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