fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include "compat.h"
  28 #include "hash.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "raid56.h"
  35 #include "locking.h"
  36 #include "free-space-cache.h"
  37 #include "math.h"
  38
  39 #undef SCRAMBLE_DELAYED_REFS
  40
  41 /*
  42  * control flags for do_chunk_alloc's force field
  43  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  44  * if we really need one.
  45  *
  46  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  47  * if we have very few chunks already allocated.  This is
  48  * used as part of the clustering code to help make sure
  49  * we have a good pool of storage to cluster in, without
  50  * filling the FS with empty chunks
  51  *
  52  * CHUNK_ALLOC_FORCE means it must try to allocate one
  53  *
  54  */
  55 enum {
  56         CHUNK_ALLOC_NO_FORCE = 0,
  57         CHUNK_ALLOC_LIMITED = 1,
  58         CHUNK_ALLOC_FORCE = 2,
  59 };
  60
  61 /*
  62  * Control how reservations are dealt with.
  63  *
  64  * RESERVE_FREE - freeing a reservation.
  65  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  66  *   ENOSPC accounting
  67  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  68  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  69  */
  70 enum {
  71         RESERVE_FREE = 0,
  72         RESERVE_ALLOC = 1,
  73         RESERVE_ALLOC_NO_ACCOUNT = 2,
  74 };
  75
  76 static int update_block_group(struct btrfs_root *root,
  77                               u64 bytenr, u64 num_bytes, int alloc);
  78 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  79                                 struct btrfs_root *root,
  80                                 u64 bytenr, u64 num_bytes, u64 parent,
  81                                 u64 root_objectid, u64 owner_objectid,
  82                                 u64 owner_offset, int refs_to_drop,
  83                                 struct btrfs_delayed_extent_op *extra_op);
  84 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  85                                     struct extent_buffer *leaf,
  86                                     struct btrfs_extent_item *ei);
  87 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  88                                       struct btrfs_root *root,
  89                                       u64 parent, u64 root_objectid,
  90                                       u64 flags, u64 owner, u64 offset,
  91                                       struct btrfs_key *ins, int ref_mod);
  92 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  93                                      struct btrfs_root *root,
  94                                      u64 parent, u64 root_objectid,
  95                                      u64 flags, struct btrfs_disk_key *key,
  96                                      int level, struct btrfs_key *ins);
  97 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  98                           struct btrfs_root *extent_root, u64 flags,
  99                           int force);
 100 static int find_next_key(struct btrfs_path *path, int level,
 101                          struct btrfs_key *key);
 102 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 103                             int dump_block_groups);
 104 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 105                                        u64 num_bytes, int reserve);
 106 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 107                                u64 num_bytes);
 108
 109 static noinline int
 110 block_group_cache_done(struct btrfs_block_group_cache *cache)
 111 {
 112         smp_mb();
 113         return cache->cached == BTRFS_CACHE_FINISHED;
 114 }
 115
 116 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 117 {
 118         return (cache->flags & bits) == bits;
 119 }
 120
 121 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         atomic_inc(&cache->count);
 124 }
 125
 126 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 127 {
 128         if (atomic_dec_and_test(&cache->count)) {
 129                 WARN_ON(cache->pinned > 0);
 130                 WARN_ON(cache->reserved > 0);
 131                 kfree(cache->free_space_ctl);
 132                 kfree(cache);
 133         }
 134 }
 135
 136 /*
 137  * this adds the block group to the fs_info rb tree for the block group
 138  * cache
 139  */
 140 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 141                                 struct btrfs_block_group_cache *block_group)
 142 {
 143         struct rb_node **p;
 144         struct rb_node *parent = NULL;
 145         struct btrfs_block_group_cache *cache;
 146
 147         spin_lock(&info->block_group_cache_lock);
 148         p = &info->block_group_cache_tree.rb_node;
 149
 150         while (*p) {
 151                 parent = *p;
 152                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 153                                  cache_node);
 154                 if (block_group->key.objectid < cache->key.objectid) {
 155                         p = &(*p)->rb_left;
 156                 } else if (block_group->key.objectid > cache->key.objectid) {
 157                         p = &(*p)->rb_right;
 158                 } else {
 159                         spin_unlock(&info->block_group_cache_lock);
 160                         return -EEXIST;
 161                 }
 162         }
 163
 164         rb_link_node(&block_group->cache_node, parent, p);
 165         rb_insert_color(&block_group->cache_node,
 166                         &info->block_group_cache_tree);
 167
 168         if (info->first_logical_byte > block_group->key.objectid)
 169                 info->first_logical_byte = block_group->key.objectid;
 170
 171         spin_unlock(&info->block_group_cache_lock);
 172
 173         return 0;
 174 }
 175
 176 /*
 177  * This will return the block group at or after bytenr if contains is 0, else
 178  * it will return the block group that contains the bytenr
 179  */
 180 static struct btrfs_block_group_cache *
 181 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 182                               int contains)
 183 {
 184         struct btrfs_block_group_cache *cache, *ret = NULL;
 185         struct rb_node *n;
 186         u64 end, start;
 187
 188         spin_lock(&info->block_group_cache_lock);
 189         n = info->block_group_cache_tree.rb_node;
 190
 191         while (n) {
 192                 cache = rb_entry(n, struct btrfs_block_group_cache,
 193                                  cache_node);
 194                 end = cache->key.objectid + cache->key.offset - 1;
 195                 start = cache->key.objectid;
 196
 197                 if (bytenr < start) {
 198                         if (!contains && (!ret || start < ret->key.objectid))
 199                                 ret = cache;
 200                         n = n->rb_left;
 201                 } else if (bytenr > start) {
 202                         if (contains && bytenr <= end) {
 203                                 ret = cache;
 204                                 break;
 205                         }
 206                         n = n->rb_right;
 207                 } else {
 208                         ret = cache;
 209                         break;
 210                 }
 211         }
 212         if (ret) {
 213                 btrfs_get_block_group(ret);
 214                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 215                         info->first_logical_byte = ret->key.objectid;
 216         }
 217         spin_unlock(&info->block_group_cache_lock);
 218
 219         return ret;
 220 }
 221
 222 static int add_excluded_extent(struct btrfs_root *root,
 223                                u64 start, u64 num_bytes)
 224 {
 225         u64 end = start + num_bytes - 1;
 226         set_extent_bits(&root->fs_info->freed_extents[0],
 227                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 228         set_extent_bits(&root->fs_info->freed_extents[1],
 229                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 230         return 0;
 231 }
 232
 233 static void free_excluded_extents(struct btrfs_root *root,
 234                                   struct btrfs_block_group_cache *cache)
 235 {
 236         u64 start, end;
 237
 238         start = cache->key.objectid;
 239         end = start + cache->key.offset - 1;
 240
 241         clear_extent_bits(&root->fs_info->freed_extents[0],
 242                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 243         clear_extent_bits(&root->fs_info->freed_extents[1],
 244                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 245 }
 246
 247 static int exclude_super_stripes(struct btrfs_root *root,
 248                                  struct btrfs_block_group_cache *cache)
 249 {
 250         u64 bytenr;
 251         u64 *logical;
 252         int stripe_len;
 253         int i, nr, ret;
 254
 255         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 256                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 257                 cache->bytes_super += stripe_len;
 258                 ret = add_excluded_extent(root, cache->key.objectid,
 259                                           stripe_len);
 260                 if (ret)
 261                         return ret;
 262         }
 263
 264         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 265                 bytenr = btrfs_sb_offset(i);
 266                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 267                                        cache->key.objectid, bytenr,
 268                                        0, &logical, &nr, &stripe_len);
 269                 if (ret)
 270                         return ret;
 271
 272                 while (nr--) {
 273                         cache->bytes_super += stripe_len;
 274                         ret = add_excluded_extent(root, logical[nr],
 275                                                   stripe_len);
 276                         if (ret) {
 277                                 kfree(logical);
 278                                 return ret;
 279                         }
 280                 }
 281
 282                 kfree(logical);
 283         }
 284         return 0;
 285 }
 286
 287 static struct btrfs_caching_control *
 288 get_caching_control(struct btrfs_block_group_cache *cache)
 289 {
 290         struct btrfs_caching_control *ctl;
 291
 292         spin_lock(&cache->lock);
 293         if (cache->cached != BTRFS_CACHE_STARTED) {
 294                 spin_unlock(&cache->lock);
 295                 return NULL;
 296         }
 297
 298         /* We're loading it the fast way, so we don't have a caching_ctl. */
 299         if (!cache->caching_ctl) {
 300                 spin_unlock(&cache->lock);
 301                 return NULL;
 302         }
 303
 304         ctl = cache->caching_ctl;
 305         atomic_inc(&ctl->count);
 306         spin_unlock(&cache->lock);
 307         return ctl;
 308 }
 309
 310 static void put_caching_control(struct btrfs_caching_control *ctl)
 311 {
 312         if (atomic_dec_and_test(&ctl->count))
 313                 kfree(ctl);
 314 }
 315
 316 /*
 317  * this is only called by cache_block_group, since we could have freed extents
 318  * we need to check the pinned_extents for any extents that can't be used yet
 319  * since their free space will be released as soon as the transaction commits.
 320  */
 321 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 322                               struct btrfs_fs_info *info, u64 start, u64 end)
 323 {
 324         u64 extent_start, extent_end, size, total_added = 0;
 325         int ret;
 326
 327         while (start < end) {
 328                 ret = find_first_extent_bit(info->pinned_extents, start,
 329                                             &extent_start, &extent_end,
 330                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 331                                             NULL);
 332                 if (ret)
 333                         break;
 334
 335                 if (extent_start <= start) {
 336                         start = extent_end + 1;
 337                 } else if (extent_start > start && extent_start < end) {
 338                         size = extent_start - start;
 339                         total_added += size;
 340                         ret = btrfs_add_free_space(block_group, start,
 341                                                    size);
 342                         BUG_ON(ret); /* -ENOMEM or logic error */
 343                         start = extent_end + 1;
 344                 } else {
 345                         break;
 346                 }
 347         }
 348
 349         if (start < end) {
 350                 size = end - start;
 351                 total_added += size;
 352                 ret = btrfs_add_free_space(block_group, start, size);
 353                 BUG_ON(ret); /* -ENOMEM or logic error */
 354         }
 355
 356         return total_added;
 357 }
 358
 359 static noinline void caching_thread(struct btrfs_work *work)
 360 {
 361         struct btrfs_block_group_cache *block_group;
 362         struct btrfs_fs_info *fs_info;
 363         struct btrfs_caching_control *caching_ctl;
 364         struct btrfs_root *extent_root;
 365         struct btrfs_path *path;
 366         struct extent_buffer *leaf;
 367         struct btrfs_key key;
 368         u64 total_found = 0;
 369         u64 last = 0;
 370         u32 nritems;
 371         int ret = 0;
 372
 373         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 374         block_group = caching_ctl->block_group;
 375         fs_info = block_group->fs_info;
 376         extent_root = fs_info->extent_root;
 377
 378         path = btrfs_alloc_path();
 379         if (!path)
 380                 goto out;
 381
 382         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 383
 384         /*
 385          * We don't want to deadlock with somebody trying to allocate a new
 386          * extent for the extent root while also trying to search the extent
 387          * root to add free space.  So we skip locking and search the commit
 388          * root, since its read-only
 389          */
 390         path->skip_locking = 1;
 391         path->search_commit_root = 1;
 392         path->reada = 1;
 393
 394         key.objectid = last;
 395         key.offset = 0;
 396         key.type = BTRFS_EXTENT_ITEM_KEY;
 397 again:
 398         mutex_lock(&caching_ctl->mutex);
 399         /* need to make sure the commit_root doesn't disappear */
 400         down_read(&fs_info->extent_commit_sem);
 401
 402         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 403         if (ret < 0)
 404                 goto err;
 405
 406         leaf = path->nodes[0];
 407         nritems = btrfs_header_nritems(leaf);
 408
 409         while (1) {
 410                 if (btrfs_fs_closing(fs_info) > 1) {
 411                         last = (u64)-1;
 412                         break;
 413                 }
 414
 415                 if (path->slots[0] < nritems) {
 416                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 417                 } else {
 418                         ret = find_next_key(path, 0, &key);
 419                         if (ret)
 420                                 break;
 421
 422                         if (need_resched() ||
 423                             btrfs_next_leaf(extent_root, path)) {
 424                                 caching_ctl->progress = last;
 425                                 btrfs_release_path(path);
 426                                 up_read(&fs_info->extent_commit_sem);
 427                                 mutex_unlock(&caching_ctl->mutex);
 428                                 cond_resched();
 429                                 goto again;
 430                         }
 431                         leaf = path->nodes[0];
 432                         nritems = btrfs_header_nritems(leaf);
 433                         continue;
 434                 }
 435
 436                 if (key.objectid < block_group->key.objectid) {
 437                         path->slots[0]++;
 438                         continue;
 439                 }
 440
 441                 if (key.objectid >= block_group->key.objectid +
 442                     block_group->key.offset)
 443                         break;
 444
 445                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 446                         total_found += add_new_free_space(block_group,
 447                                                           fs_info, last,
 448                                                           key.objectid);
 449                         last = key.objectid + key.offset;
 450
 451                         if (total_found > (1024 * 1024 * 2)) {
 452                                 total_found = 0;
 453                                 wake_up(&caching_ctl->wait);
 454                         }
 455                 }
 456                 path->slots[0]++;
 457         }
 458         ret = 0;
 459
 460         total_found += add_new_free_space(block_group, fs_info, last,
 461                                           block_group->key.objectid +
 462                                           block_group->key.offset);
 463         caching_ctl->progress = (u64)-1;
 464
 465         spin_lock(&block_group->lock);
 466         block_group->caching_ctl = NULL;
 467         block_group->cached = BTRFS_CACHE_FINISHED;
 468         spin_unlock(&block_group->lock);
 469
 470 err:
 471         btrfs_free_path(path);
 472         up_read(&fs_info->extent_commit_sem);
 473
 474         free_excluded_extents(extent_root, block_group);
 475
 476         mutex_unlock(&caching_ctl->mutex);
 477 out:
 478         wake_up(&caching_ctl->wait);
 479
 480         put_caching_control(caching_ctl);
 481         btrfs_put_block_group(block_group);
 482 }
 483
 484 static int cache_block_group(struct btrfs_block_group_cache *cache,
 485                              int load_cache_only)
 486 {
 487         DEFINE_WAIT(wait);
 488         struct btrfs_fs_info *fs_info = cache->fs_info;
 489         struct btrfs_caching_control *caching_ctl;
 490         int ret = 0;
 491
 492         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 493         if (!caching_ctl)
 494                 return -ENOMEM;
 495
 496         INIT_LIST_HEAD(&caching_ctl->list);
 497         mutex_init(&caching_ctl->mutex);
 498         init_waitqueue_head(&caching_ctl->wait);
 499         caching_ctl->block_group = cache;
 500         caching_ctl->progress = cache->key.objectid;
 501         atomic_set(&caching_ctl->count, 1);
 502         caching_ctl->work.func = caching_thread;
 503
 504         spin_lock(&cache->lock);
 505         /*
 506          * This should be a rare occasion, but this could happen I think in the
 507          * case where one thread starts to load the space cache info, and then
 508          * some other thread starts a transaction commit which tries to do an
 509          * allocation while the other thread is still loading the space cache
 510          * info.  The previous loop should have kept us from choosing this block
 511          * group, but if we've moved to the state where we will wait on caching
 512          * block groups we need to first check if we're doing a fast load here,
 513          * so we can wait for it to finish, otherwise we could end up allocating
 514          * from a block group who's cache gets evicted for one reason or
 515          * another.
 516          */
 517         while (cache->cached == BTRFS_CACHE_FAST) {
 518                 struct btrfs_caching_control *ctl;
 519
 520                 ctl = cache->caching_ctl;
 521                 atomic_inc(&ctl->count);
 522                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 523                 spin_unlock(&cache->lock);
 524
 525                 schedule();
 526
 527                 finish_wait(&ctl->wait, &wait);
 528                 put_caching_control(ctl);
 529                 spin_lock(&cache->lock);
 530         }
 531
 532         if (cache->cached != BTRFS_CACHE_NO) {
 533                 spin_unlock(&cache->lock);
 534                 kfree(caching_ctl);
 535                 return 0;
 536         }
 537         WARN_ON(cache->caching_ctl);
 538         cache->caching_ctl = caching_ctl;
 539         cache->cached = BTRFS_CACHE_FAST;
 540         spin_unlock(&cache->lock);
 541
 542         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 543                 ret = load_free_space_cache(fs_info, cache);
 544
 545                 spin_lock(&cache->lock);
 546                 if (ret == 1) {
 547                         cache->caching_ctl = NULL;
 548                         cache->cached = BTRFS_CACHE_FINISHED;
 549                         cache->last_byte_to_unpin = (u64)-1;
 550                 } else {
 551                         if (load_cache_only) {
 552                                 cache->caching_ctl = NULL;
 553                                 cache->cached = BTRFS_CACHE_NO;
 554                         } else {
 555                                 cache->cached = BTRFS_CACHE_STARTED;
 556                         }
 557                 }
 558                 spin_unlock(&cache->lock);
 559                 wake_up(&caching_ctl->wait);
 560                 if (ret == 1) {
 561                         put_caching_control(caching_ctl);
 562                         free_excluded_extents(fs_info->extent_root, cache);
 563                         return 0;
 564                 }
 565         } else {
 566                 /*
 567                  * We are not going to do the fast caching, set cached to the
 568                  * appropriate value and wakeup any waiters.
 569                  */
 570                 spin_lock(&cache->lock);
 571                 if (load_cache_only) {
 572                         cache->caching_ctl = NULL;
 573                         cache->cached = BTRFS_CACHE_NO;
 574                 } else {
 575                         cache->cached = BTRFS_CACHE_STARTED;
 576                 }
 577                 spin_unlock(&cache->lock);
 578                 wake_up(&caching_ctl->wait);
 579         }
 580
 581         if (load_cache_only) {
 582                 put_caching_control(caching_ctl);
 583                 return 0;
 584         }
 585
 586         down_write(&fs_info->extent_commit_sem);
 587         atomic_inc(&caching_ctl->count);
 588         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 589         up_write(&fs_info->extent_commit_sem);
 590
 591         btrfs_get_block_group(cache);
 592
 593         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 594
 595         return ret;
 596 }
 597
 598 /*
 599  * return the block group that starts at or after bytenr
 600  */
 601 static struct btrfs_block_group_cache *
 602 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 603 {
 604         struct btrfs_block_group_cache *cache;
 605
 606         cache = block_group_cache_tree_search(info, bytenr, 0);
 607
 608         return cache;
 609 }
 610
 611 /*
 612  * return the block group that contains the given bytenr
 613  */
 614 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 615                                                  struct btrfs_fs_info *info,
 616                                                  u64 bytenr)
 617 {
 618         struct btrfs_block_group_cache *cache;
 619
 620         cache = block_group_cache_tree_search(info, bytenr, 1);
 621
 622         return cache;
 623 }
 624
 625 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 626                                                   u64 flags)
 627 {
 628         struct list_head *head = &info->space_info;
 629         struct btrfs_space_info *found;
 630
 631         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 632
 633         rcu_read_lock();
 634         list_for_each_entry_rcu(found, head, list) {
 635                 if (found->flags & flags) {
 636                         rcu_read_unlock();
 637                         return found;
 638                 }
 639         }
 640         rcu_read_unlock();
 641         return NULL;
 642 }
 643
 644 /*
 645  * after adding space to the filesystem, we need to clear the full flags
 646  * on all the space infos.
 647  */
 648 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 649 {
 650         struct list_head *head = &info->space_info;
 651         struct btrfs_space_info *found;
 652
 653         rcu_read_lock();
 654         list_for_each_entry_rcu(found, head, list)
 655                 found->full = 0;
 656         rcu_read_unlock();
 657 }
 658
 659 u64 btrfs_find_block_group(struct btrfs_root *root,
 660                            u64 search_start, u64 search_hint, int owner)
 661 {
 662         struct btrfs_block_group_cache *cache;
 663         u64 used;
 664         u64 last = max(search_hint, search_start);
 665         u64 group_start = 0;
 666         int full_search = 0;
 667         int factor = 9;
 668         int wrapped = 0;
 669 again:
 670         while (1) {
 671                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 672                 if (!cache)
 673                         break;
 674
 675                 spin_lock(&cache->lock);
 676                 last = cache->key.objectid + cache->key.offset;
 677                 used = btrfs_block_group_used(&cache->item);
 678
 679                 if ((full_search || !cache->ro) &&
 680                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 681                         if (used + cache->pinned + cache->reserved <
 682                             div_factor(cache->key.offset, factor)) {
 683                                 group_start = cache->key.objectid;
 684                                 spin_unlock(&cache->lock);
 685                                 btrfs_put_block_group(cache);
 686                                 goto found;
 687                         }
 688                 }
 689                 spin_unlock(&cache->lock);
 690                 btrfs_put_block_group(cache);
 691                 cond_resched();
 692         }
 693         if (!wrapped) {
 694                 last = search_start;
 695                 wrapped = 1;
 696                 goto again;
 697         }
 698         if (!full_search && factor < 10) {
 699                 last = search_start;
 700                 full_search = 1;
 701                 factor = 10;
 702                 goto again;
 703         }
 704 found:
 705         return group_start;
 706 }
 707
 708 /* simple helper to search for an existing extent at a given offset */
 709 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 710 {
 711         int ret;
 712         struct btrfs_key key;
 713         struct btrfs_path *path;
 714
 715         path = btrfs_alloc_path();
 716         if (!path)
 717                 return -ENOMEM;
 718
 719         key.objectid = start;
 720         key.offset = len;
 721         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 722         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 723                                 0, 0);
 724         btrfs_free_path(path);
 725         return ret;
 726 }
 727
 728 /*
 729  * helper function to lookup reference count and flags of extent.
 730  *
 731  * the head node for delayed ref is used to store the sum of all the
 732  * reference count modifications queued up in the rbtree. the head
 733  * node may also store the extent flags to set. This way you can check
 734  * to see what the reference count and extent flags would be if all of
 735  * the delayed refs are not processed.
 736  */
 737 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 738                              struct btrfs_root *root, u64 bytenr,
 739                              u64 num_bytes, u64 *refs, u64 *flags)
 740 {
 741         struct btrfs_delayed_ref_head *head;
 742         struct btrfs_delayed_ref_root *delayed_refs;
 743         struct btrfs_path *path;
 744         struct btrfs_extent_item *ei;
 745         struct extent_buffer *leaf;
 746         struct btrfs_key key;
 747         u32 item_size;
 748         u64 num_refs;
 749         u64 extent_flags;
 750         int ret;
 751
 752         path = btrfs_alloc_path();
 753         if (!path)
 754                 return -ENOMEM;
 755
 756         key.objectid = bytenr;
 757         key.type = BTRFS_EXTENT_ITEM_KEY;
 758         key.offset = num_bytes;
 759         if (!trans) {
 760                 path->skip_locking = 1;
 761                 path->search_commit_root = 1;
 762         }
 763 again:
 764         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 765                                 &key, path, 0, 0);
 766         if (ret < 0)
 767                 goto out_free;
 768
 769         if (ret == 0) {
 770                 leaf = path->nodes[0];
 771                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 772                 if (item_size >= sizeof(*ei)) {
 773                         ei = btrfs_item_ptr(leaf, path->slots[0],
 774                                             struct btrfs_extent_item);
 775                         num_refs = btrfs_extent_refs(leaf, ei);
 776                         extent_flags = btrfs_extent_flags(leaf, ei);
 777                 } else {
 778 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 779                         struct btrfs_extent_item_v0 *ei0;
 780                         BUG_ON(item_size != sizeof(*ei0));
 781                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 782                                              struct btrfs_extent_item_v0);
 783                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 784                         /* FIXME: this isn't correct for data */
 785                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 786 #else
 787                         BUG();
 788 #endif
 789                 }
 790                 BUG_ON(num_refs == 0);
 791         } else {
 792                 num_refs = 0;
 793                 extent_flags = 0;
 794                 ret = 0;
 795         }
 796
 797         if (!trans)
 798                 goto out;
 799
 800         delayed_refs = &trans->transaction->delayed_refs;
 801         spin_lock(&delayed_refs->lock);
 802         head = btrfs_find_delayed_ref_head(trans, bytenr);
 803         if (head) {
 804                 if (!mutex_trylock(&head->mutex)) {
 805                         atomic_inc(&head->node.refs);
 806                         spin_unlock(&delayed_refs->lock);
 807
 808                         btrfs_release_path(path);
 809
 810                         /*
 811                          * Mutex was contended, block until it's released and try
 812                          * again
 813                          */
 814                         mutex_lock(&head->mutex);
 815                         mutex_unlock(&head->mutex);
 816                         btrfs_put_delayed_ref(&head->node);
 817                         goto again;
 818                 }
 819                 if (head->extent_op && head->extent_op->update_flags)
 820                         extent_flags |= head->extent_op->flags_to_set;
 821                 else
 822                         BUG_ON(num_refs == 0);
 823
 824                 num_refs += head->node.ref_mod;
 825                 mutex_unlock(&head->mutex);
 826         }
 827         spin_unlock(&delayed_refs->lock);
 828 out:
 829         WARN_ON(num_refs == 0);
 830         if (refs)
 831                 *refs = num_refs;
 832         if (flags)
 833                 *flags = extent_flags;
 834 out_free:
 835         btrfs_free_path(path);
 836         return ret;
 837 }
 838
 839 /*
 840  * Back reference rules.  Back refs have three main goals:
 841  *
 842  * 1) differentiate between all holders of references to an extent so that
 843  *    when a reference is dropped we can make sure it was a valid reference
 844  *    before freeing the extent.
 845  *
 846  * 2) Provide enough information to quickly find the holders of an extent
 847  *    if we notice a given block is corrupted or bad.
 848  *
 849  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 850  *    maintenance.  This is actually the same as #2, but with a slightly
 851  *    different use case.
 852  *
 853  * There are two kinds of back refs. The implicit back refs is optimized
 854  * for pointers in non-shared tree blocks. For a given pointer in a block,
 855  * back refs of this kind provide information about the block's owner tree
 856  * and the pointer's key. These information allow us to find the block by
 857  * b-tree searching. The full back refs is for pointers in tree blocks not
 858  * referenced by their owner trees. The location of tree block is recorded
 859  * in the back refs. Actually the full back refs is generic, and can be
 860  * used in all cases the implicit back refs is used. The major shortcoming
 861  * of the full back refs is its overhead. Every time a tree block gets
 862  * COWed, we have to update back refs entry for all pointers in it.
 863  *
 864  * For a newly allocated tree block, we use implicit back refs for
 865  * pointers in it. This means most tree related operations only involve
 866  * implicit back refs. For a tree block created in old transaction, the
 867  * only way to drop a reference to it is COW it. So we can detect the
 868  * event that tree block loses its owner tree's reference and do the
 869  * back refs conversion.
 870  *
 871  * When a tree block is COW'd through a tree, there are four cases:
 872  *
 873  * The reference count of the block is one and the tree is the block's
 874  * owner tree. Nothing to do in this case.
 875  *
 876  * The reference count of the block is one and the tree is not the
 877  * block's owner tree. In this case, full back refs is used for pointers
 878  * in the block. Remove these full back refs, add implicit back refs for
 879  * every pointers in the new block.
 880  *
 881  * The reference count of the block is greater than one and the tree is
 882  * the block's owner tree. In this case, implicit back refs is used for
 883  * pointers in the block. Add full back refs for every pointers in the
 884  * block, increase lower level extents' reference counts. The original
 885  * implicit back refs are entailed to the new block.
 886  *
 887  * The reference count of the block is greater than one and the tree is
 888  * not the block's owner tree. Add implicit back refs for every pointer in
 889  * the new block, increase lower level extents' reference count.
 890  *
 891  * Back Reference Key composing:
 892  *
 893  * The key objectid corresponds to the first byte in the extent,
 894  * The key type is used to differentiate between types of back refs.
 895  * There are different meanings of the key offset for different types
 896  * of back refs.
 897  *
 898  * File extents can be referenced by:
 899  *
 900  * - multiple snapshots, subvolumes, or different generations in one subvol
 901  * - different files inside a single subvolume
 902  * - different offsets inside a file (bookend extents in file.c)
 903  *
 904  * The extent ref structure for the implicit back refs has fields for:
 905  *
 906  * - Objectid of the subvolume root
 907  * - objectid of the file holding the reference
 908  * - original offset in the file
 909  * - how many bookend extents
 910  *
 911  * The key offset for the implicit back refs is hash of the first
 912  * three fields.
 913  *
 914  * The extent ref structure for the full back refs has field for:
 915  *
 916  * - number of pointers in the tree leaf
 917  *
 918  * The key offset for the implicit back refs is the first byte of
 919  * the tree leaf
 920  *
 921  * When a file extent is allocated, The implicit back refs is used.
 922  * the fields are filled in:
 923  *
 924  *     (root_key.objectid, inode objectid, offset in file, 1)
 925  *
 926  * When a file extent is removed file truncation, we find the
 927  * corresponding implicit back refs and check the following fields:
 928  *
 929  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 930  *
 931  * Btree extents can be referenced by:
 932  *
 933  * - Different subvolumes
 934  *
 935  * Both the implicit back refs and the full back refs for tree blocks
 936  * only consist of key. The key offset for the implicit back refs is
 937  * objectid of block's owner tree. The key offset for the full back refs
 938  * is the first byte of parent block.
 939  *
 940  * When implicit back refs is used, information about the lowest key and
 941  * level of the tree block are required. These information are stored in
 942  * tree block info structure.
 943  */
 944
 945 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 946 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 947                                   struct btrfs_root *root,
 948                                   struct btrfs_path *path,
 949                                   u64 owner, u32 extra_size)
 950 {
 951         struct btrfs_extent_item *item;
 952         struct btrfs_extent_item_v0 *ei0;
 953         struct btrfs_extent_ref_v0 *ref0;
 954         struct btrfs_tree_block_info *bi;
 955         struct extent_buffer *leaf;
 956         struct btrfs_key key;
 957         struct btrfs_key found_key;
 958         u32 new_size = sizeof(*item);
 959         u64 refs;
 960         int ret;
 961
 962         leaf = path->nodes[0];
 963         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 964
 965         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 966         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 967                              struct btrfs_extent_item_v0);
 968         refs = btrfs_extent_refs_v0(leaf, ei0);
 969
 970         if (owner == (u64)-1) {
 971                 while (1) {
 972                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 973                                 ret = btrfs_next_leaf(root, path);
 974                                 if (ret < 0)
 975                                         return ret;
 976                                 BUG_ON(ret > 0); /* Corruption */
 977                                 leaf = path->nodes[0];
 978                         }
 979                         btrfs_item_key_to_cpu(leaf, &found_key,
 980                                               path->slots[0]);
 981                         BUG_ON(key.objectid != found_key.objectid);
 982                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 983                                 path->slots[0]++;
 984                                 continue;
 985                         }
 986                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 987                                               struct btrfs_extent_ref_v0);
 988                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 989                         break;
 990                 }
 991         }
 992         btrfs_release_path(path);
 993
 994         if (owner < BTRFS_FIRST_FREE_OBJECTID)
 995                 new_size += sizeof(*bi);
 996
 997         new_size -= sizeof(*ei0);
 998         ret = btrfs_search_slot(trans, root, &key, path,
 999                                 new_size + extra_size, 1);
1000         if (ret < 0)
1001                 return ret;
1002         BUG_ON(ret); /* Corruption */
1003
1004         btrfs_extend_item(trans, root, path, new_size);
1005
1006         leaf = path->nodes[0];
1007         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1008         btrfs_set_extent_refs(leaf, item, refs);
1009         /* FIXME: get real generation */
1010         btrfs_set_extent_generation(leaf, item, 0);
1011         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1012                 btrfs_set_extent_flags(leaf, item,
1013                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1014                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1015                 bi = (struct btrfs_tree_block_info *)(item + 1);
1016                 /* FIXME: get first key of the block */
1017                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1018                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1019         } else {
1020                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1021         }
1022         btrfs_mark_buffer_dirty(leaf);
1023         return 0;
1024 }
1025 #endif
1026
1027 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1028 {
1029         u32 high_crc = ~(u32)0;
1030         u32 low_crc = ~(u32)0;
1031         __le64 lenum;
1032
1033         lenum = cpu_to_le64(root_objectid);
1034         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1035         lenum = cpu_to_le64(owner);
1036         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1037         lenum = cpu_to_le64(offset);
1038         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1039
1040         return ((u64)high_crc << 31) ^ (u64)low_crc;
1041 }
1042
1043 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1044                                      struct btrfs_extent_data_ref *ref)
1045 {
1046         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1047                                     btrfs_extent_data_ref_objectid(leaf, ref),
1048                                     btrfs_extent_data_ref_offset(leaf, ref));
1049 }
1050
1051 static int match_extent_data_ref(struct extent_buffer *leaf,
1052                                  struct btrfs_extent_data_ref *ref,
1053                                  u64 root_objectid, u64 owner, u64 offset)
1054 {
1055         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1056             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1057             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1058                 return 0;
1059         return 1;
1060 }
1061
1062 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1063                                            struct btrfs_root *root,
1064                                            struct btrfs_path *path,
1065                                            u64 bytenr, u64 parent,
1066                                            u64 root_objectid,
1067                                            u64 owner, u64 offset)
1068 {
1069         struct btrfs_key key;
1070         struct btrfs_extent_data_ref *ref;
1071         struct extent_buffer *leaf;
1072         u32 nritems;
1073         int ret;
1074         int recow;
1075         int err = -ENOENT;
1076
1077         key.objectid = bytenr;
1078         if (parent) {
1079                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1080                 key.offset = parent;
1081         } else {
1082                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1083                 key.offset = hash_extent_data_ref(root_objectid,
1084                                                   owner, offset);
1085         }
1086 again:
1087         recow = 0;
1088         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1089         if (ret < 0) {
1090                 err = ret;
1091                 goto fail;
1092         }
1093
1094         if (parent) {
1095                 if (!ret)
1096                         return 0;
1097 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1098                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1099                 btrfs_release_path(path);
1100                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1101                 if (ret < 0) {
1102                         err = ret;
1103                         goto fail;
1104                 }
1105                 if (!ret)
1106                         return 0;
1107 #endif
1108                 goto fail;
1109         }
1110
1111         leaf = path->nodes[0];
1112         nritems = btrfs_header_nritems(leaf);
1113         while (1) {
1114                 if (path->slots[0] >= nritems) {
1115                         ret = btrfs_next_leaf(root, path);
1116                         if (ret < 0)
1117                                 err = ret;
1118                         if (ret)
1119                                 goto fail;
1120
1121                         leaf = path->nodes[0];
1122                         nritems = btrfs_header_nritems(leaf);
1123                         recow = 1;
1124                 }
1125
1126                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1127                 if (key.objectid != bytenr ||
1128                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1129                         goto fail;
1130
1131                 ref = btrfs_item_ptr(leaf, path->slots[0],
1132                                      struct btrfs_extent_data_ref);
1133
1134                 if (match_extent_data_ref(leaf, ref, root_objectid,
1135                                           owner, offset)) {
1136                         if (recow) {
1137                                 btrfs_release_path(path);
1138                                 goto again;
1139                         }
1140                         err = 0;
1141                         break;
1142                 }
1143                 path->slots[0]++;
1144         }
1145 fail:
1146         return err;
1147 }
1148
1149 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1150                                            struct btrfs_root *root,
1151                                            struct btrfs_path *path,
1152                                            u64 bytenr, u64 parent,
1153                                            u64 root_objectid, u64 owner,
1154                                            u64 offset, int refs_to_add)
1155 {
1156         struct btrfs_key key;
1157         struct extent_buffer *leaf;
1158         u32 size;
1159         u32 num_refs;
1160         int ret;
1161
1162         key.objectid = bytenr;
1163         if (parent) {
1164                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1165                 key.offset = parent;
1166                 size = sizeof(struct btrfs_shared_data_ref);
1167         } else {
1168                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1169                 key.offset = hash_extent_data_ref(root_objectid,
1170                                                   owner, offset);
1171                 size = sizeof(struct btrfs_extent_data_ref);
1172         }
1173
1174         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1175         if (ret && ret != -EEXIST)
1176                 goto fail;
1177
1178         leaf = path->nodes[0];
1179         if (parent) {
1180                 struct btrfs_shared_data_ref *ref;
1181                 ref = btrfs_item_ptr(leaf, path->slots[0],
1182                                      struct btrfs_shared_data_ref);
1183                 if (ret == 0) {
1184                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1185                 } else {
1186                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1187                         num_refs += refs_to_add;
1188                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1189                 }
1190         } else {
1191                 struct btrfs_extent_data_ref *ref;
1192                 while (ret == -EEXIST) {
1193                         ref = btrfs_item_ptr(leaf, path->slots[0],
1194                                              struct btrfs_extent_data_ref);
1195                         if (match_extent_data_ref(leaf, ref, root_objectid,
1196                                                   owner, offset))
1197                                 break;
1198                         btrfs_release_path(path);
1199                         key.offset++;
1200                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1201                                                       size);
1202                         if (ret && ret != -EEXIST)
1203                                 goto fail;
1204
1205                         leaf = path->nodes[0];
1206                 }
1207                 ref = btrfs_item_ptr(leaf, path->slots[0],
1208                                      struct btrfs_extent_data_ref);
1209                 if (ret == 0) {
1210                         btrfs_set_extent_data_ref_root(leaf, ref,
1211                                                        root_objectid);
1212                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1213                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1214                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1215                 } else {
1216                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1217                         num_refs += refs_to_add;
1218                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1219                 }
1220         }
1221         btrfs_mark_buffer_dirty(leaf);
1222         ret = 0;
1223 fail:
1224         btrfs_release_path(path);
1225         return ret;
1226 }
1227
1228 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1229                                            struct btrfs_root *root,
1230                                            struct btrfs_path *path,
1231                                            int refs_to_drop)
1232 {
1233         struct btrfs_key key;
1234         struct btrfs_extent_data_ref *ref1 = NULL;
1235         struct btrfs_shared_data_ref *ref2 = NULL;
1236         struct extent_buffer *leaf;
1237         u32 num_refs = 0;
1238         int ret = 0;
1239
1240         leaf = path->nodes[0];
1241         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1242
1243         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1244                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1245                                       struct btrfs_extent_data_ref);
1246                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1247         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1248                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1249                                       struct btrfs_shared_data_ref);
1250                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1251 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1252         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1253                 struct btrfs_extent_ref_v0 *ref0;
1254                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1255                                       struct btrfs_extent_ref_v0);
1256                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1257 #endif
1258         } else {
1259                 BUG();
1260         }
1261
1262         BUG_ON(num_refs < refs_to_drop);
1263         num_refs -= refs_to_drop;
1264
1265         if (num_refs == 0) {
1266                 ret = btrfs_del_item(trans, root, path);
1267         } else {
1268                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1269                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1270                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1271                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1272 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1273                 else {
1274                         struct btrfs_extent_ref_v0 *ref0;
1275                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1276                                         struct btrfs_extent_ref_v0);
1277                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1278                 }
1279 #endif
1280                 btrfs_mark_buffer_dirty(leaf);
1281         }
1282         return ret;
1283 }
1284
1285 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1286                                           struct btrfs_path *path,
1287                                           struct btrfs_extent_inline_ref *iref)
1288 {
1289         struct btrfs_key key;
1290         struct extent_buffer *leaf;
1291         struct btrfs_extent_data_ref *ref1;
1292         struct btrfs_shared_data_ref *ref2;
1293         u32 num_refs = 0;
1294
1295         leaf = path->nodes[0];
1296         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1297         if (iref) {
1298                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1299                     BTRFS_EXTENT_DATA_REF_KEY) {
1300                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1301                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1302                 } else {
1303                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1304                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1305                 }
1306         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1307                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1308                                       struct btrfs_extent_data_ref);
1309                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1310         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1311                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1312                                       struct btrfs_shared_data_ref);
1313                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1314 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1315         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1316                 struct btrfs_extent_ref_v0 *ref0;
1317                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1318                                       struct btrfs_extent_ref_v0);
1319                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1320 #endif
1321         } else {
1322                 WARN_ON(1);
1323         }
1324         return num_refs;
1325 }
1326
1327 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1328                                           struct btrfs_root *root,
1329                                           struct btrfs_path *path,
1330                                           u64 bytenr, u64 parent,
1331                                           u64 root_objectid)
1332 {
1333         struct btrfs_key key;
1334         int ret;
1335
1336         key.objectid = bytenr;
1337         if (parent) {
1338                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1339                 key.offset = parent;
1340         } else {
1341                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1342                 key.offset = root_objectid;
1343         }
1344
1345         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1346         if (ret > 0)
1347                 ret = -ENOENT;
1348 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1349         if (ret == -ENOENT && parent) {
1350                 btrfs_release_path(path);
1351                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1352                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1353                 if (ret > 0)
1354                         ret = -ENOENT;
1355         }
1356 #endif
1357         return ret;
1358 }
1359
1360 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1361                                           struct btrfs_root *root,
1362                                           struct btrfs_path *path,
1363                                           u64 bytenr, u64 parent,
1364                                           u64 root_objectid)
1365 {
1366         struct btrfs_key key;
1367         int ret;
1368
1369         key.objectid = bytenr;
1370         if (parent) {
1371                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1372                 key.offset = parent;
1373         } else {
1374                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1375                 key.offset = root_objectid;
1376         }
1377
1378         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1379         btrfs_release_path(path);
1380         return ret;
1381 }
1382
1383 static inline int extent_ref_type(u64 parent, u64 owner)
1384 {
1385         int type;
1386         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1387                 if (parent > 0)
1388                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1389                 else
1390                         type = BTRFS_TREE_BLOCK_REF_KEY;
1391         } else {
1392                 if (parent > 0)
1393                         type = BTRFS_SHARED_DATA_REF_KEY;
1394                 else
1395                         type = BTRFS_EXTENT_DATA_REF_KEY;
1396         }
1397         return type;
1398 }
1399
1400 static int find_next_key(struct btrfs_path *path, int level,
1401                          struct btrfs_key *key)
1402
1403 {
1404         for (; level < BTRFS_MAX_LEVEL; level++) {
1405                 if (!path->nodes[level])
1406                         break;
1407                 if (path->slots[level] + 1 >=
1408                     btrfs_header_nritems(path->nodes[level]))
1409                         continue;
1410                 if (level == 0)
1411                         btrfs_item_key_to_cpu(path->nodes[level], key,
1412                                               path->slots[level] + 1);
1413                 else
1414                         btrfs_node_key_to_cpu(path->nodes[level], key,
1415                                               path->slots[level] + 1);
1416                 return 0;
1417         }
1418         return 1;
1419 }
1420
1421 /*
1422  * look for inline back ref. if back ref is found, *ref_ret is set
1423  * to the address of inline back ref, and 0 is returned.
1424  *
1425  * if back ref isn't found, *ref_ret is set to the address where it
1426  * should be inserted, and -ENOENT is returned.
1427  *
1428  * if insert is true and there are too many inline back refs, the path
1429  * points to the extent item, and -EAGAIN is returned.
1430  *
1431  * NOTE: inline back refs are ordered in the same way that back ref
1432  *       items in the tree are ordered.
1433  */
1434 static noinline_for_stack
1435 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1436                                  struct btrfs_root *root,
1437                                  struct btrfs_path *path,
1438                                  struct btrfs_extent_inline_ref **ref_ret,
1439                                  u64 bytenr, u64 num_bytes,
1440                                  u64 parent, u64 root_objectid,
1441                                  u64 owner, u64 offset, int insert)
1442 {
1443         struct btrfs_key key;
1444         struct extent_buffer *leaf;
1445         struct btrfs_extent_item *ei;
1446         struct btrfs_extent_inline_ref *iref;
1447         u64 flags;
1448         u64 item_size;
1449         unsigned long ptr;
1450         unsigned long end;
1451         int extra_size;
1452         int type;
1453         int want;
1454         int ret;
1455         int err = 0;
1456
1457         key.objectid = bytenr;
1458         key.type = BTRFS_EXTENT_ITEM_KEY;
1459         key.offset = num_bytes;
1460
1461         want = extent_ref_type(parent, owner);
1462         if (insert) {
1463                 extra_size = btrfs_extent_inline_ref_size(want);
1464                 path->keep_locks = 1;
1465         } else
1466                 extra_size = -1;
1467         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1468         if (ret < 0) {
1469                 err = ret;
1470                 goto out;
1471         }
1472         if (ret && !insert) {
1473                 err = -ENOENT;
1474                 goto out;
1475         } else if (ret) {
1476                 err = -EIO;
1477                 WARN_ON(1);
1478                 goto out;
1479         }
1480
1481         leaf = path->nodes[0];
1482         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1483 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1484         if (item_size < sizeof(*ei)) {
1485                 if (!insert) {
1486                         err = -ENOENT;
1487                         goto out;
1488                 }
1489                 ret = convert_extent_item_v0(trans, root, path, owner,
1490                                              extra_size);
1491                 if (ret < 0) {
1492                         err = ret;
1493                         goto out;
1494                 }
1495                 leaf = path->nodes[0];
1496                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1497         }
1498 #endif
1499         BUG_ON(item_size < sizeof(*ei));
1500
1501         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1502         flags = btrfs_extent_flags(leaf, ei);
1503
1504         ptr = (unsigned long)(ei + 1);
1505         end = (unsigned long)ei + item_size;
1506
1507         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1508                 ptr += sizeof(struct btrfs_tree_block_info);
1509                 BUG_ON(ptr > end);
1510         } else {
1511                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1512         }
1513
1514         err = -ENOENT;
1515         while (1) {
1516                 if (ptr >= end) {
1517                         WARN_ON(ptr > end);
1518                         break;
1519                 }
1520                 iref = (struct btrfs_extent_inline_ref *)ptr;
1521                 type = btrfs_extent_inline_ref_type(leaf, iref);
1522                 if (want < type)
1523                         break;
1524                 if (want > type) {
1525                         ptr += btrfs_extent_inline_ref_size(type);
1526                         continue;
1527                 }
1528
1529                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1530                         struct btrfs_extent_data_ref *dref;
1531                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1532                         if (match_extent_data_ref(leaf, dref, root_objectid,
1533                                                   owner, offset)) {
1534                                 err = 0;
1535                                 break;
1536                         }
1537                         if (hash_extent_data_ref_item(leaf, dref) <
1538                             hash_extent_data_ref(root_objectid, owner, offset))
1539                                 break;
1540                 } else {
1541                         u64 ref_offset;
1542                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1543                         if (parent > 0) {
1544                                 if (parent == ref_offset) {
1545                                         err = 0;
1546                                         break;
1547                                 }
1548                                 if (ref_offset < parent)
1549                                         break;
1550                         } else {
1551                                 if (root_objectid == ref_offset) {
1552                                         err = 0;
1553                                         break;
1554                                 }
1555                                 if (ref_offset < root_objectid)
1556                                         break;
1557                         }
1558                 }
1559                 ptr += btrfs_extent_inline_ref_size(type);
1560         }
1561         if (err == -ENOENT && insert) {
1562                 if (item_size + extra_size >=
1563                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1564                         err = -EAGAIN;
1565                         goto out;
1566                 }
1567                 /*
1568                  * To add new inline back ref, we have to make sure
1569                  * there is no corresponding back ref item.
1570                  * For simplicity, we just do not add new inline back
1571                  * ref if there is any kind of item for this block
1572                  */
1573                 if (find_next_key(path, 0, &key) == 0 &&
1574                     key.objectid == bytenr &&
1575                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1576                         err = -EAGAIN;
1577                         goto out;
1578                 }
1579         }
1580         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1581 out:
1582         if (insert) {
1583                 path->keep_locks = 0;
1584                 btrfs_unlock_up_safe(path, 1);
1585         }
1586         return err;
1587 }
1588
1589 /*
1590  * helper to add new inline back ref
1591  */
1592 static noinline_for_stack
1593 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1594                                  struct btrfs_root *root,
1595                                  struct btrfs_path *path,
1596                                  struct btrfs_extent_inline_ref *iref,
1597                                  u64 parent, u64 root_objectid,
1598                                  u64 owner, u64 offset, int refs_to_add,
1599                                  struct btrfs_delayed_extent_op *extent_op)
1600 {
1601         struct extent_buffer *leaf;
1602         struct btrfs_extent_item *ei;
1603         unsigned long ptr;
1604         unsigned long end;
1605         unsigned long item_offset;
1606         u64 refs;
1607         int size;
1608         int type;
1609
1610         leaf = path->nodes[0];
1611         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1612         item_offset = (unsigned long)iref - (unsigned long)ei;
1613
1614         type = extent_ref_type(parent, owner);
1615         size = btrfs_extent_inline_ref_size(type);
1616
1617         btrfs_extend_item(trans, root, path, size);
1618
1619         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1620         refs = btrfs_extent_refs(leaf, ei);
1621         refs += refs_to_add;
1622         btrfs_set_extent_refs(leaf, ei, refs);
1623         if (extent_op)
1624                 __run_delayed_extent_op(extent_op, leaf, ei);
1625
1626         ptr = (unsigned long)ei + item_offset;
1627         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1628         if (ptr < end - size)
1629                 memmove_extent_buffer(leaf, ptr + size, ptr,
1630                                       end - size - ptr);
1631
1632         iref = (struct btrfs_extent_inline_ref *)ptr;
1633         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1634         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1635                 struct btrfs_extent_data_ref *dref;
1636                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1637                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1638                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1639                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1640                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1641         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1642                 struct btrfs_shared_data_ref *sref;
1643                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1644                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1645                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1646         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1647                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1648         } else {
1649                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1650         }
1651         btrfs_mark_buffer_dirty(leaf);
1652 }
1653
1654 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1655                                  struct btrfs_root *root,
1656                                  struct btrfs_path *path,
1657                                  struct btrfs_extent_inline_ref **ref_ret,
1658                                  u64 bytenr, u64 num_bytes, u64 parent,
1659                                  u64 root_objectid, u64 owner, u64 offset)
1660 {
1661         int ret;
1662
1663         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1664                                            bytenr, num_bytes, parent,
1665                                            root_objectid, owner, offset, 0);
1666         if (ret != -ENOENT)
1667                 return ret;
1668
1669         btrfs_release_path(path);
1670         *ref_ret = NULL;
1671
1672         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1673                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1674                                             root_objectid);
1675         } else {
1676                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1677                                              root_objectid, owner, offset);
1678         }
1679         return ret;
1680 }
1681
1682 /*
1683  * helper to update/remove inline back ref
1684  */
1685 static noinline_for_stack
1686 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1687                                   struct btrfs_root *root,
1688                                   struct btrfs_path *path,
1689                                   struct btrfs_extent_inline_ref *iref,
1690                                   int refs_to_mod,
1691                                   struct btrfs_delayed_extent_op *extent_op)
1692 {
1693         struct extent_buffer *leaf;
1694         struct btrfs_extent_item *ei;
1695         struct btrfs_extent_data_ref *dref = NULL;
1696         struct btrfs_shared_data_ref *sref = NULL;
1697         unsigned long ptr;
1698         unsigned long end;
1699         u32 item_size;
1700         int size;
1701         int type;
1702         u64 refs;
1703
1704         leaf = path->nodes[0];
1705         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1706         refs = btrfs_extent_refs(leaf, ei);
1707         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1708         refs += refs_to_mod;
1709         btrfs_set_extent_refs(leaf, ei, refs);
1710         if (extent_op)
1711                 __run_delayed_extent_op(extent_op, leaf, ei);
1712
1713         type = btrfs_extent_inline_ref_type(leaf, iref);
1714
1715         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1716                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1717                 refs = btrfs_extent_data_ref_count(leaf, dref);
1718         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1719                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1720                 refs = btrfs_shared_data_ref_count(leaf, sref);
1721         } else {
1722                 refs = 1;
1723                 BUG_ON(refs_to_mod != -1);
1724         }
1725
1726         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1727         refs += refs_to_mod;
1728
1729         if (refs > 0) {
1730                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1731                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1732                 else
1733                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1734         } else {
1735                 size =  btrfs_extent_inline_ref_size(type);
1736                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1737                 ptr = (unsigned long)iref;
1738                 end = (unsigned long)ei + item_size;
1739                 if (ptr + size < end)
1740                         memmove_extent_buffer(leaf, ptr, ptr + size,
1741                                               end - ptr - size);
1742                 item_size -= size;
1743                 btrfs_truncate_item(trans, root, path, item_size, 1);
1744         }
1745         btrfs_mark_buffer_dirty(leaf);
1746 }
1747
1748 static noinline_for_stack
1749 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1750                                  struct btrfs_root *root,
1751                                  struct btrfs_path *path,
1752                                  u64 bytenr, u64 num_bytes, u64 parent,
1753                                  u64 root_objectid, u64 owner,
1754                                  u64 offset, int refs_to_add,
1755                                  struct btrfs_delayed_extent_op *extent_op)
1756 {
1757         struct btrfs_extent_inline_ref *iref;
1758         int ret;
1759
1760         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1761                                            bytenr, num_bytes, parent,
1762                                            root_objectid, owner, offset, 1);
1763         if (ret == 0) {
1764                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1765                 update_inline_extent_backref(trans, root, path, iref,
1766                                              refs_to_add, extent_op);
1767         } else if (ret == -ENOENT) {
1768                 setup_inline_extent_backref(trans, root, path, iref, parent,
1769                                             root_objectid, owner, offset,
1770                                             refs_to_add, extent_op);
1771                 ret = 0;
1772         }
1773         return ret;
1774 }
1775
1776 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1777                                  struct btrfs_root *root,
1778                                  struct btrfs_path *path,
1779                                  u64 bytenr, u64 parent, u64 root_objectid,
1780                                  u64 owner, u64 offset, int refs_to_add)
1781 {
1782         int ret;
1783         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1784                 BUG_ON(refs_to_add != 1);
1785                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1786                                             parent, root_objectid);
1787         } else {
1788                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1789                                              parent, root_objectid,
1790                                              owner, offset, refs_to_add);
1791         }
1792         return ret;
1793 }
1794
1795 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1796                                  struct btrfs_root *root,
1797                                  struct btrfs_path *path,
1798                                  struct btrfs_extent_inline_ref *iref,
1799                                  int refs_to_drop, int is_data)
1800 {
1801         int ret = 0;
1802
1803         BUG_ON(!is_data && refs_to_drop != 1);
1804         if (iref) {
1805                 update_inline_extent_backref(trans, root, path, iref,
1806                                              -refs_to_drop, NULL);
1807         } else if (is_data) {
1808                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1809         } else {
1810                 ret = btrfs_del_item(trans, root, path);
1811         }
1812         return ret;
1813 }
1814
1815 static int btrfs_issue_discard(struct block_device *bdev,
1816                                 u64 start, u64 len)
1817 {
1818         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1819 }
1820
1821 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1822                                 u64 num_bytes, u64 *actual_bytes)
1823 {
1824         int ret;
1825         u64 discarded_bytes = 0;
1826         struct btrfs_bio *bbio = NULL;
1827
1828
1829         /* Tell the block device(s) that the sectors can be discarded */
1830         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1831                               bytenr, &num_bytes, &bbio, 0);
1832         /* Error condition is -ENOMEM */
1833         if (!ret) {
1834                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1835                 int i;
1836
1837
1838                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1839                         if (!stripe->dev->can_discard)
1840                                 continue;
1841
1842                         ret = btrfs_issue_discard(stripe->dev->bdev,
1843                                                   stripe->physical,
1844                                                   stripe->length);
1845                         if (!ret)
1846                                 discarded_bytes += stripe->length;
1847                         else if (ret != -EOPNOTSUPP)
1848                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1849
1850                         /*
1851                          * Just in case we get back EOPNOTSUPP for some reason,
1852                          * just ignore the return value so we don't screw up
1853                          * people calling discard_extent.
1854                          */
1855                         ret = 0;
1856                 }
1857                 kfree(bbio);
1858         }
1859
1860         if (actual_bytes)
1861                 *actual_bytes = discarded_bytes;
1862
1863
1864         if (ret == -EOPNOTSUPP)
1865                 ret = 0;
1866         return ret;
1867 }
1868
1869 /* Can return -ENOMEM */
1870 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1871                          struct btrfs_root *root,
1872                          u64 bytenr, u64 num_bytes, u64 parent,
1873                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1874 {
1875         int ret;
1876         struct btrfs_fs_info *fs_info = root->fs_info;
1877
1878         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1879                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1880
1881         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1882                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1883                                         num_bytes,
1884                                         parent, root_objectid, (int)owner,
1885                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1886         } else {
1887                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1888                                         num_bytes,
1889                                         parent, root_objectid, owner, offset,
1890                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1891         }
1892         return ret;
1893 }
1894
1895 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1896                                   struct btrfs_root *root,
1897                                   u64 bytenr, u64 num_bytes,
1898                                   u64 parent, u64 root_objectid,
1899                                   u64 owner, u64 offset, int refs_to_add,
1900                                   struct btrfs_delayed_extent_op *extent_op)
1901 {
1902         struct btrfs_path *path;
1903         struct extent_buffer *leaf;
1904         struct btrfs_extent_item *item;
1905         u64 refs;
1906         int ret;
1907         int err = 0;
1908
1909         path = btrfs_alloc_path();
1910         if (!path)
1911                 return -ENOMEM;
1912
1913         path->reada = 1;
1914         path->leave_spinning = 1;
1915         /* this will setup the path even if it fails to insert the back ref */
1916         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1917                                            path, bytenr, num_bytes, parent,
1918                                            root_objectid, owner, offset,
1919                                            refs_to_add, extent_op);
1920         if (ret == 0)
1921                 goto out;
1922
1923         if (ret != -EAGAIN) {
1924                 err = ret;
1925                 goto out;
1926         }
1927
1928         leaf = path->nodes[0];
1929         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1930         refs = btrfs_extent_refs(leaf, item);
1931         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1932         if (extent_op)
1933                 __run_delayed_extent_op(extent_op, leaf, item);
1934
1935         btrfs_mark_buffer_dirty(leaf);
1936         btrfs_release_path(path);
1937
1938         path->reada = 1;
1939         path->leave_spinning = 1;
1940
1941         /* now insert the actual backref */
1942         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1943                                     path, bytenr, parent, root_objectid,
1944                                     owner, offset, refs_to_add);
1945         if (ret)
1946                 btrfs_abort_transaction(trans, root, ret);
1947 out:
1948         btrfs_free_path(path);
1949         return err;
1950 }
1951
1952 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1953                                 struct btrfs_root *root,
1954                                 struct btrfs_delayed_ref_node *node,
1955                                 struct btrfs_delayed_extent_op *extent_op,
1956                                 int insert_reserved)
1957 {
1958         int ret = 0;
1959         struct btrfs_delayed_data_ref *ref;
1960         struct btrfs_key ins;
1961         u64 parent = 0;
1962         u64 ref_root = 0;
1963         u64 flags = 0;
1964
1965         ins.objectid = node->bytenr;
1966         ins.offset = node->num_bytes;
1967         ins.type = BTRFS_EXTENT_ITEM_KEY;
1968
1969         ref = btrfs_delayed_node_to_data_ref(node);
1970         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1971                 parent = ref->parent;
1972         else
1973                 ref_root = ref->root;
1974
1975         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1976                 if (extent_op) {
1977                         BUG_ON(extent_op->update_key);
1978                         flags |= extent_op->flags_to_set;
1979                 }
1980                 ret = alloc_reserved_file_extent(trans, root,
1981                                                  parent, ref_root, flags,
1982                                                  ref->objectid, ref->offset,
1983                                                  &ins, node->ref_mod);
1984         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1985                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1986                                              node->num_bytes, parent,
1987                                              ref_root, ref->objectid,
1988                                              ref->offset, node->ref_mod,
1989                                              extent_op);
1990         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1991                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1992                                           node->num_bytes, parent,
1993                                           ref_root, ref->objectid,
1994                                           ref->offset, node->ref_mod,
1995                                           extent_op);
1996         } else {
1997                 BUG();
1998         }
1999         return ret;
2000 }
2001
2002 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2003                                     struct extent_buffer *leaf,
2004                                     struct btrfs_extent_item *ei)
2005 {
2006         u64 flags = btrfs_extent_flags(leaf, ei);
2007         if (extent_op->update_flags) {
2008                 flags |= extent_op->flags_to_set;
2009                 btrfs_set_extent_flags(leaf, ei, flags);
2010         }
2011
2012         if (extent_op->update_key) {
2013                 struct btrfs_tree_block_info *bi;
2014                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2015                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2016                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2017         }
2018 }
2019
2020 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2021                                  struct btrfs_root *root,
2022                                  struct btrfs_delayed_ref_node *node,
2023                                  struct btrfs_delayed_extent_op *extent_op)
2024 {
2025         struct btrfs_key key;
2026         struct btrfs_path *path;
2027         struct btrfs_extent_item *ei;
2028         struct extent_buffer *leaf;
2029         u32 item_size;
2030         int ret;
2031         int err = 0;
2032
2033         if (trans->aborted)
2034                 return 0;
2035
2036         path = btrfs_alloc_path();
2037         if (!path)
2038                 return -ENOMEM;
2039
2040         key.objectid = node->bytenr;
2041         key.type = BTRFS_EXTENT_ITEM_KEY;
2042         key.offset = node->num_bytes;
2043
2044         path->reada = 1;
2045         path->leave_spinning = 1;
2046         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2047                                 path, 0, 1);
2048         if (ret < 0) {
2049                 err = ret;
2050                 goto out;
2051         }
2052         if (ret > 0) {
2053                 err = -EIO;
2054                 goto out;
2055         }
2056
2057         leaf = path->nodes[0];
2058         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2059 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2060         if (item_size < sizeof(*ei)) {
2061                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2062                                              path, (u64)-1, 0);
2063                 if (ret < 0) {
2064                         err = ret;
2065                         goto out;
2066                 }
2067                 leaf = path->nodes[0];
2068                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2069         }
2070 #endif
2071         BUG_ON(item_size < sizeof(*ei));
2072         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2073         __run_delayed_extent_op(extent_op, leaf, ei);
2074
2075         btrfs_mark_buffer_dirty(leaf);
2076 out:
2077         btrfs_free_path(path);
2078         return err;
2079 }
2080
2081 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2082                                 struct btrfs_root *root,
2083                                 struct btrfs_delayed_ref_node *node,
2084                                 struct btrfs_delayed_extent_op *extent_op,
2085                                 int insert_reserved)
2086 {
2087         int ret = 0;
2088         struct btrfs_delayed_tree_ref *ref;
2089         struct btrfs_key ins;
2090         u64 parent = 0;
2091         u64 ref_root = 0;
2092
2093         ins.objectid = node->bytenr;
2094         ins.offset = node->num_bytes;
2095         ins.type = BTRFS_EXTENT_ITEM_KEY;
2096
2097         ref = btrfs_delayed_node_to_tree_ref(node);
2098         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2099                 parent = ref->parent;
2100         else
2101                 ref_root = ref->root;
2102
2103         BUG_ON(node->ref_mod != 1);
2104         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2105                 BUG_ON(!extent_op || !extent_op->update_flags ||
2106                        !extent_op->update_key);
2107                 ret = alloc_reserved_tree_block(trans, root,
2108                                                 parent, ref_root,
2109                                                 extent_op->flags_to_set,
2110                                                 &extent_op->key,
2111                                                 ref->level, &ins);
2112         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2113                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2114                                              node->num_bytes, parent, ref_root,
2115                                              ref->level, 0, 1, extent_op);
2116         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2117                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2118                                           node->num_bytes, parent, ref_root,
2119                                           ref->level, 0, 1, extent_op);
2120         } else {
2121                 BUG();
2122         }
2123         return ret;
2124 }
2125
2126 /* helper function to actually process a single delayed ref entry */
2127 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2128                                struct btrfs_root *root,
2129                                struct btrfs_delayed_ref_node *node,
2130                                struct btrfs_delayed_extent_op *extent_op,
2131                                int insert_reserved)
2132 {
2133         int ret = 0;
2134
2135         if (trans->aborted)
2136                 return 0;
2137
2138         if (btrfs_delayed_ref_is_head(node)) {
2139                 struct btrfs_delayed_ref_head *head;
2140                 /*
2141                  * we've hit the end of the chain and we were supposed
2142                  * to insert this extent into the tree.  But, it got
2143                  * deleted before we ever needed to insert it, so all
2144                  * we have to do is clean up the accounting
2145                  */
2146                 BUG_ON(extent_op);
2147                 head = btrfs_delayed_node_to_head(node);
2148                 if (insert_reserved) {
2149                         btrfs_pin_extent(root, node->bytenr,
2150                                          node->num_bytes, 1);
2151                         if (head->is_data) {
2152                                 ret = btrfs_del_csums(trans, root,
2153                                                       node->bytenr,
2154                                                       node->num_bytes);
2155                         }
2156                 }
2157                 return ret;
2158         }
2159
2160         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2161             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2162                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2163                                            insert_reserved);
2164         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2165                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2166                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2167                                            insert_reserved);
2168         else
2169                 BUG();
2170         return ret;
2171 }
2172
2173 static noinline struct btrfs_delayed_ref_node *
2174 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2175 {
2176         struct rb_node *node;
2177         struct btrfs_delayed_ref_node *ref;
2178         int action = BTRFS_ADD_DELAYED_REF;
2179 again:
2180         /*
2181          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2182          * this prevents ref count from going down to zero when
2183          * there still are pending delayed ref.
2184          */
2185         node = rb_prev(&head->node.rb_node);
2186         while (1) {
2187                 if (!node)
2188                         break;
2189                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2190                                 rb_node);
2191                 if (ref->bytenr != head->node.bytenr)
2192                         break;
2193                 if (ref->action == action)
2194                         return ref;
2195                 node = rb_prev(node);
2196         }
2197         if (action == BTRFS_ADD_DELAYED_REF) {
2198                 action = BTRFS_DROP_DELAYED_REF;
2199                 goto again;
2200         }
2201         return NULL;
2202 }
2203
2204 /*
2205  * Returns 0 on success or if called with an already aborted transaction.
2206  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2207  */
2208 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2209                                        struct btrfs_root *root,
2210                                        struct list_head *cluster)
2211 {
2212         struct btrfs_delayed_ref_root *delayed_refs;
2213         struct btrfs_delayed_ref_node *ref;
2214         struct btrfs_delayed_ref_head *locked_ref = NULL;
2215         struct btrfs_delayed_extent_op *extent_op;
2216         struct btrfs_fs_info *fs_info = root->fs_info;
2217         int ret;
2218         int count = 0;
2219         int must_insert_reserved = 0;
2220
2221         delayed_refs = &trans->transaction->delayed_refs;
2222         while (1) {
2223                 if (!locked_ref) {
2224                         /* pick a new head ref from the cluster list */
2225                         if (list_empty(cluster))
2226                                 break;
2227
2228                         locked_ref = list_entry(cluster->next,
2229                                      struct btrfs_delayed_ref_head, cluster);
2230
2231                         /* grab the lock that says we are going to process
2232                          * all the refs for this head */
2233                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2234
2235                         /*
2236                          * we may have dropped the spin lock to get the head
2237                          * mutex lock, and that might have given someone else
2238                          * time to free the head.  If that's true, it has been
2239                          * removed from our list and we can move on.
2240                          */
2241                         if (ret == -EAGAIN) {
2242                                 locked_ref = NULL;
2243                                 count++;
2244                                 continue;
2245                         }
2246                 }
2247
2248                 /*
2249                  * We need to try and merge add/drops of the same ref since we
2250                  * can run into issues with relocate dropping the implicit ref
2251                  * and then it being added back again before the drop can
2252                  * finish.  If we merged anything we need to re-loop so we can
2253                  * get a good ref.
2254                  */
2255                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2256                                          locked_ref);
2257
2258                 /*
2259                  * locked_ref is the head node, so we have to go one
2260                  * node back for any delayed ref updates
2261                  */
2262                 ref = select_delayed_ref(locked_ref);
2263
2264                 if (ref && ref->seq &&
2265                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2266                         /*
2267                          * there are still refs with lower seq numbers in the
2268                          * process of being added. Don't run this ref yet.
2269                          */
2270                         list_del_init(&locked_ref->cluster);
2271                         btrfs_delayed_ref_unlock(locked_ref);
2272                         locked_ref = NULL;
2273                         delayed_refs->num_heads_ready++;
2274                         spin_unlock(&delayed_refs->lock);
2275                         cond_resched();
2276                         spin_lock(&delayed_refs->lock);
2277                         continue;
2278                 }
2279
2280                 /*
2281                  * record the must insert reserved flag before we
2282                  * drop the spin lock.
2283                  */
2284                 must_insert_reserved = locked_ref->must_insert_reserved;
2285                 locked_ref->must_insert_reserved = 0;
2286
2287                 extent_op = locked_ref->extent_op;
2288                 locked_ref->extent_op = NULL;
2289
2290                 if (!ref) {
2291                         /* All delayed refs have been processed, Go ahead
2292                          * and send the head node to run_one_delayed_ref,
2293                          * so that any accounting fixes can happen
2294                          */
2295                         ref = &locked_ref->node;
2296
2297                         if (extent_op && must_insert_reserved) {
2298                                 btrfs_free_delayed_extent_op(extent_op);
2299                                 extent_op = NULL;
2300                         }
2301
2302                         if (extent_op) {
2303                                 spin_unlock(&delayed_refs->lock);
2304
2305                                 ret = run_delayed_extent_op(trans, root,
2306                                                             ref, extent_op);
2307                                 btrfs_free_delayed_extent_op(extent_op);
2308
2309                                 if (ret) {
2310                                         printk(KERN_DEBUG
2311                                                "btrfs: run_delayed_extent_op "
2312                                                "returned %d\n", ret);
2313                                         spin_lock(&delayed_refs->lock);
2314                                         btrfs_delayed_ref_unlock(locked_ref);
2315                                         return ret;
2316                                 }
2317
2318                                 goto next;
2319                         }
2320                 }
2321
2322                 ref->in_tree = 0;
2323                 rb_erase(&ref->rb_node, &delayed_refs->root);
2324                 delayed_refs->num_entries--;
2325                 if (!btrfs_delayed_ref_is_head(ref)) {
2326                         /*
2327                          * when we play the delayed ref, also correct the
2328                          * ref_mod on head
2329                          */
2330                         switch (ref->action) {
2331                         case BTRFS_ADD_DELAYED_REF:
2332                         case BTRFS_ADD_DELAYED_EXTENT:
2333                                 locked_ref->node.ref_mod -= ref->ref_mod;
2334                                 break;
2335                         case BTRFS_DROP_DELAYED_REF:
2336                                 locked_ref->node.ref_mod += ref->ref_mod;
2337                                 break;
2338                         default:
2339                                 WARN_ON(1);
2340                         }
2341                 }
2342                 spin_unlock(&delayed_refs->lock);
2343
2344                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2345                                           must_insert_reserved);
2346
2347                 btrfs_free_delayed_extent_op(extent_op);
2348                 if (ret) {
2349                         btrfs_delayed_ref_unlock(locked_ref);
2350                         btrfs_put_delayed_ref(ref);
2351                         printk(KERN_DEBUG
2352                                "btrfs: run_one_delayed_ref returned %d\n", ret);
2353                         spin_lock(&delayed_refs->lock);
2354                         return ret;
2355                 }
2356
2357                 /*
2358                  * If this node is a head, that means all the refs in this head
2359                  * have been dealt with, and we will pick the next head to deal
2360                  * with, so we must unlock the head and drop it from the cluster
2361                  * list before we release it.
2362                  */
2363                 if (btrfs_delayed_ref_is_head(ref)) {
2364                         list_del_init(&locked_ref->cluster);
2365                         btrfs_delayed_ref_unlock(locked_ref);
2366                         locked_ref = NULL;
2367                 }
2368                 btrfs_put_delayed_ref(ref);
2369                 count++;
2370 next:
2371                 cond_resched();
2372                 spin_lock(&delayed_refs->lock);
2373         }
2374         return count;
2375 }
2376
2377 #ifdef SCRAMBLE_DELAYED_REFS
2378 /*
2379  * Normally delayed refs get processed in ascending bytenr order. This
2380  * correlates in most cases to the order added. To expose dependencies on this
2381  * order, we start to process the tree in the middle instead of the beginning
2382  */
2383 static u64 find_middle(struct rb_root *root)
2384 {
2385         struct rb_node *n = root->rb_node;
2386         struct btrfs_delayed_ref_node *entry;
2387         int alt = 1;
2388         u64 middle;
2389         u64 first = 0, last = 0;
2390
2391         n = rb_first(root);
2392         if (n) {
2393                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2394                 first = entry->bytenr;
2395         }
2396         n = rb_last(root);
2397         if (n) {
2398                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2399                 last = entry->bytenr;
2400         }
2401         n = root->rb_node;
2402
2403         while (n) {
2404                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2405                 WARN_ON(!entry->in_tree);
2406
2407                 middle = entry->bytenr;
2408
2409                 if (alt)
2410                         n = n->rb_left;
2411                 else
2412                         n = n->rb_right;
2413
2414                 alt = 1 - alt;
2415         }
2416         return middle;
2417 }
2418 #endif
2419
2420 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2421                                          struct btrfs_fs_info *fs_info)
2422 {
2423         struct qgroup_update *qgroup_update;
2424         int ret = 0;
2425
2426         if (list_empty(&trans->qgroup_ref_list) !=
2427             !trans->delayed_ref_elem.seq) {
2428                 /* list without seq or seq without list */
2429                 printk(KERN_ERR "btrfs: qgroup accounting update error, list is%s empty, seq is %llu\n",
2430                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2431                         trans->delayed_ref_elem.seq);
2432                 BUG();
2433         }
2434
2435         if (!trans->delayed_ref_elem.seq)
2436                 return 0;
2437
2438         while (!list_empty(&trans->qgroup_ref_list)) {
2439                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2440                                                  struct qgroup_update, list);
2441                 list_del(&qgroup_update->list);
2442                 if (!ret)
2443                         ret = btrfs_qgroup_account_ref(
2444                                         trans, fs_info, qgroup_update->node,
2445                                         qgroup_update->extent_op);
2446                 kfree(qgroup_update);
2447         }
2448
2449         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2450
2451         return ret;
2452 }
2453
2454 static int refs_newer(struct btrfs_delayed_ref_root *delayed_refs, int seq,
2455                       int count)
2456 {
2457         int val = atomic_read(&delayed_refs->ref_seq);
2458
2459         if (val < seq || val >= seq + count)
2460                 return 1;
2461         return 0;
2462 }
2463
2464 /*
2465  * this starts processing the delayed reference count updates and
2466  * extent insertions we have queued up so far.  count can be
2467  * 0, which means to process everything in the tree at the start
2468  * of the run (but not newly added entries), or it can be some target
2469  * number you'd like to process.
2470  *
2471  * Returns 0 on success or if called with an aborted transaction
2472  * Returns <0 on error and aborts the transaction
2473  */
2474 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2475                            struct btrfs_root *root, unsigned long count)
2476 {
2477         struct rb_node *node;
2478         struct btrfs_delayed_ref_root *delayed_refs;
2479         struct btrfs_delayed_ref_node *ref;
2480         struct list_head cluster;
2481         int ret;
2482         u64 delayed_start;
2483         int run_all = count == (unsigned long)-1;
2484         int run_most = 0;
2485         int loops;
2486
2487         /* We'll clean this up in btrfs_cleanup_transaction */
2488         if (trans->aborted)
2489                 return 0;
2490
2491         if (root == root->fs_info->extent_root)
2492                 root = root->fs_info->tree_root;
2493
2494         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2495
2496         delayed_refs = &trans->transaction->delayed_refs;
2497         INIT_LIST_HEAD(&cluster);
2498         if (count == 0) {
2499                 count = delayed_refs->num_entries * 2;
2500                 run_most = 1;
2501         }
2502
2503         if (!run_all && !run_most) {
2504                 int old;
2505                 int seq = atomic_read(&delayed_refs->ref_seq);
2506
2507 progress:
2508                 old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2509                 if (old) {
2510                         DEFINE_WAIT(__wait);
2511                         if (delayed_refs->num_entries < 16348)
2512                                 return 0;
2513
2514                         prepare_to_wait(&delayed_refs->wait, &__wait,
2515                                         TASK_UNINTERRUPTIBLE);
2516
2517                         old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2518                         if (old) {
2519                                 schedule();
2520                                 finish_wait(&delayed_refs->wait, &__wait);
2521
2522                                 if (!refs_newer(delayed_refs, seq, 256))
2523                                         goto progress;
2524                                 else
2525                                         return 0;
2526                         } else {
2527                                 finish_wait(&delayed_refs->wait, &__wait);
2528                                 goto again;
2529                         }
2530                 }
2531
2532         } else {
2533                 atomic_inc(&delayed_refs->procs_running_refs);
2534         }
2535
2536 again:
2537         loops = 0;
2538         spin_lock(&delayed_refs->lock);
2539
2540 #ifdef SCRAMBLE_DELAYED_REFS
2541         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2542 #endif
2543
2544         while (1) {
2545                 if (!(run_all || run_most) &&
2546                     delayed_refs->num_heads_ready < 64)
2547                         break;
2548
2549                 /*
2550                  * go find something we can process in the rbtree.  We start at
2551                  * the beginning of the tree, and then build a cluster
2552                  * of refs to process starting at the first one we are able to
2553                  * lock
2554                  */
2555                 delayed_start = delayed_refs->run_delayed_start;
2556                 ret = btrfs_find_ref_cluster(trans, &cluster,
2557                                              delayed_refs->run_delayed_start);
2558                 if (ret)
2559                         break;
2560
2561                 ret = run_clustered_refs(trans, root, &cluster);
2562                 if (ret < 0) {
2563                         btrfs_release_ref_cluster(&cluster);
2564                         spin_unlock(&delayed_refs->lock);
2565                         btrfs_abort_transaction(trans, root, ret);
2566                         atomic_dec(&delayed_refs->procs_running_refs);
2567                         return ret;
2568                 }
2569
2570                 atomic_add(ret, &delayed_refs->ref_seq);
2571
2572                 count -= min_t(unsigned long, ret, count);
2573
2574                 if (count == 0)
2575                         break;
2576
2577                 if (delayed_start >= delayed_refs->run_delayed_start) {
2578                         if (loops == 0) {
2579                                 /*
2580                                  * btrfs_find_ref_cluster looped. let's do one
2581                                  * more cycle. if we don't run any delayed ref
2582                                  * during that cycle (because we can't because
2583                                  * all of them are blocked), bail out.
2584                                  */
2585                                 loops = 1;
2586                         } else {
2587                                 /*
2588                                  * no runnable refs left, stop trying
2589                                  */
2590                                 BUG_ON(run_all);
2591                                 break;
2592                         }
2593                 }
2594                 if (ret) {
2595                         /* refs were run, let's reset staleness detection */
2596                         loops = 0;
2597                 }
2598         }
2599
2600         if (run_all) {
2601                 if (!list_empty(&trans->new_bgs)) {
2602                         spin_unlock(&delayed_refs->lock);
2603                         btrfs_create_pending_block_groups(trans, root);
2604                         spin_lock(&delayed_refs->lock);
2605                 }
2606
2607                 node = rb_first(&delayed_refs->root);
2608                 if (!node)
2609                         goto out;
2610                 count = (unsigned long)-1;
2611
2612                 while (node) {
2613                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2614                                        rb_node);
2615                         if (btrfs_delayed_ref_is_head(ref)) {
2616                                 struct btrfs_delayed_ref_head *head;
2617
2618                                 head = btrfs_delayed_node_to_head(ref);
2619                                 atomic_inc(&ref->refs);
2620
2621                                 spin_unlock(&delayed_refs->lock);
2622                                 /*
2623                                  * Mutex was contended, block until it's
2624                                  * released and try again
2625                                  */
2626                                 mutex_lock(&head->mutex);
2627                                 mutex_unlock(&head->mutex);
2628
2629                                 btrfs_put_delayed_ref(ref);
2630                                 cond_resched();
2631                                 goto again;
2632                         }
2633                         node = rb_next(node);
2634                 }
2635                 spin_unlock(&delayed_refs->lock);
2636                 schedule_timeout(1);
2637                 goto again;
2638         }
2639 out:
2640         atomic_dec(&delayed_refs->procs_running_refs);
2641         smp_mb();
2642         if (waitqueue_active(&delayed_refs->wait))
2643                 wake_up(&delayed_refs->wait);
2644
2645         spin_unlock(&delayed_refs->lock);
2646         assert_qgroups_uptodate(trans);
2647         return 0;
2648 }
2649
2650 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2651                                 struct btrfs_root *root,
2652                                 u64 bytenr, u64 num_bytes, u64 flags,
2653                                 int is_data)
2654 {
2655         struct btrfs_delayed_extent_op *extent_op;
2656         int ret;
2657
2658         extent_op = btrfs_alloc_delayed_extent_op();
2659         if (!extent_op)
2660                 return -ENOMEM;
2661
2662         extent_op->flags_to_set = flags;
2663         extent_op->update_flags = 1;
2664         extent_op->update_key = 0;
2665         extent_op->is_data = is_data ? 1 : 0;
2666
2667         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2668                                           num_bytes, extent_op);
2669         if (ret)
2670                 btrfs_free_delayed_extent_op(extent_op);
2671         return ret;
2672 }
2673
2674 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2675                                       struct btrfs_root *root,
2676                                       struct btrfs_path *path,
2677                                       u64 objectid, u64 offset, u64 bytenr)
2678 {
2679         struct btrfs_delayed_ref_head *head;
2680         struct btrfs_delayed_ref_node *ref;
2681         struct btrfs_delayed_data_ref *data_ref;
2682         struct btrfs_delayed_ref_root *delayed_refs;
2683         struct rb_node *node;
2684         int ret = 0;
2685
2686         ret = -ENOENT;
2687         delayed_refs = &trans->transaction->delayed_refs;
2688         spin_lock(&delayed_refs->lock);
2689         head = btrfs_find_delayed_ref_head(trans, bytenr);
2690         if (!head)
2691                 goto out;
2692
2693         if (!mutex_trylock(&head->mutex)) {
2694                 atomic_inc(&head->node.refs);
2695                 spin_unlock(&delayed_refs->lock);
2696
2697                 btrfs_release_path(path);
2698
2699                 /*
2700                  * Mutex was contended, block until it's released and let
2701                  * caller try again
2702                  */
2703                 mutex_lock(&head->mutex);
2704                 mutex_unlock(&head->mutex);
2705                 btrfs_put_delayed_ref(&head->node);
2706                 return -EAGAIN;
2707         }
2708
2709         node = rb_prev(&head->node.rb_node);
2710         if (!node)
2711                 goto out_unlock;
2712
2713         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2714
2715         if (ref->bytenr != bytenr)
2716                 goto out_unlock;
2717
2718         ret = 1;
2719         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2720                 goto out_unlock;
2721
2722         data_ref = btrfs_delayed_node_to_data_ref(ref);
2723
2724         node = rb_prev(node);
2725         if (node) {
2726                 int seq = ref->seq;
2727
2728                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2729                 if (ref->bytenr == bytenr && ref->seq == seq)
2730                         goto out_unlock;
2731         }
2732
2733         if (data_ref->root != root->root_key.objectid ||
2734             data_ref->objectid != objectid || data_ref->offset != offset)
2735                 goto out_unlock;
2736
2737         ret = 0;
2738 out_unlock:
2739         mutex_unlock(&head->mutex);
2740 out:
2741         spin_unlock(&delayed_refs->lock);
2742         return ret;
2743 }
2744
2745 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2746                                         struct btrfs_root *root,
2747                                         struct btrfs_path *path,
2748                                         u64 objectid, u64 offset, u64 bytenr)
2749 {
2750         struct btrfs_root *extent_root = root->fs_info->extent_root;
2751         struct extent_buffer *leaf;
2752         struct btrfs_extent_data_ref *ref;
2753         struct btrfs_extent_inline_ref *iref;
2754         struct btrfs_extent_item *ei;
2755         struct btrfs_key key;
2756         u32 item_size;
2757         int ret;
2758
2759         key.objectid = bytenr;
2760         key.offset = (u64)-1;
2761         key.type = BTRFS_EXTENT_ITEM_KEY;
2762
2763         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2764         if (ret < 0)
2765                 goto out;
2766         BUG_ON(ret == 0); /* Corruption */
2767
2768         ret = -ENOENT;
2769         if (path->slots[0] == 0)
2770                 goto out;
2771
2772         path->slots[0]--;
2773         leaf = path->nodes[0];
2774         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2775
2776         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2777                 goto out;
2778
2779         ret = 1;
2780         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2781 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2782         if (item_size < sizeof(*ei)) {
2783                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2784                 goto out;
2785         }
2786 #endif
2787         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2788
2789         if (item_size != sizeof(*ei) +
2790             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2791                 goto out;
2792
2793         if (btrfs_extent_generation(leaf, ei) <=
2794             btrfs_root_last_snapshot(&root->root_item))
2795                 goto out;
2796
2797         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2798         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2799             BTRFS_EXTENT_DATA_REF_KEY)
2800                 goto out;
2801
2802         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2803         if (btrfs_extent_refs(leaf, ei) !=
2804             btrfs_extent_data_ref_count(leaf, ref) ||
2805             btrfs_extent_data_ref_root(leaf, ref) !=
2806             root->root_key.objectid ||
2807             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2808             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2809                 goto out;
2810
2811         ret = 0;
2812 out:
2813         return ret;
2814 }
2815
2816 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2817                           struct btrfs_root *root,
2818                           u64 objectid, u64 offset, u64 bytenr)
2819 {
2820         struct btrfs_path *path;
2821         int ret;
2822         int ret2;
2823
2824         path = btrfs_alloc_path();
2825         if (!path)
2826                 return -ENOENT;
2827
2828         do {
2829                 ret = check_committed_ref(trans, root, path, objectid,
2830                                           offset, bytenr);
2831                 if (ret && ret != -ENOENT)
2832                         goto out;
2833
2834                 ret2 = check_delayed_ref(trans, root, path, objectid,
2835                                          offset, bytenr);
2836         } while (ret2 == -EAGAIN);
2837
2838         if (ret2 && ret2 != -ENOENT) {
2839                 ret = ret2;
2840                 goto out;
2841         }
2842
2843         if (ret != -ENOENT || ret2 != -ENOENT)
2844                 ret = 0;
2845 out:
2846         btrfs_free_path(path);
2847         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2848                 WARN_ON(ret > 0);
2849         return ret;
2850 }
2851
2852 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2853                            struct btrfs_root *root,
2854                            struct extent_buffer *buf,
2855                            int full_backref, int inc, int for_cow)
2856 {
2857         u64 bytenr;
2858         u64 num_bytes;
2859         u64 parent;
2860         u64 ref_root;
2861         u32 nritems;
2862         struct btrfs_key key;
2863         struct btrfs_file_extent_item *fi;
2864         int i;
2865         int level;
2866         int ret = 0;
2867         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2868                             u64, u64, u64, u64, u64, u64, int);
2869
2870         ref_root = btrfs_header_owner(buf);
2871         nritems = btrfs_header_nritems(buf);
2872         level = btrfs_header_level(buf);
2873
2874         if (!root->ref_cows && level == 0)
2875                 return 0;
2876
2877         if (inc)
2878                 process_func = btrfs_inc_extent_ref;
2879         else
2880                 process_func = btrfs_free_extent;
2881
2882         if (full_backref)
2883                 parent = buf->start;
2884         else
2885                 parent = 0;
2886
2887         for (i = 0; i < nritems; i++) {
2888                 if (level == 0) {
2889                         btrfs_item_key_to_cpu(buf, &key, i);
2890                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2891                                 continue;
2892                         fi = btrfs_item_ptr(buf, i,
2893                                             struct btrfs_file_extent_item);
2894                         if (btrfs_file_extent_type(buf, fi) ==
2895                             BTRFS_FILE_EXTENT_INLINE)
2896                                 continue;
2897                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2898                         if (bytenr == 0)
2899                                 continue;
2900
2901                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2902                         key.offset -= btrfs_file_extent_offset(buf, fi);
2903                         ret = process_func(trans, root, bytenr, num_bytes,
2904                                            parent, ref_root, key.objectid,
2905                                            key.offset, for_cow);
2906                         if (ret)
2907                                 goto fail;
2908                 } else {
2909                         bytenr = btrfs_node_blockptr(buf, i);
2910                         num_bytes = btrfs_level_size(root, level - 1);
2911                         ret = process_func(trans, root, bytenr, num_bytes,
2912                                            parent, ref_root, level - 1, 0,
2913                                            for_cow);
2914                         if (ret)
2915                                 goto fail;
2916                 }
2917         }
2918         return 0;
2919 fail:
2920         return ret;
2921 }
2922
2923 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2924                   struct extent_buffer *buf, int full_backref, int for_cow)
2925 {
2926         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2927 }
2928
2929 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2930                   struct extent_buffer *buf, int full_backref, int for_cow)
2931 {
2932         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2933 }
2934
2935 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2936                                  struct btrfs_root *root,
2937                                  struct btrfs_path *path,
2938                                  struct btrfs_block_group_cache *cache)
2939 {
2940         int ret;
2941         struct btrfs_root *extent_root = root->fs_info->extent_root;
2942         unsigned long bi;
2943         struct extent_buffer *leaf;
2944
2945         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2946         if (ret < 0)
2947                 goto fail;
2948         BUG_ON(ret); /* Corruption */
2949
2950         leaf = path->nodes[0];
2951         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2952         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2953         btrfs_mark_buffer_dirty(leaf);
2954         btrfs_release_path(path);
2955 fail:
2956         if (ret) {
2957                 btrfs_abort_transaction(trans, root, ret);
2958                 return ret;
2959         }
2960         return 0;
2961
2962 }
2963
2964 static struct btrfs_block_group_cache *
2965 next_block_group(struct btrfs_root *root,
2966                  struct btrfs_block_group_cache *cache)
2967 {
2968         struct rb_node *node;
2969         spin_lock(&root->fs_info->block_group_cache_lock);
2970         node = rb_next(&cache->cache_node);
2971         btrfs_put_block_group(cache);
2972         if (node) {
2973                 cache = rb_entry(node, struct btrfs_block_group_cache,
2974                                  cache_node);
2975                 btrfs_get_block_group(cache);
2976         } else
2977                 cache = NULL;
2978         spin_unlock(&root->fs_info->block_group_cache_lock);
2979         return cache;
2980 }
2981
2982 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2983                             struct btrfs_trans_handle *trans,
2984                             struct btrfs_path *path)
2985 {
2986         struct btrfs_root *root = block_group->fs_info->tree_root;
2987         struct inode *inode = NULL;
2988         u64 alloc_hint = 0;
2989         int dcs = BTRFS_DC_ERROR;
2990         int num_pages = 0;
2991         int retries = 0;
2992         int ret = 0;
2993
2994         /*
2995          * If this block group is smaller than 100 megs don't bother caching the
2996          * block group.
2997          */
2998         if (block_group->key.offset < (100 * 1024 * 1024)) {
2999                 spin_lock(&block_group->lock);
3000                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3001                 spin_unlock(&block_group->lock);
3002                 return 0;
3003         }
3004
3005 again:
3006         inode = lookup_free_space_inode(root, block_group, path);
3007         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3008                 ret = PTR_ERR(inode);
3009                 btrfs_release_path(path);
3010                 goto out;
3011         }
3012
3013         if (IS_ERR(inode)) {
3014                 BUG_ON(retries);
3015                 retries++;
3016
3017                 if (block_group->ro)
3018                         goto out_free;
3019
3020                 ret = create_free_space_inode(root, trans, block_group, path);
3021                 if (ret)
3022                         goto out_free;
3023                 goto again;
3024         }
3025
3026         /* We've already setup this transaction, go ahead and exit */
3027         if (block_group->cache_generation == trans->transid &&
3028             i_size_read(inode)) {
3029                 dcs = BTRFS_DC_SETUP;
3030                 goto out_put;
3031         }
3032
3033         /*
3034          * We want to set the generation to 0, that way if anything goes wrong
3035          * from here on out we know not to trust this cache when we load up next
3036          * time.
3037          */
3038         BTRFS_I(inode)->generation = 0;
3039         ret = btrfs_update_inode(trans, root, inode);
3040         WARN_ON(ret);
3041
3042         if (i_size_read(inode) > 0) {
3043                 ret = btrfs_truncate_free_space_cache(root, trans, path,
3044                                                       inode);
3045                 if (ret)
3046                         goto out_put;
3047         }
3048
3049         spin_lock(&block_group->lock);
3050         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3051             !btrfs_test_opt(root, SPACE_CACHE)) {
3052                 /*
3053                  * don't bother trying to write stuff out _if_
3054                  * a) we're not cached,
3055                  * b) we're with nospace_cache mount option.
3056                  */
3057                 dcs = BTRFS_DC_WRITTEN;
3058                 spin_unlock(&block_group->lock);
3059                 goto out_put;
3060         }
3061         spin_unlock(&block_group->lock);
3062
3063         /*
3064          * Try to preallocate enough space based on how big the block group is.
3065          * Keep in mind this has to include any pinned space which could end up
3066          * taking up quite a bit since it's not folded into the other space
3067          * cache.
3068          */
3069         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3070         if (!num_pages)
3071                 num_pages = 1;
3072
3073         num_pages *= 16;
3074         num_pages *= PAGE_CACHE_SIZE;
3075
3076         ret = btrfs_check_data_free_space(inode, num_pages);
3077         if (ret)
3078                 goto out_put;
3079
3080         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3081                                               num_pages, num_pages,
3082                                               &alloc_hint);
3083         if (!ret)
3084                 dcs = BTRFS_DC_SETUP;
3085         btrfs_free_reserved_data_space(inode, num_pages);
3086
3087 out_put:
3088         iput(inode);
3089 out_free:
3090         btrfs_release_path(path);
3091 out:
3092         spin_lock(&block_group->lock);
3093         if (!ret && dcs == BTRFS_DC_SETUP)
3094                 block_group->cache_generation = trans->transid;
3095         block_group->disk_cache_state = dcs;
3096         spin_unlock(&block_group->lock);
3097
3098         return ret;
3099 }
3100
3101 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3102                                    struct btrfs_root *root)
3103 {
3104         struct btrfs_block_group_cache *cache;
3105         int err = 0;
3106         struct btrfs_path *path;
3107         u64 last = 0;
3108
3109         path = btrfs_alloc_path();
3110         if (!path)
3111                 return -ENOMEM;
3112
3113 again:
3114         while (1) {
3115                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3116                 while (cache) {
3117                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3118                                 break;
3119                         cache = next_block_group(root, cache);
3120                 }
3121                 if (!cache) {
3122                         if (last == 0)
3123                                 break;
3124                         last = 0;
3125                         continue;
3126                 }
3127                 err = cache_save_setup(cache, trans, path);
3128                 last = cache->key.objectid + cache->key.offset;
3129                 btrfs_put_block_group(cache);
3130         }
3131
3132         while (1) {
3133                 if (last == 0) {
3134                         err = btrfs_run_delayed_refs(trans, root,
3135                                                      (unsigned long)-1);
3136                         if (err) /* File system offline */
3137                                 goto out;
3138                 }
3139
3140                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3141                 while (cache) {
3142                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3143                                 btrfs_put_block_group(cache);
3144                                 goto again;
3145                         }
3146
3147                         if (cache->dirty)
3148                                 break;
3149                         cache = next_block_group(root, cache);
3150                 }
3151                 if (!cache) {
3152                         if (last == 0)
3153                                 break;
3154                         last = 0;
3155                         continue;
3156                 }
3157
3158                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3159                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3160                 cache->dirty = 0;
3161                 last = cache->key.objectid + cache->key.offset;
3162
3163                 err = write_one_cache_group(trans, root, path, cache);
3164                 if (err) /* File system offline */
3165                         goto out;
3166
3167                 btrfs_put_block_group(cache);
3168         }
3169
3170         while (1) {
3171                 /*
3172                  * I don't think this is needed since we're just marking our
3173                  * preallocated extent as written, but just in case it can't
3174                  * hurt.
3175                  */
3176                 if (last == 0) {
3177                         err = btrfs_run_delayed_refs(trans, root,
3178                                                      (unsigned long)-1);
3179                         if (err) /* File system offline */
3180                                 goto out;
3181                 }
3182
3183                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3184                 while (cache) {
3185                         /*
3186                          * Really this shouldn't happen, but it could if we
3187                          * couldn't write the entire preallocated extent and
3188                          * splitting the extent resulted in a new block.
3189                          */
3190                         if (cache->dirty) {
3191                                 btrfs_put_block_group(cache);
3192                                 goto again;
3193                         }
3194                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3195                                 break;
3196                         cache = next_block_group(root, cache);
3197                 }
3198                 if (!cache) {
3199                         if (last == 0)
3200                                 break;
3201                         last = 0;
3202                         continue;
3203                 }
3204
3205                 err = btrfs_write_out_cache(root, trans, cache, path);
3206
3207                 /*
3208                  * If we didn't have an error then the cache state is still
3209                  * NEED_WRITE, so we can set it to WRITTEN.
3210                  */
3211                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3212                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3213                 last = cache->key.objectid + cache->key.offset;
3214                 btrfs_put_block_group(cache);
3215         }
3216 out:
3217
3218         btrfs_free_path(path);
3219         return err;
3220 }
3221
3222 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3223 {
3224         struct btrfs_block_group_cache *block_group;
3225         int readonly = 0;
3226
3227         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3228         if (!block_group || block_group->ro)
3229                 readonly = 1;
3230         if (block_group)
3231                 btrfs_put_block_group(block_group);
3232         return readonly;
3233 }
3234
3235 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3236                              u64 total_bytes, u64 bytes_used,
3237                              struct btrfs_space_info **space_info)
3238 {
3239         struct btrfs_space_info *found;
3240         int i;
3241         int factor;
3242
3243         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3244                      BTRFS_BLOCK_GROUP_RAID10))
3245                 factor = 2;
3246         else
3247                 factor = 1;
3248
3249         found = __find_space_info(info, flags);
3250         if (found) {
3251                 spin_lock(&found->lock);
3252                 found->total_bytes += total_bytes;
3253                 found->disk_total += total_bytes * factor;
3254                 found->bytes_used += bytes_used;
3255                 found->disk_used += bytes_used * factor;
3256                 found->full = 0;
3257                 spin_unlock(&found->lock);
3258                 *space_info = found;
3259                 return 0;
3260         }
3261         found = kzalloc(sizeof(*found), GFP_NOFS);
3262         if (!found)
3263                 return -ENOMEM;
3264
3265         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3266                 INIT_LIST_HEAD(&found->block_groups[i]);
3267         init_rwsem(&found->groups_sem);
3268         spin_lock_init(&found->lock);
3269         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3270         found->total_bytes = total_bytes;
3271         found->disk_total = total_bytes * factor;
3272         found->bytes_used = bytes_used;
3273         found->disk_used = bytes_used * factor;
3274         found->bytes_pinned = 0;
3275         found->bytes_reserved = 0;
3276         found->bytes_readonly = 0;
3277         found->bytes_may_use = 0;
3278         found->full = 0;
3279         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3280         found->chunk_alloc = 0;
3281         found->flush = 0;
3282         init_waitqueue_head(&found->wait);
3283         *space_info = found;
3284         list_add_rcu(&found->list, &info->space_info);
3285         if (flags & BTRFS_BLOCK_GROUP_DATA)
3286                 info->data_sinfo = found;
3287         return 0;
3288 }
3289
3290 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3291 {
3292         u64 extra_flags = chunk_to_extended(flags) &
3293                                 BTRFS_EXTENDED_PROFILE_MASK;
3294
3295         write_seqlock(&fs_info->profiles_lock);
3296         if (flags & BTRFS_BLOCK_GROUP_DATA)
3297                 fs_info->avail_data_alloc_bits |= extra_flags;
3298         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3299                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3300         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3301                 fs_info->avail_system_alloc_bits |= extra_flags;
3302         write_sequnlock(&fs_info->profiles_lock);
3303 }
3304
3305 /*
3306  * returns target flags in extended format or 0 if restripe for this
3307  * chunk_type is not in progress
3308  *
3309  * should be called with either volume_mutex or balance_lock held
3310  */
3311 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3312 {
3313         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3314         u64 target = 0;
3315
3316         if (!bctl)
3317                 return 0;
3318
3319         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3320             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3321                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3322         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3323                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3324                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3325         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3326                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3327                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3328         }
3329
3330         return target;
3331 }
3332
3333 /*
3334  * @flags: available profiles in extended format (see ctree.h)
3335  *
3336  * Returns reduced profile in chunk format.  If profile changing is in
3337  * progress (either running or paused) picks the target profile (if it's
3338  * already available), otherwise falls back to plain reducing.
3339  */
3340 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3341 {
3342         /*
3343          * we add in the count of missing devices because we want
3344          * to make sure that any RAID levels on a degraded FS
3345          * continue to be honored.
3346          */
3347         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3348                 root->fs_info->fs_devices->missing_devices;
3349         u64 target;
3350         u64 tmp;
3351
3352         /*
3353          * see if restripe for this chunk_type is in progress, if so
3354          * try to reduce to the target profile
3355          */
3356         spin_lock(&root->fs_info->balance_lock);
3357         target = get_restripe_target(root->fs_info, flags);
3358         if (target) {
3359                 /* pick target profile only if it's already available */
3360                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3361                         spin_unlock(&root->fs_info->balance_lock);
3362                         return extended_to_chunk(target);
3363                 }
3364         }
3365         spin_unlock(&root->fs_info->balance_lock);
3366
3367         /* First, mask out the RAID levels which aren't possible */
3368         if (num_devices == 1)
3369                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3370                            BTRFS_BLOCK_GROUP_RAID5);
3371         if (num_devices < 3)
3372                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3373         if (num_devices < 4)
3374                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3375
3376         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3377                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3378                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3379         flags &= ~tmp;
3380
3381         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3382                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3383         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3384                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3385         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3386                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3387         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3388                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3389         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3390                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3391
3392         return extended_to_chunk(flags | tmp);
3393 }
3394
3395 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3396 {
3397         unsigned seq;
3398
3399         do {
3400                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3401
3402                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3403                         flags |= root->fs_info->avail_data_alloc_bits;
3404                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3405                         flags |= root->fs_info->avail_system_alloc_bits;
3406                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3407                         flags |= root->fs_info->avail_metadata_alloc_bits;
3408         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3409
3410         return btrfs_reduce_alloc_profile(root, flags);
3411 }
3412
3413 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3414 {
3415         u64 flags;
3416         u64 ret;
3417
3418         if (data)
3419                 flags = BTRFS_BLOCK_GROUP_DATA;
3420         else if (root == root->fs_info->chunk_root)
3421                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3422         else
3423                 flags = BTRFS_BLOCK_GROUP_METADATA;
3424
3425         ret = get_alloc_profile(root, flags);
3426         return ret;
3427 }
3428
3429 /*
3430  * This will check the space that the inode allocates from to make sure we have
3431  * enough space for bytes.
3432  */
3433 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3434 {
3435         struct btrfs_space_info *data_sinfo;
3436         struct btrfs_root *root = BTRFS_I(inode)->root;
3437         struct btrfs_fs_info *fs_info = root->fs_info;
3438         u64 used;
3439         int ret = 0, committed = 0, alloc_chunk = 1;
3440
3441         /* make sure bytes are sectorsize aligned */
3442         bytes = ALIGN(bytes, root->sectorsize);
3443
3444         if (root == root->fs_info->tree_root ||
3445             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3446                 alloc_chunk = 0;
3447                 committed = 1;
3448         }
3449
3450         data_sinfo = fs_info->data_sinfo;
3451         if (!data_sinfo)
3452                 goto alloc;
3453
3454 again:
3455         /* make sure we have enough space to handle the data first */
3456         spin_lock(&data_sinfo->lock);
3457         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3458                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3459                 data_sinfo->bytes_may_use;
3460
3461         if (used + bytes > data_sinfo->total_bytes) {
3462                 struct btrfs_trans_handle *trans;
3463
3464                 /*
3465                  * if we don't have enough free bytes in this space then we need
3466                  * to alloc a new chunk.
3467                  */
3468                 if (!data_sinfo->full && alloc_chunk) {
3469                         u64 alloc_target;
3470
3471                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3472                         spin_unlock(&data_sinfo->lock);
3473 alloc:
3474                         alloc_target = btrfs_get_alloc_profile(root, 1);
3475                         trans = btrfs_join_transaction(root);
3476                         if (IS_ERR(trans))
3477                                 return PTR_ERR(trans);
3478
3479                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3480                                              alloc_target,
3481                                              CHUNK_ALLOC_NO_FORCE);
3482                         btrfs_end_transaction(trans, root);
3483                         if (ret < 0) {
3484                                 if (ret != -ENOSPC)
3485                                         return ret;
3486                                 else
3487                                         goto commit_trans;
3488                         }
3489
3490                         if (!data_sinfo)
3491                                 data_sinfo = fs_info->data_sinfo;
3492
3493                         goto again;
3494                 }
3495
3496                 /*
3497                  * If we have less pinned bytes than we want to allocate then
3498                  * don't bother committing the transaction, it won't help us.
3499                  */
3500                 if (data_sinfo->bytes_pinned < bytes)
3501                         committed = 1;
3502                 spin_unlock(&data_sinfo->lock);
3503
3504                 /* commit the current transaction and try again */
3505 commit_trans:
3506                 if (!committed &&
3507                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3508                         committed = 1;
3509                         trans = btrfs_join_transaction(root);
3510                         if (IS_ERR(trans))
3511                                 return PTR_ERR(trans);
3512                         ret = btrfs_commit_transaction(trans, root);
3513                         if (ret)
3514                                 return ret;
3515                         goto again;
3516                 }
3517
3518                 return -ENOSPC;
3519         }
3520         data_sinfo->bytes_may_use += bytes;
3521         trace_btrfs_space_reservation(root->fs_info, "space_info",
3522                                       data_sinfo->flags, bytes, 1);
3523         spin_unlock(&data_sinfo->lock);
3524
3525         return 0;
3526 }
3527
3528 /*
3529  * Called if we need to clear a data reservation for this inode.
3530  */
3531 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3532 {
3533         struct btrfs_root *root = BTRFS_I(inode)->root;
3534         struct btrfs_space_info *data_sinfo;
3535
3536         /* make sure bytes are sectorsize aligned */
3537         bytes = ALIGN(bytes, root->sectorsize);
3538
3539         data_sinfo = root->fs_info->data_sinfo;
3540         spin_lock(&data_sinfo->lock);
3541         data_sinfo->bytes_may_use -= bytes;
3542         trace_btrfs_space_reservation(root->fs_info, "space_info",
3543                                       data_sinfo->flags, bytes, 0);
3544         spin_unlock(&data_sinfo->lock);
3545 }
3546
3547 static void force_metadata_allocation(struct btrfs_fs_info *info)
3548 {
3549         struct list_head *head = &info->space_info;
3550         struct btrfs_space_info *found;
3551
3552         rcu_read_lock();
3553         list_for_each_entry_rcu(found, head, list) {
3554                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3555                         found->force_alloc = CHUNK_ALLOC_FORCE;
3556         }
3557         rcu_read_unlock();
3558 }
3559
3560 static int should_alloc_chunk(struct btrfs_root *root,
3561                               struct btrfs_space_info *sinfo, int force)
3562 {
3563         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3564         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3565         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3566         u64 thresh;
3567
3568         if (force == CHUNK_ALLOC_FORCE)
3569                 return 1;
3570
3571         /*
3572          * We need to take into account the global rsv because for all intents
3573          * and purposes it's used space.  Don't worry about locking the
3574          * global_rsv, it doesn't change except when the transaction commits.
3575          */
3576         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3577                 num_allocated += global_rsv->size;
3578
3579         /*
3580          * in limited mode, we want to have some free space up to
3581          * about 1% of the FS size.
3582          */
3583         if (force == CHUNK_ALLOC_LIMITED) {
3584                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3585                 thresh = max_t(u64, 64 * 1024 * 1024,
3586                                div_factor_fine(thresh, 1));
3587
3588                 if (num_bytes - num_allocated < thresh)
3589                         return 1;
3590         }
3591
3592         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3593                 return 0;
3594         return 1;
3595 }
3596
3597 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3598 {
3599         u64 num_dev;
3600
3601         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3602                     BTRFS_BLOCK_GROUP_RAID0 |
3603                     BTRFS_BLOCK_GROUP_RAID5 |
3604                     BTRFS_BLOCK_GROUP_RAID6))
3605                 num_dev = root->fs_info->fs_devices->rw_devices;
3606         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3607                 num_dev = 2;
3608         else
3609                 num_dev = 1;    /* DUP or single */
3610
3611         /* metadata for updaing devices and chunk tree */
3612         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3613 }
3614
3615 static void check_system_chunk(struct btrfs_trans_handle *trans,
3616                                struct btrfs_root *root, u64 type)
3617 {
3618         struct btrfs_space_info *info;
3619         u64 left;
3620         u64 thresh;
3621
3622         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3623         spin_lock(&info->lock);
3624         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3625                 info->bytes_reserved - info->bytes_readonly;
3626         spin_unlock(&info->lock);
3627
3628         thresh = get_system_chunk_thresh(root, type);
3629         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3630                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3631                        left, thresh, type);
3632                 dump_space_info(info, 0, 0);
3633         }
3634
3635         if (left < thresh) {
3636                 u64 flags;
3637
3638                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3639                 btrfs_alloc_chunk(trans, root, flags);
3640         }
3641 }
3642
3643 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3644                           struct btrfs_root *extent_root, u64 flags, int force)
3645 {
3646         struct btrfs_space_info *space_info;
3647         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3648         int wait_for_alloc = 0;
3649         int ret = 0;
3650
3651         /* Don't re-enter if we're already allocating a chunk */
3652         if (trans->allocating_chunk)
3653                 return -ENOSPC;
3654
3655         space_info = __find_space_info(extent_root->fs_info, flags);
3656         if (!space_info) {
3657                 ret = update_space_info(extent_root->fs_info, flags,
3658                                         0, 0, &space_info);
3659                 BUG_ON(ret); /* -ENOMEM */
3660         }
3661         BUG_ON(!space_info); /* Logic error */
3662
3663 again:
3664         spin_lock(&space_info->lock);
3665         if (force < space_info->force_alloc)
3666                 force = space_info->force_alloc;
3667         if (space_info->full) {
3668                 spin_unlock(&space_info->lock);
3669                 return 0;
3670         }
3671
3672         if (!should_alloc_chunk(extent_root, space_info, force)) {
3673                 spin_unlock(&space_info->lock);
3674                 return 0;
3675         } else if (space_info->chunk_alloc) {
3676                 wait_for_alloc = 1;
3677         } else {
3678                 space_info->chunk_alloc = 1;
3679         }
3680
3681         spin_unlock(&space_info->lock);
3682
3683         mutex_lock(&fs_info->chunk_mutex);
3684
3685         /*
3686          * The chunk_mutex is held throughout the entirety of a chunk
3687          * allocation, so once we've acquired the chunk_mutex we know that the
3688          * other guy is done and we need to recheck and see if we should
3689          * allocate.
3690          */
3691         if (wait_for_alloc) {
3692                 mutex_unlock(&fs_info->chunk_mutex);
3693                 wait_for_alloc = 0;
3694                 goto again;
3695         }
3696
3697         trans->allocating_chunk = true;
3698
3699         /*
3700          * If we have mixed data/metadata chunks we want to make sure we keep
3701          * allocating mixed chunks instead of individual chunks.
3702          */
3703         if (btrfs_mixed_space_info(space_info))
3704                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3705
3706         /*
3707          * if we're doing a data chunk, go ahead and make sure that
3708          * we keep a reasonable number of metadata chunks allocated in the
3709          * FS as well.
3710          */
3711         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3712                 fs_info->data_chunk_allocations++;
3713                 if (!(fs_info->data_chunk_allocations %
3714                       fs_info->metadata_ratio))
3715                         force_metadata_allocation(fs_info);
3716         }
3717
3718         /*
3719          * Check if we have enough space in SYSTEM chunk because we may need
3720          * to update devices.
3721          */
3722         check_system_chunk(trans, extent_root, flags);
3723
3724         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3725         trans->allocating_chunk = false;
3726
3727         spin_lock(&space_info->lock);
3728         if (ret < 0 && ret != -ENOSPC)
3729                 goto out;
3730         if (ret)
3731                 space_info->full = 1;
3732         else
3733                 ret = 1;
3734
3735         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3736 out:
3737         space_info->chunk_alloc = 0;
3738         spin_unlock(&space_info->lock);
3739         mutex_unlock(&fs_info->chunk_mutex);
3740         return ret;
3741 }
3742
3743 static int can_overcommit(struct btrfs_root *root,
3744                           struct btrfs_space_info *space_info, u64 bytes,
3745                           enum btrfs_reserve_flush_enum flush)
3746 {
3747         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3748         u64 profile = btrfs_get_alloc_profile(root, 0);
3749         u64 rsv_size = 0;
3750         u64 avail;
3751         u64 used;
3752         u64 to_add;
3753
3754         used = space_info->bytes_used + space_info->bytes_reserved +
3755                 space_info->bytes_pinned + space_info->bytes_readonly;
3756
3757         spin_lock(&global_rsv->lock);
3758         rsv_size = global_rsv->size;
3759         spin_unlock(&global_rsv->lock);
3760
3761         /*
3762          * We only want to allow over committing if we have lots of actual space
3763          * free, but if we don't have enough space to handle the global reserve
3764          * space then we could end up having a real enospc problem when trying
3765          * to allocate a chunk or some other such important allocation.
3766          */
3767         rsv_size <<= 1;
3768         if (used + rsv_size >= space_info->total_bytes)
3769                 return 0;
3770
3771         used += space_info->bytes_may_use;
3772
3773         spin_lock(&root->fs_info->free_chunk_lock);
3774         avail = root->fs_info->free_chunk_space;
3775         spin_unlock(&root->fs_info->free_chunk_lock);
3776
3777         /*
3778          * If we have dup, raid1 or raid10 then only half of the free
3779          * space is actually useable.  For raid56, the space info used
3780          * doesn't include the parity drive, so we don't have to
3781          * change the math
3782          */
3783         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3784                        BTRFS_BLOCK_GROUP_RAID1 |
3785                        BTRFS_BLOCK_GROUP_RAID10))
3786                 avail >>= 1;
3787
3788         to_add = space_info->total_bytes;
3789
3790         /*
3791          * If we aren't flushing all things, let us overcommit up to
3792          * 1/2th of the space. If we can flush, don't let us overcommit
3793          * too much, let it overcommit up to 1/8 of the space.
3794          */
3795         if (flush == BTRFS_RESERVE_FLUSH_ALL)
3796                 to_add >>= 3;
3797         else
3798                 to_add >>= 1;
3799
3800         /*
3801          * Limit the overcommit to the amount of free space we could possibly
3802          * allocate for chunks.
3803          */
3804         to_add = min(avail, to_add);
3805
3806         if (used + bytes < space_info->total_bytes + to_add)
3807                 return 1;
3808         return 0;
3809 }
3810
3811 void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3812                                   unsigned long nr_pages)
3813 {
3814         struct super_block *sb = root->fs_info->sb;
3815         int started;
3816
3817         /* If we can not start writeback, just sync all the delalloc file. */
3818         started = try_to_writeback_inodes_sb_nr(sb, nr_pages,
3819                                                       WB_REASON_FS_FREE_SPACE);
3820         if (!started) {
3821                 /*
3822                  * We needn't worry the filesystem going from r/w to r/o though
3823                  * we don't acquire ->s_umount mutex, because the filesystem
3824                  * should guarantee the delalloc inodes list be empty after
3825                  * the filesystem is readonly(all dirty pages are written to
3826                  * the disk).
3827                  */
3828                 btrfs_start_delalloc_inodes(root, 0);
3829                 btrfs_wait_ordered_extents(root, 0);
3830         }
3831 }
3832
3833 /*
3834  * shrink metadata reservation for delalloc
3835  */
3836 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3837                             bool wait_ordered)
3838 {
3839         struct btrfs_block_rsv *block_rsv;
3840         struct btrfs_space_info *space_info;
3841         struct btrfs_trans_handle *trans;
3842         u64 delalloc_bytes;
3843         u64 max_reclaim;
3844         long time_left;
3845         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3846         int loops = 0;
3847         enum btrfs_reserve_flush_enum flush;
3848
3849         trans = (struct btrfs_trans_handle *)current->journal_info;
3850         block_rsv = &root->fs_info->delalloc_block_rsv;
3851         space_info = block_rsv->space_info;
3852
3853         smp_mb();
3854         delalloc_bytes = percpu_counter_sum_positive(
3855                                                 &root->fs_info->delalloc_bytes);
3856         if (delalloc_bytes == 0) {
3857                 if (trans)
3858                         return;
3859                 btrfs_wait_ordered_extents(root, 0);
3860                 return;
3861         }
3862
3863         while (delalloc_bytes && loops < 3) {
3864                 max_reclaim = min(delalloc_bytes, to_reclaim);
3865                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3866                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
3867                 /*
3868                  * We need to wait for the async pages to actually start before
3869                  * we do anything.
3870                  */
3871                 wait_event(root->fs_info->async_submit_wait,
3872                            !atomic_read(&root->fs_info->async_delalloc_pages));
3873
3874                 if (!trans)
3875                         flush = BTRFS_RESERVE_FLUSH_ALL;
3876                 else
3877                         flush = BTRFS_RESERVE_NO_FLUSH;
3878                 spin_lock(&space_info->lock);
3879                 if (can_overcommit(root, space_info, orig, flush)) {
3880                         spin_unlock(&space_info->lock);
3881                         break;
3882                 }
3883                 spin_unlock(&space_info->lock);
3884
3885                 loops++;
3886                 if (wait_ordered && !trans) {
3887                         btrfs_wait_ordered_extents(root, 0);
3888                 } else {
3889                         time_left = schedule_timeout_killable(1);
3890                         if (time_left)
3891                                 break;
3892                 }
3893                 smp_mb();
3894                 delalloc_bytes = percpu_counter_sum_positive(
3895                                                 &root->fs_info->delalloc_bytes);
3896         }
3897 }
3898
3899 /**
3900  * maybe_commit_transaction - possibly commit the transaction if its ok to
3901  * @root - the root we're allocating for
3902  * @bytes - the number of bytes we want to reserve
3903  * @force - force the commit
3904  *
3905  * This will check to make sure that committing the transaction will actually
3906  * get us somewhere and then commit the transaction if it does.  Otherwise it
3907  * will return -ENOSPC.
3908  */
3909 static int may_commit_transaction(struct btrfs_root *root,
3910                                   struct btrfs_space_info *space_info,
3911                                   u64 bytes, int force)
3912 {
3913         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3914         struct btrfs_trans_handle *trans;
3915
3916         trans = (struct btrfs_trans_handle *)current->journal_info;
3917         if (trans)
3918                 return -EAGAIN;
3919
3920         if (force)
3921                 goto commit;
3922
3923         /* See if there is enough pinned space to make this reservation */
3924         spin_lock(&space_info->lock);
3925         if (space_info->bytes_pinned >= bytes) {
3926                 spin_unlock(&space_info->lock);
3927                 goto commit;
3928         }
3929         spin_unlock(&space_info->lock);
3930
3931         /*
3932          * See if there is some space in the delayed insertion reservation for
3933          * this reservation.
3934          */
3935         if (space_info != delayed_rsv->space_info)
3936                 return -ENOSPC;
3937
3938         spin_lock(&space_info->lock);
3939         spin_lock(&delayed_rsv->lock);
3940         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3941                 spin_unlock(&delayed_rsv->lock);
3942                 spin_unlock(&space_info->lock);
3943                 return -ENOSPC;
3944         }
3945         spin_unlock(&delayed_rsv->lock);
3946         spin_unlock(&space_info->lock);
3947
3948 commit:
3949         trans = btrfs_join_transaction(root);
3950         if (IS_ERR(trans))
3951                 return -ENOSPC;
3952
3953         return btrfs_commit_transaction(trans, root);
3954 }
3955
3956 enum flush_state {
3957         FLUSH_DELAYED_ITEMS_NR  =       1,
3958         FLUSH_DELAYED_ITEMS     =       2,
3959         FLUSH_DELALLOC          =       3,
3960         FLUSH_DELALLOC_WAIT     =       4,
3961         ALLOC_CHUNK             =       5,
3962         COMMIT_TRANS            =       6,
3963 };
3964
3965 static int flush_space(struct btrfs_root *root,
3966                        struct btrfs_space_info *space_info, u64 num_bytes,
3967                        u64 orig_bytes, int state)
3968 {
3969         struct btrfs_trans_handle *trans;
3970         int nr;
3971         int ret = 0;
3972
3973         switch (state) {
3974         case FLUSH_DELAYED_ITEMS_NR:
3975         case FLUSH_DELAYED_ITEMS:
3976                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3977                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3978
3979                         nr = (int)div64_u64(num_bytes, bytes);
3980                         if (!nr)
3981                                 nr = 1;
3982                         nr *= 2;
3983                 } else {
3984                         nr = -1;
3985                 }
3986                 trans = btrfs_join_transaction(root);
3987                 if (IS_ERR(trans)) {
3988                         ret = PTR_ERR(trans);
3989                         break;
3990                 }
3991                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3992                 btrfs_end_transaction(trans, root);
3993                 break;
3994         case FLUSH_DELALLOC:
3995         case FLUSH_DELALLOC_WAIT:
3996                 shrink_delalloc(root, num_bytes, orig_bytes,
3997                                 state == FLUSH_DELALLOC_WAIT);
3998                 break;
3999         case ALLOC_CHUNK:
4000                 trans = btrfs_join_transaction(root);
4001                 if (IS_ERR(trans)) {
4002                         ret = PTR_ERR(trans);
4003                         break;
4004                 }
4005                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4006                                      btrfs_get_alloc_profile(root, 0),
4007                                      CHUNK_ALLOC_NO_FORCE);
4008                 btrfs_end_transaction(trans, root);
4009                 if (ret == -ENOSPC)
4010                         ret = 0;
4011                 break;
4012         case COMMIT_TRANS:
4013                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4014                 break;
4015         default:
4016                 ret = -ENOSPC;
4017                 break;
4018         }
4019
4020         return ret;
4021 }
4022 /**
4023  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4024  * @root - the root we're allocating for
4025  * @block_rsv - the block_rsv we're allocating for
4026  * @orig_bytes - the number of bytes we want
4027  * @flush - whether or not we can flush to make our reservation
4028  *
4029  * This will reserve orgi_bytes number of bytes from the space info associated
4030  * with the block_rsv.  If there is not enough space it will make an attempt to
4031  * flush out space to make room.  It will do this by flushing delalloc if
4032  * possible or committing the transaction.  If flush is 0 then no attempts to
4033  * regain reservations will be made and this will fail if there is not enough
4034  * space already.
4035  */
4036 static int reserve_metadata_bytes(struct btrfs_root *root,
4037                                   struct btrfs_block_rsv *block_rsv,
4038                                   u64 orig_bytes,
4039                                   enum btrfs_reserve_flush_enum flush)
4040 {
4041         struct btrfs_space_info *space_info = block_rsv->space_info;
4042         u64 used;
4043         u64 num_bytes = orig_bytes;
4044         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4045         int ret = 0;
4046         bool flushing = false;
4047
4048 again:
4049         ret = 0;
4050         spin_lock(&space_info->lock);
4051         /*
4052          * We only want to wait if somebody other than us is flushing and we
4053          * are actually allowed to flush all things.
4054          */
4055         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4056                space_info->flush) {
4057                 spin_unlock(&space_info->lock);
4058                 /*
4059                  * If we have a trans handle we can't wait because the flusher
4060                  * may have to commit the transaction, which would mean we would
4061                  * deadlock since we are waiting for the flusher to finish, but
4062                  * hold the current transaction open.
4063                  */
4064                 if (current->journal_info)
4065                         return -EAGAIN;
4066                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4067                 /* Must have been killed, return */
4068                 if (ret)
4069                         return -EINTR;
4070
4071                 spin_lock(&space_info->lock);
4072         }
4073
4074         ret = -ENOSPC;
4075         used = space_info->bytes_used + space_info->bytes_reserved +
4076                 space_info->bytes_pinned + space_info->bytes_readonly +
4077                 space_info->bytes_may_use;
4078
4079         /*
4080          * The idea here is that we've not already over-reserved the block group
4081          * then we can go ahead and save our reservation first and then start
4082          * flushing if we need to.  Otherwise if we've already overcommitted
4083          * lets start flushing stuff first and then come back and try to make
4084          * our reservation.
4085          */
4086         if (used <= space_info->total_bytes) {
4087                 if (used + orig_bytes <= space_info->total_bytes) {
4088                         space_info->bytes_may_use += orig_bytes;
4089                         trace_btrfs_space_reservation(root->fs_info,
4090                                 "space_info", space_info->flags, orig_bytes, 1);
4091                         ret = 0;
4092                 } else {
4093                         /*
4094                          * Ok set num_bytes to orig_bytes since we aren't
4095                          * overocmmitted, this way we only try and reclaim what
4096                          * we need.
4097                          */
4098                         num_bytes = orig_bytes;
4099                 }
4100         } else {
4101                 /*
4102                  * Ok we're over committed, set num_bytes to the overcommitted
4103                  * amount plus the amount of bytes that we need for this
4104                  * reservation.
4105                  */
4106                 num_bytes = used - space_info->total_bytes +
4107                         (orig_bytes * 2);
4108         }
4109
4110         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4111                 space_info->bytes_may_use += orig_bytes;
4112                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4113                                               space_info->flags, orig_bytes,
4114                                               1);
4115                 ret = 0;
4116         }
4117
4118         /*
4119          * Couldn't make our reservation, save our place so while we're trying
4120          * to reclaim space we can actually use it instead of somebody else
4121          * stealing it from us.
4122          *
4123          * We make the other tasks wait for the flush only when we can flush
4124          * all things.
4125          */
4126         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4127                 flushing = true;
4128                 space_info->flush = 1;
4129         }
4130
4131         spin_unlock(&space_info->lock);
4132
4133         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4134                 goto out;
4135
4136         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4137                           flush_state);
4138         flush_state++;
4139
4140         /*
4141          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4142          * would happen. So skip delalloc flush.
4143          */
4144         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4145             (flush_state == FLUSH_DELALLOC ||
4146              flush_state == FLUSH_DELALLOC_WAIT))
4147                 flush_state = ALLOC_CHUNK;
4148
4149         if (!ret)
4150                 goto again;
4151         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4152                  flush_state < COMMIT_TRANS)
4153                 goto again;
4154         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4155                  flush_state <= COMMIT_TRANS)
4156                 goto again;
4157
4158 out:
4159         if (ret == -ENOSPC &&
4160             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4161                 struct btrfs_block_rsv *global_rsv =
4162                         &root->fs_info->global_block_rsv;
4163
4164                 if (block_rsv != global_rsv &&
4165                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4166                         ret = 0;
4167         }
4168         if (flushing) {
4169                 spin_lock(&space_info->lock);
4170                 space_info->flush = 0;
4171                 wake_up_all(&space_info->wait);
4172                 spin_unlock(&space_info->lock);
4173         }
4174         return ret;
4175 }
4176
4177 static struct btrfs_block_rsv *get_block_rsv(
4178                                         const struct btrfs_trans_handle *trans,
4179                                         const struct btrfs_root *root)
4180 {
4181         struct btrfs_block_rsv *block_rsv = NULL;
4182
4183         if (root->ref_cows)
4184                 block_rsv = trans->block_rsv;
4185
4186         if (root == root->fs_info->csum_root && trans->adding_csums)
4187                 block_rsv = trans->block_rsv;
4188
4189         if (!block_rsv)
4190                 block_rsv = root->block_rsv;
4191
4192         if (!block_rsv)
4193                 block_rsv = &root->fs_info->empty_block_rsv;
4194
4195         return block_rsv;
4196 }
4197
4198 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4199                                u64 num_bytes)
4200 {
4201         int ret = -ENOSPC;
4202         spin_lock(&block_rsv->lock);
4203         if (block_rsv->reserved >= num_bytes) {
4204                 block_rsv->reserved -= num_bytes;
4205                 if (block_rsv->reserved < block_rsv->size)
4206                         block_rsv->full = 0;
4207                 ret = 0;
4208         }
4209         spin_unlock(&block_rsv->lock);
4210         return ret;
4211 }
4212
4213 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4214                                 u64 num_bytes, int update_size)
4215 {
4216         spin_lock(&block_rsv->lock);
4217         block_rsv->reserved += num_bytes;
4218         if (update_size)
4219                 block_rsv->size += num_bytes;
4220         else if (block_rsv->reserved >= block_rsv->size)
4221                 block_rsv->full = 1;
4222         spin_unlock(&block_rsv->lock);
4223 }
4224
4225 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4226                                     struct btrfs_block_rsv *block_rsv,
4227                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4228 {
4229         struct btrfs_space_info *space_info = block_rsv->space_info;
4230
4231         spin_lock(&block_rsv->lock);
4232         if (num_bytes == (u64)-1)
4233                 num_bytes = block_rsv->size;
4234         block_rsv->size -= num_bytes;
4235         if (block_rsv->reserved >= block_rsv->size) {
4236                 num_bytes = block_rsv->reserved - block_rsv->size;
4237                 block_rsv->reserved = block_rsv->size;
4238                 block_rsv->full = 1;
4239         } else {
4240                 num_bytes = 0;
4241         }
4242         spin_unlock(&block_rsv->lock);
4243
4244         if (num_bytes > 0) {
4245                 if (dest) {
4246                         spin_lock(&dest->lock);
4247                         if (!dest->full) {
4248                                 u64 bytes_to_add;
4249
4250                                 bytes_to_add = dest->size - dest->reserved;
4251                                 bytes_to_add = min(num_bytes, bytes_to_add);
4252                                 dest->reserved += bytes_to_add;
4253                                 if (dest->reserved >= dest->size)
4254                                         dest->full = 1;
4255                                 num_bytes -= bytes_to_add;
4256                         }
4257                         spin_unlock(&dest->lock);
4258                 }
4259                 if (num_bytes) {
4260                         spin_lock(&space_info->lock);
4261                         space_info->bytes_may_use -= num_bytes;
4262                         trace_btrfs_space_reservation(fs_info, "space_info",
4263                                         space_info->flags, num_bytes, 0);
4264                         space_info->reservation_progress++;
4265                         spin_unlock(&space_info->lock);
4266                 }
4267         }
4268 }
4269
4270 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4271                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4272 {
4273         int ret;
4274
4275         ret = block_rsv_use_bytes(src, num_bytes);
4276         if (ret)
4277                 return ret;
4278
4279         block_rsv_add_bytes(dst, num_bytes, 1);
4280         return 0;
4281 }
4282
4283 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4284 {
4285         memset(rsv, 0, sizeof(*rsv));
4286         spin_lock_init(&rsv->lock);
4287         rsv->type = type;
4288 }
4289
4290 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4291                                               unsigned short type)
4292 {
4293         struct btrfs_block_rsv *block_rsv;
4294         struct btrfs_fs_info *fs_info = root->fs_info;
4295
4296         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4297         if (!block_rsv)
4298                 return NULL;
4299
4300         btrfs_init_block_rsv(block_rsv, type);
4301         block_rsv->space_info = __find_space_info(fs_info,
4302                                                   BTRFS_BLOCK_GROUP_METADATA);
4303         return block_rsv;
4304 }
4305
4306 void btrfs_free_block_rsv(struct btrfs_root *root,
4307                           struct btrfs_block_rsv *rsv)
4308 {
4309         if (!rsv)
4310                 return;
4311         btrfs_block_rsv_release(root, rsv, (u64)-1);
4312         kfree(rsv);
4313 }
4314
4315 int btrfs_block_rsv_add(struct btrfs_root *root,
4316                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4317                         enum btrfs_reserve_flush_enum flush)
4318 {
4319         int ret;
4320
4321         if (num_bytes == 0)
4322                 return 0;
4323
4324         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4325         if (!ret) {
4326                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4327                 return 0;
4328         }
4329
4330         return ret;
4331 }
4332
4333 int btrfs_block_rsv_check(struct btrfs_root *root,
4334                           struct btrfs_block_rsv *block_rsv, int min_factor)
4335 {
4336         u64 num_bytes = 0;
4337         int ret = -ENOSPC;
4338
4339         if (!block_rsv)
4340                 return 0;
4341
4342         spin_lock(&block_rsv->lock);
4343         num_bytes = div_factor(block_rsv->size, min_factor);
4344         if (block_rsv->reserved >= num_bytes)
4345                 ret = 0;
4346         spin_unlock(&block_rsv->lock);
4347
4348         return ret;
4349 }
4350
4351 int btrfs_block_rsv_refill(struct btrfs_root *root,
4352                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4353                            enum btrfs_reserve_flush_enum flush)
4354 {
4355         u64 num_bytes = 0;
4356         int ret = -ENOSPC;
4357
4358         if (!block_rsv)
4359                 return 0;
4360
4361         spin_lock(&block_rsv->lock);
4362         num_bytes = min_reserved;
4363         if (block_rsv->reserved >= num_bytes)
4364                 ret = 0;
4365         else
4366                 num_bytes -= block_rsv->reserved;
4367         spin_unlock(&block_rsv->lock);
4368
4369         if (!ret)
4370                 return 0;
4371
4372         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4373         if (!ret) {
4374                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4375                 return 0;
4376         }
4377
4378         return ret;
4379 }
4380
4381 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4382                             struct btrfs_block_rsv *dst_rsv,
4383                             u64 num_bytes)
4384 {
4385         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4386 }
4387
4388 void btrfs_block_rsv_release(struct btrfs_root *root,
4389                              struct btrfs_block_rsv *block_rsv,
4390                              u64 num_bytes)
4391 {
4392         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4393         if (global_rsv->full || global_rsv == block_rsv ||
4394             block_rsv->space_info != global_rsv->space_info)
4395                 global_rsv = NULL;
4396         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4397                                 num_bytes);
4398 }
4399
4400 /*
4401  * helper to calculate size of global block reservation.
4402  * the desired value is sum of space used by extent tree,
4403  * checksum tree and root tree
4404  */
4405 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4406 {
4407         struct btrfs_space_info *sinfo;
4408         u64 num_bytes;
4409         u64 meta_used;
4410         u64 data_used;
4411         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4412
4413         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4414         spin_lock(&sinfo->lock);
4415         data_used = sinfo->bytes_used;
4416         spin_unlock(&sinfo->lock);
4417
4418         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4419         spin_lock(&sinfo->lock);
4420         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4421                 data_used = 0;
4422         meta_used = sinfo->bytes_used;
4423         spin_unlock(&sinfo->lock);
4424
4425         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4426                     csum_size * 2;
4427         num_bytes += div64_u64(data_used + meta_used, 50);
4428
4429         if (num_bytes * 3 > meta_used)
4430                 num_bytes = div64_u64(meta_used, 3);
4431
4432         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4433 }
4434
4435 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4436 {
4437         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4438         struct btrfs_space_info *sinfo = block_rsv->space_info;
4439         u64 num_bytes;
4440
4441         num_bytes = calc_global_metadata_size(fs_info);
4442
4443         spin_lock(&sinfo->lock);
4444         spin_lock(&block_rsv->lock);
4445
4446         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4447
4448         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4449                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4450                     sinfo->bytes_may_use;
4451
4452         if (sinfo->total_bytes > num_bytes) {
4453                 num_bytes = sinfo->total_bytes - num_bytes;
4454                 block_rsv->reserved += num_bytes;
4455                 sinfo->bytes_may_use += num_bytes;
4456                 trace_btrfs_space_reservation(fs_info, "space_info",
4457                                       sinfo->flags, num_bytes, 1);
4458         }
4459
4460         if (block_rsv->reserved >= block_rsv->size) {
4461                 num_bytes = block_rsv->reserved - block_rsv->size;
4462                 sinfo->bytes_may_use -= num_bytes;
4463                 trace_btrfs_space_reservation(fs_info, "space_info",
4464                                       sinfo->flags, num_bytes, 0);
4465                 sinfo->reservation_progress++;
4466                 block_rsv->reserved = block_rsv->size;
4467                 block_rsv->full = 1;
4468         }
4469
4470         spin_unlock(&block_rsv->lock);
4471         spin_unlock(&sinfo->lock);
4472 }
4473
4474 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4475 {
4476         struct btrfs_space_info *space_info;
4477
4478         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4479         fs_info->chunk_block_rsv.space_info = space_info;
4480
4481         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4482         fs_info->global_block_rsv.space_info = space_info;
4483         fs_info->delalloc_block_rsv.space_info = space_info;
4484         fs_info->trans_block_rsv.space_info = space_info;
4485         fs_info->empty_block_rsv.space_info = space_info;
4486         fs_info->delayed_block_rsv.space_info = space_info;
4487
4488         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4489         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4490         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4491         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4492         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4493
4494         update_global_block_rsv(fs_info);
4495 }
4496
4497 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4498 {
4499         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4500                                 (u64)-1);
4501         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4502         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4503         WARN_ON(fs_info->trans_block_rsv.size > 0);
4504         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4505         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4506         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4507         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4508         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4509 }
4510
4511 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4512                                   struct btrfs_root *root)
4513 {
4514         if (!trans->block_rsv)
4515                 return;
4516
4517         if (!trans->bytes_reserved)
4518                 return;
4519
4520         trace_btrfs_space_reservation(root->fs_info, "transaction",
4521                                       trans->transid, trans->bytes_reserved, 0);
4522         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4523         trans->bytes_reserved = 0;
4524 }
4525
4526 /* Can only return 0 or -ENOSPC */
4527 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4528                                   struct inode *inode)
4529 {
4530         struct btrfs_root *root = BTRFS_I(inode)->root;
4531         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4532         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4533
4534         /*
4535          * We need to hold space in order to delete our orphan item once we've
4536          * added it, so this takes the reservation so we can release it later
4537          * when we are truly done with the orphan item.
4538          */
4539         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4540         trace_btrfs_space_reservation(root->fs_info, "orphan",
4541                                       btrfs_ino(inode), num_bytes, 1);
4542         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4543 }
4544
4545 void btrfs_orphan_release_metadata(struct inode *inode)
4546 {
4547         struct btrfs_root *root = BTRFS_I(inode)->root;
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, 0);
4551         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4552 }
4553
4554 /*
4555  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4556  * root: the root of the parent directory
4557  * rsv: block reservation
4558  * items: the number of items that we need do reservation
4559  * qgroup_reserved: used to return the reserved size in qgroup
4560  *
4561  * This function is used to reserve the space for snapshot/subvolume
4562  * creation and deletion. Those operations are different with the
4563  * common file/directory operations, they change two fs/file trees
4564  * and root tree, the number of items that the qgroup reserves is
4565  * different with the free space reservation. So we can not use
4566  * the space reseravtion mechanism in start_transaction().
4567  */
4568 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4569                                      struct btrfs_block_rsv *rsv,
4570                                      int items,
4571                                      u64 *qgroup_reserved)
4572 {
4573         u64 num_bytes;
4574         int ret;
4575
4576         if (root->fs_info->quota_enabled) {
4577                 /* One for parent inode, two for dir entries */
4578                 num_bytes = 3 * root->leafsize;
4579                 ret = btrfs_qgroup_reserve(root, num_bytes);
4580                 if (ret)
4581                         return ret;
4582         } else {
4583                 num_bytes = 0;
4584         }
4585
4586         *qgroup_reserved = num_bytes;
4587
4588         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4589         rsv->space_info = __find_space_info(root->fs_info,
4590                                             BTRFS_BLOCK_GROUP_METADATA);
4591         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4592                                   BTRFS_RESERVE_FLUSH_ALL);
4593         if (ret) {
4594                 if (*qgroup_reserved)
4595                         btrfs_qgroup_free(root, *qgroup_reserved);
4596         }
4597
4598         return ret;
4599 }
4600
4601 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4602                                       struct btrfs_block_rsv *rsv,
4603                                       u64 qgroup_reserved)
4604 {
4605         btrfs_block_rsv_release(root, rsv, (u64)-1);
4606         if (qgroup_reserved)
4607                 btrfs_qgroup_free(root, qgroup_reserved);
4608 }
4609
4610 /**
4611  * drop_outstanding_extent - drop an outstanding extent
4612  * @inode: the inode we're dropping the extent for
4613  *
4614  * This is called when we are freeing up an outstanding extent, either called
4615  * after an error or after an extent is written.  This will return the number of
4616  * reserved extents that need to be freed.  This must be called with
4617  * BTRFS_I(inode)->lock held.
4618  */
4619 static unsigned drop_outstanding_extent(struct inode *inode)
4620 {
4621         unsigned drop_inode_space = 0;
4622         unsigned dropped_extents = 0;
4623
4624         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4625         BTRFS_I(inode)->outstanding_extents--;
4626
4627         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4628             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4629                                &BTRFS_I(inode)->runtime_flags))
4630                 drop_inode_space = 1;
4631
4632         /*
4633          * If we have more or the same amount of outsanding extents than we have
4634          * reserved then we need to leave the reserved extents count alone.
4635          */
4636         if (BTRFS_I(inode)->outstanding_extents >=
4637             BTRFS_I(inode)->reserved_extents)
4638                 return drop_inode_space;
4639
4640         dropped_extents = BTRFS_I(inode)->reserved_extents -
4641                 BTRFS_I(inode)->outstanding_extents;
4642         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4643         return dropped_extents + drop_inode_space;
4644 }
4645
4646 /**
4647  * calc_csum_metadata_size - return the amount of metada space that must be
4648  *      reserved/free'd for the given bytes.
4649  * @inode: the inode we're manipulating
4650  * @num_bytes: the number of bytes in question
4651  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4652  *
4653  * This adjusts the number of csum_bytes in the inode and then returns the
4654  * correct amount of metadata that must either be reserved or freed.  We
4655  * calculate how many checksums we can fit into one leaf and then divide the
4656  * number of bytes that will need to be checksumed by this value to figure out
4657  * how many checksums will be required.  If we are adding bytes then the number
4658  * may go up and we will return the number of additional bytes that must be
4659  * reserved.  If it is going down we will return the number of bytes that must
4660  * be freed.
4661  *
4662  * This must be called with BTRFS_I(inode)->lock held.
4663  */
4664 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4665                                    int reserve)
4666 {
4667         struct btrfs_root *root = BTRFS_I(inode)->root;
4668         u64 csum_size;
4669         int num_csums_per_leaf;
4670         int num_csums;
4671         int old_csums;
4672
4673         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4674             BTRFS_I(inode)->csum_bytes == 0)
4675                 return 0;
4676
4677         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4678         if (reserve)
4679                 BTRFS_I(inode)->csum_bytes += num_bytes;
4680         else
4681                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4682         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4683         num_csums_per_leaf = (int)div64_u64(csum_size,
4684                                             sizeof(struct btrfs_csum_item) +
4685                                             sizeof(struct btrfs_disk_key));
4686         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4687         num_csums = num_csums + num_csums_per_leaf - 1;
4688         num_csums = num_csums / num_csums_per_leaf;
4689
4690         old_csums = old_csums + num_csums_per_leaf - 1;
4691         old_csums = old_csums / num_csums_per_leaf;
4692
4693         /* No change, no need to reserve more */
4694         if (old_csums == num_csums)
4695                 return 0;
4696
4697         if (reserve)
4698                 return btrfs_calc_trans_metadata_size(root,
4699                                                       num_csums - old_csums);
4700
4701         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4702 }
4703
4704 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4705 {
4706         struct btrfs_root *root = BTRFS_I(inode)->root;
4707         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4708         u64 to_reserve = 0;
4709         u64 csum_bytes;
4710         unsigned nr_extents = 0;
4711         int extra_reserve = 0;
4712         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
4713         int ret = 0;
4714         bool delalloc_lock = true;
4715         u64 to_free = 0;
4716         unsigned dropped;
4717
4718         /* If we are a free space inode we need to not flush since we will be in
4719          * the middle of a transaction commit.  We also don't need the delalloc
4720          * mutex since we won't race with anybody.  We need this mostly to make
4721          * lockdep shut its filthy mouth.
4722          */
4723         if (btrfs_is_free_space_inode(inode)) {
4724                 flush = BTRFS_RESERVE_NO_FLUSH;
4725                 delalloc_lock = false;
4726         }
4727
4728         if (flush != BTRFS_RESERVE_NO_FLUSH &&
4729             btrfs_transaction_in_commit(root->fs_info))
4730                 schedule_timeout(1);
4731
4732         if (delalloc_lock)
4733                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4734
4735         num_bytes = ALIGN(num_bytes, root->sectorsize);
4736
4737         spin_lock(&BTRFS_I(inode)->lock);
4738         BTRFS_I(inode)->outstanding_extents++;
4739
4740         if (BTRFS_I(inode)->outstanding_extents >
4741             BTRFS_I(inode)->reserved_extents)
4742                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4743                         BTRFS_I(inode)->reserved_extents;
4744
4745         /*
4746          * Add an item to reserve for updating the inode when we complete the
4747          * delalloc io.
4748          */
4749         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4750                       &BTRFS_I(inode)->runtime_flags)) {
4751                 nr_extents++;
4752                 extra_reserve = 1;
4753         }
4754
4755         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4756         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4757         csum_bytes = BTRFS_I(inode)->csum_bytes;
4758         spin_unlock(&BTRFS_I(inode)->lock);
4759
4760         if (root->fs_info->quota_enabled) {
4761                 ret = btrfs_qgroup_reserve(root, num_bytes +
4762                                            nr_extents * root->leafsize);
4763                 if (ret)
4764                         goto out_fail;
4765         }
4766
4767         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4768         if (unlikely(ret)) {
4769                 if (root->fs_info->quota_enabled)
4770                         btrfs_qgroup_free(root, num_bytes +
4771                                                 nr_extents * root->leafsize);
4772                 goto out_fail;
4773         }
4774
4775         spin_lock(&BTRFS_I(inode)->lock);
4776         if (extra_reserve) {
4777                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4778                         &BTRFS_I(inode)->runtime_flags);
4779                 nr_extents--;
4780         }
4781         BTRFS_I(inode)->reserved_extents += nr_extents;
4782         spin_unlock(&BTRFS_I(inode)->lock);
4783
4784         if (delalloc_lock)
4785                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4786
4787         if (to_reserve)
4788                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4789                                               btrfs_ino(inode), to_reserve, 1);
4790         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4791
4792         return 0;
4793
4794 out_fail:
4795         spin_lock(&BTRFS_I(inode)->lock);
4796         dropped = drop_outstanding_extent(inode);
4797         /*
4798          * If the inodes csum_bytes is the same as the original
4799          * csum_bytes then we know we haven't raced with any free()ers
4800          * so we can just reduce our inodes csum bytes and carry on.
4801          */
4802         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
4803                 calc_csum_metadata_size(inode, num_bytes, 0);
4804         } else {
4805                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
4806                 u64 bytes;
4807
4808                 /*
4809                  * This is tricky, but first we need to figure out how much we
4810                  * free'd from any free-ers that occured during this
4811                  * reservation, so we reset ->csum_bytes to the csum_bytes
4812                  * before we dropped our lock, and then call the free for the
4813                  * number of bytes that were freed while we were trying our
4814                  * reservation.
4815                  */
4816                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
4817                 BTRFS_I(inode)->csum_bytes = csum_bytes;
4818                 to_free = calc_csum_metadata_size(inode, bytes, 0);
4819
4820
4821                 /*
4822                  * Now we need to see how much we would have freed had we not
4823                  * been making this reservation and our ->csum_bytes were not
4824                  * artificially inflated.
4825                  */
4826                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
4827                 bytes = csum_bytes - orig_csum_bytes;
4828                 bytes = calc_csum_metadata_size(inode, bytes, 0);
4829
4830                 /*
4831                  * Now reset ->csum_bytes to what it should be.  If bytes is
4832                  * more than to_free then we would have free'd more space had we
4833                  * not had an artificially high ->csum_bytes, so we need to free
4834                  * the remainder.  If bytes is the same or less then we don't
4835                  * need to do anything, the other free-ers did the correct
4836                  * thing.
4837                  */
4838                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
4839                 if (bytes > to_free)
4840                         to_free = bytes - to_free;
4841                 else
4842                         to_free = 0;
4843         }
4844         spin_unlock(&BTRFS_I(inode)->lock);
4845         if (dropped)
4846                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4847
4848         if (to_free) {
4849                 btrfs_block_rsv_release(root, block_rsv, to_free);
4850                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
4851                                               btrfs_ino(inode), to_free, 0);
4852         }
4853         if (delalloc_lock)
4854                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4855         return ret;
4856 }
4857
4858 /**
4859  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4860  * @inode: the inode to release the reservation for
4861  * @num_bytes: the number of bytes we're releasing
4862  *
4863  * This will release the metadata reservation for an inode.  This can be called
4864  * once we complete IO for a given set of bytes to release their metadata
4865  * reservations.
4866  */
4867 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4868 {
4869         struct btrfs_root *root = BTRFS_I(inode)->root;
4870         u64 to_free = 0;
4871         unsigned dropped;
4872
4873         num_bytes = ALIGN(num_bytes, root->sectorsize);
4874         spin_lock(&BTRFS_I(inode)->lock);
4875         dropped = drop_outstanding_extent(inode);
4876
4877         if (num_bytes)
4878                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4879         spin_unlock(&BTRFS_I(inode)->lock);
4880         if (dropped > 0)
4881                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4882
4883         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4884                                       btrfs_ino(inode), to_free, 0);
4885         if (root->fs_info->quota_enabled) {
4886                 btrfs_qgroup_free(root, num_bytes +
4887                                         dropped * root->leafsize);
4888         }
4889
4890         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4891                                 to_free);
4892 }
4893
4894 /**
4895  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4896  * @inode: inode we're writing to
4897  * @num_bytes: the number of bytes we want to allocate
4898  *
4899  * This will do the following things
4900  *
4901  * o reserve space in the data space info for num_bytes
4902  * o reserve space in the metadata space info based on number of outstanding
4903  *   extents and how much csums will be needed
4904  * o add to the inodes ->delalloc_bytes
4905  * o add it to the fs_info's delalloc inodes list.
4906  *
4907  * This will return 0 for success and -ENOSPC if there is no space left.
4908  */
4909 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4910 {
4911         int ret;
4912
4913         ret = btrfs_check_data_free_space(inode, num_bytes);
4914         if (ret)
4915                 return ret;
4916
4917         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4918         if (ret) {
4919                 btrfs_free_reserved_data_space(inode, num_bytes);
4920                 return ret;
4921         }
4922
4923         return 0;
4924 }
4925
4926 /**
4927  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4928  * @inode: inode we're releasing space for
4929  * @num_bytes: the number of bytes we want to free up
4930  *
4931  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4932  * called in the case that we don't need the metadata AND data reservations
4933  * anymore.  So if there is an error or we insert an inline extent.
4934  *
4935  * This function will release the metadata space that was not used and will
4936  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4937  * list if there are no delalloc bytes left.
4938  */
4939 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4940 {
4941         btrfs_delalloc_release_metadata(inode, num_bytes);
4942         btrfs_free_reserved_data_space(inode, num_bytes);
4943 }
4944
4945 static int update_block_group(struct btrfs_root *root,
4946                               u64 bytenr, u64 num_bytes, int alloc)
4947 {
4948         struct btrfs_block_group_cache *cache = NULL;
4949         struct btrfs_fs_info *info = root->fs_info;
4950         u64 total = num_bytes;
4951         u64 old_val;
4952         u64 byte_in_group;
4953         int factor;
4954
4955         /* block accounting for super block */
4956         spin_lock(&info->delalloc_lock);
4957         old_val = btrfs_super_bytes_used(info->super_copy);
4958         if (alloc)
4959                 old_val += num_bytes;
4960         else
4961                 old_val -= num_bytes;
4962         btrfs_set_super_bytes_used(info->super_copy, old_val);
4963         spin_unlock(&info->delalloc_lock);
4964
4965         while (total) {
4966                 cache = btrfs_lookup_block_group(info, bytenr);
4967                 if (!cache)
4968                         return -ENOENT;
4969                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4970                                     BTRFS_BLOCK_GROUP_RAID1 |
4971                                     BTRFS_BLOCK_GROUP_RAID10))
4972                         factor = 2;
4973                 else
4974                         factor = 1;
4975                 /*
4976                  * If this block group has free space cache written out, we
4977                  * need to make sure to load it if we are removing space.  This
4978                  * is because we need the unpinning stage to actually add the
4979                  * space back to the block group, otherwise we will leak space.
4980                  */
4981                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4982                         cache_block_group(cache, 1);
4983
4984                 byte_in_group = bytenr - cache->key.objectid;
4985                 WARN_ON(byte_in_group > cache->key.offset);
4986
4987                 spin_lock(&cache->space_info->lock);
4988                 spin_lock(&cache->lock);
4989
4990                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4991                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4992                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4993
4994                 cache->dirty = 1;
4995                 old_val = btrfs_block_group_used(&cache->item);
4996                 num_bytes = min(total, cache->key.offset - byte_in_group);
4997                 if (alloc) {
4998                         old_val += num_bytes;
4999                         btrfs_set_block_group_used(&cache->item, old_val);
5000                         cache->reserved -= num_bytes;
5001                         cache->space_info->bytes_reserved -= num_bytes;
5002                         cache->space_info->bytes_used += num_bytes;
5003                         cache->space_info->disk_used += num_bytes * factor;
5004                         spin_unlock(&cache->lock);
5005                         spin_unlock(&cache->space_info->lock);
5006                 } else {
5007                         old_val -= num_bytes;
5008                         btrfs_set_block_group_used(&cache->item, old_val);
5009                         cache->pinned += num_bytes;
5010                         cache->space_info->bytes_pinned += num_bytes;
5011                         cache->space_info->bytes_used -= num_bytes;
5012                         cache->space_info->disk_used -= num_bytes * factor;
5013                         spin_unlock(&cache->lock);
5014                         spin_unlock(&cache->space_info->lock);
5015
5016                         set_extent_dirty(info->pinned_extents,
5017                                          bytenr, bytenr + num_bytes - 1,
5018                                          GFP_NOFS | __GFP_NOFAIL);
5019                 }
5020                 btrfs_put_block_group(cache);
5021                 total -= num_bytes;
5022                 bytenr += num_bytes;
5023         }
5024         return 0;
5025 }
5026
5027 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5028 {
5029         struct btrfs_block_group_cache *cache;
5030         u64 bytenr;
5031
5032         spin_lock(&root->fs_info->block_group_cache_lock);
5033         bytenr = root->fs_info->first_logical_byte;
5034         spin_unlock(&root->fs_info->block_group_cache_lock);
5035
5036         if (bytenr < (u64)-1)
5037                 return bytenr;
5038
5039         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5040         if (!cache)
5041                 return 0;
5042
5043         bytenr = cache->key.objectid;
5044         btrfs_put_block_group(cache);
5045
5046         return bytenr;
5047 }
5048
5049 static int pin_down_extent(struct btrfs_root *root,
5050                            struct btrfs_block_group_cache *cache,
5051                            u64 bytenr, u64 num_bytes, int reserved)
5052 {
5053         spin_lock(&cache->space_info->lock);
5054         spin_lock(&cache->lock);
5055         cache->pinned += num_bytes;
5056         cache->space_info->bytes_pinned += num_bytes;
5057         if (reserved) {
5058                 cache->reserved -= num_bytes;
5059                 cache->space_info->bytes_reserved -= num_bytes;
5060         }
5061         spin_unlock(&cache->lock);
5062         spin_unlock(&cache->space_info->lock);
5063
5064         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5065                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5066         return 0;
5067 }
5068
5069 /*
5070  * this function must be called within transaction
5071  */
5072 int btrfs_pin_extent(struct btrfs_root *root,
5073                      u64 bytenr, u64 num_bytes, int reserved)
5074 {
5075         struct btrfs_block_group_cache *cache;
5076
5077         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5078         BUG_ON(!cache); /* Logic error */
5079
5080         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5081
5082         btrfs_put_block_group(cache);
5083         return 0;
5084 }
5085
5086 /*
5087  * this function must be called within transaction
5088  */
5089 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5090                                     u64 bytenr, u64 num_bytes)
5091 {
5092         struct btrfs_block_group_cache *cache;
5093
5094         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5095         BUG_ON(!cache); /* Logic error */
5096
5097         /*
5098          * pull in the free space cache (if any) so that our pin
5099          * removes the free space from the cache.  We have load_only set
5100          * to one because the slow code to read in the free extents does check
5101          * the pinned extents.
5102          */
5103         cache_block_group(cache, 1);
5104
5105         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5106
5107         /* remove us from the free space cache (if we're there at all) */
5108         btrfs_remove_free_space(cache, bytenr, num_bytes);
5109         btrfs_put_block_group(cache);
5110         return 0;
5111 }
5112
5113 /**
5114  * btrfs_update_reserved_bytes - update the block_group and space info counters
5115  * @cache:      The cache we are manipulating
5116  * @num_bytes:  The number of bytes in question
5117  * @reserve:    One of the reservation enums
5118  *
5119  * This is called by the allocator when it reserves space, or by somebody who is
5120  * freeing space that was never actually used on disk.  For example if you
5121  * reserve some space for a new leaf in transaction A and before transaction A
5122  * commits you free that leaf, you call this with reserve set to 0 in order to
5123  * clear the reservation.
5124  *
5125  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5126  * ENOSPC accounting.  For data we handle the reservation through clearing the
5127  * delalloc bits in the io_tree.  We have to do this since we could end up
5128  * allocating less disk space for the amount of data we have reserved in the
5129  * case of compression.
5130  *
5131  * If this is a reservation and the block group has become read only we cannot
5132  * make the reservation and return -EAGAIN, otherwise this function always
5133  * succeeds.
5134  */
5135 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5136                                        u64 num_bytes, int reserve)
5137 {
5138         struct btrfs_space_info *space_info = cache->space_info;
5139         int ret = 0;
5140
5141         spin_lock(&space_info->lock);
5142         spin_lock(&cache->lock);
5143         if (reserve != RESERVE_FREE) {
5144                 if (cache->ro) {
5145                         ret = -EAGAIN;
5146                 } else {
5147                         cache->reserved += num_bytes;
5148                         space_info->bytes_reserved += num_bytes;
5149                         if (reserve == RESERVE_ALLOC) {
5150                                 trace_btrfs_space_reservation(cache->fs_info,
5151                                                 "space_info", space_info->flags,
5152                                                 num_bytes, 0);
5153                                 space_info->bytes_may_use -= num_bytes;
5154                         }
5155                 }
5156         } else {
5157                 if (cache->ro)
5158                         space_info->bytes_readonly += num_bytes;
5159                 cache->reserved -= num_bytes;
5160                 space_info->bytes_reserved -= num_bytes;
5161                 space_info->reservation_progress++;
5162         }
5163         spin_unlock(&cache->lock);
5164         spin_unlock(&space_info->lock);
5165         return ret;
5166 }
5167
5168 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5169                                 struct btrfs_root *root)
5170 {
5171         struct btrfs_fs_info *fs_info = root->fs_info;
5172         struct btrfs_caching_control *next;
5173         struct btrfs_caching_control *caching_ctl;
5174         struct btrfs_block_group_cache *cache;
5175
5176         down_write(&fs_info->extent_commit_sem);
5177
5178         list_for_each_entry_safe(caching_ctl, next,
5179                                  &fs_info->caching_block_groups, list) {
5180                 cache = caching_ctl->block_group;
5181                 if (block_group_cache_done(cache)) {
5182                         cache->last_byte_to_unpin = (u64)-1;
5183                         list_del_init(&caching_ctl->list);
5184                         put_caching_control(caching_ctl);
5185                 } else {
5186                         cache->last_byte_to_unpin = caching_ctl->progress;
5187                 }
5188         }
5189
5190         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5191                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5192         else
5193                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5194
5195         up_write(&fs_info->extent_commit_sem);
5196
5197         update_global_block_rsv(fs_info);
5198 }
5199
5200 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5201 {
5202         struct btrfs_fs_info *fs_info = root->fs_info;
5203         struct btrfs_block_group_cache *cache = NULL;
5204         struct btrfs_space_info *space_info;
5205         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5206         u64 len;
5207         bool readonly;
5208
5209         while (start <= end) {
5210                 readonly = false;
5211                 if (!cache ||
5212                     start >= cache->key.objectid + cache->key.offset) {
5213                         if (cache)
5214                                 btrfs_put_block_group(cache);
5215                         cache = btrfs_lookup_block_group(fs_info, start);
5216                         BUG_ON(!cache); /* Logic error */
5217                 }
5218
5219                 len = cache->key.objectid + cache->key.offset - start;
5220                 len = min(len, end + 1 - start);
5221
5222                 if (start < cache->last_byte_to_unpin) {
5223                         len = min(len, cache->last_byte_to_unpin - start);
5224                         btrfs_add_free_space(cache, start, len);
5225                 }
5226
5227                 start += len;
5228                 space_info = cache->space_info;
5229
5230                 spin_lock(&space_info->lock);
5231                 spin_lock(&cache->lock);
5232                 cache->pinned -= len;
5233                 space_info->bytes_pinned -= len;
5234                 if (cache->ro) {
5235                         space_info->bytes_readonly += len;
5236                         readonly = true;
5237                 }
5238                 spin_unlock(&cache->lock);
5239                 if (!readonly && global_rsv->space_info == space_info) {
5240                         spin_lock(&global_rsv->lock);
5241                         if (!global_rsv->full) {
5242                                 len = min(len, global_rsv->size -
5243                                           global_rsv->reserved);
5244                                 global_rsv->reserved += len;
5245                                 space_info->bytes_may_use += len;
5246                                 if (global_rsv->reserved >= global_rsv->size)
5247                                         global_rsv->full = 1;
5248                         }
5249                         spin_unlock(&global_rsv->lock);
5250                 }
5251                 spin_unlock(&space_info->lock);
5252         }
5253
5254         if (cache)
5255                 btrfs_put_block_group(cache);
5256         return 0;
5257 }
5258
5259 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5260                                struct btrfs_root *root)
5261 {
5262         struct btrfs_fs_info *fs_info = root->fs_info;
5263         struct extent_io_tree *unpin;
5264         u64 start;
5265         u64 end;
5266         int ret;
5267
5268         if (trans->aborted)
5269                 return 0;
5270
5271         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5272                 unpin = &fs_info->freed_extents[1];
5273         else
5274                 unpin = &fs_info->freed_extents[0];
5275
5276         while (1) {
5277                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5278                                             EXTENT_DIRTY, NULL);
5279                 if (ret)
5280                         break;
5281
5282                 if (btrfs_test_opt(root, DISCARD))
5283                         ret = btrfs_discard_extent(root, start,
5284                                                    end + 1 - start, NULL);
5285
5286                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5287                 unpin_extent_range(root, start, end);
5288                 cond_resched();
5289         }
5290
5291         return 0;
5292 }
5293
5294 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5295                                 struct btrfs_root *root,
5296                                 u64 bytenr, u64 num_bytes, u64 parent,
5297                                 u64 root_objectid, u64 owner_objectid,
5298                                 u64 owner_offset, int refs_to_drop,
5299                                 struct btrfs_delayed_extent_op *extent_op)
5300 {
5301         struct btrfs_key key;
5302         struct btrfs_path *path;
5303         struct btrfs_fs_info *info = root->fs_info;
5304         struct btrfs_root *extent_root = info->extent_root;
5305         struct extent_buffer *leaf;
5306         struct btrfs_extent_item *ei;
5307         struct btrfs_extent_inline_ref *iref;
5308         int ret;
5309         int is_data;
5310         int extent_slot = 0;
5311         int found_extent = 0;
5312         int num_to_del = 1;
5313         u32 item_size;
5314         u64 refs;
5315
5316         path = btrfs_alloc_path();
5317         if (!path)
5318                 return -ENOMEM;
5319
5320         path->reada = 1;
5321         path->leave_spinning = 1;
5322
5323         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5324         BUG_ON(!is_data && refs_to_drop != 1);
5325
5326         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5327                                     bytenr, num_bytes, parent,
5328                                     root_objectid, owner_objectid,
5329                                     owner_offset);
5330         if (ret == 0) {
5331                 extent_slot = path->slots[0];
5332                 while (extent_slot >= 0) {
5333                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5334                                               extent_slot);
5335                         if (key.objectid != bytenr)
5336                                 break;
5337                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5338                             key.offset == num_bytes) {
5339                                 found_extent = 1;
5340                                 break;
5341                         }
5342                         if (path->slots[0] - extent_slot > 5)
5343                                 break;
5344                         extent_slot--;
5345                 }
5346 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5347                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5348                 if (found_extent && item_size < sizeof(*ei))
5349                         found_extent = 0;
5350 #endif
5351                 if (!found_extent) {
5352                         BUG_ON(iref);
5353                         ret = remove_extent_backref(trans, extent_root, path,
5354                                                     NULL, refs_to_drop,
5355                                                     is_data);
5356                         if (ret) {
5357                                 btrfs_abort_transaction(trans, extent_root, ret);
5358                                 goto out;
5359                         }
5360                         btrfs_release_path(path);
5361                         path->leave_spinning = 1;
5362
5363                         key.objectid = bytenr;
5364                         key.type = BTRFS_EXTENT_ITEM_KEY;
5365                         key.offset = num_bytes;
5366
5367                         ret = btrfs_search_slot(trans, extent_root,
5368                                                 &key, path, -1, 1);
5369                         if (ret) {
5370                                 printk(KERN_ERR "umm, got %d back from search"
5371                                        ", was looking for %llu\n", ret,
5372                                        (unsigned long long)bytenr);
5373                                 if (ret > 0)
5374                                         btrfs_print_leaf(extent_root,
5375                                                          path->nodes[0]);
5376                         }
5377                         if (ret < 0) {
5378                                 btrfs_abort_transaction(trans, extent_root, ret);
5379                                 goto out;
5380                         }
5381                         extent_slot = path->slots[0];
5382                 }
5383         } else if (ret == -ENOENT) {
5384                 btrfs_print_leaf(extent_root, path->nodes[0]);
5385                 WARN_ON(1);
5386                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5387                        "parent %llu root %llu  owner %llu offset %llu\n",
5388                        (unsigned long long)bytenr,
5389                        (unsigned long long)parent,
5390                        (unsigned long long)root_objectid,
5391                        (unsigned long long)owner_objectid,
5392                        (unsigned long long)owner_offset);
5393         } else {
5394                 btrfs_abort_transaction(trans, extent_root, ret);
5395                 goto out;
5396         }
5397
5398         leaf = path->nodes[0];
5399         item_size = btrfs_item_size_nr(leaf, extent_slot);
5400 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5401         if (item_size < sizeof(*ei)) {
5402                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5403                 ret = convert_extent_item_v0(trans, extent_root, path,
5404                                              owner_objectid, 0);
5405                 if (ret < 0) {
5406                         btrfs_abort_transaction(trans, extent_root, ret);
5407                         goto out;
5408                 }
5409
5410                 btrfs_release_path(path);
5411                 path->leave_spinning = 1;
5412
5413                 key.objectid = bytenr;
5414                 key.type = BTRFS_EXTENT_ITEM_KEY;
5415                 key.offset = num_bytes;
5416
5417                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5418                                         -1, 1);
5419                 if (ret) {
5420                         printk(KERN_ERR "umm, got %d back from search"
5421                                ", was looking for %llu\n", ret,
5422                                (unsigned long long)bytenr);
5423                         btrfs_print_leaf(extent_root, path->nodes[0]);
5424                 }
5425                 if (ret < 0) {
5426                         btrfs_abort_transaction(trans, extent_root, ret);
5427                         goto out;
5428                 }
5429
5430                 extent_slot = path->slots[0];
5431                 leaf = path->nodes[0];
5432                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5433         }
5434 #endif
5435         BUG_ON(item_size < sizeof(*ei));
5436         ei = btrfs_item_ptr(leaf, extent_slot,
5437                             struct btrfs_extent_item);
5438         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5439                 struct btrfs_tree_block_info *bi;
5440                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5441                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5442                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5443         }
5444
5445         refs = btrfs_extent_refs(leaf, ei);
5446         BUG_ON(refs < refs_to_drop);
5447         refs -= refs_to_drop;
5448
5449         if (refs > 0) {
5450                 if (extent_op)
5451                         __run_delayed_extent_op(extent_op, leaf, ei);
5452                 /*
5453                  * In the case of inline back ref, reference count will
5454                  * be updated by remove_extent_backref
5455                  */
5456                 if (iref) {
5457                         BUG_ON(!found_extent);
5458                 } else {
5459                         btrfs_set_extent_refs(leaf, ei, refs);
5460                         btrfs_mark_buffer_dirty(leaf);
5461                 }
5462                 if (found_extent) {
5463                         ret = remove_extent_backref(trans, extent_root, path,
5464                                                     iref, refs_to_drop,
5465                                                     is_data);
5466                         if (ret) {
5467                                 btrfs_abort_transaction(trans, extent_root, ret);
5468                                 goto out;
5469                         }
5470                 }
5471         } else {
5472                 if (found_extent) {
5473                         BUG_ON(is_data && refs_to_drop !=
5474                                extent_data_ref_count(root, path, iref));
5475                         if (iref) {
5476                                 BUG_ON(path->slots[0] != extent_slot);
5477                         } else {
5478                                 BUG_ON(path->slots[0] != extent_slot + 1);
5479                                 path->slots[0] = extent_slot;
5480                                 num_to_del = 2;
5481                         }
5482                 }
5483
5484                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5485                                       num_to_del);
5486                 if (ret) {
5487                         btrfs_abort_transaction(trans, extent_root, ret);
5488                         goto out;
5489                 }
5490                 btrfs_release_path(path);
5491
5492                 if (is_data) {
5493                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5494                         if (ret) {
5495                                 btrfs_abort_transaction(trans, extent_root, ret);
5496                                 goto out;
5497                         }
5498                 }
5499
5500                 ret = update_block_group(root, bytenr, num_bytes, 0);
5501                 if (ret) {
5502                         btrfs_abort_transaction(trans, extent_root, ret);
5503                         goto out;
5504                 }
5505         }
5506 out:
5507         btrfs_free_path(path);
5508         return ret;
5509 }
5510
5511 /*
5512  * when we free an block, it is possible (and likely) that we free the last
5513  * delayed ref for that extent as well.  This searches the delayed ref tree for
5514  * a given extent, and if there are no other delayed refs to be processed, it
5515  * removes it from the tree.
5516  */
5517 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5518                                       struct btrfs_root *root, u64 bytenr)
5519 {
5520         struct btrfs_delayed_ref_head *head;
5521         struct btrfs_delayed_ref_root *delayed_refs;
5522         struct btrfs_delayed_ref_node *ref;
5523         struct rb_node *node;
5524         int ret = 0;
5525
5526         delayed_refs = &trans->transaction->delayed_refs;
5527         spin_lock(&delayed_refs->lock);
5528         head = btrfs_find_delayed_ref_head(trans, bytenr);
5529         if (!head)
5530                 goto out;
5531
5532         node = rb_prev(&head->node.rb_node);
5533         if (!node)
5534                 goto out;
5535
5536         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5537
5538         /* there are still entries for this ref, we can't drop it */
5539         if (ref->bytenr == bytenr)
5540                 goto out;
5541
5542         if (head->extent_op) {
5543                 if (!head->must_insert_reserved)
5544                         goto out;
5545                 btrfs_free_delayed_extent_op(head->extent_op);
5546                 head->extent_op = NULL;
5547         }
5548
5549         /*
5550          * waiting for the lock here would deadlock.  If someone else has it
5551          * locked they are already in the process of dropping it anyway
5552          */
5553         if (!mutex_trylock(&head->mutex))
5554                 goto out;
5555
5556         /*
5557          * at this point we have a head with no other entries.  Go
5558          * ahead and process it.
5559          */
5560         head->node.in_tree = 0;
5561         rb_erase(&head->node.rb_node, &delayed_refs->root);
5562
5563         delayed_refs->num_entries--;
5564
5565         /*
5566          * we don't take a ref on the node because we're removing it from the
5567          * tree, so we just steal the ref the tree was holding.
5568          */
5569         delayed_refs->num_heads--;
5570         if (list_empty(&head->cluster))
5571                 delayed_refs->num_heads_ready--;
5572
5573         list_del_init(&head->cluster);
5574         spin_unlock(&delayed_refs->lock);
5575
5576         BUG_ON(head->extent_op);
5577         if (head->must_insert_reserved)
5578                 ret = 1;
5579
5580         mutex_unlock(&head->mutex);
5581         btrfs_put_delayed_ref(&head->node);
5582         return ret;
5583 out:
5584         spin_unlock(&delayed_refs->lock);
5585         return 0;
5586 }
5587
5588 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5589                            struct btrfs_root *root,
5590                            struct extent_buffer *buf,
5591                            u64 parent, int last_ref)
5592 {
5593         struct btrfs_block_group_cache *cache = NULL;
5594         int ret;
5595
5596         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5597                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5598                                         buf->start, buf->len,
5599                                         parent, root->root_key.objectid,
5600                                         btrfs_header_level(buf),
5601                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5602                 BUG_ON(ret); /* -ENOMEM */
5603         }
5604
5605         if (!last_ref)
5606                 return;
5607
5608         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5609
5610         if (btrfs_header_generation(buf) == trans->transid) {
5611                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5612                         ret = check_ref_cleanup(trans, root, buf->start);
5613                         if (!ret)
5614                                 goto out;
5615                 }
5616
5617                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5618                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5619                         goto out;
5620                 }
5621
5622                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5623
5624                 btrfs_add_free_space(cache, buf->start, buf->len);
5625                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5626         }
5627 out:
5628         /*
5629          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5630          * anymore.
5631          */
5632         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5633         btrfs_put_block_group(cache);
5634 }
5635
5636 /* Can return -ENOMEM */
5637 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5638                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5639                       u64 owner, u64 offset, int for_cow)
5640 {
5641         int ret;
5642         struct btrfs_fs_info *fs_info = root->fs_info;
5643
5644         /*
5645          * tree log blocks never actually go into the extent allocation
5646          * tree, just update pinning info and exit early.
5647          */
5648         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5649                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5650                 /* unlocks the pinned mutex */
5651                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5652                 ret = 0;
5653         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5654                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5655                                         num_bytes,
5656                                         parent, root_objectid, (int)owner,
5657                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5658         } else {
5659                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5660                                                 num_bytes,
5661                                                 parent, root_objectid, owner,
5662                                                 offset, BTRFS_DROP_DELAYED_REF,
5663                                                 NULL, for_cow);
5664         }
5665         return ret;
5666 }
5667
5668 static u64 stripe_align(struct btrfs_root *root,
5669                         struct btrfs_block_group_cache *cache,
5670                         u64 val, u64 num_bytes)
5671 {
5672         u64 ret = ALIGN(val, root->stripesize);
5673         return ret;
5674 }
5675
5676 /*
5677  * when we wait for progress in the block group caching, its because
5678  * our allocation attempt failed at least once.  So, we must sleep
5679  * and let some progress happen before we try again.
5680  *
5681  * This function will sleep at least once waiting for new free space to
5682  * show up, and then it will check the block group free space numbers
5683  * for our min num_bytes.  Another option is to have it go ahead
5684  * and look in the rbtree for a free extent of a given size, but this
5685  * is a good start.
5686  */
5687 static noinline int
5688 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5689                                 u64 num_bytes)
5690 {
5691         struct btrfs_caching_control *caching_ctl;
5692
5693         caching_ctl = get_caching_control(cache);
5694         if (!caching_ctl)
5695                 return 0;
5696
5697         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5698                    (cache->free_space_ctl->free_space >= num_bytes));
5699
5700         put_caching_control(caching_ctl);
5701         return 0;
5702 }
5703
5704 static noinline int
5705 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5706 {
5707         struct btrfs_caching_control *caching_ctl;
5708
5709         caching_ctl = get_caching_control(cache);
5710         if (!caching_ctl)
5711                 return 0;
5712
5713         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5714
5715         put_caching_control(caching_ctl);
5716         return 0;
5717 }
5718
5719 int __get_raid_index(u64 flags)
5720 {
5721         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5722                 return BTRFS_RAID_RAID10;
5723         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5724                 return BTRFS_RAID_RAID1;
5725         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5726                 return BTRFS_RAID_DUP;
5727         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5728                 return BTRFS_RAID_RAID0;
5729         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
5730                 return BTRFS_RAID_RAID5;
5731         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
5732                 return BTRFS_RAID_RAID6;
5733
5734         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
5735 }
5736
5737 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5738 {
5739         return __get_raid_index(cache->flags);
5740 }
5741
5742 enum btrfs_loop_type {
5743         LOOP_CACHING_NOWAIT = 0,
5744         LOOP_CACHING_WAIT = 1,
5745         LOOP_ALLOC_CHUNK = 2,
5746         LOOP_NO_EMPTY_SIZE = 3,
5747 };
5748
5749 /*
5750  * walks the btree of allocated extents and find a hole of a given size.
5751  * The key ins is changed to record the hole:
5752  * ins->objectid == block start
5753  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5754  * ins->offset == number of blocks
5755  * Any available blocks before search_start are skipped.
5756  */
5757 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5758                                      struct btrfs_root *orig_root,
5759                                      u64 num_bytes, u64 empty_size,
5760                                      u64 hint_byte, struct btrfs_key *ins,
5761                                      u64 data)
5762 {
5763         int ret = 0;
5764         struct btrfs_root *root = orig_root->fs_info->extent_root;
5765         struct btrfs_free_cluster *last_ptr = NULL;
5766         struct btrfs_block_group_cache *block_group = NULL;
5767         struct btrfs_block_group_cache *used_block_group;
5768         u64 search_start = 0;
5769         int empty_cluster = 2 * 1024 * 1024;
5770         struct btrfs_space_info *space_info;
5771         int loop = 0;
5772         int index = __get_raid_index(data);
5773         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5774                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5775         bool found_uncached_bg = false;
5776         bool failed_cluster_refill = false;
5777         bool failed_alloc = false;
5778         bool use_cluster = true;
5779         bool have_caching_bg = false;
5780
5781         WARN_ON(num_bytes < root->sectorsize);
5782         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5783         ins->objectid = 0;
5784         ins->offset = 0;
5785
5786         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5787
5788         space_info = __find_space_info(root->fs_info, data);
5789         if (!space_info) {
5790                 printk(KERN_ERR "No space info for %llu\n", data);
5791                 return -ENOSPC;
5792         }
5793
5794         /*
5795          * If the space info is for both data and metadata it means we have a
5796          * small filesystem and we can't use the clustering stuff.
5797          */
5798         if (btrfs_mixed_space_info(space_info))
5799                 use_cluster = false;
5800
5801         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5802                 last_ptr = &root->fs_info->meta_alloc_cluster;
5803                 if (!btrfs_test_opt(root, SSD))
5804                         empty_cluster = 64 * 1024;
5805         }
5806
5807         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5808             btrfs_test_opt(root, SSD)) {
5809                 last_ptr = &root->fs_info->data_alloc_cluster;
5810         }
5811
5812         if (last_ptr) {
5813                 spin_lock(&last_ptr->lock);
5814                 if (last_ptr->block_group)
5815                         hint_byte = last_ptr->window_start;
5816                 spin_unlock(&last_ptr->lock);
5817         }
5818
5819         search_start = max(search_start, first_logical_byte(root, 0));
5820         search_start = max(search_start, hint_byte);
5821
5822         if (!last_ptr)
5823                 empty_cluster = 0;
5824
5825         if (search_start == hint_byte) {
5826                 block_group = btrfs_lookup_block_group(root->fs_info,
5827                                                        search_start);
5828                 used_block_group = block_group;
5829                 /*
5830                  * we don't want to use the block group if it doesn't match our
5831                  * allocation bits, or if its not cached.
5832                  *
5833                  * However if we are re-searching with an ideal block group
5834                  * picked out then we don't care that the block group is cached.
5835                  */
5836                 if (block_group && block_group_bits(block_group, data) &&
5837                     block_group->cached != BTRFS_CACHE_NO) {
5838                         down_read(&space_info->groups_sem);
5839                         if (list_empty(&block_group->list) ||
5840                             block_group->ro) {
5841                                 /*
5842                                  * someone is removing this block group,
5843                                  * we can't jump into the have_block_group
5844                                  * target because our list pointers are not
5845                                  * valid
5846                                  */
5847                                 btrfs_put_block_group(block_group);
5848                                 up_read(&space_info->groups_sem);
5849                         } else {
5850                                 index = get_block_group_index(block_group);
5851                                 goto have_block_group;
5852                         }
5853                 } else if (block_group) {
5854                         btrfs_put_block_group(block_group);
5855                 }
5856         }
5857 search:
5858         have_caching_bg = false;
5859         down_read(&space_info->groups_sem);
5860         list_for_each_entry(block_group, &space_info->block_groups[index],
5861                             list) {
5862                 u64 offset;
5863                 int cached;
5864
5865                 used_block_group = block_group;
5866                 btrfs_get_block_group(block_group);
5867                 search_start = block_group->key.objectid;
5868
5869                 /*
5870                  * this can happen if we end up cycling through all the
5871                  * raid types, but we want to make sure we only allocate
5872                  * for the proper type.
5873                  */
5874                 if (!block_group_bits(block_group, data)) {
5875                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5876                                 BTRFS_BLOCK_GROUP_RAID1 |
5877                                 BTRFS_BLOCK_GROUP_RAID5 |
5878                                 BTRFS_BLOCK_GROUP_RAID6 |
5879                                 BTRFS_BLOCK_GROUP_RAID10;
5880
5881                         /*
5882                          * if they asked for extra copies and this block group
5883                          * doesn't provide them, bail.  This does allow us to
5884                          * fill raid0 from raid1.
5885                          */
5886                         if ((data & extra) && !(block_group->flags & extra))
5887                                 goto loop;
5888                 }
5889
5890 have_block_group:
5891                 cached = block_group_cache_done(block_group);
5892                 if (unlikely(!cached)) {
5893                         found_uncached_bg = true;
5894                         ret = cache_block_group(block_group, 0);
5895                         BUG_ON(ret < 0);
5896                         ret = 0;
5897                 }
5898
5899                 if (unlikely(block_group->ro))
5900                         goto loop;
5901
5902                 /*
5903                  * Ok we want to try and use the cluster allocator, so
5904                  * lets look there
5905                  */
5906                 if (last_ptr) {
5907                         unsigned long aligned_cluster;
5908                         /*
5909                          * the refill lock keeps out other
5910                          * people trying to start a new cluster
5911                          */
5912                         spin_lock(&last_ptr->refill_lock);
5913                         used_block_group = last_ptr->block_group;
5914                         if (used_block_group != block_group &&
5915                             (!used_block_group ||
5916                              used_block_group->ro ||
5917                              !block_group_bits(used_block_group, data))) {
5918                                 used_block_group = block_group;
5919                                 goto refill_cluster;
5920                         }
5921
5922                         if (used_block_group != block_group)
5923                                 btrfs_get_block_group(used_block_group);
5924
5925                         offset = btrfs_alloc_from_cluster(used_block_group,
5926                           last_ptr, num_bytes, used_block_group->key.objectid);
5927                         if (offset) {
5928                                 /* we have a block, we're done */
5929                                 spin_unlock(&last_ptr->refill_lock);
5930                                 trace_btrfs_reserve_extent_cluster(root,
5931                                         block_group, search_start, num_bytes);
5932                                 goto checks;
5933                         }
5934
5935                         WARN_ON(last_ptr->block_group != used_block_group);
5936                         if (used_block_group != block_group) {
5937                                 btrfs_put_block_group(used_block_group);
5938                                 used_block_group = block_group;
5939                         }
5940 refill_cluster:
5941                         BUG_ON(used_block_group != block_group);
5942                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5943                          * set up a new clusters, so lets just skip it
5944                          * and let the allocator find whatever block
5945                          * it can find.  If we reach this point, we
5946                          * will have tried the cluster allocator
5947                          * plenty of times and not have found
5948                          * anything, so we are likely way too
5949                          * fragmented for the clustering stuff to find
5950                          * anything.
5951                          *
5952                          * However, if the cluster is taken from the
5953                          * current block group, release the cluster
5954                          * first, so that we stand a better chance of
5955                          * succeeding in the unclustered
5956                          * allocation.  */
5957                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5958                             last_ptr->block_group != block_group) {
5959                                 spin_unlock(&last_ptr->refill_lock);
5960                                 goto unclustered_alloc;
5961                         }
5962
5963                         /*
5964                          * this cluster didn't work out, free it and
5965                          * start over
5966                          */
5967                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5968
5969                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5970                                 spin_unlock(&last_ptr->refill_lock);
5971                                 goto unclustered_alloc;
5972                         }
5973
5974                         aligned_cluster = max_t(unsigned long,
5975                                                 empty_cluster + empty_size,
5976                                               block_group->full_stripe_len);
5977
5978                         /* allocate a cluster in this block group */
5979                         ret = btrfs_find_space_cluster(trans, root,
5980                                                block_group, last_ptr,
5981                                                search_start, num_bytes,
5982                                                aligned_cluster);
5983                         if (ret == 0) {
5984                                 /*
5985                                  * now pull our allocation out of this
5986                                  * cluster
5987                                  */
5988                                 offset = btrfs_alloc_from_cluster(block_group,
5989                                                   last_ptr, num_bytes,
5990                                                   search_start);
5991                                 if (offset) {
5992                                         /* we found one, proceed */
5993                                         spin_unlock(&last_ptr->refill_lock);
5994                                         trace_btrfs_reserve_extent_cluster(root,
5995                                                 block_group, search_start,
5996                                                 num_bytes);
5997                                         goto checks;
5998                                 }
5999                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6000                                    && !failed_cluster_refill) {
6001                                 spin_unlock(&last_ptr->refill_lock);
6002
6003                                 failed_cluster_refill = true;
6004                                 wait_block_group_cache_progress(block_group,
6005                                        num_bytes + empty_cluster + empty_size);
6006                                 goto have_block_group;
6007                         }
6008
6009                         /*
6010                          * at this point we either didn't find a cluster
6011                          * or we weren't able to allocate a block from our
6012                          * cluster.  Free the cluster we've been trying
6013                          * to use, and go to the next block group
6014                          */
6015                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6016                         spin_unlock(&last_ptr->refill_lock);
6017                         goto loop;
6018                 }
6019
6020 unclustered_alloc:
6021                 spin_lock(&block_group->free_space_ctl->tree_lock);
6022                 if (cached &&
6023                     block_group->free_space_ctl->free_space <
6024                     num_bytes + empty_cluster + empty_size) {
6025                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6026                         goto loop;
6027                 }
6028                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6029
6030                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6031                                                     num_bytes, empty_size);
6032                 /*
6033                  * If we didn't find a chunk, and we haven't failed on this
6034                  * block group before, and this block group is in the middle of
6035                  * caching and we are ok with waiting, then go ahead and wait
6036                  * for progress to be made, and set failed_alloc to true.
6037                  *
6038                  * If failed_alloc is true then we've already waited on this
6039                  * block group once and should move on to the next block group.
6040                  */
6041                 if (!offset && !failed_alloc && !cached &&
6042                     loop > LOOP_CACHING_NOWAIT) {
6043                         wait_block_group_cache_progress(block_group,
6044                                                 num_bytes + empty_size);
6045                         failed_alloc = true;
6046                         goto have_block_group;
6047                 } else if (!offset) {
6048                         if (!cached)
6049                                 have_caching_bg = true;
6050                         goto loop;
6051                 }
6052 checks:
6053                 search_start = stripe_align(root, used_block_group,
6054                                             offset, num_bytes);
6055
6056                 /* move on to the next group */
6057                 if (search_start + num_bytes >
6058                     used_block_group->key.objectid + used_block_group->key.offset) {
6059                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6060                         goto loop;
6061                 }
6062
6063                 if (offset < search_start)
6064                         btrfs_add_free_space(used_block_group, offset,
6065                                              search_start - offset);
6066                 BUG_ON(offset > search_start);
6067
6068                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
6069                                                   alloc_type);
6070                 if (ret == -EAGAIN) {
6071                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6072                         goto loop;
6073                 }
6074
6075                 /* we are all good, lets return */
6076                 ins->objectid = search_start;
6077                 ins->offset = num_bytes;
6078
6079                 trace_btrfs_reserve_extent(orig_root, block_group,
6080                                            search_start, num_bytes);
6081                 if (used_block_group != block_group)
6082                         btrfs_put_block_group(used_block_group);
6083                 btrfs_put_block_group(block_group);
6084                 break;
6085 loop:
6086                 failed_cluster_refill = false;
6087                 failed_alloc = false;
6088                 BUG_ON(index != get_block_group_index(block_group));
6089                 if (used_block_group != block_group)
6090                         btrfs_put_block_group(used_block_group);
6091                 btrfs_put_block_group(block_group);
6092         }
6093         up_read(&space_info->groups_sem);
6094
6095         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6096                 goto search;
6097
6098         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6099                 goto search;
6100
6101         /*
6102          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6103          *                      caching kthreads as we move along
6104          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6105          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6106          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6107          *                      again
6108          */
6109         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6110                 index = 0;
6111                 loop++;
6112                 if (loop == LOOP_ALLOC_CHUNK) {
6113                         ret = do_chunk_alloc(trans, root, data,
6114                                              CHUNK_ALLOC_FORCE);
6115                         /*
6116                          * Do not bail out on ENOSPC since we
6117                          * can do more things.
6118                          */
6119                         if (ret < 0 && ret != -ENOSPC) {
6120                                 btrfs_abort_transaction(trans,
6121                                                         root, ret);
6122                                 goto out;
6123                         }
6124                 }
6125
6126                 if (loop == LOOP_NO_EMPTY_SIZE) {
6127                         empty_size = 0;
6128                         empty_cluster = 0;
6129                 }
6130
6131                 goto search;
6132         } else if (!ins->objectid) {
6133                 ret = -ENOSPC;
6134         } else if (ins->objectid) {
6135                 ret = 0;
6136         }
6137 out:
6138
6139         return ret;
6140 }
6141
6142 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6143                             int dump_block_groups)
6144 {
6145         struct btrfs_block_group_cache *cache;
6146         int index = 0;
6147
6148         spin_lock(&info->lock);
6149         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
6150                (unsigned long long)info->flags,
6151                (unsigned long long)(info->total_bytes - info->bytes_used -
6152                                     info->bytes_pinned - info->bytes_reserved -
6153                                     info->bytes_readonly),
6154                (info->full) ? "" : "not ");
6155         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
6156                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6157                (unsigned long long)info->total_bytes,
6158                (unsigned long long)info->bytes_used,
6159                (unsigned long long)info->bytes_pinned,
6160                (unsigned long long)info->bytes_reserved,
6161                (unsigned long long)info->bytes_may_use,
6162                (unsigned long long)info->bytes_readonly);
6163         spin_unlock(&info->lock);
6164
6165         if (!dump_block_groups)
6166                 return;
6167
6168         down_read(&info->groups_sem);
6169 again:
6170         list_for_each_entry(cache, &info->block_groups[index], list) {
6171                 spin_lock(&cache->lock);
6172                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
6173                        (unsigned long long)cache->key.objectid,
6174                        (unsigned long long)cache->key.offset,
6175                        (unsigned long long)btrfs_block_group_used(&cache->item),
6176                        (unsigned long long)cache->pinned,
6177                        (unsigned long long)cache->reserved,
6178                        cache->ro ? "[readonly]" : "");
6179                 btrfs_dump_free_space(cache, bytes);
6180                 spin_unlock(&cache->lock);
6181         }
6182         if (++index < BTRFS_NR_RAID_TYPES)
6183                 goto again;
6184         up_read(&info->groups_sem);
6185 }
6186
6187 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
6188                          struct btrfs_root *root,
6189                          u64 num_bytes, u64 min_alloc_size,
6190                          u64 empty_size, u64 hint_byte,
6191                          struct btrfs_key *ins, u64 data)
6192 {
6193         bool final_tried = false;
6194         int ret;
6195
6196         data = btrfs_get_alloc_profile(root, data);
6197 again:
6198         WARN_ON(num_bytes < root->sectorsize);
6199         ret = find_free_extent(trans, root, num_bytes, empty_size,
6200                                hint_byte, ins, data);
6201
6202         if (ret == -ENOSPC) {
6203                 if (!final_tried) {
6204                         num_bytes = num_bytes >> 1;
6205                         num_bytes = round_down(num_bytes, root->sectorsize);
6206                         num_bytes = max(num_bytes, min_alloc_size);
6207                         if (num_bytes == min_alloc_size)
6208                                 final_tried = true;
6209                         goto again;
6210                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6211                         struct btrfs_space_info *sinfo;
6212
6213                         sinfo = __find_space_info(root->fs_info, data);
6214                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
6215                                "wanted %llu\n", (unsigned long long)data,
6216                                (unsigned long long)num_bytes);
6217                         if (sinfo)
6218                                 dump_space_info(sinfo, num_bytes, 1);
6219                 }
6220         }
6221
6222         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6223
6224         return ret;
6225 }
6226
6227 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6228                                         u64 start, u64 len, int pin)
6229 {
6230         struct btrfs_block_group_cache *cache;
6231         int ret = 0;
6232
6233         cache = btrfs_lookup_block_group(root->fs_info, start);
6234         if (!cache) {
6235                 printk(KERN_ERR "Unable to find block group for %llu\n",
6236                        (unsigned long long)start);
6237                 return -ENOSPC;
6238         }
6239
6240         if (btrfs_test_opt(root, DISCARD))
6241                 ret = btrfs_discard_extent(root, start, len, NULL);
6242
6243         if (pin)
6244                 pin_down_extent(root, cache, start, len, 1);
6245         else {
6246                 btrfs_add_free_space(cache, start, len);
6247                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6248         }
6249         btrfs_put_block_group(cache);
6250
6251         trace_btrfs_reserved_extent_free(root, start, len);
6252
6253         return ret;
6254 }
6255
6256 int btrfs_free_reserved_extent(struct btrfs_root *root,
6257                                         u64 start, u64 len)
6258 {
6259         return __btrfs_free_reserved_extent(root, start, len, 0);
6260 }
6261
6262 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6263                                        u64 start, u64 len)
6264 {
6265         return __btrfs_free_reserved_extent(root, start, len, 1);
6266 }
6267
6268 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6269                                       struct btrfs_root *root,
6270                                       u64 parent, u64 root_objectid,
6271                                       u64 flags, u64 owner, u64 offset,
6272                                       struct btrfs_key *ins, int ref_mod)
6273 {
6274         int ret;
6275         struct btrfs_fs_info *fs_info = root->fs_info;
6276         struct btrfs_extent_item *extent_item;
6277         struct btrfs_extent_inline_ref *iref;
6278         struct btrfs_path *path;
6279         struct extent_buffer *leaf;
6280         int type;
6281         u32 size;
6282
6283         if (parent > 0)
6284                 type = BTRFS_SHARED_DATA_REF_KEY;
6285         else
6286                 type = BTRFS_EXTENT_DATA_REF_KEY;
6287
6288         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6289
6290         path = btrfs_alloc_path();
6291         if (!path)
6292                 return -ENOMEM;
6293
6294         path->leave_spinning = 1;
6295         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6296                                       ins, size);
6297         if (ret) {
6298                 btrfs_free_path(path);
6299                 return ret;
6300         }
6301
6302         leaf = path->nodes[0];
6303         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6304                                      struct btrfs_extent_item);
6305         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6306         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6307         btrfs_set_extent_flags(leaf, extent_item,
6308                                flags | BTRFS_EXTENT_FLAG_DATA);
6309
6310         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6311         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6312         if (parent > 0) {
6313                 struct btrfs_shared_data_ref *ref;
6314                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6315                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6316                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6317         } else {
6318                 struct btrfs_extent_data_ref *ref;
6319                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6320                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6321                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6322                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6323                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6324         }
6325
6326         btrfs_mark_buffer_dirty(path->nodes[0]);
6327         btrfs_free_path(path);
6328
6329         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6330         if (ret) { /* -ENOENT, logic error */
6331                 printk(KERN_ERR "btrfs update block group failed for %llu "
6332                        "%llu\n", (unsigned long long)ins->objectid,
6333                        (unsigned long long)ins->offset);
6334                 BUG();
6335         }
6336         return ret;
6337 }
6338
6339 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6340                                      struct btrfs_root *root,
6341                                      u64 parent, u64 root_objectid,
6342                                      u64 flags, struct btrfs_disk_key *key,
6343                                      int level, struct btrfs_key *ins)
6344 {
6345         int ret;
6346         struct btrfs_fs_info *fs_info = root->fs_info;
6347         struct btrfs_extent_item *extent_item;
6348         struct btrfs_tree_block_info *block_info;
6349         struct btrfs_extent_inline_ref *iref;
6350         struct btrfs_path *path;
6351         struct extent_buffer *leaf;
6352         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6353
6354         path = btrfs_alloc_path();
6355         if (!path)
6356                 return -ENOMEM;
6357
6358         path->leave_spinning = 1;
6359         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6360                                       ins, size);
6361         if (ret) {
6362                 btrfs_free_path(path);
6363                 return ret;
6364         }
6365
6366         leaf = path->nodes[0];
6367         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6368                                      struct btrfs_extent_item);
6369         btrfs_set_extent_refs(leaf, extent_item, 1);
6370         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6371         btrfs_set_extent_flags(leaf, extent_item,
6372                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6373         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6374
6375         btrfs_set_tree_block_key(leaf, block_info, key);
6376         btrfs_set_tree_block_level(leaf, block_info, level);
6377
6378         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6379         if (parent > 0) {
6380                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6381                 btrfs_set_extent_inline_ref_type(leaf, iref,
6382                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6383                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6384         } else {
6385                 btrfs_set_extent_inline_ref_type(leaf, iref,
6386                                                  BTRFS_TREE_BLOCK_REF_KEY);
6387                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6388         }
6389
6390         btrfs_mark_buffer_dirty(leaf);
6391         btrfs_free_path(path);
6392
6393         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6394         if (ret) { /* -ENOENT, logic error */
6395                 printk(KERN_ERR "btrfs update block group failed for %llu "
6396                        "%llu\n", (unsigned long long)ins->objectid,
6397                        (unsigned long long)ins->offset);
6398                 BUG();
6399         }
6400         return ret;
6401 }
6402
6403 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6404                                      struct btrfs_root *root,
6405                                      u64 root_objectid, u64 owner,
6406                                      u64 offset, struct btrfs_key *ins)
6407 {
6408         int ret;
6409
6410         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6411
6412         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6413                                          ins->offset, 0,
6414                                          root_objectid, owner, offset,
6415                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6416         return ret;
6417 }
6418
6419 /*
6420  * this is used by the tree logging recovery code.  It records that
6421  * an extent has been allocated and makes sure to clear the free
6422  * space cache bits as well
6423  */
6424 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6425                                    struct btrfs_root *root,
6426                                    u64 root_objectid, u64 owner, u64 offset,
6427                                    struct btrfs_key *ins)
6428 {
6429         int ret;
6430         struct btrfs_block_group_cache *block_group;
6431         struct btrfs_caching_control *caching_ctl;
6432         u64 start = ins->objectid;
6433         u64 num_bytes = ins->offset;
6434
6435         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6436         cache_block_group(block_group, 0);
6437         caching_ctl = get_caching_control(block_group);
6438
6439         if (!caching_ctl) {
6440                 BUG_ON(!block_group_cache_done(block_group));
6441                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6442                 BUG_ON(ret); /* -ENOMEM */
6443         } else {
6444                 mutex_lock(&caching_ctl->mutex);
6445
6446                 if (start >= caching_ctl->progress) {
6447                         ret = add_excluded_extent(root, start, num_bytes);
6448                         BUG_ON(ret); /* -ENOMEM */
6449                 } else if (start + num_bytes <= caching_ctl->progress) {
6450                         ret = btrfs_remove_free_space(block_group,
6451                                                       start, num_bytes);
6452                         BUG_ON(ret); /* -ENOMEM */
6453                 } else {
6454                         num_bytes = caching_ctl->progress - start;
6455                         ret = btrfs_remove_free_space(block_group,
6456                                                       start, num_bytes);
6457                         BUG_ON(ret); /* -ENOMEM */
6458
6459                         start = caching_ctl->progress;
6460                         num_bytes = ins->objectid + ins->offset -
6461                                     caching_ctl->progress;
6462                         ret = add_excluded_extent(root, start, num_bytes);
6463                         BUG_ON(ret); /* -ENOMEM */
6464                 }
6465
6466                 mutex_unlock(&caching_ctl->mutex);
6467                 put_caching_control(caching_ctl);
6468         }
6469
6470         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6471                                           RESERVE_ALLOC_NO_ACCOUNT);
6472         BUG_ON(ret); /* logic error */
6473         btrfs_put_block_group(block_group);
6474         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6475                                          0, owner, offset, ins, 1);
6476         return ret;
6477 }
6478
6479 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6480                                             struct btrfs_root *root,
6481                                             u64 bytenr, u32 blocksize,
6482                                             int level)
6483 {
6484         struct extent_buffer *buf;
6485
6486         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6487         if (!buf)
6488                 return ERR_PTR(-ENOMEM);
6489         btrfs_set_header_generation(buf, trans->transid);
6490         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6491         btrfs_tree_lock(buf);
6492         clean_tree_block(trans, root, buf);
6493         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6494
6495         btrfs_set_lock_blocking(buf);
6496         btrfs_set_buffer_uptodate(buf);
6497
6498         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6499                 /*
6500                  * we allow two log transactions at a time, use different
6501                  * EXENT bit to differentiate dirty pages.
6502                  */
6503                 if (root->log_transid % 2 == 0)
6504                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6505                                         buf->start + buf->len - 1, GFP_NOFS);
6506                 else
6507                         set_extent_new(&root->dirty_log_pages, buf->start,
6508                                         buf->start + buf->len - 1, GFP_NOFS);
6509         } else {
6510                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6511                          buf->start + buf->len - 1, GFP_NOFS);
6512         }
6513         trans->blocks_used++;
6514         /* this returns a buffer locked for blocking */
6515         return buf;
6516 }
6517
6518 static struct btrfs_block_rsv *
6519 use_block_rsv(struct btrfs_trans_handle *trans,
6520               struct btrfs_root *root, u32 blocksize)
6521 {
6522         struct btrfs_block_rsv *block_rsv;
6523         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6524         int ret;
6525
6526         block_rsv = get_block_rsv(trans, root);
6527
6528         if (block_rsv->size == 0) {
6529                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6530                                              BTRFS_RESERVE_NO_FLUSH);
6531                 /*
6532                  * If we couldn't reserve metadata bytes try and use some from
6533                  * the global reserve.
6534                  */
6535                 if (ret && block_rsv != global_rsv) {
6536                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6537                         if (!ret)
6538                                 return global_rsv;
6539                         return ERR_PTR(ret);
6540                 } else if (ret) {
6541                         return ERR_PTR(ret);
6542                 }
6543                 return block_rsv;
6544         }
6545
6546         ret = block_rsv_use_bytes(block_rsv, blocksize);
6547         if (!ret)
6548                 return block_rsv;
6549         if (ret && !block_rsv->failfast) {
6550                 if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6551                         static DEFINE_RATELIMIT_STATE(_rs,
6552                                         DEFAULT_RATELIMIT_INTERVAL * 10,
6553                                         /*DEFAULT_RATELIMIT_BURST*/ 1);
6554                         if (__ratelimit(&_rs))
6555                                 WARN(1, KERN_DEBUG
6556                                         "btrfs: block rsv returned %d\n", ret);
6557                 }
6558                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6559                                              BTRFS_RESERVE_NO_FLUSH);
6560                 if (!ret) {
6561                         return block_rsv;
6562                 } else if (ret && block_rsv != global_rsv) {
6563                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6564                         if (!ret)
6565                                 return global_rsv;
6566                 }
6567         }
6568
6569         return ERR_PTR(-ENOSPC);
6570 }
6571
6572 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6573                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6574 {
6575         block_rsv_add_bytes(block_rsv, blocksize, 0);
6576         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6577 }
6578
6579 /*
6580  * finds a free extent and does all the dirty work required for allocation
6581  * returns the key for the extent through ins, and a tree buffer for
6582  * the first block of the extent through buf.
6583  *
6584  * returns the tree buffer or NULL.
6585  */
6586 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6587                                         struct btrfs_root *root, u32 blocksize,
6588                                         u64 parent, u64 root_objectid,
6589                                         struct btrfs_disk_key *key, int level,
6590                                         u64 hint, u64 empty_size)
6591 {
6592         struct btrfs_key ins;
6593         struct btrfs_block_rsv *block_rsv;
6594         struct extent_buffer *buf;
6595         u64 flags = 0;
6596         int ret;
6597
6598
6599         block_rsv = use_block_rsv(trans, root, blocksize);
6600         if (IS_ERR(block_rsv))
6601                 return ERR_CAST(block_rsv);
6602
6603         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6604                                    empty_size, hint, &ins, 0);
6605         if (ret) {
6606                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6607                 return ERR_PTR(ret);
6608         }
6609
6610         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6611                                     blocksize, level);
6612         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6613
6614         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6615                 if (parent == 0)
6616                         parent = ins.objectid;
6617                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6618         } else
6619                 BUG_ON(parent > 0);
6620
6621         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6622                 struct btrfs_delayed_extent_op *extent_op;
6623                 extent_op = btrfs_alloc_delayed_extent_op();
6624                 BUG_ON(!extent_op); /* -ENOMEM */
6625                 if (key)
6626                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6627                 else
6628                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6629                 extent_op->flags_to_set = flags;
6630                 extent_op->update_key = 1;
6631                 extent_op->update_flags = 1;
6632                 extent_op->is_data = 0;
6633
6634                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6635                                         ins.objectid,
6636                                         ins.offset, parent, root_objectid,
6637                                         level, BTRFS_ADD_DELAYED_EXTENT,
6638                                         extent_op, 0);
6639                 BUG_ON(ret); /* -ENOMEM */
6640         }
6641         return buf;
6642 }
6643
6644 struct walk_control {
6645         u64 refs[BTRFS_MAX_LEVEL];
6646         u64 flags[BTRFS_MAX_LEVEL];
6647         struct btrfs_key update_progress;
6648         int stage;
6649         int level;
6650         int shared_level;
6651         int update_ref;
6652         int keep_locks;
6653         int reada_slot;
6654         int reada_count;
6655         int for_reloc;
6656 };
6657
6658 #define DROP_REFERENCE  1
6659 #define UPDATE_BACKREF  2
6660
6661 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6662                                      struct btrfs_root *root,
6663                                      struct walk_control *wc,
6664                                      struct btrfs_path *path)
6665 {
6666         u64 bytenr;
6667         u64 generation;
6668         u64 refs;
6669         u64 flags;
6670         u32 nritems;
6671         u32 blocksize;
6672         struct btrfs_key key;
6673         struct extent_buffer *eb;
6674         int ret;
6675         int slot;
6676         int nread = 0;
6677
6678         if (path->slots[wc->level] < wc->reada_slot) {
6679                 wc->reada_count = wc->reada_count * 2 / 3;
6680                 wc->reada_count = max(wc->reada_count, 2);
6681         } else {
6682                 wc->reada_count = wc->reada_count * 3 / 2;
6683                 wc->reada_count = min_t(int, wc->reada_count,
6684                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6685         }
6686
6687         eb = path->nodes[wc->level];
6688         nritems = btrfs_header_nritems(eb);
6689         blocksize = btrfs_level_size(root, wc->level - 1);
6690
6691         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6692                 if (nread >= wc->reada_count)
6693                         break;
6694
6695                 cond_resched();
6696                 bytenr = btrfs_node_blockptr(eb, slot);
6697                 generation = btrfs_node_ptr_generation(eb, slot);
6698
6699                 if (slot == path->slots[wc->level])
6700                         goto reada;
6701
6702                 if (wc->stage == UPDATE_BACKREF &&
6703                     generation <= root->root_key.offset)
6704                         continue;
6705
6706                 /* We don't lock the tree block, it's OK to be racy here */
6707                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6708                                                &refs, &flags);
6709                 /* We don't care about errors in readahead. */
6710                 if (ret < 0)
6711                         continue;
6712                 BUG_ON(refs == 0);
6713
6714                 if (wc->stage == DROP_REFERENCE) {
6715                         if (refs == 1)
6716                                 goto reada;
6717
6718                         if (wc->level == 1 &&
6719                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6720                                 continue;
6721                         if (!wc->update_ref ||
6722                             generation <= root->root_key.offset)
6723                                 continue;
6724                         btrfs_node_key_to_cpu(eb, &key, slot);
6725                         ret = btrfs_comp_cpu_keys(&key,
6726                                                   &wc->update_progress);
6727                         if (ret < 0)
6728                                 continue;
6729                 } else {
6730                         if (wc->level == 1 &&
6731                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6732                                 continue;
6733                 }
6734 reada:
6735                 ret = readahead_tree_block(root, bytenr, blocksize,
6736                                            generation);
6737                 if (ret)
6738                         break;
6739                 nread++;
6740         }
6741         wc->reada_slot = slot;
6742 }
6743
6744 /*
6745  * helper to process tree block while walking down the tree.
6746  *
6747  * when wc->stage == UPDATE_BACKREF, this function updates
6748  * back refs for pointers in the block.
6749  *
6750  * NOTE: return value 1 means we should stop walking down.
6751  */
6752 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6753                                    struct btrfs_root *root,
6754                                    struct btrfs_path *path,
6755                                    struct walk_control *wc, int lookup_info)
6756 {
6757         int level = wc->level;
6758         struct extent_buffer *eb = path->nodes[level];
6759         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6760         int ret;
6761
6762         if (wc->stage == UPDATE_BACKREF &&
6763             btrfs_header_owner(eb) != root->root_key.objectid)
6764                 return 1;
6765
6766         /*
6767          * when reference count of tree block is 1, it won't increase
6768          * again. once full backref flag is set, we never clear it.
6769          */
6770         if (lookup_info &&
6771             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6772              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6773                 BUG_ON(!path->locks[level]);
6774                 ret = btrfs_lookup_extent_info(trans, root,
6775                                                eb->start, eb->len,
6776                                                &wc->refs[level],
6777                                                &wc->flags[level]);
6778                 BUG_ON(ret == -ENOMEM);
6779                 if (ret)
6780                         return ret;
6781                 BUG_ON(wc->refs[level] == 0);
6782         }
6783
6784         if (wc->stage == DROP_REFERENCE) {
6785                 if (wc->refs[level] > 1)
6786                         return 1;
6787
6788                 if (path->locks[level] && !wc->keep_locks) {
6789                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6790                         path->locks[level] = 0;
6791                 }
6792                 return 0;
6793         }
6794
6795         /* wc->stage == UPDATE_BACKREF */
6796         if (!(wc->flags[level] & flag)) {
6797                 BUG_ON(!path->locks[level]);
6798                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6799                 BUG_ON(ret); /* -ENOMEM */
6800                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6801                 BUG_ON(ret); /* -ENOMEM */
6802                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6803                                                   eb->len, flag, 0);
6804                 BUG_ON(ret); /* -ENOMEM */
6805                 wc->flags[level] |= flag;
6806         }
6807
6808         /*
6809          * the block is shared by multiple trees, so it's not good to
6810          * keep the tree lock
6811          */
6812         if (path->locks[level] && level > 0) {
6813                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6814                 path->locks[level] = 0;
6815         }
6816         return 0;
6817 }
6818
6819 /*
6820  * helper to process tree block pointer.
6821  *
6822  * when wc->stage == DROP_REFERENCE, this function checks
6823  * reference count of the block pointed to. if the block
6824  * is shared and we need update back refs for the subtree
6825  * rooted at the block, this function changes wc->stage to
6826  * UPDATE_BACKREF. if the block is shared and there is no
6827  * need to update back, this function drops the reference
6828  * to the block.
6829  *
6830  * NOTE: return value 1 means we should stop walking down.
6831  */
6832 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6833                                  struct btrfs_root *root,
6834                                  struct btrfs_path *path,
6835                                  struct walk_control *wc, int *lookup_info)
6836 {
6837         u64 bytenr;
6838         u64 generation;
6839         u64 parent;
6840         u32 blocksize;
6841         struct btrfs_key key;
6842         struct extent_buffer *next;
6843         int level = wc->level;
6844         int reada = 0;
6845         int ret = 0;
6846
6847         generation = btrfs_node_ptr_generation(path->nodes[level],
6848                                                path->slots[level]);
6849         /*
6850          * if the lower level block was created before the snapshot
6851          * was created, we know there is no need to update back refs
6852          * for the subtree
6853          */
6854         if (wc->stage == UPDATE_BACKREF &&
6855             generation <= root->root_key.offset) {
6856                 *lookup_info = 1;
6857                 return 1;
6858         }
6859
6860         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6861         blocksize = btrfs_level_size(root, level - 1);
6862
6863         next = btrfs_find_tree_block(root, bytenr, blocksize);
6864         if (!next) {
6865                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6866                 if (!next)
6867                         return -ENOMEM;
6868                 reada = 1;
6869         }
6870         btrfs_tree_lock(next);
6871         btrfs_set_lock_blocking(next);
6872
6873         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6874                                        &wc->refs[level - 1],
6875                                        &wc->flags[level - 1]);
6876         if (ret < 0) {
6877                 btrfs_tree_unlock(next);
6878                 return ret;
6879         }
6880
6881         BUG_ON(wc->refs[level - 1] == 0);
6882         *lookup_info = 0;
6883
6884         if (wc->stage == DROP_REFERENCE) {
6885                 if (wc->refs[level - 1] > 1) {
6886                         if (level == 1 &&
6887                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6888                                 goto skip;
6889
6890                         if (!wc->update_ref ||
6891                             generation <= root->root_key.offset)
6892                                 goto skip;
6893
6894                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6895                                               path->slots[level]);
6896                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6897                         if (ret < 0)
6898                                 goto skip;
6899
6900                         wc->stage = UPDATE_BACKREF;
6901                         wc->shared_level = level - 1;
6902                 }
6903         } else {
6904                 if (level == 1 &&
6905                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6906                         goto skip;
6907         }
6908
6909         if (!btrfs_buffer_uptodate(next, generation, 0)) {
6910                 btrfs_tree_unlock(next);
6911                 free_extent_buffer(next);
6912                 next = NULL;
6913                 *lookup_info = 1;
6914         }
6915
6916         if (!next) {
6917                 if (reada && level == 1)
6918                         reada_walk_down(trans, root, wc, path);
6919                 next = read_tree_block(root, bytenr, blocksize, generation);
6920                 if (!next)
6921                         return -EIO;
6922                 btrfs_tree_lock(next);
6923                 btrfs_set_lock_blocking(next);
6924         }
6925
6926         level--;
6927         BUG_ON(level != btrfs_header_level(next));
6928         path->nodes[level] = next;
6929         path->slots[level] = 0;
6930         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6931         wc->level = level;
6932         if (wc->level == 1)
6933                 wc->reada_slot = 0;
6934         return 0;
6935 skip:
6936         wc->refs[level - 1] = 0;
6937         wc->flags[level - 1] = 0;
6938         if (wc->stage == DROP_REFERENCE) {
6939                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6940                         parent = path->nodes[level]->start;
6941                 } else {
6942                         BUG_ON(root->root_key.objectid !=
6943                                btrfs_header_owner(path->nodes[level]));
6944                         parent = 0;
6945                 }
6946
6947                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6948                                 root->root_key.objectid, level - 1, 0, 0);
6949                 BUG_ON(ret); /* -ENOMEM */
6950         }
6951         btrfs_tree_unlock(next);
6952         free_extent_buffer(next);
6953         *lookup_info = 1;
6954         return 1;
6955 }
6956
6957 /*
6958  * helper to process tree block while walking up the tree.
6959  *
6960  * when wc->stage == DROP_REFERENCE, this function drops
6961  * reference count on the block.
6962  *
6963  * when wc->stage == UPDATE_BACKREF, this function changes
6964  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6965  * to UPDATE_BACKREF previously while processing the block.
6966  *
6967  * NOTE: return value 1 means we should stop walking up.
6968  */
6969 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6970                                  struct btrfs_root *root,
6971                                  struct btrfs_path *path,
6972                                  struct walk_control *wc)
6973 {
6974         int ret;
6975         int level = wc->level;
6976         struct extent_buffer *eb = path->nodes[level];
6977         u64 parent = 0;
6978
6979         if (wc->stage == UPDATE_BACKREF) {
6980                 BUG_ON(wc->shared_level < level);
6981                 if (level < wc->shared_level)
6982                         goto out;
6983
6984                 ret = find_next_key(path, level + 1, &wc->update_progress);
6985                 if (ret > 0)
6986                         wc->update_ref = 0;
6987
6988                 wc->stage = DROP_REFERENCE;
6989                 wc->shared_level = -1;
6990                 path->slots[level] = 0;
6991
6992                 /*
6993                  * check reference count again if the block isn't locked.
6994                  * we should start walking down the tree again if reference
6995                  * count is one.
6996                  */
6997                 if (!path->locks[level]) {
6998                         BUG_ON(level == 0);
6999                         btrfs_tree_lock(eb);
7000                         btrfs_set_lock_blocking(eb);
7001                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7002
7003                         ret = btrfs_lookup_extent_info(trans, root,
7004                                                        eb->start, eb->len,
7005                                                        &wc->refs[level],
7006                                                        &wc->flags[level]);
7007                         if (ret < 0) {
7008                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7009                                 path->locks[level] = 0;
7010                                 return ret;
7011                         }
7012                         BUG_ON(wc->refs[level] == 0);
7013                         if (wc->refs[level] == 1) {
7014                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7015                                 path->locks[level] = 0;
7016                                 return 1;
7017                         }
7018                 }
7019         }
7020
7021         /* wc->stage == DROP_REFERENCE */
7022         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7023
7024         if (wc->refs[level] == 1) {
7025                 if (level == 0) {
7026                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7027                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7028                                                     wc->for_reloc);
7029                         else
7030                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7031                                                     wc->for_reloc);
7032                         BUG_ON(ret); /* -ENOMEM */
7033                 }
7034                 /* make block locked assertion in clean_tree_block happy */
7035                 if (!path->locks[level] &&
7036                     btrfs_header_generation(eb) == trans->transid) {
7037                         btrfs_tree_lock(eb);
7038                         btrfs_set_lock_blocking(eb);
7039                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7040                 }
7041                 clean_tree_block(trans, root, eb);
7042         }
7043
7044         if (eb == root->node) {
7045                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7046                         parent = eb->start;
7047                 else
7048                         BUG_ON(root->root_key.objectid !=
7049                                btrfs_header_owner(eb));
7050         } else {
7051                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7052                         parent = path->nodes[level + 1]->start;
7053                 else
7054                         BUG_ON(root->root_key.objectid !=
7055                                btrfs_header_owner(path->nodes[level + 1]));
7056         }
7057
7058         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7059 out:
7060         wc->refs[level] = 0;
7061         wc->flags[level] = 0;
7062         return 0;
7063 }
7064
7065 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7066                                    struct btrfs_root *root,
7067                                    struct btrfs_path *path,
7068                                    struct walk_control *wc)
7069 {
7070         int level = wc->level;
7071         int lookup_info = 1;
7072         int ret;
7073
7074         while (level >= 0) {
7075                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7076                 if (ret > 0)
7077                         break;
7078
7079                 if (level == 0)
7080                         break;
7081
7082                 if (path->slots[level] >=
7083                     btrfs_header_nritems(path->nodes[level]))
7084                         break;
7085
7086                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7087                 if (ret > 0) {
7088                         path->slots[level]++;
7089                         continue;
7090                 } else if (ret < 0)
7091                         return ret;
7092                 level = wc->level;
7093         }
7094         return 0;
7095 }
7096
7097 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7098                                  struct btrfs_root *root,
7099                                  struct btrfs_path *path,
7100                                  struct walk_control *wc, int max_level)
7101 {
7102         int level = wc->level;
7103         int ret;
7104
7105         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7106         while (level < max_level && path->nodes[level]) {
7107                 wc->level = level;
7108                 if (path->slots[level] + 1 <
7109                     btrfs_header_nritems(path->nodes[level])) {
7110                         path->slots[level]++;
7111                         return 0;
7112                 } else {
7113                         ret = walk_up_proc(trans, root, path, wc);
7114                         if (ret > 0)
7115                                 return 0;
7116
7117                         if (path->locks[level]) {
7118                                 btrfs_tree_unlock_rw(path->nodes[level],
7119                                                      path->locks[level]);
7120                                 path->locks[level] = 0;
7121                         }
7122                         free_extent_buffer(path->nodes[level]);
7123                         path->nodes[level] = NULL;
7124                         level++;
7125                 }
7126         }
7127         return 1;
7128 }
7129
7130 /*
7131  * drop a subvolume tree.
7132  *
7133  * this function traverses the tree freeing any blocks that only
7134  * referenced by the tree.
7135  *
7136  * when a shared tree block is found. this function decreases its
7137  * reference count by one. if update_ref is true, this function
7138  * also make sure backrefs for the shared block and all lower level
7139  * blocks are properly updated.
7140  */
7141 int btrfs_drop_snapshot(struct btrfs_root *root,
7142                          struct btrfs_block_rsv *block_rsv, int update_ref,
7143                          int for_reloc)
7144 {
7145         struct btrfs_path *path;
7146         struct btrfs_trans_handle *trans;
7147         struct btrfs_root *tree_root = root->fs_info->tree_root;
7148         struct btrfs_root_item *root_item = &root->root_item;
7149         struct walk_control *wc;
7150         struct btrfs_key key;
7151         int err = 0;
7152         int ret;
7153         int level;
7154
7155         path = btrfs_alloc_path();
7156         if (!path) {
7157                 err = -ENOMEM;
7158                 goto out;
7159         }
7160
7161         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7162         if (!wc) {
7163                 btrfs_free_path(path);
7164                 err = -ENOMEM;
7165                 goto out;
7166         }
7167
7168         trans = btrfs_start_transaction(tree_root, 0);
7169         if (IS_ERR(trans)) {
7170                 err = PTR_ERR(trans);
7171                 goto out_free;
7172         }
7173
7174         if (block_rsv)
7175                 trans->block_rsv = block_rsv;
7176
7177         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7178                 level = btrfs_header_level(root->node);
7179                 path->nodes[level] = btrfs_lock_root_node(root);
7180                 btrfs_set_lock_blocking(path->nodes[level]);
7181                 path->slots[level] = 0;
7182                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7183                 memset(&wc->update_progress, 0,
7184                        sizeof(wc->update_progress));
7185         } else {
7186                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7187                 memcpy(&wc->update_progress, &key,
7188                        sizeof(wc->update_progress));
7189
7190                 level = root_item->drop_level;
7191                 BUG_ON(level == 0);
7192                 path->lowest_level = level;
7193                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7194                 path->lowest_level = 0;
7195                 if (ret < 0) {
7196                         err = ret;
7197                         goto out_end_trans;
7198                 }
7199                 WARN_ON(ret > 0);
7200
7201                 /*
7202                  * unlock our path, this is safe because only this
7203                  * function is allowed to delete this snapshot
7204                  */
7205                 btrfs_unlock_up_safe(path, 0);
7206
7207                 level = btrfs_header_level(root->node);
7208                 while (1) {
7209                         btrfs_tree_lock(path->nodes[level]);
7210                         btrfs_set_lock_blocking(path->nodes[level]);
7211
7212                         ret = btrfs_lookup_extent_info(trans, root,
7213                                                 path->nodes[level]->start,
7214                                                 path->nodes[level]->len,
7215                                                 &wc->refs[level],
7216                                                 &wc->flags[level]);
7217                         if (ret < 0) {
7218                                 err = ret;
7219                                 goto out_end_trans;
7220                         }
7221                         BUG_ON(wc->refs[level] == 0);
7222
7223                         if (level == root_item->drop_level)
7224                                 break;
7225
7226                         btrfs_tree_unlock(path->nodes[level]);
7227                         WARN_ON(wc->refs[level] != 1);
7228                         level--;
7229                 }
7230         }
7231
7232         wc->level = level;
7233         wc->shared_level = -1;
7234         wc->stage = DROP_REFERENCE;
7235         wc->update_ref = update_ref;
7236         wc->keep_locks = 0;
7237         wc->for_reloc = for_reloc;
7238         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7239
7240         while (1) {
7241                 ret = walk_down_tree(trans, root, path, wc);
7242                 if (ret < 0) {
7243                         err = ret;
7244                         break;
7245                 }
7246
7247                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7248                 if (ret < 0) {
7249                         err = ret;
7250                         break;
7251                 }
7252
7253                 if (ret > 0) {
7254                         BUG_ON(wc->stage != DROP_REFERENCE);
7255                         break;
7256                 }
7257
7258                 if (wc->stage == DROP_REFERENCE) {
7259                         level = wc->level;
7260                         btrfs_node_key(path->nodes[level],
7261                                        &root_item->drop_progress,
7262                                        path->slots[level]);
7263                         root_item->drop_level = level;
7264                 }
7265
7266                 BUG_ON(wc->level == 0);
7267                 if (btrfs_should_end_transaction(trans, tree_root)) {
7268                         ret = btrfs_update_root(trans, tree_root,
7269                                                 &root->root_key,
7270                                                 root_item);
7271                         if (ret) {
7272                                 btrfs_abort_transaction(trans, tree_root, ret);
7273                                 err = ret;
7274                                 goto out_end_trans;
7275                         }
7276
7277                         btrfs_end_transaction_throttle(trans, tree_root);
7278                         trans = btrfs_start_transaction(tree_root, 0);
7279                         if (IS_ERR(trans)) {
7280                                 err = PTR_ERR(trans);
7281                                 goto out_free;
7282                         }
7283                         if (block_rsv)
7284                                 trans->block_rsv = block_rsv;
7285                 }
7286         }
7287         btrfs_release_path(path);
7288         if (err)
7289                 goto out_end_trans;
7290
7291         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7292         if (ret) {
7293                 btrfs_abort_transaction(trans, tree_root, ret);
7294                 goto out_end_trans;
7295         }
7296
7297         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7298                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
7299                                            NULL, NULL);
7300                 if (ret < 0) {
7301                         btrfs_abort_transaction(trans, tree_root, ret);
7302                         err = ret;
7303                         goto out_end_trans;
7304                 } else if (ret > 0) {
7305                         /* if we fail to delete the orphan item this time
7306                          * around, it'll get picked up the next time.
7307                          *
7308                          * The most common failure here is just -ENOENT.
7309                          */
7310                         btrfs_del_orphan_item(trans, tree_root,
7311                                               root->root_key.objectid);
7312                 }
7313         }
7314
7315         if (root->in_radix) {
7316                 btrfs_free_fs_root(tree_root->fs_info, root);
7317         } else {
7318                 free_extent_buffer(root->node);
7319                 free_extent_buffer(root->commit_root);
7320                 kfree(root);
7321         }
7322 out_end_trans:
7323         btrfs_end_transaction_throttle(trans, tree_root);
7324 out_free:
7325         kfree(wc);
7326         btrfs_free_path(path);
7327 out:
7328         if (err)
7329                 btrfs_std_error(root->fs_info, err);
7330         return err;
7331 }
7332
7333 /*
7334  * drop subtree rooted at tree block 'node'.
7335  *
7336  * NOTE: this function will unlock and release tree block 'node'
7337  * only used by relocation code
7338  */
7339 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7340                         struct btrfs_root *root,
7341                         struct extent_buffer *node,
7342                         struct extent_buffer *parent)
7343 {
7344         struct btrfs_path *path;
7345         struct walk_control *wc;
7346         int level;
7347         int parent_level;
7348         int ret = 0;
7349         int wret;
7350
7351         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7352
7353         path = btrfs_alloc_path();
7354         if (!path)
7355                 return -ENOMEM;
7356
7357         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7358         if (!wc) {
7359                 btrfs_free_path(path);
7360                 return -ENOMEM;
7361         }
7362
7363         btrfs_assert_tree_locked(parent);
7364         parent_level = btrfs_header_level(parent);
7365         extent_buffer_get(parent);
7366         path->nodes[parent_level] = parent;
7367         path->slots[parent_level] = btrfs_header_nritems(parent);
7368
7369         btrfs_assert_tree_locked(node);
7370         level = btrfs_header_level(node);
7371         path->nodes[level] = node;
7372         path->slots[level] = 0;
7373         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7374
7375         wc->refs[parent_level] = 1;
7376         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7377         wc->level = level;
7378         wc->shared_level = -1;
7379         wc->stage = DROP_REFERENCE;
7380         wc->update_ref = 0;
7381         wc->keep_locks = 1;
7382         wc->for_reloc = 1;
7383         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7384
7385         while (1) {
7386                 wret = walk_down_tree(trans, root, path, wc);
7387                 if (wret < 0) {
7388                         ret = wret;
7389                         break;
7390                 }
7391
7392                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7393                 if (wret < 0)
7394                         ret = wret;
7395                 if (wret != 0)
7396                         break;
7397         }
7398
7399         kfree(wc);
7400         btrfs_free_path(path);
7401         return ret;
7402 }
7403
7404 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7405 {
7406         u64 num_devices;
7407         u64 stripped;
7408
7409         /*
7410          * if restripe for this chunk_type is on pick target profile and
7411          * return, otherwise do the usual balance
7412          */
7413         stripped = get_restripe_target(root->fs_info, flags);
7414         if (stripped)
7415                 return extended_to_chunk(stripped);
7416
7417         /*
7418          * we add in the count of missing devices because we want
7419          * to make sure that any RAID levels on a degraded FS
7420          * continue to be honored.
7421          */
7422         num_devices = root->fs_info->fs_devices->rw_devices +
7423                 root->fs_info->fs_devices->missing_devices;
7424
7425         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7426                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7427                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7428
7429         if (num_devices == 1) {
7430                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7431                 stripped = flags & ~stripped;
7432
7433                 /* turn raid0 into single device chunks */
7434                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7435                         return stripped;
7436
7437                 /* turn mirroring into duplication */
7438                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7439                              BTRFS_BLOCK_GROUP_RAID10))
7440                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7441         } else {
7442                 /* they already had raid on here, just return */
7443                 if (flags & stripped)
7444                         return flags;
7445
7446                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7447                 stripped = flags & ~stripped;
7448
7449                 /* switch duplicated blocks with raid1 */
7450                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7451                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7452
7453                 /* this is drive concat, leave it alone */
7454         }
7455
7456         return flags;
7457 }
7458
7459 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7460 {
7461         struct btrfs_space_info *sinfo = cache->space_info;
7462         u64 num_bytes;
7463         u64 min_allocable_bytes;
7464         int ret = -ENOSPC;
7465
7466
7467         /*
7468          * We need some metadata space and system metadata space for
7469          * allocating chunks in some corner cases until we force to set
7470          * it to be readonly.
7471          */
7472         if ((sinfo->flags &
7473              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7474             !force)
7475                 min_allocable_bytes = 1 * 1024 * 1024;
7476         else
7477                 min_allocable_bytes = 0;
7478
7479         spin_lock(&sinfo->lock);
7480         spin_lock(&cache->lock);
7481
7482         if (cache->ro) {
7483                 ret = 0;
7484                 goto out;
7485         }
7486
7487         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7488                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7489
7490         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7491             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7492             min_allocable_bytes <= sinfo->total_bytes) {
7493                 sinfo->bytes_readonly += num_bytes;
7494                 cache->ro = 1;
7495                 ret = 0;
7496         }
7497 out:
7498         spin_unlock(&cache->lock);
7499         spin_unlock(&sinfo->lock);
7500         return ret;
7501 }
7502
7503 int btrfs_set_block_group_ro(struct btrfs_root *root,
7504                              struct btrfs_block_group_cache *cache)
7505
7506 {
7507         struct btrfs_trans_handle *trans;
7508         u64 alloc_flags;
7509         int ret;
7510
7511         BUG_ON(cache->ro);
7512
7513         trans = btrfs_join_transaction(root);
7514         if (IS_ERR(trans))
7515                 return PTR_ERR(trans);
7516
7517         alloc_flags = update_block_group_flags(root, cache->flags);
7518         if (alloc_flags != cache->flags) {
7519                 ret = do_chunk_alloc(trans, root, alloc_flags,
7520                                      CHUNK_ALLOC_FORCE);
7521                 if (ret < 0)
7522                         goto out;
7523         }
7524
7525         ret = set_block_group_ro(cache, 0);
7526         if (!ret)
7527                 goto out;
7528         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7529         ret = do_chunk_alloc(trans, root, alloc_flags,
7530                              CHUNK_ALLOC_FORCE);
7531         if (ret < 0)
7532                 goto out;
7533         ret = set_block_group_ro(cache, 0);
7534 out:
7535         btrfs_end_transaction(trans, root);
7536         return ret;
7537 }
7538
7539 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7540                             struct btrfs_root *root, u64 type)
7541 {
7542         u64 alloc_flags = get_alloc_profile(root, type);
7543         return do_chunk_alloc(trans, root, alloc_flags,
7544                               CHUNK_ALLOC_FORCE);
7545 }
7546
7547 /*
7548  * helper to account the unused space of all the readonly block group in the
7549  * list. takes mirrors into account.
7550  */
7551 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7552 {
7553         struct btrfs_block_group_cache *block_group;
7554         u64 free_bytes = 0;
7555         int factor;
7556
7557         list_for_each_entry(block_group, groups_list, list) {
7558                 spin_lock(&block_group->lock);
7559
7560                 if (!block_group->ro) {
7561                         spin_unlock(&block_group->lock);
7562                         continue;
7563                 }
7564
7565                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7566                                           BTRFS_BLOCK_GROUP_RAID10 |
7567                                           BTRFS_BLOCK_GROUP_DUP))
7568                         factor = 2;
7569                 else
7570                         factor = 1;
7571
7572                 free_bytes += (block_group->key.offset -
7573                                btrfs_block_group_used(&block_group->item)) *
7574                                factor;
7575
7576                 spin_unlock(&block_group->lock);
7577         }
7578
7579         return free_bytes;
7580 }
7581
7582 /*
7583  * helper to account the unused space of all the readonly block group in the
7584  * space_info. takes mirrors into account.
7585  */
7586 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7587 {
7588         int i;
7589         u64 free_bytes = 0;
7590
7591         spin_lock(&sinfo->lock);
7592
7593         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7594                 if (!list_empty(&sinfo->block_groups[i]))
7595                         free_bytes += __btrfs_get_ro_block_group_free_space(
7596                                                 &sinfo->block_groups[i]);
7597
7598         spin_unlock(&sinfo->lock);
7599
7600         return free_bytes;
7601 }
7602
7603 void btrfs_set_block_group_rw(struct btrfs_root *root,
7604                               struct btrfs_block_group_cache *cache)
7605 {
7606         struct btrfs_space_info *sinfo = cache->space_info;
7607         u64 num_bytes;
7608
7609         BUG_ON(!cache->ro);
7610
7611         spin_lock(&sinfo->lock);
7612         spin_lock(&cache->lock);
7613         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7614                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7615         sinfo->bytes_readonly -= num_bytes;
7616         cache->ro = 0;
7617         spin_unlock(&cache->lock);
7618         spin_unlock(&sinfo->lock);
7619 }
7620
7621 /*
7622  * checks to see if its even possible to relocate this block group.
7623  *
7624  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7625  * ok to go ahead and try.
7626  */
7627 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7628 {
7629         struct btrfs_block_group_cache *block_group;
7630         struct btrfs_space_info *space_info;
7631         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7632         struct btrfs_device *device;
7633         u64 min_free;
7634         u64 dev_min = 1;
7635         u64 dev_nr = 0;
7636         u64 target;
7637         int index;
7638         int full = 0;
7639         int ret = 0;
7640
7641         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7642
7643         /* odd, couldn't find the block group, leave it alone */
7644         if (!block_group)
7645                 return -1;
7646
7647         min_free = btrfs_block_group_used(&block_group->item);
7648
7649         /* no bytes used, we're good */
7650         if (!min_free)
7651                 goto out;
7652
7653         space_info = block_group->space_info;
7654         spin_lock(&space_info->lock);
7655
7656         full = space_info->full;
7657
7658         /*
7659          * if this is the last block group we have in this space, we can't
7660          * relocate it unless we're able to allocate a new chunk below.
7661          *
7662          * Otherwise, we need to make sure we have room in the space to handle
7663          * all of the extents from this block group.  If we can, we're good
7664          */
7665         if ((space_info->total_bytes != block_group->key.offset) &&
7666             (space_info->bytes_used + space_info->bytes_reserved +
7667              space_info->bytes_pinned + space_info->bytes_readonly +
7668              min_free < space_info->total_bytes)) {
7669                 spin_unlock(&space_info->lock);
7670                 goto out;
7671         }
7672         spin_unlock(&space_info->lock);
7673
7674         /*
7675          * ok we don't have enough space, but maybe we have free space on our
7676          * devices to allocate new chunks for relocation, so loop through our
7677          * alloc devices and guess if we have enough space.  if this block
7678          * group is going to be restriped, run checks against the target
7679          * profile instead of the current one.
7680          */
7681         ret = -1;
7682
7683         /*
7684          * index:
7685          *      0: raid10
7686          *      1: raid1
7687          *      2: dup
7688          *      3: raid0
7689          *      4: single
7690          */
7691         target = get_restripe_target(root->fs_info, block_group->flags);
7692         if (target) {
7693                 index = __get_raid_index(extended_to_chunk(target));
7694         } else {
7695                 /*
7696                  * this is just a balance, so if we were marked as full
7697                  * we know there is no space for a new chunk
7698                  */
7699                 if (full)
7700                         goto out;
7701
7702                 index = get_block_group_index(block_group);
7703         }
7704
7705         if (index == BTRFS_RAID_RAID10) {
7706                 dev_min = 4;
7707                 /* Divide by 2 */
7708                 min_free >>= 1;
7709         } else if (index == BTRFS_RAID_RAID1) {
7710                 dev_min = 2;
7711         } else if (index == BTRFS_RAID_DUP) {
7712                 /* Multiply by 2 */
7713                 min_free <<= 1;
7714         } else if (index == BTRFS_RAID_RAID0) {
7715                 dev_min = fs_devices->rw_devices;
7716                 do_div(min_free, dev_min);
7717         }
7718
7719         mutex_lock(&root->fs_info->chunk_mutex);
7720         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7721                 u64 dev_offset;
7722
7723                 /*
7724                  * check to make sure we can actually find a chunk with enough
7725                  * space to fit our block group in.
7726                  */
7727                 if (device->total_bytes > device->bytes_used + min_free &&
7728                     !device->is_tgtdev_for_dev_replace) {
7729                         ret = find_free_dev_extent(device, min_free,
7730                                                    &dev_offset, NULL);
7731                         if (!ret)
7732                                 dev_nr++;
7733
7734                         if (dev_nr >= dev_min)
7735                                 break;
7736
7737                         ret = -1;
7738                 }
7739         }
7740         mutex_unlock(&root->fs_info->chunk_mutex);
7741 out:
7742         btrfs_put_block_group(block_group);
7743         return ret;
7744 }
7745
7746 static int find_first_block_group(struct btrfs_root *root,
7747                 struct btrfs_path *path, struct btrfs_key *key)
7748 {
7749         int ret = 0;
7750         struct btrfs_key found_key;
7751         struct extent_buffer *leaf;
7752         int slot;
7753
7754         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7755         if (ret < 0)
7756                 goto out;
7757
7758         while (1) {
7759                 slot = path->slots[0];
7760                 leaf = path->nodes[0];
7761                 if (slot >= btrfs_header_nritems(leaf)) {
7762                         ret = btrfs_next_leaf(root, path);
7763                         if (ret == 0)
7764                                 continue;
7765                         if (ret < 0)
7766                                 goto out;
7767                         break;
7768                 }
7769                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7770
7771                 if (found_key.objectid >= key->objectid &&
7772                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7773                         ret = 0;
7774                         goto out;
7775                 }
7776                 path->slots[0]++;
7777         }
7778 out:
7779         return ret;
7780 }
7781
7782 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7783 {
7784         struct btrfs_block_group_cache *block_group;
7785         u64 last = 0;
7786
7787         while (1) {
7788                 struct inode *inode;
7789
7790                 block_group = btrfs_lookup_first_block_group(info, last);
7791                 while (block_group) {
7792                         spin_lock(&block_group->lock);
7793                         if (block_group->iref)
7794                                 break;
7795                         spin_unlock(&block_group->lock);
7796                         block_group = next_block_group(info->tree_root,
7797                                                        block_group);
7798                 }
7799                 if (!block_group) {
7800                         if (last == 0)
7801                                 break;
7802                         last = 0;
7803                         continue;
7804                 }
7805
7806                 inode = block_group->inode;
7807                 block_group->iref = 0;
7808                 block_group->inode = NULL;
7809                 spin_unlock(&block_group->lock);
7810                 iput(inode);
7811                 last = block_group->key.objectid + block_group->key.offset;
7812                 btrfs_put_block_group(block_group);
7813         }
7814 }
7815
7816 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7817 {
7818         struct btrfs_block_group_cache *block_group;
7819         struct btrfs_space_info *space_info;
7820         struct btrfs_caching_control *caching_ctl;
7821         struct rb_node *n;
7822
7823         down_write(&info->extent_commit_sem);
7824         while (!list_empty(&info->caching_block_groups)) {
7825                 caching_ctl = list_entry(info->caching_block_groups.next,
7826                                          struct btrfs_caching_control, list);
7827                 list_del(&caching_ctl->list);
7828                 put_caching_control(caching_ctl);
7829         }
7830         up_write(&info->extent_commit_sem);
7831
7832         spin_lock(&info->block_group_cache_lock);
7833         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7834                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7835                                        cache_node);
7836                 rb_erase(&block_group->cache_node,
7837                          &info->block_group_cache_tree);
7838                 spin_unlock(&info->block_group_cache_lock);
7839
7840                 down_write(&block_group->space_info->groups_sem);
7841                 list_del(&block_group->list);
7842                 up_write(&block_group->space_info->groups_sem);
7843
7844                 if (block_group->cached == BTRFS_CACHE_STARTED)
7845                         wait_block_group_cache_done(block_group);
7846
7847                 /*
7848                  * We haven't cached this block group, which means we could
7849                  * possibly have excluded extents on this block group.
7850                  */
7851                 if (block_group->cached == BTRFS_CACHE_NO)
7852                         free_excluded_extents(info->extent_root, block_group);
7853
7854                 btrfs_remove_free_space_cache(block_group);
7855                 btrfs_put_block_group(block_group);
7856
7857                 spin_lock(&info->block_group_cache_lock);
7858         }
7859         spin_unlock(&info->block_group_cache_lock);
7860
7861         /* now that all the block groups are freed, go through and
7862          * free all the space_info structs.  This is only called during
7863          * the final stages of unmount, and so we know nobody is
7864          * using them.  We call synchronize_rcu() once before we start,
7865          * just to be on the safe side.
7866          */
7867         synchronize_rcu();
7868
7869         release_global_block_rsv(info);
7870
7871         while(!list_empty(&info->space_info)) {
7872                 space_info = list_entry(info->space_info.next,
7873                                         struct btrfs_space_info,
7874                                         list);
7875                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
7876                         if (space_info->bytes_pinned > 0 ||
7877                             space_info->bytes_reserved > 0 ||
7878                             space_info->bytes_may_use > 0) {
7879                                 WARN_ON(1);
7880                                 dump_space_info(space_info, 0, 0);
7881                         }
7882                 }
7883                 list_del(&space_info->list);
7884                 kfree(space_info);
7885         }
7886         return 0;
7887 }
7888
7889 static void __link_block_group(struct btrfs_space_info *space_info,
7890                                struct btrfs_block_group_cache *cache)
7891 {
7892         int index = get_block_group_index(cache);
7893
7894         down_write(&space_info->groups_sem);
7895         list_add_tail(&cache->list, &space_info->block_groups[index]);
7896         up_write(&space_info->groups_sem);
7897 }
7898
7899 int btrfs_read_block_groups(struct btrfs_root *root)
7900 {
7901         struct btrfs_path *path;
7902         int ret;
7903         struct btrfs_block_group_cache *cache;
7904         struct btrfs_fs_info *info = root->fs_info;
7905         struct btrfs_space_info *space_info;
7906         struct btrfs_key key;
7907         struct btrfs_key found_key;
7908         struct extent_buffer *leaf;
7909         int need_clear = 0;
7910         u64 cache_gen;
7911
7912         root = info->extent_root;
7913         key.objectid = 0;
7914         key.offset = 0;
7915         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7916         path = btrfs_alloc_path();
7917         if (!path)
7918                 return -ENOMEM;
7919         path->reada = 1;
7920
7921         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7922         if (btrfs_test_opt(root, SPACE_CACHE) &&
7923             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7924                 need_clear = 1;
7925         if (btrfs_test_opt(root, CLEAR_CACHE))
7926                 need_clear = 1;
7927
7928         while (1) {
7929                 ret = find_first_block_group(root, path, &key);
7930                 if (ret > 0)
7931                         break;
7932                 if (ret != 0)
7933                         goto error;
7934                 leaf = path->nodes[0];
7935                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7936                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7937                 if (!cache) {
7938                         ret = -ENOMEM;
7939                         goto error;
7940                 }
7941                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7942                                                 GFP_NOFS);
7943                 if (!cache->free_space_ctl) {
7944                         kfree(cache);
7945                         ret = -ENOMEM;
7946                         goto error;
7947                 }
7948
7949                 atomic_set(&cache->count, 1);
7950                 spin_lock_init(&cache->lock);
7951                 cache->fs_info = info;
7952                 INIT_LIST_HEAD(&cache->list);
7953                 INIT_LIST_HEAD(&cache->cluster_list);
7954
7955                 if (need_clear) {
7956                         /*
7957                          * When we mount with old space cache, we need to
7958                          * set BTRFS_DC_CLEAR and set dirty flag.
7959                          *
7960                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
7961                          *    truncate the old free space cache inode and
7962                          *    setup a new one.
7963                          * b) Setting 'dirty flag' makes sure that we flush
7964                          *    the new space cache info onto disk.
7965                          */
7966                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7967                         if (btrfs_test_opt(root, SPACE_CACHE))
7968                                 cache->dirty = 1;
7969                 }
7970
7971                 read_extent_buffer(leaf, &cache->item,
7972                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7973                                    sizeof(cache->item));
7974                 memcpy(&cache->key, &found_key, sizeof(found_key));
7975
7976                 key.objectid = found_key.objectid + found_key.offset;
7977                 btrfs_release_path(path);
7978                 cache->flags = btrfs_block_group_flags(&cache->item);
7979                 cache->sectorsize = root->sectorsize;
7980                 cache->full_stripe_len = btrfs_full_stripe_len(root,
7981                                                &root->fs_info->mapping_tree,
7982                                                found_key.objectid);
7983                 btrfs_init_free_space_ctl(cache);
7984
7985                 /*
7986                  * We need to exclude the super stripes now so that the space
7987                  * info has super bytes accounted for, otherwise we'll think
7988                  * we have more space than we actually do.
7989                  */
7990                 ret = exclude_super_stripes(root, cache);
7991                 if (ret) {
7992                         /*
7993                          * We may have excluded something, so call this just in
7994                          * case.
7995                          */
7996                         free_excluded_extents(root, cache);
7997                         kfree(cache->free_space_ctl);
7998                         kfree(cache);
7999                         goto error;
8000                 }
8001
8002                 /*
8003                  * check for two cases, either we are full, and therefore
8004                  * don't need to bother with the caching work since we won't
8005                  * find any space, or we are empty, and we can just add all
8006                  * the space in and be done with it.  This saves us _alot_ of
8007                  * time, particularly in the full case.
8008                  */
8009                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8010                         cache->last_byte_to_unpin = (u64)-1;
8011                         cache->cached = BTRFS_CACHE_FINISHED;
8012                         free_excluded_extents(root, cache);
8013                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8014                         cache->last_byte_to_unpin = (u64)-1;
8015                         cache->cached = BTRFS_CACHE_FINISHED;
8016                         add_new_free_space(cache, root->fs_info,
8017                                            found_key.objectid,
8018                                            found_key.objectid +
8019                                            found_key.offset);
8020                         free_excluded_extents(root, cache);
8021                 }
8022
8023                 ret = update_space_info(info, cache->flags, found_key.offset,
8024                                         btrfs_block_group_used(&cache->item),
8025                                         &space_info);
8026                 BUG_ON(ret); /* -ENOMEM */
8027                 cache->space_info = space_info;
8028                 spin_lock(&cache->space_info->lock);
8029                 cache->space_info->bytes_readonly += cache->bytes_super;
8030                 spin_unlock(&cache->space_info->lock);
8031
8032                 __link_block_group(space_info, cache);
8033
8034                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8035                 BUG_ON(ret); /* Logic error */
8036
8037                 set_avail_alloc_bits(root->fs_info, cache->flags);
8038                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8039                         set_block_group_ro(cache, 1);
8040         }
8041
8042         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8043                 if (!(get_alloc_profile(root, space_info->flags) &
8044                       (BTRFS_BLOCK_GROUP_RAID10 |
8045                        BTRFS_BLOCK_GROUP_RAID1 |
8046                        BTRFS_BLOCK_GROUP_RAID5 |
8047                        BTRFS_BLOCK_GROUP_RAID6 |
8048                        BTRFS_BLOCK_GROUP_DUP)))
8049                         continue;
8050                 /*
8051                  * avoid allocating from un-mirrored block group if there are
8052                  * mirrored block groups.
8053                  */
8054                 list_for_each_entry(cache, &space_info->block_groups[3], list)
8055                         set_block_group_ro(cache, 1);
8056                 list_for_each_entry(cache, &space_info->block_groups[4], list)
8057                         set_block_group_ro(cache, 1);
8058         }
8059
8060         init_global_block_rsv(info);
8061         ret = 0;
8062 error:
8063         btrfs_free_path(path);
8064         return ret;
8065 }
8066
8067 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8068                                        struct btrfs_root *root)
8069 {
8070         struct btrfs_block_group_cache *block_group, *tmp;
8071         struct btrfs_root *extent_root = root->fs_info->extent_root;
8072         struct btrfs_block_group_item item;
8073         struct btrfs_key key;
8074         int ret = 0;
8075
8076         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8077                                  new_bg_list) {
8078                 list_del_init(&block_group->new_bg_list);
8079
8080                 if (ret)
8081                         continue;
8082
8083                 spin_lock(&block_group->lock);
8084                 memcpy(&item, &block_group->item, sizeof(item));
8085                 memcpy(&key, &block_group->key, sizeof(key));
8086                 spin_unlock(&block_group->lock);
8087
8088                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8089                                         sizeof(item));
8090                 if (ret)
8091                         btrfs_abort_transaction(trans, extent_root, ret);
8092         }
8093 }
8094
8095 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8096                            struct btrfs_root *root, u64 bytes_used,
8097                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8098                            u64 size)
8099 {
8100         int ret;
8101         struct btrfs_root *extent_root;
8102         struct btrfs_block_group_cache *cache;
8103
8104         extent_root = root->fs_info->extent_root;
8105
8106         root->fs_info->last_trans_log_full_commit = trans->transid;
8107
8108         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8109         if (!cache)
8110                 return -ENOMEM;
8111         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8112                                         GFP_NOFS);
8113         if (!cache->free_space_ctl) {
8114                 kfree(cache);
8115                 return -ENOMEM;
8116         }
8117
8118         cache->key.objectid = chunk_offset;
8119         cache->key.offset = size;
8120         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8121         cache->sectorsize = root->sectorsize;
8122         cache->fs_info = root->fs_info;
8123         cache->full_stripe_len = btrfs_full_stripe_len(root,
8124                                                &root->fs_info->mapping_tree,
8125                                                chunk_offset);
8126
8127         atomic_set(&cache->count, 1);
8128         spin_lock_init(&cache->lock);
8129         INIT_LIST_HEAD(&cache->list);
8130         INIT_LIST_HEAD(&cache->cluster_list);
8131         INIT_LIST_HEAD(&cache->new_bg_list);
8132
8133         btrfs_init_free_space_ctl(cache);
8134
8135         btrfs_set_block_group_used(&cache->item, bytes_used);
8136         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8137         cache->flags = type;
8138         btrfs_set_block_group_flags(&cache->item, type);
8139
8140         cache->last_byte_to_unpin = (u64)-1;
8141         cache->cached = BTRFS_CACHE_FINISHED;
8142         ret = exclude_super_stripes(root, cache);
8143         if (ret) {
8144                 /*
8145                  * We may have excluded something, so call this just in
8146                  * case.
8147                  */
8148                 free_excluded_extents(root, cache);
8149                 kfree(cache->free_space_ctl);
8150                 kfree(cache);
8151                 return ret;
8152         }
8153
8154         add_new_free_space(cache, root->fs_info, chunk_offset,
8155                            chunk_offset + size);
8156
8157         free_excluded_extents(root, cache);
8158
8159         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8160                                 &cache->space_info);
8161         BUG_ON(ret); /* -ENOMEM */
8162         update_global_block_rsv(root->fs_info);
8163
8164         spin_lock(&cache->space_info->lock);
8165         cache->space_info->bytes_readonly += cache->bytes_super;
8166         spin_unlock(&cache->space_info->lock);
8167
8168         __link_block_group(cache->space_info, cache);
8169
8170         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8171         BUG_ON(ret); /* Logic error */
8172
8173         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8174
8175         set_avail_alloc_bits(extent_root->fs_info, type);
8176
8177         return 0;
8178 }
8179
8180 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8181 {
8182         u64 extra_flags = chunk_to_extended(flags) &
8183                                 BTRFS_EXTENDED_PROFILE_MASK;
8184
8185         write_seqlock(&fs_info->profiles_lock);
8186         if (flags & BTRFS_BLOCK_GROUP_DATA)
8187                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8188         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8189                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8190         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8191                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8192         write_sequnlock(&fs_info->profiles_lock);
8193 }
8194
8195 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8196                              struct btrfs_root *root, u64 group_start)
8197 {
8198         struct btrfs_path *path;
8199         struct btrfs_block_group_cache *block_group;
8200         struct btrfs_free_cluster *cluster;
8201         struct btrfs_root *tree_root = root->fs_info->tree_root;
8202         struct btrfs_key key;
8203         struct inode *inode;
8204         int ret;
8205         int index;
8206         int factor;
8207
8208         root = root->fs_info->extent_root;
8209
8210         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8211         BUG_ON(!block_group);
8212         BUG_ON(!block_group->ro);
8213
8214         /*
8215          * Free the reserved super bytes from this block group before
8216          * remove it.
8217          */
8218         free_excluded_extents(root, block_group);
8219
8220         memcpy(&key, &block_group->key, sizeof(key));
8221         index = get_block_group_index(block_group);
8222         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8223                                   BTRFS_BLOCK_GROUP_RAID1 |
8224                                   BTRFS_BLOCK_GROUP_RAID10))
8225                 factor = 2;
8226         else
8227                 factor = 1;
8228
8229         /* make sure this block group isn't part of an allocation cluster */
8230         cluster = &root->fs_info->data_alloc_cluster;
8231         spin_lock(&cluster->refill_lock);
8232         btrfs_return_cluster_to_free_space(block_group, cluster);
8233         spin_unlock(&cluster->refill_lock);
8234
8235         /*
8236          * make sure this block group isn't part of a metadata
8237          * allocation cluster
8238          */
8239         cluster = &root->fs_info->meta_alloc_cluster;
8240         spin_lock(&cluster->refill_lock);
8241         btrfs_return_cluster_to_free_space(block_group, cluster);
8242         spin_unlock(&cluster->refill_lock);
8243
8244         path = btrfs_alloc_path();
8245         if (!path) {
8246                 ret = -ENOMEM;
8247                 goto out;
8248         }
8249
8250         inode = lookup_free_space_inode(tree_root, block_group, path);
8251         if (!IS_ERR(inode)) {
8252                 ret = btrfs_orphan_add(trans, inode);
8253                 if (ret) {
8254                         btrfs_add_delayed_iput(inode);
8255                         goto out;
8256                 }
8257                 clear_nlink(inode);
8258                 /* One for the block groups ref */
8259                 spin_lock(&block_group->lock);
8260                 if (block_group->iref) {
8261                         block_group->iref = 0;
8262                         block_group->inode = NULL;
8263                         spin_unlock(&block_group->lock);
8264                         iput(inode);
8265                 } else {
8266                         spin_unlock(&block_group->lock);
8267                 }
8268                 /* One for our lookup ref */
8269                 btrfs_add_delayed_iput(inode);
8270         }
8271
8272         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8273         key.offset = block_group->key.objectid;
8274         key.type = 0;
8275
8276         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8277         if (ret < 0)
8278                 goto out;
8279         if (ret > 0)
8280                 btrfs_release_path(path);
8281         if (ret == 0) {
8282                 ret = btrfs_del_item(trans, tree_root, path);
8283                 if (ret)
8284                         goto out;
8285                 btrfs_release_path(path);
8286         }
8287
8288         spin_lock(&root->fs_info->block_group_cache_lock);
8289         rb_erase(&block_group->cache_node,
8290                  &root->fs_info->block_group_cache_tree);
8291
8292         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8293                 root->fs_info->first_logical_byte = (u64)-1;
8294         spin_unlock(&root->fs_info->block_group_cache_lock);
8295
8296         down_write(&block_group->space_info->groups_sem);
8297         /*
8298          * we must use list_del_init so people can check to see if they
8299          * are still on the list after taking the semaphore
8300          */
8301         list_del_init(&block_group->list);
8302         if (list_empty(&block_group->space_info->block_groups[index]))
8303                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8304         up_write(&block_group->space_info->groups_sem);
8305
8306         if (block_group->cached == BTRFS_CACHE_STARTED)
8307                 wait_block_group_cache_done(block_group);
8308
8309         btrfs_remove_free_space_cache(block_group);
8310
8311         spin_lock(&block_group->space_info->lock);
8312         block_group->space_info->total_bytes -= block_group->key.offset;
8313         block_group->space_info->bytes_readonly -= block_group->key.offset;
8314         block_group->space_info->disk_total -= block_group->key.offset * factor;
8315         spin_unlock(&block_group->space_info->lock);
8316
8317         memcpy(&key, &block_group->key, sizeof(key));
8318
8319         btrfs_clear_space_info_full(root->fs_info);
8320
8321         btrfs_put_block_group(block_group);
8322         btrfs_put_block_group(block_group);
8323
8324         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8325         if (ret > 0)
8326                 ret = -EIO;
8327         if (ret < 0)
8328                 goto out;
8329
8330         ret = btrfs_del_item(trans, root, path);
8331 out:
8332         btrfs_free_path(path);
8333         return ret;
8334 }
8335
8336 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8337 {
8338         struct btrfs_space_info *space_info;
8339         struct btrfs_super_block *disk_super;
8340         u64 features;
8341         u64 flags;
8342         int mixed = 0;
8343         int ret;
8344
8345         disk_super = fs_info->super_copy;
8346         if (!btrfs_super_root(disk_super))
8347                 return 1;
8348
8349         features = btrfs_super_incompat_flags(disk_super);
8350         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8351                 mixed = 1;
8352
8353         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8354         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8355         if (ret)
8356                 goto out;
8357
8358         if (mixed) {
8359                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8360                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8361         } else {
8362                 flags = BTRFS_BLOCK_GROUP_METADATA;
8363                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8364                 if (ret)
8365                         goto out;
8366
8367                 flags = BTRFS_BLOCK_GROUP_DATA;
8368                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8369         }
8370 out:
8371         return ret;
8372 }
8373
8374 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8375 {
8376         return unpin_extent_range(root, start, end);
8377 }
8378
8379 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8380                                u64 num_bytes, u64 *actual_bytes)
8381 {
8382         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8383 }
8384
8385 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8386 {
8387         struct btrfs_fs_info *fs_info = root->fs_info;
8388         struct btrfs_block_group_cache *cache = NULL;
8389         u64 group_trimmed;
8390         u64 start;
8391         u64 end;
8392         u64 trimmed = 0;
8393         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8394         int ret = 0;
8395
8396         /*
8397          * try to trim all FS space, our block group may start from non-zero.
8398          */
8399         if (range->len == total_bytes)
8400                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8401         else
8402                 cache = btrfs_lookup_block_group(fs_info, range->start);
8403
8404         while (cache) {
8405                 if (cache->key.objectid >= (range->start + range->len)) {
8406                         btrfs_put_block_group(cache);
8407                         break;
8408                 }
8409
8410                 start = max(range->start, cache->key.objectid);
8411                 end = min(range->start + range->len,
8412                                 cache->key.objectid + cache->key.offset);
8413
8414                 if (end - start >= range->minlen) {
8415                         if (!block_group_cache_done(cache)) {
8416                                 ret = cache_block_group(cache, 0);
8417                                 if (!ret)
8418                                         wait_block_group_cache_done(cache);
8419                         }
8420                         ret = btrfs_trim_block_group(cache,
8421                                                      &group_trimmed,
8422                                                      start,
8423                                                      end,
8424                                                      range->minlen);
8425
8426                         trimmed += group_trimmed;
8427                         if (ret) {
8428                                 btrfs_put_block_group(cache);
8429                                 break;
8430                         }
8431                 }
8432
8433                 cache = next_block_group(fs_info->tree_root, cache);
8434         }
8435
8436         range->len = trimmed;
8437         return ret;
8438 }