606
7607                 /* this is drive concat, leave it alone */
7608         }
7609
7610         return flags;
7611 }
7612
7613 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7614 {
7615         struct btrfs_space_info *sinfo = cache->space_info;
7616         u64 num_bytes;
7617         u64 min_allocable_bytes;
7618         int ret = -ENOSPC;
7619
7620
7621         /*
7622          * We need some metadata space and system metadata space for
7623          * allocating chunks in some corner cases until we force to set
7624          * it to be readonly.
7625          */
7626         if ((sinfo->flags &
7627              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7628             !force)
7629                 min_allocable_bytes = 1 * 1024 * 1024;
7630         else
7631                 min_allocable_bytes = 0;
7632
7633         spin_lock(&sinfo->lock);
7634         spin_lock(&cache->lock);
7635
7636         if (cache->ro) {
7637                 ret = 0;
7638                 goto out;
7639         }
7640
7641         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7642                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7643
7644         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7645             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7646             min_allocable_bytes <= sinfo->total_bytes) {
7647                 sinfo->bytes_readonly += num_bytes;
7648                 cache->ro = 1;
7649                 ret = 0;
7650         }
7651 out:
7652         spin_unlock(&cache->lock);
7653         spin_unlock(&sinfo->lock);
7654         return ret;
7655 }
7656
7657 int btrfs_set_block_group_ro(struct btrfs_root *root,
7658                              struct btrfs_block_group_cache *cache)
7659
7660 {
7661         struct btrfs_trans_handle *trans;
7662         u64 alloc_flags;
7663         int ret;
7664
7665         BUG_ON(cache->ro);
7666
7667         trans = btrfs_join_transaction(root);
7668         if (IS_ERR(trans))
7669                 return PTR_ERR(trans);
7670
7671         alloc_flags = update_block_group_flags(root, cache->flags);
7672         if (alloc_flags != cache->flags) {
7673                 ret = do_chunk_alloc(trans, root, alloc_flags,
7674                                      CHUNK_ALLOC_FORCE);
7675                 if (ret < 0)
7676                         goto out;
7677         }
7678
7679         ret = set_block_group_ro(cache, 0);
7680         if (!ret)
7681                 goto out;
7682         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7683         ret = do_chunk_alloc(trans, root, alloc_flags,
7684                              CHUNK_ALLOC_FORCE);
7685         if (ret < 0)
7686                 goto out;
7687         ret = set_block_group_ro(cache, 0);
7688 out:
7689         btrfs_end_transaction(trans, root);
7690         return ret;
7691 }
7692
7693 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7694                             struct btrfs_root *root, u64 type)
7695 {
7696         u64 alloc_flags = get_alloc_profile(root, type);
7697         return do_chunk_alloc(trans, root, alloc_flags,
7698                               CHUNK_ALLOC_FORCE);
7699 }
7700
7701 /*
7702  * helper to account the unused space of all the readonly block group in the
7703  * list. takes mirrors into account.
7704  */
7705 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7706 {
7707         struct btrfs_block_group_cache *block_group;
7708         u64 free_bytes = 0;
7709         int factor;
7710
7711         list_for_each_entry(block_group, groups_list, list) {
7712                 spin_lock(&block_group->lock);
7713
7714                 if (!block_group->ro) {
7715                         spin_unlock(&block_group->lock);
7716                         continue;
7717                 }
7718
7719                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7720                                           BTRFS_BLOCK_GROUP_RAID10 |
7721                                           BTRFS_BLOCK_GROUP_DUP))
7722                         factor = 2;
7723                 else
7724                         factor = 1;
7725
7726                 free_bytes += (block_group->key.offset -
7727                                btrfs_block_group_used(&block_group->item)) *
7728                                factor;
7729
7730                 spin_unlock(&block_group->lock);
7731         }
7732
7733         return free_bytes;
7734 }
7735
7736 /*
7737  * helper to account the unused space of all the readonly block group in the
7738  * space_info. takes mirrors into account.
7739  */
7740 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7741 {
7742         int i;
7743         u64 free_bytes = 0;
7744
7745         spin_lock(&sinfo->lock);
7746
7747         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7748                 if (!list_empty(&sinfo->block_groups[i]))
7749                         free_bytes += __btrfs_get_ro_block_group_free_space(
7750                                                 &sinfo->block_groups[i]);
7751
7752         spin_unlock(&sinfo->lock);
7753
7754         return free_bytes;
7755 }
7756
7757 void btrfs_set_block_group_rw(struct btrfs_root *root,
7758                               struct btrfs_block_group_cache *cache)
7759 {
7760         struct btrfs_space_info *sinfo = cache->space_info;
7761         u64 num_bytes;
7762
7763         BUG_ON(!cache->ro);
7764
7765         spin_lock(&sinfo->lock);
7766         spin_lock(&cache->lock);
7767         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7768                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7769         sinfo->bytes_readonly -= num_bytes;
7770         cache->ro = 0;
7771         spin_unlock(&cache->lock);
7772         spin_unlock(&sinfo->lock);
7773 }
7774
7775 /*
7776  * checks to see if its even possible to relocate this block group.
7777  *
7778  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7779  * ok to go ahead and try.
7780  */
7781 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7782 {
7783         struct btrfs_block_group_cache *block_group;
7784         struct btrfs_space_info *space_info;
7785         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7786         struct btrfs_device *device;
7787         u64 min_free;
7788         u64 dev_min = 1;
7789         u64 dev_nr = 0;
7790         u64 target;
7791         int index;
7792         int full = 0;
7793         int ret = 0;
7794
7795         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7796
7797         /* odd, couldn't find the block group, leave it alone */
7798         if (!block_group)
7799                 return -1;
7800
7801         min_free = btrfs_block_group_used(&block_group->item);
7802
7803         /* no bytes used, we're good */
7804         if (!min_free)
7805                 goto out;
7806
7807         space_info = block_group->space_info;
7808         spin_lock(&space_info->lock);
7809
7810         full = space_info->full;
7811
7812         /*
7813          * if this is the last block group we have in this space, we can't
7814          * relocate it unless we're able to allocate a new chunk below.
7815          *
7816          * Otherwise, we need to make sure we have room in the space to handle
7817          * all of the extents from this block group.  If we can, we're good
7818          */
7819         if ((space_info->total_bytes != block_group->key.offset) &&
7820             (space_info->bytes_used + space_info->bytes_reserved +
7821              space_info->bytes_pinned + space_info->bytes_readonly +
7822              min_free < space_info->total_bytes)) {
7823                 spin_unlock(&space_info->lock);
7824                 goto out;
7825         }
7826         spin_unlock(&space_info->lock);
7827
7828         /*
7829          * ok we don't have enough space, but maybe we have free space on our
7830          * devices to allocate new chunks for relocation, so loop through our
7831          * alloc devices and guess if we have enough space.  if this block
7832          * group is going to be restriped, run checks against the target
7833          * profile instead of the current one.
7834          */
7835         ret = -1;
7836
7837         /*
7838          * index:
7839          *      0: raid10
7840          *      1: raid1
7841          *      2: dup
7842          *      3: raid0
7843          *      4: single
7844          */
7845         target = get_restripe_target(root->fs_info, block_group->flags);
7846         if (target) {
7847                 index = __get_raid_index(extended_to_chunk(target));
7848         } else {
7849                 /*
7850                  * this is just a balance, so if we were marked as full
7851                  * we know there is no space for a new chunk
7852                  */
7853                 if (full)
7854                         goto out;
7855
7856                 index = get_block_group_index(block_group);
7857         }
7858
7859         if (index == BTRFS_RAID_RAID10) {
7860                 dev_min = 4;
7861                 /* Divide by 2 */
7862                 min_free >>= 1;
7863         } else if (index == BTRFS_RAID_RAID1) {
7864                 dev_min = 2;
7865         } else if (index == BTRFS_RAID_DUP) {
7866                 /* Multiply by 2 */
7867                 min_free <<= 1;
7868         } else if (index == BTRFS_RAID_RAID0) {
7869                 dev_min = fs_devices->rw_devices;
7870                 do_div(min_free, dev_min);
7871         }
7872
7873         mutex_lock(&root->fs_info->chunk_mutex);
7874         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7875                 u64 dev_offset;
7876
7877                 /*
7878                  * check to make sure we can actually find a chunk with enough
7879                  * space to fit our block group in.
7880                  */
7881                 if (device->total_bytes > device->bytes_used + min_free &&
7882                     !device->is_tgtdev_for_dev_replace) {
7883                         ret = find_free_dev_extent(device, min_free,
7884                                                    &dev_offset, NULL);
7885                         if (!ret)
7886                                 dev_nr++;
7887
7888                         if (dev_nr >= dev_min)
7889                                 break;
7890
7891                         ret = -1;
7892                 }
7893         }
7894         mutex_unlock(&root->fs_info->chunk_mutex);
7895 out:
7896         btrfs_put_block_group(block_group);
7897         return ret;
7898 }
7899
7900 static int find_first_block_group(struct btrfs_root *root,
7901                 struct btrfs_path *path, struct btrfs_key *key)
7902 {
7903         int ret = 0;
7904         struct btrfs_key found_key;
7905         struct extent_buffer *leaf;
7906         int slot;
7907
7908         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7909         if (ret < 0)
7910                 goto out;
7911
7912         while (1) {
7913                 slot = path->slots[0];
7914                 leaf = path->nodes[0];
7915                 if (slot >= btrfs_header_nritems(leaf)) {
7916                         ret = btrfs_next_leaf(root, path);
7917                         if (ret == 0)
7918                                 continue;
7919                         if (ret < 0)
7920                                 goto out;
7921                         break;
7922                 }
7923                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7924
7925                 if (found_key.objectid >= key->objectid &&
7926                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7927                         ret = 0;
7928                         goto out;
7929                 }
7930                 path->slots[0]++;
7931         }
7932 out:
7933         return ret;
7934 }
7935
7936 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7937 {
7938         struct btrfs_block_group_cache *block_group;
7939         u64 last = 0;
7940
7941         while (1) {
7942                 struct inode *inode;
7943
7944                 block_group = btrfs_lookup_first_block_group(info, last);
7945                 while (block_group) {
7946                         spin_lock(&block_group->lock);
7947                         if (block_group->iref)
7948                                 break;
7949                         spin_unlock(&block_group->lock);
7950                         block_group = next_block_group(info->tree_root,
7951                                                        block_group);
7952                 }
7953                 if (!block_group) {
7954                         if (last == 0)
7955                                 break;
7956                         last = 0;
7957                         continue;
7958                 }
7959
7960                 inode = block_group->inode;
7961                 block_group->iref = 0;
7962                 block_group->inode = NULL;
7963                 spin_unlock(&block_group->lock);
7964                 iput(inode);
7965                 last = block_group->key.objectid + block_group->key.offset;
7966                 btrfs_put_block_group(block_group);
7967         }
7968 }
7969
7970 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7971 {
7972         struct btrfs_block_group_cache *block_group;
7973         struct btrfs_space_info *space_info;
7974         struct btrfs_caching_control *caching_ctl;
7975         struct rb_node *n;
7976
7977         down_write(&info->extent_commit_sem);
7978         while (!list_empty(&info->caching_block_groups)) {
7979                 caching_ctl = list_entry(info->caching_block_groups.next,
7980                                          struct btrfs_caching_control, list);
7981                 list_del(&caching_ctl->list);
7982                 put_caching_control(caching_ctl);
7983         }
7984         up_write(&info->extent_commit_sem);
7985
7986         spin_lock(&info->block_group_cache_lock);
7987         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7988                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7989                                        cache_node);
7990                 rb_erase(&block_group->cache_node,
7991                          &info->block_group_cache_tree);
7992                 spin_unlock(&info->block_group_cache_lock);
7993
7994                 down_write(&block_group->space_info->groups_sem);
7995                 list_del(&block_group->list);
7996                 up_write(&block_group->space_info->groups_sem);
7997
7998                 if (block_group->cached == BTRFS_CACHE_STARTED)
7999                         wait_block_group_cache_done(block_group);
8000
8001                 /*
8002                  * We haven't cached this block group, which means we could
8003                  * possibly have excluded extents on this block group.
8004                  */
8005                 if (block_group->cached == BTRFS_CACHE_NO)
8006                         free_excluded_extents(info->extent_root, block_group);
8007
8008                 btrfs_remove_free_space_cache(block_group);
8009                 btrfs_put_block_group(block_group);
8010
8011                 spin_lock(&info->block_group_cache_lock);
8012         }
8013         spin_unlock(&info->block_group_cache_lock);
8014
8015         /* now that all the block groups are freed, go through and
8016          * free all the space_info structs.  This is only called during
8017          * the final stages of unmount, and so we know nobody is
8018          * using them.  We call synchronize_rcu() once before we start,
8019          * just to be on the safe side.
8020          */
8021         synchronize_rcu();
8022
8023         release_global_block_rsv(info);
8024
8025         while(!list_empty(&info->space_info)) {
8026                 space_info = list_entry(info->space_info.next,
8027                                         struct btrfs_space_info,
8028                                         list);
8029                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8030                         if (space_info->bytes_pinned > 0 ||
8031                             space_info->bytes_reserved > 0 ||
8032                             space_info->bytes_may_use > 0) {
8033                                 WARN_ON(1);
8034                                 dump_space_info(space_info, 0, 0);
8035                         }
8036                 }
8037                 list_del(&space_info->list);
8038                 kfree(space_info);
8039         }
8040         return 0;
8041 }
8042
8043 static void __link_block_group(struct btrfs_space_info *space_info,
8044                                struct btrfs_block_group_cache *cache)
8045 {
8046         int index = get_block_group_index(cache);
8047
8048         down_write(&space_info->groups_sem);
8049         list_add_tail(&cache->list, &space_info->block_groups[index]);
8050         up_write(&space_info->groups_sem);
8051 }
8052
8053 int btrfs_read_block_groups(struct btrfs_root *root)
8054 {
8055         struct btrfs_path *path;
8056         int ret;
8057         struct btrfs_block_group_cache *cache;
8058         struct btrfs_fs_info *info = root->fs_info;
8059         struct btrfs_space_info *space_info;
8060         struct btrfs_key key;
8061         struct btrfs_key found_key;
8062         struct extent_buffer *leaf;
8063         int need_clear = 0;
8064         u64 cache_gen;
8065
8066         root = info->extent_root;
8067         key.objectid = 0;
8068         key.offset = 0;
8069         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8070         path = btrfs_alloc_path();
8071         if (!path)
8072                 return -ENOMEM;
8073         path->reada = 1;
8074
8075         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
8076         if (btrfs_test_opt(root, SPACE_CACHE) &&
8077             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
8078                 need_clear = 1;
8079         if (btrfs_test_opt(root, CLEAR_CACHE))
8080                 need_clear = 1;
8081
8082         while (1) {
8083                 ret = find_first_block_group(root, path, &key);
8084                 if (ret > 0)
8085                         break;
8086                 if (ret != 0)
8087                         goto error;
8088                 leaf = path->nodes[0];
8089                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8090                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
8091                 if (!cache) {
8092                         ret = -ENOMEM;
8093                         goto error;
8094                 }
8095                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8096                                                 GFP_NOFS);
8097                 if (!cache->free_space_ctl) {
8098                         kfree(cache);
8099                         ret = -ENOMEM;
8100                         goto error;
8101                 }
8102
8103                 atomic_set(&cache->count, 1);
8104                 spin_lock_init(&cache->lock);
8105                 cache->fs_info = info;
8106                 INIT_LIST_HEAD(&cache->list);
8107                 INIT_LIST_HEAD(&cache->cluster_list);
8108
8109                 if (need_clear) {
8110                         /*
8111                          * When we mount with old space cache, we need to
8112                          * set BTRFS_DC_CLEAR and set dirty flag.
8113                          *
8114                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8115                          *    truncate the old free space cache inode and
8116                          *    setup a new one.
8117                          * b) Setting 'dirty flag' makes sure that we flush
8118                          *    the new space cache info onto disk.
8119                          */
8120                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8121                         if (btrfs_test_opt(root, SPACE_CACHE))
8122                                 cache->dirty = 1;
8123                 }
8124
8125                 read_extent_buffer(leaf, &cache->item,
8126                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8127                                    sizeof(cache->item));
8128                 memcpy(&cache->key, &found_key, sizeof(found_key));
8129
8130                 key.objectid = found_key.objectid + found_key.offset;
8131                 btrfs_release_path(path);
8132                 cache->flags = btrfs_block_group_flags(&cache->item);
8133                 cache->sectorsize = root->sectorsize;
8134                 cache->full_stripe_len = btrfs_full_stripe_len(root,
8135                                                &root->fs_info->mapping_tree,
8136                                                found_key.objectid);
8137                 btrfs_init_free_space_ctl(cache);
8138
8139                 /*
8140                  * We need to exclude the super stripes now so that the space
8141                  * info has super bytes accounted for, otherwise we'll think
8142                  * we have more space than we actually do.
8143                  */
8144                 ret = exclude_super_stripes(root, cache);
8145                 if (ret) {
8146                         /*
8147                          * We may have excluded something, so call this just in
8148                          * case.
8149                          */
8150                         free_excluded_extents(root, cache);
8151                         kfree(cache->free_space_ctl);
8152                         kfree(cache);
8153                         goto error;
8154                 }
8155
8156                 /*
8157                  * check for two cases, either we are full, and therefore
8158                  * don't need to bother with the caching work since we won't
8159                  * find any space, or we are empty, and we can just add all
8160                  * the space in and be done with it.  This saves us _alot_ of
8161                  * time, particularly in the full case.
8162                  */
8163                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8164                         cache->last_byte_to_unpin = (u64)-1;
8165                         cache->cached = BTRFS_CACHE_FINISHED;
8166                         free_excluded_extents(root, cache);
8167                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8168                         cache->last_byte_to_unpin = (u64)-1;
8169                         cache->cached = BTRFS_CACHE_FINISHED;
8170                         add_new_free_space(cache, root->fs_info,
8171                                            found_key.objectid,
8172                                            found_key.objectid +
8173                                            found_key.offset);
8174                         free_excluded_extents(root, cache);
8175                 }
8176
8177                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8178                 if (ret) {
8179                         btrfs_remove_free_space_cache(cache);
8180                         btrfs_put_block_group(cache);
8181                         goto error;
8182                 }
8183
8184                 ret = update_space_info(info, cache->flags, found_key.offset,
8185                                         btrfs_block_group_used(&cache->item),
8186                                         &space_info);
8187                 if (ret) {
8188                         btrfs_remove_free_space_cache(cache);
8189                         spin_lock(&info->block_group_cache_lock);
8190                         rb_erase(&cache->cache_node,
8191                                  &info->block_group_cache_tree);
8192                         spin_unlock(&info->block_group_cache_lock);
8193                         btrfs_put_block_group(cache);
8194                         goto error;
8195                 }
8196
8197                 cache->space_info = space_info;
8198                 spin_lock(&cache->space_info->lock);
8199                 cache->space_info->bytes_readonly += cache->bytes_super;
8200                 spin_unlock(&cache->space_info->lock);
8201
8202                 __link_block_group(space_info, cache);
8203
8204                 set_avail_alloc_bits(root->fs_info, cache->flags);
8205                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8206                         set_block_group_ro(cache, 1);
8207         }
8208
8209         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8210                 if (!(get_alloc_profile(root, space_info->flags) &
8211                       (BTRFS_BLOCK_GROUP_RAID10 |
8212                        BTRFS_BLOCK_GROUP_RAID1 |
8213                        BTRFS_BLOCK_GROUP_RAID5 |
8214                        BTRFS_BLOCK_GROUP_RAID6 |
8215                        BTRFS_BLOCK_GROUP_DUP)))
8216                         continue;
8217                 /*
8218                  * avoid allocating from un-mirrored block group if there are
8219                  * mirrored block groups.
8220                  */
8221                 list_for_each_entry(cache, &space_info->block_groups[3], list)
8222                         set_block_group_ro(cache, 1);
8223                 list_for_each_entry(cache, &space_info->block_groups[4], list)
8224                         set_block_group_ro(cache, 1);
8225         }
8226
8227         init_global_block_rsv(info);
8228         ret = 0;
8229 error:
8230         btrfs_free_path(path);
8231         return ret;
8232 }
8233
8234 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8235                                        struct btrfs_root *root)
8236 {
8237         struct btrfs_block_group_cache *block_group, *tmp;
8238         struct btrfs_root *extent_root = root->fs_info->extent_root;
8239         struct btrfs_block_group_item item;
8240         struct btrfs_key key;
8241         int ret = 0;
8242
8243         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8244                                  new_bg_list) {
8245                 list_del_init(&block_group->new_bg_list);
8246
8247                 if (ret)
8248                         continue;
8249
8250                 spin_lock(&block_group->lock);
8251                 memcpy(&item, &block_group->item, sizeof(item));
8252                 memcpy(&key, &block_group->key, sizeof(key));
8253                 spin_unlock(&block_group->lock);
8254
8255                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8256                                         sizeof(item));
8257                 if (ret)
8258                         btrfs_abort_transaction(trans, extent_root, ret);
8259         }
8260 }
8261
8262 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8263                            struct btrfs_root *root, u64 bytes_used,
8264                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8265                            u64 size)
8266 {
8267         int ret;
8268         struct btrfs_root *extent_root;
8269         struct btrfs_block_group_cache *cache;
8270
8271         extent_root = root->fs_info->extent_root;
8272
8273         root->fs_info->last_trans_log_full_commit = trans->transid;
8274
8275         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8276         if (!cache)
8277                 return -ENOMEM;
8278         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8279                                         GFP_NOFS);
8280         if (!cache->free_space_ctl) {
8281                 kfree(cache);
8282                 return -ENOMEM;
8283         }
8284
8285         cache->key.objectid = chunk_offset;
8286         cache->key.offset = size;
8287         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8288         cache->sectorsize = root->sectorsize;
8289         cache->fs_info = root->fs_info;
8290         cache->full_stripe_len = btrfs_full_stripe_len(root,
8291                                                &root->fs_info->mapping_tree,
8292                                                chunk_offset);
8293
8294         atomic_set(&cache->count, 1);
8295         spin_lock_init(&cache->lock);
8296         INIT_LIST_HEAD(&cache->list);
8297         INIT_LIST_HEAD(&cache->cluster_list);
8298         INIT_LIST_HEAD(&cache->new_bg_list);
8299
8300         btrfs_init_free_space_ctl(cache);
8301
8302         btrfs_set_block_group_used(&cache->item, bytes_used);
8303         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8304         cache->flags = type;
8305         btrfs_set_block_group_flags(&cache->item, type);
8306
8307         cache->last_byte_to_unpin = (u64)-1;
8308         cache->cached = BTRFS_CACHE_FINISHED;
8309         ret = exclude_super_stripes(root, cache);
8310         if (ret) {
8311                 /*
8312                  * We may have excluded something, so call this just in
8313                  * case.
8314                  */
8315                 free_excluded_extents(root, cache);
8316                 kfree(cache->free_space_ctl);
8317                 kfree(cache);
8318                 return ret;
8319         }
8320
8321         add_new_free_space(cache, root->fs_info, chunk_offset,
8322                            chunk_offset + size);
8323
8324         free_excluded_extents(root, cache);
8325
8326         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8327         if (ret) {
8328                 btrfs_remove_free_space_cache(cache);
8329                 btrfs_put_block_group(cache);
8330                 return ret;
8331         }
8332
8333         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8334                                 &cache->space_info);
8335         if (ret) {
8336                 btrfs_remove_free_space_cache(cache);
8337                 spin_lock(&root->fs_info->block_group_cache_lock);
8338                 rb_erase(&cache->cache_node,
8339                          &root->fs_info->block_group_cache_tree);
8340                 spin_unlock(&root->fs_info->block_group_cache_lock);
8341                 btrfs_put_block_group(cache);
8342                 return ret;
8343         }
8344         update_global_block_rsv(root->fs_info);
8345
8346         spin_lock(&cache->space_info->lock);
8347         cache->space_info->bytes_readonly += cache->bytes_super;
8348         spin_unlock(&cache->space_info->lock);
8349
8350         __link_block_group(cache->space_info, cache);
8351
8352         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8353
8354         set_avail_alloc_bits(extent_root->fs_info, type);
8355
8356         return 0;
8357 }
8358
8359 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8360 {
8361         u64 extra_flags = chunk_to_extended(flags) &
8362                                 BTRFS_EXTENDED_PROFILE_MASK;
8363
8364         write_seqlock(&fs_info->profiles_lock);
8365         if (flags & BTRFS_BLOCK_GROUP_DATA)
8366                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8367         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8368                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8369         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8370                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8371         write_sequnlock(&fs_info->profiles_lock);
8372 }
8373
8374 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8375                              struct btrfs_root *root, u64 group_start)
8376 {
8377         struct btrfs_path *path;
8378         struct btrfs_block_group_cache *block_group;
8379         struct btrfs_free_cluster *cluster;
8380         struct btrfs_root *tree_root = root->fs_info->tree_root;
8381         struct btrfs_key key;
8382         struct inode *inode;
8383         int ret;
8384         int index;
8385         int factor;
8386
8387         root = root->fs_info->extent_root;
8388
8389         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8390         BUG_ON(!block_group);
8391         BUG_ON(!block_group->ro);
8392
8393         /*
8394          * Free the reserved super bytes from this block group before
8395          * remove it.
8396          */
8397         free_excluded_extents(root, block_group);
8398
8399         memcpy(&key, &block_group->key, sizeof(key));
8400         index = get_block_group_index(block_group);
8401         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8402                                   BTRFS_BLOCK_GROUP_RAID1 |
8403                                   BTRFS_BLOCK_GROUP_RAID10))
8404                 factor = 2;
8405         else
8406                 factor = 1;
8407
8408         /* make sure this block group isn't part of an allocation cluster */
8409         cluster = &root->fs_info->data_alloc_cluster;
8410         spin_lock(&cluster->refill_lock);
8411         btrfs_return_cluster_to_free_space(block_group, cluster);
8412         spin_unlock(&cluster->refill_lock);
8413
8414         /*
8415          * make sure this block group isn't part of a metadata
8416          * allocation cluster
8417          */
8418         cluster = &root->fs_info->meta_alloc_cluster;
8419         spin_lock(&cluster->refill_lock);
8420         btrfs_return_cluster_to_free_space(block_group, cluster);
8421         spin_unlock(&cluster->refill_lock);
8422
8423         path = btrfs_alloc_path();
8424         if (!path) {
8425                 ret = -ENOMEM;
8426                 goto out;
8427         }
8428
8429         inode = lookup_free_space_inode(tree_root, block_group, path);
8430         if (!IS_ERR(inode)) {
8431                 ret = btrfs_orphan_add(trans, inode);
8432                 if (ret) {
8433                         btrfs_add_delayed_iput(inode);
8434                         goto out;
8435                 }
8436                 clear_nlink(inode);
8437                 /* One for the block groups ref */
8438                 spin_lock(&block_group->lock);
8439                 if (block_group->iref) {
8440                         block_group->iref = 0;
8441                         block_group->inode = NULL;
8442                         spin_unlock(&block_group->lock);
8443                         iput(inode);
8444                 } else {
8445                         spin_unlock(&block_group->lock);
8446                 }
8447                 /* One for our lookup ref */
8448                 btrfs_add_delayed_iput(inode);
8449         }
8450
8451         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8452         key.offset = block_group->key.objectid;
8453         key.type = 0;
8454
8455         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8456         if (ret < 0)
8457                 goto out;
8458         if (ret > 0)
8459                 btrfs_release_path(path);
8460         if (ret == 0) {
8461                 ret = btrfs_del_item(trans, tree_root, path);
8462                 if (ret)
8463                         goto out;
8464                 btrfs_release_path(path);
8465         }
8466
8467         spin_lock(&root->fs_info->block_group_cache_lock);
8468         rb_erase(&block_group->cache_node,
8469                  &root->fs_info->block_group_cache_tree);
8470
8471         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8472                 root->fs_info->first_logical_byte = (u64)-1;
8473         spin_unlock(&root->fs_info->block_group_cache_lock);
8474
8475         down_write(&block_group->space_info->groups_sem);
8476         /*
8477          * we must use list_del_init so people can check to see if they
8478          * are still on the list after taking the semaphore
8479          */
8480         list_del_init(&block_group->list);
8481         if (list_empty(&block_group->space_info->block_groups[index]))
8482                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8483         up_write(&block_group->space_info->groups_sem);
8484
8485         if (block_group->cached == BTRFS_CACHE_STARTED)
8486                 wait_block_group_cache_done(block_group);
8487
8488         btrfs_remove_free_space_cache(block_group);
8489
8490         spin_lock(&block_group->space_info->lock);
8491         block_group->space_info->total_bytes -= block_group->key.offset;
8492         block_group->space_info->bytes_readonly -= block_group->key.offset;
8493         block_group->space_info->disk_total -= block_group->key.offset * factor;
8494         spin_unlock(&block_group->space_info->lock);
8495
8496         memcpy(&key, &block_group->key, sizeof(key));
8497
8498         btrfs_clear_space_info_full(root->fs_info);
8499
8500         btrfs_put_block_group(block_group);
8501         btrfs_put_block_group(block_group);
8502
8503         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8504         if (ret > 0)
8505                 ret = -EIO;
8506         if (ret < 0)
8507                 goto out;
8508
8509         ret = btrfs_del_item(trans, root, path);
8510 out:
8511         btrfs_free_path(path);
8512         return ret;
8513 }
8514
8515 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8516 {
8517         struct btrfs_space_info *space_info;
8518         struct btrfs_super_block *disk_super;
8519         u64 features;
8520         u64 flags;
8521         int mixed = 0;
8522         int ret;
8523
8524         disk_super = fs_info->super_copy;
8525         if (!btrfs_super_root(disk_super))
8526                 return 1;
8527
8528         features = btrfs_super_incompat_flags(disk_super);
8529         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8530                 mixed = 1;
8531
8532         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8533         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8534         if (ret)
8535                 goto out;
8536
8537         if (mixed) {
8538                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8539                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8540         } else {
8541                 flags = BTRFS_BLOCK_GROUP_METADATA;
8542                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8543                 if (ret)
8544                         goto out;
8545
8546                 flags = BTRFS_BLOCK_GROUP_DATA;
8547                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8548         }
8549 out:
8550         return ret;
8551 }
8552
8553 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8554 {
8555         return unpin_extent_range(root, start, end);
8556 }
8557
8558 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8559                                u64 num_bytes, u64 *actual_bytes)
8560 {
8561         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8562 }
8563
8564 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8565 {
8566         struct btrfs_fs_info *fs_info = root->fs_info;
8567         struct btrfs_block_group_cache *cache = NULL;
8568         u64 group_trimmed;
8569         u64 start;
8570         u64 end;
8571         u64 trimmed = 0;
8572         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8573         int ret = 0;
8574
8575         /*
8576          * try to trim all FS space, our block group may start from non-zero.
8577          */
8578         if (range->len == total_bytes)
8579                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8580         else
8581                 cache = btrfs_lookup_block_group(fs_info, range->start);
8582
8583         while (cache) {
8584                 if (cache->key.objectid >= (range->start + range->len)) {
8585                         btrfs_put_block_group(cache);
8586                         break;
8587                 }
8588
8589                 start = max(range->start, cache->key.objectid);
8590                 end = min(range->start + range->len,
8591                                 cache->key.objectid + cache->key.offset);
8592
8593                 if (end - start >= range->minlen) {
8594                         if (!block_group_cache_done(cache)) {
8595                                 ret = cache_block_group(cache, 0);
8596                                 if (!ret)
8597                                         wait_block_group_cache_done(cache);
8598                         }
8599                         ret = btrfs_trim_block_group(cache,
8600                                                      &group_trimmed,
8601                                                      start,
8602                                                      end,
8603                                                      range->minlen);
8604
8605                         trimmed += group_trimmed;
8606                         if (ret) {
8607                                 btrfs_put_block_group(cache);
8608                                 break;
8609                         }
8610                 }
8611
8612                 cache = next_block_group(fs_info->tree_root, cache);
8613         }
8614
8615         range->len = trimmed;
8616         return ret;
8617 }