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 static inline int writeback_inodes_sb_nr_if_idle_safe(struct super_block *sb,
3812                                                       unsigned long nr_pages,
3813                                                       enum wb_reason reason)
3814 {
3815         /* the flusher is dealing with the dirty inodes now. */
3816         if (writeback_in_progress(sb->s_bdi))
3817                 return 1;
3818
3819         if (down_read_trylock(&sb->s_umount)) {
3820                 writeback_inodes_sb_nr(sb, nr_pages, reason);
3821                 up_read(&sb->s_umount);
3822                 return 1;
3823         }
3824
3825         return 0;
3826 }
3827
3828 void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3829                                   unsigned long nr_pages)
3830 {
3831         struct super_block *sb = root->fs_info->sb;
3832         int started;
3833
3834         /* If we can not start writeback, just sync all the delalloc file. */
3835         started = writeback_inodes_sb_nr_if_idle_safe(sb, nr_pages,
3836                                                       WB_REASON_FS_FREE_SPACE);
3837         if (!started) {
3838                 /*
3839                  * We needn't worry the filesystem going from r/w to r/o though
3840                  * we don't acquire ->s_umount mutex, because the filesystem
3841                  * should guarantee the delalloc inodes list be empty after
3842                  * the filesystem is readonly(all dirty pages are written to
3843                  * the disk).
3844                  */
3845                 btrfs_start_delalloc_inodes(root, 0);
3846                 btrfs_wait_ordered_extents(root, 0);
3847         }
3848 }
3849
3850 /*
3851  * shrink metadata reservation for delalloc
3852  */
3853 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3854                             bool wait_ordered)
3855 {
3856         struct btrfs_block_rsv *block_rsv;
3857         struct btrfs_space_info *space_info;
3858         struct btrfs_trans_handle *trans;
3859         u64 delalloc_bytes;
3860         u64 max_reclaim;
3861         long time_left;
3862         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3863         int loops = 0;
3864         enum btrfs_reserve_flush_enum flush;
3865
3866         trans = (struct btrfs_trans_handle *)current->journal_info;
3867         block_rsv = &root->fs_info->delalloc_block_rsv;
3868         space_info = block_rsv->space_info;
3869
3870         smp_mb();
3871         delalloc_bytes = percpu_counter_sum_positive(
3872                                                 &root->fs_info->delalloc_bytes);
3873         if (delalloc_bytes == 0) {
3874                 if (trans)
3875                         return;
3876                 btrfs_wait_ordered_extents(root, 0);
3877                 return;
3878         }
3879
3880         while (delalloc_bytes && loops < 3) {
3881                 max_reclaim = min(delalloc_bytes, to_reclaim);
3882                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3883                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
3884                 /*
3885                  * We need to wait for the async pages to actually start before
3886                  * we do anything.
3887                  */
3888                 wait_event(root->fs_info->async_submit_wait,
3889                            !atomic_read(&root->fs_info->async_delalloc_pages));
3890
3891                 if (!trans)
3892                         flush = BTRFS_RESERVE_FLUSH_ALL;
3893                 else
3894                         flush = BTRFS_RESERVE_NO_FLUSH;
3895                 spin_lock(&space_info->lock);
3896                 if (can_overcommit(root, space_info, orig, flush)) {
3897                         spin_unlock(&space_info->lock);
3898                         break;
3899                 }
3900                 spin_unlock(&space_info->lock);
3901
3902                 loops++;
3903                 if (wait_ordered && !trans) {
3904                         btrfs_wait_ordered_extents(root, 0);
3905                 } else {
3906                         time_left = schedule_timeout_killable(1);
3907                         if (time_left)
3908                                 break;
3909                 }
3910                 smp_mb();
3911                 delalloc_bytes = percpu_counter_sum_positive(
3912                                                 &root->fs_info->delalloc_bytes);
3913         }
3914 }
3915
3916 /**
3917  * maybe_commit_transaction - possibly commit the transaction if its ok to
3918  * @root - the root we're allocating for
3919  * @bytes - the number of bytes we want to reserve
3920  * @force - force the commit
3921  *
3922  * This will check to make sure that committing the transaction will actually
3923  * get us somewhere and then commit the transaction if it does.  Otherwise it
3924  * will return -ENOSPC.
3925  */
3926 static int may_commit_transaction(struct btrfs_root *root,
3927                                   struct btrfs_space_info *space_info,
3928                                   u64 bytes, int force)
3929 {
3930         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3931         struct btrfs_trans_handle *trans;
3932
3933         trans = (struct btrfs_trans_handle *)current->journal_info;
3934         if (trans)
3935                 return -EAGAIN;
3936
3937         if (force)
3938                 goto commit;
3939
3940         /* See if there is enough pinned space to make this reservation */
3941         spin_lock(&space_info->lock);
3942         if (space_info->bytes_pinned >= bytes) {
3943                 spin_unlock(&space_info->lock);
3944                 goto commit;
3945         }
3946         spin_unlock(&space_info->lock);
3947
3948         /*
3949          * See if there is some space in the delayed insertion reservation for
3950          * this reservation.
3951          */
3952         if (space_info != delayed_rsv->space_info)
3953                 return -ENOSPC;
3954
3955         spin_lock(&space_info->lock);
3956         spin_lock(&delayed_rsv->lock);
3957         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3958                 spin_unlock(&delayed_rsv->lock);
3959                 spin_unlock(&space_info->lock);
3960                 return -ENOSPC;
3961         }
3962         spin_unlock(&delayed_rsv->lock);
3963         spin_unlock(&space_info->lock);
3964
3965 commit:
3966         trans = btrfs_join_transaction(root);
3967         if (IS_ERR(trans))
3968                 return -ENOSPC;
3969
3970         return btrfs_commit_transaction(trans, root);
3971 }
3972
3973 enum flush_state {
3974         FLUSH_DELAYED_ITEMS_NR  =       1,
3975         FLUSH_DELAYED_ITEMS     =       2,
3976         FLUSH_DELALLOC          =       3,
3977         FLUSH_DELALLOC_WAIT     =       4,
3978         ALLOC_CHUNK             =       5,
3979         COMMIT_TRANS            =       6,
3980 };
3981
3982 static int flush_space(struct btrfs_root *root,
3983                        struct btrfs_space_info *space_info, u64 num_bytes,
3984                        u64 orig_bytes, int state)
3985 {
3986         struct btrfs_trans_handle *trans;
3987         int nr;
3988         int ret = 0;
3989
3990         switch (state) {
3991         case FLUSH_DELAYED_ITEMS_NR:
3992         case FLUSH_DELAYED_ITEMS:
3993                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3994                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3995
3996                         nr = (int)div64_u64(num_bytes, bytes);
3997                         if (!nr)
3998                                 nr = 1;
3999                         nr *= 2;
4000                 } else {
4001                         nr = -1;
4002                 }
4003                 trans = btrfs_join_transaction(root);
4004                 if (IS_ERR(trans)) {
4005                         ret = PTR_ERR(trans);
4006                         break;
4007                 }
4008                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4009                 btrfs_end_transaction(trans, root);
4010                 break;
4011         case FLUSH_DELALLOC:
4012         case FLUSH_DELALLOC_WAIT:
4013                 shrink_delalloc(root, num_bytes, orig_bytes,
4014                                 state == FLUSH_DELALLOC_WAIT);
4015                 break;
4016         case ALLOC_CHUNK:
4017                 trans = btrfs_join_transaction(root);
4018                 if (IS_ERR(trans)) {
4019                         ret = PTR_ERR(trans);
4020                         break;
4021                 }
4022                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4023                                      btrfs_get_alloc_profile(root, 0),
4024                                      CHUNK_ALLOC_NO_FORCE);
4025                 btrfs_end_transaction(trans, root);
4026                 if (ret == -ENOSPC)
4027                         ret = 0;
4028                 break;
4029         case COMMIT_TRANS:
4030                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4031                 break;
4032         default:
4033                 ret = -ENOSPC;
4034                 break;
4035         }
4036
4037         return ret;
4038 }
4039 /**
4040  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4041  * @root - the root we're allocating for
4042  * @block_rsv - the block_rsv we're allocating for
4043  * @orig_bytes - the number of bytes we want
4044  * @flush - whether or not we can flush to make our reservation
4045  *
4046  * This will reserve orgi_bytes number of bytes from the space info associated
4047  * with the block_rsv.  If there is not enough space it will make an attempt to
4048  * flush out space to make room.  It will do this by flushing delalloc if
4049  * possible or committing the transaction.  If flush is 0 then no attempts to
4050  * regain reservations will be made and this will fail if there is not enough
4051  * space already.
4052  */
4053 static int reserve_metadata_bytes(struct btrfs_root *root,
4054                                   struct btrfs_block_rsv *block_rsv,
4055                                   u64 orig_bytes,
4056                                   enum btrfs_reserve_flush_enum flush)
4057 {
4058         struct btrfs_space_info *space_info = block_rsv->space_info;
4059         u64 used;
4060         u64 num_bytes = orig_bytes;
4061         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4062         int ret = 0;
4063         bool flushing = false;
4064
4065 again:
4066         ret = 0;
4067         spin_lock(&space_info->lock);
4068         /*
4069          * We only want to wait if somebody other than us is flushing and we
4070          * are actually allowed to flush all things.
4071          */
4072         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4073                space_info->flush) {
4074                 spin_unlock(&space_info->lock);
4075                 /*
4076                  * If we have a trans handle we can't wait because the flusher
4077                  * may have to commit the transaction, which would mean we would
4078                  * deadlock since we are waiting for the flusher to finish, but
4079                  * hold the current transaction open.
4080                  */
4081                 if (current->journal_info)
4082                         return -EAGAIN;
4083                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4084                 /* Must have been killed, return */
4085                 if (ret)
4086                         return -EINTR;
4087
4088                 spin_lock(&space_info->lock);
4089         }
4090
4091         ret = -ENOSPC;
4092         used = space_info->bytes_used + space_info->bytes_reserved +
4093                 space_info->bytes_pinned + space_info->bytes_readonly +
4094                 space_info->bytes_may_use;
4095
4096         /*
4097          * The idea here is that we've not already over-reserved the block group
4098          * then we can go ahead and save our reservation first and then start
4099          * flushing if we need to.  Otherwise if we've already overcommitted
4100          * lets start flushing stuff first and then come back and try to make
4101          * our reservation.
4102          */
4103         if (used <= space_info->total_bytes) {
4104                 if (used + orig_bytes <= space_info->total_bytes) {
4105                         space_info->bytes_may_use += orig_bytes;
4106                         trace_btrfs_space_reservation(root->fs_info,
4107                                 "space_info", space_info->flags, orig_bytes, 1);
4108                         ret = 0;
4109                 } else {
4110                         /*
4111                          * Ok set num_bytes to orig_bytes since we aren't
4112                          * overocmmitted, this way we only try and reclaim what
4113                          * we need.
4114                          */
4115                         num_bytes = orig_bytes;
4116                 }
4117         } else {
4118                 /*
4119                  * Ok we're over committed, set num_bytes to the overcommitted
4120                  * amount plus the amount of bytes that we need for this
4121                  * reservation.
4122                  */
4123                 num_bytes = used - space_info->total_bytes +
4124                         (orig_bytes * 2);
4125         }
4126
4127         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4128                 space_info->bytes_may_use += orig_bytes;
4129                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4130                                               space_info->flags, orig_bytes,
4131                                               1);
4132                 ret = 0;
4133         }
4134
4135         /*
4136          * Couldn't make our reservation, save our place so while we're trying
4137          * to reclaim space we can actually use it instead of somebody else
4138          * stealing it from us.
4139          *
4140          * We make the other tasks wait for the flush only when we can flush
4141          * all things.
4142          */
4143         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4144                 flushing = true;
4145                 space_info->flush = 1;
4146         }
4147
4148         spin_unlock(&space_info->lock);
4149
4150         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4151                 goto out;
4152
4153         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4154                           flush_state);
4155         flush_state++;
4156
4157         /*
4158          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4159          * would happen. So skip delalloc flush.
4160          */
4161         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4162             (flush_state == FLUSH_DELALLOC ||
4163              flush_state == FLUSH_DELALLOC_WAIT))
4164                 flush_state = ALLOC_CHUNK;
4165
4166         if (!ret)
4167                 goto again;
4168         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4169                  flush_state < COMMIT_TRANS)
4170                 goto again;
4171         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4172                  flush_state <= COMMIT_TRANS)
4173                 goto again;
4174
4175 out:
4176         if (ret == -ENOSPC &&
4177             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4178                 struct btrfs_block_rsv *global_rsv =
4179                         &root->fs_info->global_block_rsv;
4180
4181                 if (block_rsv != global_rsv &&
4182                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4183                         ret = 0;
4184         }
4185         if (flushing) {
4186                 spin_lock(&space_info->lock);
4187                 space_info->flush = 0;
4188                 wake_up_all(&space_info->wait);
4189                 spin_unlock(&space_info->lock);
4190         }
4191         return ret;
4192 }
4193
4194 static struct btrfs_block_rsv *get_block_rsv(
4195                                         const struct btrfs_trans_handle *trans,
4196                                         const struct btrfs_root *root)
4197 {
4198         struct btrfs_block_rsv *block_rsv = NULL;
4199
4200         if (root->ref_cows)
4201                 block_rsv = trans->block_rsv;
4202
4203         if (root == root->fs_info->csum_root && trans->adding_csums)
4204                 block_rsv = trans->block_rsv;
4205
4206         if (!block_rsv)
4207                 block_rsv = root->block_rsv;
4208
4209         if (!block_rsv)
4210                 block_rsv = &root->fs_info->empty_block_rsv;
4211
4212         return block_rsv;
4213 }
4214
4215 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4216                                u64 num_bytes)
4217 {
4218         int ret = -ENOSPC;
4219         spin_lock(&block_rsv->lock);
4220         if (block_rsv->reserved >= num_bytes) {
4221                 block_rsv->reserved -= num_bytes;
4222                 if (block_rsv->reserved < block_rsv->size)
4223                         block_rsv->full = 0;
4224                 ret = 0;
4225         }
4226         spin_unlock(&block_rsv->lock);
4227         return ret;
4228 }
4229
4230 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4231                                 u64 num_bytes, int update_size)
4232 {
4233         spin_lock(&block_rsv->lock);
4234         block_rsv->reserved += num_bytes;
4235         if (update_size)
4236                 block_rsv->size += num_bytes;
4237         else if (block_rsv->reserved >= block_rsv->size)
4238                 block_rsv->full = 1;
4239         spin_unlock(&block_rsv->lock);
4240 }
4241
4242 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4243                                     struct btrfs_block_rsv *block_rsv,
4244                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4245 {
4246         struct btrfs_space_info *space_info = block_rsv->space_info;
4247
4248         spin_lock(&block_rsv->lock);
4249         if (num_bytes == (u64)-1)
4250                 num_bytes = block_rsv->size;
4251         block_rsv->size -= num_bytes;
4252         if (block_rsv->reserved >= block_rsv->size) {
4253                 num_bytes = block_rsv->reserved - block_rsv->size;
4254                 block_rsv->reserved = block_rsv->size;
4255                 block_rsv->full = 1;
4256         } else {
4257                 num_bytes = 0;
4258         }
4259         spin_unlock(&block_rsv->lock);
4260
4261         if (num_bytes > 0) {
4262                 if (dest) {
4263                         spin_lock(&dest->lock);
4264                         if (!dest->full) {
4265                                 u64 bytes_to_add;
4266
4267                                 bytes_to_add = dest->size - dest->reserved;
4268                                 bytes_to_add = min(num_bytes, bytes_to_add);
4269                                 dest->reserved += bytes_to_add;
4270                                 if (dest->reserved >= dest->size)
4271                                         dest->full = 1;
4272                                 num_bytes -= bytes_to_add;
4273                         }
4274                         spin_unlock(&dest->lock);
4275                 }
4276                 if (num_bytes) {
4277                         spin_lock(&space_info->lock);
4278                         space_info->bytes_may_use -= num_bytes;
4279                         trace_btrfs_space_reservation(fs_info, "space_info",
4280                                         space_info->flags, num_bytes, 0);
4281                         space_info->reservation_progress++;
4282                         spin_unlock(&space_info->lock);
4283                 }
4284         }
4285 }
4286
4287 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4288                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4289 {
4290         int ret;
4291
4292         ret = block_rsv_use_bytes(src, num_bytes);
4293         if (ret)
4294                 return ret;
4295
4296         block_rsv_add_bytes(dst, num_bytes, 1);
4297         return 0;
4298 }
4299
4300 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4301 {
4302         memset(rsv, 0, sizeof(*rsv));
4303         spin_lock_init(&rsv->lock);
4304         rsv->type = type;
4305 }
4306
4307 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4308                                               unsigned short type)
4309 {
4310         struct btrfs_block_rsv *block_rsv;
4311         struct btrfs_fs_info *fs_info = root->fs_info;
4312
4313         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4314         if (!block_rsv)
4315                 return NULL;
4316
4317         btrfs_init_block_rsv(block_rsv, type);
4318         block_rsv->space_info = __find_space_info(fs_info,
4319                                                   BTRFS_BLOCK_GROUP_METADATA);
4320         return block_rsv;
4321 }
4322
4323 void btrfs_free_block_rsv(struct btrfs_root *root,
4324                           struct btrfs_block_rsv *rsv)
4325 {
4326         if (!rsv)
4327                 return;
4328         btrfs_block_rsv_release(root, rsv, (u64)-1);
4329         kfree(rsv);
4330 }
4331
4332 int btrfs_block_rsv_add(struct btrfs_root *root,
4333                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4334                         enum btrfs_reserve_flush_enum flush)
4335 {
4336         int ret;
4337
4338         if (num_bytes == 0)
4339                 return 0;
4340
4341         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4342         if (!ret) {
4343                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4344                 return 0;
4345         }
4346
4347         return ret;
4348 }
4349
4350 int btrfs_block_rsv_check(struct btrfs_root *root,
4351                           struct btrfs_block_rsv *block_rsv, int min_factor)
4352 {
4353         u64 num_bytes = 0;
4354         int ret = -ENOSPC;
4355
4356         if (!block_rsv)
4357                 return 0;
4358
4359         spin_lock(&block_rsv->lock);
4360         num_bytes = div_factor(block_rsv->size, min_factor);
4361         if (block_rsv->reserved >= num_bytes)
4362                 ret = 0;
4363         spin_unlock(&block_rsv->lock);
4364
4365         return ret;
4366 }
4367
4368 int btrfs_block_rsv_refill(struct btrfs_root *root,
4369                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4370                            enum btrfs_reserve_flush_enum flush)
4371 {
4372         u64 num_bytes = 0;
4373         int ret = -ENOSPC;
4374
4375         if (!block_rsv)
4376                 return 0;
4377
4378         spin_lock(&block_rsv->lock);
4379         num_bytes = min_reserved;
4380         if (block_rsv->reserved >= num_bytes)
4381                 ret = 0;
4382         else
4383                 num_bytes -= block_rsv->reserved;
4384         spin_unlock(&block_rsv->lock);
4385
4386         if (!ret)
4387                 return 0;
4388
4389         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4390         if (!ret) {
4391                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4392                 return 0;
4393         }
4394
4395         return ret;
4396 }
4397
4398 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4399                             struct btrfs_block_rsv *dst_rsv,
4400                             u64 num_bytes)
4401 {
4402         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4403 }
4404
4405 void btrfs_block_rsv_release(struct btrfs_root *root,
4406                              struct btrfs_block_rsv *block_rsv,
4407                              u64 num_bytes)
4408 {
4409         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4410         if (global_rsv->full || global_rsv == block_rsv ||
4411             block_rsv->space_info != global_rsv->space_info)
4412                 global_rsv = NULL;
4413         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4414                                 num_bytes);
4415 }
4416
4417 /*
4418  * helper to calculate size of global block reservation.
4419  * the desired value is sum of space used by extent tree,
4420  * checksum tree and root tree
4421  */
4422 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4423 {
4424         struct btrfs_space_info *sinfo;
4425         u64 num_bytes;
4426         u64 meta_used;
4427         u64 data_used;
4428         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4429
4430         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4431         spin_lock(&sinfo->lock);
4432         data_used = sinfo->bytes_used;
4433         spin_unlock(&sinfo->lock);
4434
4435         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4436         spin_lock(&sinfo->lock);
4437         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4438                 data_used = 0;
4439         meta_used = sinfo->bytes_used;
4440         spin_unlock(&sinfo->lock);
4441
4442         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4443                     csum_size * 2;
4444         num_bytes += div64_u64(data_used + meta_used, 50);
4445
4446         if (num_bytes * 3 > meta_used)
4447                 num_bytes = div64_u64(meta_used, 3);
4448
4449         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4450 }
4451
4452 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4453 {
4454         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4455         struct btrfs_space_info *sinfo = block_rsv->space_info;
4456         u64 num_bytes;
4457
4458         num_bytes = calc_global_metadata_size(fs_info);
4459
4460         spin_lock(&sinfo->lock);
4461         spin_lock(&block_rsv->lock);
4462
4463         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4464
4465         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4466                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4467                     sinfo->bytes_may_use;
4468
4469         if (sinfo->total_bytes > num_bytes) {
4470                 num_bytes = sinfo->total_bytes - num_bytes;
4471                 block_rsv->reserved += num_bytes;
4472                 sinfo->bytes_may_use += num_bytes;
4473                 trace_btrfs_space_reservation(fs_info, "space_info",
4474                                       sinfo->flags, num_bytes, 1);
4475         }
4476
4477         if (block_rsv->reserved >= block_rsv->size) {
4478                 num_bytes = block_rsv->reserved - block_rsv->size;
4479                 sinfo->bytes_may_use -= num_bytes;
4480                 trace_btrfs_space_reservation(fs_info, "space_info",
4481                                       sinfo->flags, num_bytes, 0);
4482                 sinfo->reservation_progress++;
4483                 block_rsv->reserved = block_rsv->size;
4484                 block_rsv->full = 1;
4485         }
4486
4487         spin_unlock(&block_rsv->lock);
4488         spin_unlock(&sinfo->lock);
4489 }
4490
4491 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4492 {
4493         struct btrfs_space_info *space_info;
4494
4495         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4496         fs_info->chunk_block_rsv.space_info = space_info;
4497
4498         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4499         fs_info->global_block_rsv.space_info = space_info;
4500         fs_info->delalloc_block_rsv.space_info = space_info;
4501         fs_info->trans_block_rsv.space_info = space_info;
4502         fs_info->empty_block_rsv.space_info = space_info;
4503         fs_info->delayed_block_rsv.space_info = space_info;
4504
4505         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4506         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4507         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4508         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4509         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4510
4511         update_global_block_rsv(fs_info);
4512 }
4513
4514 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4515 {
4516         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4517                                 (u64)-1);
4518         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4519         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4520         WARN_ON(fs_info->trans_block_rsv.size > 0);
4521         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4522         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4523         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4524         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4525         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4526 }
4527
4528 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4529                                   struct btrfs_root *root)
4530 {
4531         if (!trans->block_rsv)
4532                 return;
4533
4534         if (!trans->bytes_reserved)
4535                 return;
4536
4537         trace_btrfs_space_reservation(root->fs_info, "transaction",
4538                                       trans->transid, trans->bytes_reserved, 0);
4539         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4540         trans->bytes_reserved = 0;
4541 }
4542
4543 /* Can only return 0 or -ENOSPC */
4544 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4545                                   struct inode *inode)
4546 {
4547         struct btrfs_root *root = BTRFS_I(inode)->root;
4548         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4549         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4550
4551         /*
4552          * We need to hold space in order to delete our orphan item once we've
4553          * added it, so this takes the reservation so we can release it later
4554          * when we are truly done with the orphan item.
4555          */
4556         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4557         trace_btrfs_space_reservation(root->fs_info, "orphan",
4558                                       btrfs_ino(inode), num_bytes, 1);
4559         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4560 }
4561
4562 void btrfs_orphan_release_metadata(struct inode *inode)
4563 {
4564         struct btrfs_root *root = BTRFS_I(inode)->root;
4565         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4566         trace_btrfs_space_reservation(root->fs_info, "orphan",
4567                                       btrfs_ino(inode), num_bytes, 0);
4568         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4569 }
4570
4571 /*
4572  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4573  * root: the root of the parent directory
4574  * rsv: block reservation
4575  * items: the number of items that we need do reservation
4576  * qgroup_reserved: used to return the reserved size in qgroup
4577  *
4578  * This function is used to reserve the space for snapshot/subvolume
4579  * creation and deletion. Those operations are different with the
4580  * common file/directory operations, they change two fs/file trees
4581  * and root tree, the number of items that the qgroup reserves is
4582  * different with the free space reservation. So we can not use
4583  * the space reseravtion mechanism in start_transaction().
4584  */
4585 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4586                                      struct btrfs_block_rsv *rsv,
4587                                      int items,
4588                                      u64 *qgroup_reserved)
4589 {
4590         u64 num_bytes;
4591         int ret;
4592
4593         if (root->fs_info->quota_enabled) {
4594                 /* One for parent inode, two for dir entries */
4595                 num_bytes = 3 * root->leafsize;
4596                 ret = btrfs_qgroup_reserve(root, num_bytes);
4597                 if (ret)
4598                         return ret;
4599         } else {
4600                 num_bytes = 0;
4601         }
4602
4603         *qgroup_reserved = num_bytes;
4604
4605         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4606         rsv->space_info = __find_space_info(root->fs_info,
4607                                             BTRFS_BLOCK_GROUP_METADATA);
4608         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4609                                   BTRFS_RESERVE_FLUSH_ALL);
4610         if (ret) {
4611                 if (*qgroup_reserved)
4612                         btrfs_qgroup_free(root, *qgroup_reserved);
4613         }
4614
4615         return ret;
4616 }
4617
4618 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4619                                       struct btrfs_block_rsv *rsv,
4620                                       u64 qgroup_reserved)
4621 {
4622         btrfs_block_rsv_release(root, rsv, (u64)-1);
4623         if (qgroup_reserved)
4624                 btrfs_qgroup_free(root, qgroup_reserved);
4625 }
4626
4627 /**
4628  * drop_outstanding_extent - drop an outstanding extent
4629  * @inode: the inode we're dropping the extent for
4630  *
4631  * This is called when we are freeing up an outstanding extent, either called
4632  * after an error or after an extent is written.  This will return the number of
4633  * reserved extents that need to be freed.  This must be called with
4634  * BTRFS_I(inode)->lock held.
4635  */
4636 static unsigned drop_outstanding_extent(struct inode *inode)
4637 {
4638         unsigned drop_inode_space = 0;
4639         unsigned dropped_extents = 0;
4640
4641         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4642         BTRFS_I(inode)->outstanding_extents--;
4643
4644         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4645             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4646                                &BTRFS_I(inode)->runtime_flags))
4647                 drop_inode_space = 1;
4648
4649         /*
4650          * If we have more or the same amount of outsanding extents than we have
4651          * reserved then we need to leave the reserved extents count alone.
4652          */
4653         if (BTRFS_I(inode)->outstanding_extents >=
4654             BTRFS_I(inode)->reserved_extents)
4655                 return drop_inode_space;
4656
4657         dropped_extents = BTRFS_I(inode)->reserved_extents -
4658                 BTRFS_I(inode)->outstanding_extents;
4659         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4660         return dropped_extents + drop_inode_space;
4661 }
4662
4663 /**
4664  * calc_csum_metadata_size - return the amount of metada space that must be
4665  *      reserved/free'd for the given bytes.
4666  * @inode: the inode we're manipulating
4667  * @num_bytes: the number of bytes in question
4668  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4669  *
4670  * This adjusts the number of csum_bytes in the inode and then returns the
4671  * correct amount of metadata that must either be reserved or freed.  We
4672  * calculate how many checksums we can fit into one leaf and then divide the
4673  * number of bytes that will need to be checksumed by this value to figure out
4674  * how many checksums will be required.  If we are adding bytes then the number
4675  * may go up and we will return the number of additional bytes that must be
4676  * reserved.  If it is going down we will return the number of bytes that must
4677  * be freed.
4678  *
4679  * This must be called with BTRFS_I(inode)->lock held.
4680  */
4681 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4682                                    int reserve)
4683 {
4684         struct btrfs_root *root = BTRFS_I(inode)->root;
4685         u64 csum_size;
4686         int num_csums_per_leaf;
4687         int num_csums;
4688         int old_csums;
4689
4690         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4691             BTRFS_I(inode)->csum_bytes == 0)
4692                 return 0;
4693
4694         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4695         if (reserve)
4696                 BTRFS_I(inode)->csum_bytes += num_bytes;
4697         else
4698                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4699         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4700         num_csums_per_leaf = (int)div64_u64(csum_size,
4701                                             sizeof(struct btrfs_csum_item) +
4702                                             sizeof(struct btrfs_disk_key));
4703         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4704         num_csums = num_csums + num_csums_per_leaf - 1;
4705         num_csums = num_csums / num_csums_per_leaf;
4706
4707         old_csums = old_csums + num_csums_per_leaf - 1;
4708         old_csums = old_csums / num_csums_per_leaf;
4709
4710         /* No change, no need to reserve more */
4711         if (old_csums == num_csums)
4712                 return 0;
4713
4714         if (reserve)
4715                 return btrfs_calc_trans_metadata_size(root,
4716                                                       num_csums - old_csums);
4717
4718         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4719 }
4720
4721 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4722 {
4723         struct btrfs_root *root = BTRFS_I(inode)->root;
4724         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4725         u64 to_reserve = 0;
4726         u64 csum_bytes;
4727         unsigned nr_extents = 0;
4728         int extra_reserve = 0;
4729         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
4730         int ret = 0;
4731         bool delalloc_lock = true;
4732         u64 to_free = 0;
4733         unsigned dropped;
4734
4735         /* If we are a free space inode we need to not flush since we will be in
4736          * the middle of a transaction commit.  We also don't need the delalloc
4737          * mutex since we won't race with anybody.  We need this mostly to make
4738          * lockdep shut its filthy mouth.
4739          */
4740         if (btrfs_is_free_space_inode(inode)) {
4741                 flush = BTRFS_RESERVE_NO_FLUSH;
4742                 delalloc_lock = false;
4743         }
4744
4745         if (flush != BTRFS_RESERVE_NO_FLUSH &&
4746             btrfs_transaction_in_commit(root->fs_info))
4747                 schedule_timeout(1);
4748
4749         if (delalloc_lock)
4750                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4751
4752         num_bytes = ALIGN(num_bytes, root->sectorsize);
4753
4754         spin_lock(&BTRFS_I(inode)->lock);
4755         BTRFS_I(inode)->outstanding_extents++;
4756
4757         if (BTRFS_I(inode)->outstanding_extents >
4758             BTRFS_I(inode)->reserved_extents)
4759                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4760                         BTRFS_I(inode)->reserved_extents;
4761
4762         /*
4763          * Add an item to reserve for updating the inode when we complete the
4764          * delalloc io.
4765          */
4766         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4767                       &BTRFS_I(inode)->runtime_flags)) {
4768                 nr_extents++;
4769                 extra_reserve = 1;
4770         }
4771
4772         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4773         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4774         csum_bytes = BTRFS_I(inode)->csum_bytes;
4775         spin_unlock(&BTRFS_I(inode)->lock);
4776
4777         if (root->fs_info->quota_enabled) {
4778                 ret = btrfs_qgroup_reserve(root, num_bytes +
4779                                            nr_extents * root->leafsize);
4780                 if (ret)
4781                         goto out_fail;
4782         }
4783
4784         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4785         if (unlikely(ret)) {
4786                 if (root->fs_info->quota_enabled)
4787                         btrfs_qgroup_free(root, num_bytes +
4788                                                 nr_extents * root->leafsize);
4789                 goto out_fail;
4790         }
4791
4792         spin_lock(&BTRFS_I(inode)->lock);
4793         if (extra_reserve) {
4794                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4795                         &BTRFS_I(inode)->runtime_flags);
4796                 nr_extents--;
4797         }
4798         BTRFS_I(inode)->reserved_extents += nr_extents;
4799         spin_unlock(&BTRFS_I(inode)->lock);
4800
4801         if (delalloc_lock)
4802                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4803
4804         if (to_reserve)
4805                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4806                                               btrfs_ino(inode), to_reserve, 1);
4807         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4808
4809         return 0;
4810
4811 out_fail:
4812         spin_lock(&BTRFS_I(inode)->lock);
4813         dropped = drop_outstanding_extent(inode);
4814         /*
4815          * If the inodes csum_bytes is the same as the original
4816          * csum_bytes then we know we haven't raced with any free()ers
4817          * so we can just reduce our inodes csum bytes and carry on.
4818          */
4819         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
4820                 calc_csum_metadata_size(inode, num_bytes, 0);
4821         } else {
4822                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
4823                 u64 bytes;
4824
4825                 /*
4826                  * This is tricky, but first we need to figure out how much we
4827                  * free'd from any free-ers that occured during this
4828                  * reservation, so we reset ->csum_bytes to the csum_bytes
4829                  * before we dropped our lock, and then call the free for the
4830                  * number of bytes that were freed while we were trying our
4831                  * reservation.
4832                  */
4833                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
4834                 BTRFS_I(inode)->csum_bytes = csum_bytes;
4835                 to_free = calc_csum_metadata_size(inode, bytes, 0);
4836
4837
4838                 /*
4839                  * Now we need to see how much we would have freed had we not
4840                  * been making this reservation and our ->csum_bytes were not
4841                  * artificially inflated.
4842                  */
4843                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
4844                 bytes = csum_bytes - orig_csum_bytes;
4845                 bytes = calc_csum_metadata_size(inode, bytes, 0);
4846
4847                 /*
4848                  * Now reset ->csum_bytes to what it should be.  If bytes is
4849                  * more than to_free then we would have free'd more space had we
4850                  * not had an artificially high ->csum_bytes, so we need to free
4851                  * the remainder.  If bytes is the same or less then we don't
4852                  * need to do anything, the other free-ers did the correct
4853                  * thing.
4854                  */
4855                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
4856                 if (bytes > to_free)
4857                         to_free = bytes - to_free;
4858                 else
4859                         to_free = 0;
4860         }
4861         spin_unlock(&BTRFS_I(inode)->lock);
4862         if (dropped)
4863                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4864
4865         if (to_free) {
4866                 btrfs_block_rsv_release(root, block_rsv, to_free);
4867                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
4868                                               btrfs_ino(inode), to_free, 0);
4869         }
4870         if (delalloc_lock)
4871                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4872         return ret;
4873 }
4874
4875 /**
4876  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4877  * @inode: the inode to release the reservation for
4878  * @num_bytes: the number of bytes we're releasing
4879  *
4880  * This will release the metadata reservation for an inode.  This can be called
4881  * once we complete IO for a given set of bytes to release their metadata
4882  * reservations.
4883  */
4884 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4885 {
4886         struct btrfs_root *root = BTRFS_I(inode)->root;
4887         u64 to_free = 0;
4888         unsigned dropped;
4889
4890         num_bytes = ALIGN(num_bytes, root->sectorsize);
4891         spin_lock(&BTRFS_I(inode)->lock);
4892         dropped = drop_outstanding_extent(inode);
4893
4894         if (num_bytes)
4895                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4896         spin_unlock(&BTRFS_I(inode)->lock);
4897         if (dropped > 0)
4898                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4899
4900         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4901                                       btrfs_ino(inode), to_free, 0);
4902         if (root->fs_info->quota_enabled) {
4903                 btrfs_qgroup_free(root, num_bytes +
4904                                         dropped * root->leafsize);
4905         }
4906
4907         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4908                                 to_free);
4909 }
4910
4911 /**
4912  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4913  * @inode: inode we're writing to
4914  * @num_bytes: the number of bytes we want to allocate
4915  *
4916  * This will do the following things
4917  *
4918  * o reserve space in the data space info for num_bytes
4919  * o reserve space in the metadata space info based on number of outstanding
4920  *   extents and how much csums will be needed
4921  * o add to the inodes ->delalloc_bytes
4922  * o add it to the fs_info's delalloc inodes list.
4923  *
4924  * This will return 0 for success and -ENOSPC if there is no space left.
4925  */
4926 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4927 {
4928         int ret;
4929
4930         ret = btrfs_check_data_free_space(inode, num_bytes);
4931         if (ret)
4932                 return ret;
4933
4934         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4935         if (ret) {
4936                 btrfs_free_reserved_data_space(inode, num_bytes);
4937                 return ret;
4938         }
4939
4940         return 0;
4941 }
4942
4943 /**
4944  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4945  * @inode: inode we're releasing space for
4946  * @num_bytes: the number of bytes we want to free up
4947  *
4948  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4949  * called in the case that we don't need the metadata AND data reservations
4950  * anymore.  So if there is an error or we insert an inline extent.
4951  *
4952  * This function will release the metadata space that was not used and will
4953  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4954  * list if there are no delalloc bytes left.
4955  */
4956 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4957 {
4958         btrfs_delalloc_release_metadata(inode, num_bytes);
4959         btrfs_free_reserved_data_space(inode, num_bytes);
4960 }
4961
4962 static int update_block_group(struct btrfs_root *root,
4963                               u64 bytenr, u64 num_bytes, int alloc)
4964 {
4965         struct btrfs_block_group_cache *cache = NULL;
4966         struct btrfs_fs_info *info = root->fs_info;
4967         u64 total = num_bytes;
4968         u64 old_val;
4969         u64 byte_in_group;
4970         int factor;
4971
4972         /* block accounting for super block */
4973         spin_lock(&info->delalloc_lock);
4974         old_val = btrfs_super_bytes_used(info->super_copy);
4975         if (alloc)
4976                 old_val += num_bytes;
4977         else
4978                 old_val -= num_bytes;
4979         btrfs_set_super_bytes_used(info->super_copy, old_val);
4980         spin_unlock(&info->delalloc_lock);
4981
4982         while (total) {
4983                 cache = btrfs_lookup_block_group(info, bytenr);
4984                 if (!cache)
4985                         return -ENOENT;
4986                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4987                                     BTRFS_BLOCK_GROUP_RAID1 |
4988                                     BTRFS_BLOCK_GROUP_RAID10))
4989                         factor = 2;
4990                 else
4991                         factor = 1;
4992                 /*
4993                  * If this block group has free space cache written out, we
4994                  * need to make sure to load it if we are removing space.  This
4995                  * is because we need the unpinning stage to actually add the
4996                  * space back to the block group, otherwise we will leak space.
4997                  */
4998                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4999                         cache_block_group(cache, 1);
5000
5001                 byte_in_group = bytenr - cache->key.objectid;
5002                 WARN_ON(byte_in_group > cache->key.offset);
5003
5004                 spin_lock(&cache->space_info->lock);
5005                 spin_lock(&cache->lock);
5006
5007                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5008                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5009                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5010
5011                 cache->dirty = 1;
5012                 old_val = btrfs_block_group_used(&cache->item);
5013                 num_bytes = min(total, cache->key.offset - byte_in_group);
5014                 if (alloc) {
5015                         old_val += num_bytes;
5016                         btrfs_set_block_group_used(&cache->item, old_val);
5017                         cache->reserved -= num_bytes;
5018                         cache->space_info->bytes_reserved -= num_bytes;
5019                         cache->space_info->bytes_used += num_bytes;
5020                         cache->space_info->disk_used += num_bytes * factor;
5021                         spin_unlock(&cache->lock);
5022                         spin_unlock(&cache->space_info->lock);
5023                 } else {
5024                         old_val -= num_bytes;
5025                         btrfs_set_block_group_used(&cache->item, old_val);
5026                         cache->pinned += num_bytes;
5027                         cache->space_info->bytes_pinned += num_bytes;
5028                         cache->space_info->bytes_used -= num_bytes;
5029                         cache->space_info->disk_used -= num_bytes * factor;
5030                         spin_unlock(&cache->lock);
5031                         spin_unlock(&cache->space_info->lock);
5032
5033                         set_extent_dirty(info->pinned_extents,
5034                                          bytenr, bytenr + num_bytes - 1,
5035                                          GFP_NOFS | __GFP_NOFAIL);
5036                 }
5037                 btrfs_put_block_group(cache);
5038                 total -= num_bytes;
5039                 bytenr += num_bytes;
5040         }
5041         return 0;
5042 }
5043
5044 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5045 {
5046         struct btrfs_block_group_cache *cache;
5047         u64 bytenr;
5048
5049         spin_lock(&root->fs_info->block_group_cache_lock);
5050         bytenr = root->fs_info->first_logical_byte;
5051         spin_unlock(&root->fs_info->block_group_cache_lock);
5052
5053         if (bytenr < (u64)-1)
5054                 return bytenr;
5055
5056         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5057         if (!cache)
5058                 return 0;
5059
5060         bytenr = cache->key.objectid;
5061         btrfs_put_block_group(cache);
5062
5063         return bytenr;
5064 }
5065
5066 static int pin_down_extent(struct btrfs_root *root,
5067                            struct btrfs_block_group_cache *cache,
5068                            u64 bytenr, u64 num_bytes, int reserved)
5069 {
5070         spin_lock(&cache->space_info->lock);
5071         spin_lock(&cache->lock);
5072         cache->pinned += num_bytes;
5073         cache->space_info->bytes_pinned += num_bytes;
5074         if (reserved) {
5075                 cache->reserved -= num_bytes;
5076                 cache->space_info->bytes_reserved -= num_bytes;
5077         }
5078         spin_unlock(&cache->lock);
5079         spin_unlock(&cache->space_info->lock);
5080
5081         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5082                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5083         return 0;
5084 }
5085
5086 /*
5087  * this function must be called within transaction
5088  */
5089 int btrfs_pin_extent(struct btrfs_root *root,
5090                      u64 bytenr, u64 num_bytes, int reserved)
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         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5098
5099         btrfs_put_block_group(cache);
5100         return 0;
5101 }
5102
5103 /*
5104  * this function must be called within transaction
5105  */
5106 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5107                                     u64 bytenr, u64 num_bytes)
5108 {
5109         struct btrfs_block_group_cache *cache;
5110
5111         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5112         BUG_ON(!cache); /* Logic error */
5113
5114         /*
5115          * pull in the free space cache (if any) so that our pin
5116          * removes the free space from the cache.  We have load_only set
5117          * to one because the slow code to read in the free extents does check
5118          * the pinned extents.
5119          */
5120         cache_block_group(cache, 1);
5121
5122         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5123
5124         /* remove us from the free space cache (if we're there at all) */
5125         btrfs_remove_free_space(cache, bytenr, num_bytes);
5126         btrfs_put_block_group(cache);
5127         return 0;
5128 }
5129
5130 /**
5131  * btrfs_update_reserved_bytes - update the block_group and space info counters
5132  * @cache:      The cache we are manipulating
5133  * @num_bytes:  The number of bytes in question
5134  * @reserve:    One of the reservation enums
5135  *
5136  * This is called by the allocator when it reserves space, or by somebody who is
5137  * freeing space that was never actually used on disk.  For example if you
5138  * reserve some space for a new leaf in transaction A and before transaction A
5139  * commits you free that leaf, you call this with reserve set to 0 in order to
5140  * clear the reservation.
5141  *
5142  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5143  * ENOSPC accounting.  For data we handle the reservation through clearing the
5144  * delalloc bits in the io_tree.  We have to do this since we could end up
5145  * allocating less disk space for the amount of data we have reserved in the
5146  * case of compression.
5147  *
5148  * If this is a reservation and the block group has become read only we cannot
5149  * make the reservation and return -EAGAIN, otherwise this function always
5150  * succeeds.
5151  */
5152 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5153                                        u64 num_bytes, int reserve)
5154 {
5155         struct btrfs_space_info *space_info = cache->space_info;
5156         int ret = 0;
5157
5158         spin_lock(&space_info->lock);
5159         spin_lock(&cache->lock);
5160         if (reserve != RESERVE_FREE) {
5161                 if (cache->ro) {
5162                         ret = -EAGAIN;
5163                 } else {
5164                         cache->reserved += num_bytes;
5165                         space_info->bytes_reserved += num_bytes;
5166                         if (reserve == RESERVE_ALLOC) {
5167                                 trace_btrfs_space_reservation(cache->fs_info,
5168                                                 "space_info", space_info->flags,
5169                                                 num_bytes, 0);
5170                                 space_info->bytes_may_use -= num_bytes;
5171                         }
5172                 }
5173         } else {
5174                 if (cache->ro)
5175                         space_info->bytes_readonly += num_bytes;
5176                 cache->reserved -= num_bytes;
5177                 space_info->bytes_reserved -= num_bytes;
5178                 space_info->reservation_progress++;
5179         }
5180         spin_unlock(&cache->lock);
5181         spin_unlock(&space_info->lock);
5182         return ret;
5183 }
5184
5185 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5186                                 struct btrfs_root *root)
5187 {
5188         struct btrfs_fs_info *fs_info = root->fs_info;
5189         struct btrfs_caching_control *next;
5190         struct btrfs_caching_control *caching_ctl;
5191         struct btrfs_block_group_cache *cache;
5192
5193         down_write(&fs_info->extent_commit_sem);
5194
5195         list_for_each_entry_safe(caching_ctl, next,
5196                                  &fs_info->caching_block_groups, list) {
5197                 cache = caching_ctl->block_group;
5198                 if (block_group_cache_done(cache)) {
5199                         cache->last_byte_to_unpin = (u64)-1;
5200                         list_del_init(&caching_ctl->list);
5201                         put_caching_control(caching_ctl);
5202                 } else {
5203                         cache->last_byte_to_unpin = caching_ctl->progress;
5204                 }
5205         }
5206
5207         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5208                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5209         else
5210                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5211
5212         up_write(&fs_info->extent_commit_sem);
5213
5214         update_global_block_rsv(fs_info);
5215 }
5216
5217 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5218 {
5219         struct btrfs_fs_info *fs_info = root->fs_info;
5220         struct btrfs_block_group_cache *cache = NULL;
5221         struct btrfs_space_info *space_info;
5222         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5223         u64 len;
5224         bool readonly;
5225
5226         while (start <= end) {
5227                 readonly = false;
5228                 if (!cache ||
5229                     start >= cache->key.objectid + cache->key.offset) {
5230                         if (cache)
5231                                 btrfs_put_block_group(cache);
5232                         cache = btrfs_lookup_block_group(fs_info, start);
5233                         BUG_ON(!cache); /* Logic error */
5234                 }
5235
5236                 len = cache->key.objectid + cache->key.offset - start;
5237                 len = min(len, end + 1 - start);
5238
5239                 if (start < cache->last_byte_to_unpin) {
5240                         len = min(len, cache->last_byte_to_unpin - start);
5241                         btrfs_add_free_space(cache, start, len);
5242                 }
5243
5244                 start += len;
5245                 space_info = cache->space_info;
5246
5247                 spin_lock(&space_info->lock);
5248                 spin_lock(&cache->lock);
5249                 cache->pinned -= len;
5250                 space_info->bytes_pinned -= len;
5251                 if (cache->ro) {
5252                         space_info->bytes_readonly += len;
5253                         readonly = true;
5254                 }
5255                 spin_unlock(&cache->lock);
5256                 if (!readonly && global_rsv->space_info == space_info) {
5257                         spin_lock(&global_rsv->lock);
5258                         if (!global_rsv->full) {
5259                                 len = min(len, global_rsv->size -
5260                                           global_rsv->reserved);
5261                                 global_rsv->reserved += len;
5262                                 space_info->bytes_may_use += len;
5263                                 if (global_rsv->reserved >= global_rsv->size)
5264                                         global_rsv->full = 1;
5265                         }
5266                         spin_unlock(&global_rsv->lock);
5267                 }
5268                 spin_unlock(&space_info->lock);
5269         }
5270
5271         if (cache)
5272                 btrfs_put_block_group(cache);
5273         return 0;
5274 }
5275
5276 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5277                                struct btrfs_root *root)
5278 {
5279         struct btrfs_fs_info *fs_info = root->fs_info;
5280         struct extent_io_tree *unpin;
5281         u64 start;
5282         u64 end;
5283         int ret;
5284
5285         if (trans->aborted)
5286                 return 0;
5287
5288         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5289                 unpin = &fs_info->freed_extents[1];
5290         else
5291                 unpin = &fs_info->freed_extents[0];
5292
5293         while (1) {
5294                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5295                                             EXTENT_DIRTY, NULL);
5296                 if (ret)
5297                         break;
5298
5299                 if (btrfs_test_opt(root, DISCARD))
5300                         ret = btrfs_discard_extent(root, start,
5301                                                    end + 1 - start, NULL);
5302
5303                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5304                 unpin_extent_range(root, start, end);
5305                 cond_resched();
5306         }
5307
5308         return 0;
5309 }
5310
5311 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5312                                 struct btrfs_root *root,
5313                                 u64 bytenr, u64 num_bytes, u64 parent,
5314                                 u64 root_objectid, u64 owner_objectid,
5315                                 u64 owner_offset, int refs_to_drop,
5316                                 struct btrfs_delayed_extent_op *extent_op)
5317 {
5318         struct btrfs_key key;
5319         struct btrfs_path *path;
5320         struct btrfs_fs_info *info = root->fs_info;
5321         struct btrfs_root *extent_root = info->extent_root;
5322         struct extent_buffer *leaf;
5323         struct btrfs_extent_item *ei;
5324         struct btrfs_extent_inline_ref *iref;
5325         int ret;
5326         int is_data;
5327         int extent_slot = 0;
5328         int found_extent = 0;
5329         int num_to_del = 1;
5330         u32 item_size;
5331         u64 refs;
5332
5333         path = btrfs_alloc_path();
5334         if (!path)
5335                 return -ENOMEM;
5336
5337         path->reada = 1;
5338         path->leave_spinning = 1;
5339
5340         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5341         BUG_ON(!is_data && refs_to_drop != 1);
5342
5343         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5344                                     bytenr, num_bytes, parent,
5345                                     root_objectid, owner_objectid,
5346                                     owner_offset);
5347         if (ret == 0) {
5348                 extent_slot = path->slots[0];
5349                 while (extent_slot >= 0) {
5350                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5351                                               extent_slot);
5352                         if (key.objectid != bytenr)
5353                                 break;
5354                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5355                             key.offset == num_bytes) {
5356                                 found_extent = 1;
5357                                 break;
5358                         }
5359                         if (path->slots[0] - extent_slot > 5)
5360                                 break;
5361                         extent_slot--;
5362                 }
5363 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5364                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5365                 if (found_extent && item_size < sizeof(*ei))
5366                         found_extent = 0;
5367 #endif
5368                 if (!found_extent) {
5369                         BUG_ON(iref);
5370                         ret = remove_extent_backref(trans, extent_root, path,
5371                                                     NULL, refs_to_drop,
5372                                                     is_data);
5373                         if (ret) {
5374                                 btrfs_abort_transaction(trans, extent_root, ret);
5375                                 goto out;
5376                         }
5377                         btrfs_release_path(path);
5378                         path->leave_spinning = 1;
5379
5380                         key.objectid = bytenr;
5381                         key.type = BTRFS_EXTENT_ITEM_KEY;
5382                         key.offset = num_bytes;
5383
5384                         ret = btrfs_search_slot(trans, extent_root,
5385                                                 &key, path, -1, 1);
5386                         if (ret) {
5387                                 printk(KERN_ERR "umm, got %d back from search"
5388                                        ", was looking for %llu\n", ret,
5389                                        (unsigned long long)bytenr);
5390                                 if (ret > 0)
5391                                         btrfs_print_leaf(extent_root,
5392                                                          path->nodes[0]);
5393                         }
5394                         if (ret < 0) {
5395                                 btrfs_abort_transaction(trans, extent_root, ret);
5396                                 goto out;
5397                         }
5398                         extent_slot = path->slots[0];
5399                 }
5400         } else if (ret == -ENOENT) {
5401                 btrfs_print_leaf(extent_root, path->nodes[0]);
5402                 WARN_ON(1);
5403                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5404                        "parent %llu root %llu  owner %llu offset %llu\n",
5405                        (unsigned long long)bytenr,
5406                        (unsigned long long)parent,
5407                        (unsigned long long)root_objectid,
5408                        (unsigned long long)owner_objectid,
5409                        (unsigned long long)owner_offset);
5410         } else {
5411                 btrfs_abort_transaction(trans, extent_root, ret);
5412                 goto out;
5413         }
5414
5415         leaf = path->nodes[0];
5416         item_size = btrfs_item_size_nr(leaf, extent_slot);
5417 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5418         if (item_size < sizeof(*ei)) {
5419                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5420                 ret = convert_extent_item_v0(trans, extent_root, path,
5421                                              owner_objectid, 0);
5422                 if (ret < 0) {
5423                         btrfs_abort_transaction(trans, extent_root, ret);
5424                         goto out;
5425                 }
5426
5427                 btrfs_release_path(path);
5428                 path->leave_spinning = 1;
5429
5430                 key.objectid = bytenr;
5431                 key.type = BTRFS_EXTENT_ITEM_KEY;
5432                 key.offset = num_bytes;
5433
5434                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5435                                         -1, 1);
5436                 if (ret) {
5437                         printk(KERN_ERR "umm, got %d back from search"
5438                                ", was looking for %llu\n", ret,
5439                                (unsigned long long)bytenr);
5440                         btrfs_print_leaf(extent_root, path->nodes[0]);
5441                 }
5442                 if (ret < 0) {
5443                         btrfs_abort_transaction(trans, extent_root, ret);
5444                         goto out;
5445                 }
5446
5447                 extent_slot = path->slots[0];
5448                 leaf = path->nodes[0];
5449                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5450         }
5451 #endif
5452         BUG_ON(item_size < sizeof(*ei));
5453         ei = btrfs_item_ptr(leaf, extent_slot,
5454                             struct btrfs_extent_item);
5455         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5456                 struct btrfs_tree_block_info *bi;
5457                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5458                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5459                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5460         }
5461
5462         refs = btrfs_extent_refs(leaf, ei);
5463         BUG_ON(refs < refs_to_drop);
5464         refs -= refs_to_drop;
5465
5466         if (refs > 0) {
5467                 if (extent_op)
5468                         __run_delayed_extent_op(extent_op, leaf, ei);
5469                 /*
5470                  * In the case of inline back ref, reference count will
5471                  * be updated by remove_extent_backref
5472                  */
5473                 if (iref) {
5474                         BUG_ON(!found_extent);
5475                 } else {
5476                         btrfs_set_extent_refs(leaf, ei, refs);
5477                         btrfs_mark_buffer_dirty(leaf);
5478                 }
5479                 if (found_extent) {
5480                         ret = remove_extent_backref(trans, extent_root, path,
5481                                                     iref, refs_to_drop,
5482                                                     is_data);
5483                         if (ret) {
5484                                 btrfs_abort_transaction(trans, extent_root, ret);
5485                                 goto out;
5486                         }
5487                 }
5488         } else {
5489                 if (found_extent) {
5490                         BUG_ON(is_data && refs_to_drop !=
5491                                extent_data_ref_count(root, path, iref));
5492                         if (iref) {
5493                                 BUG_ON(path->slots[0] != extent_slot);
5494                         } else {
5495                                 BUG_ON(path->slots[0] != extent_slot + 1);
5496                                 path->slots[0] = extent_slot;
5497                                 num_to_del = 2;
5498                         }
5499                 }
5500
5501                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5502                                       num_to_del);
5503                 if (ret) {
5504                         btrfs_abort_transaction(trans, extent_root, ret);
5505                         goto out;
5506                 }
5507                 btrfs_release_path(path);
5508
5509                 if (is_data) {
5510                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5511                         if (ret) {
5512                                 btrfs_abort_transaction(trans, extent_root, ret);
5513                                 goto out;
5514                         }
5515                 }
5516
5517                 ret = update_block_group(root, bytenr, num_bytes, 0);
5518                 if (ret) {
5519                         btrfs_abort_transaction(trans, extent_root, ret);
5520                         goto out;
5521                 }
5522         }
5523 out:
5524         btrfs_free_path(path);
5525         return ret;
5526 }
5527
5528 /*
5529  * when we free an block, it is possible (and likely) that we free the last
5530  * delayed ref for that extent as well.  This searches the delayed ref tree for
5531  * a given extent, and if there are no other delayed refs to be processed, it
5532  * removes it from the tree.
5533  */
5534 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5535                                       struct btrfs_root *root, u64 bytenr)
5536 {
5537         struct btrfs_delayed_ref_head *head;
5538         struct btrfs_delayed_ref_root *delayed_refs;
5539         struct btrfs_delayed_ref_node *ref;
5540         struct rb_node *node;
5541         int ret = 0;
5542
5543         delayed_refs = &trans->transaction->delayed_refs;
5544         spin_lock(&delayed_refs->lock);
5545         head = btrfs_find_delayed_ref_head(trans, bytenr);
5546         if (!head)
5547                 goto out;
5548
5549         node = rb_prev(&head->node.rb_node);
5550         if (!node)
5551                 goto out;
5552
5553         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5554
5555         /* there are still entries for this ref, we can't drop it */
5556         if (ref->bytenr == bytenr)
5557                 goto out;
5558
5559         if (head->extent_op) {
5560                 if (!head->must_insert_reserved)
5561                         goto out;
5562                 btrfs_free_delayed_extent_op(head->extent_op);
5563                 head->extent_op = NULL;
5564         }
5565
5566         /*
5567          * waiting for the lock here would deadlock.  If someone else has it
5568          * locked they are already in the process of dropping it anyway
5569          */
5570         if (!mutex_trylock(&head->mutex))
5571                 goto out;
5572
5573         /*
5574          * at this point we have a head with no other entries.  Go
5575          * ahead and process it.
5576          */
5577         head->node.in_tree = 0;
5578         rb_erase(&head->node.rb_node, &delayed_refs->root);
5579
5580         delayed_refs->num_entries--;
5581
5582         /*
5583          * we don't take a ref on the node because we're removing it from the
5584          * tree, so we just steal the ref the tree was holding.
5585          */
5586         delayed_refs->num_heads--;
5587         if (list_empty(&head->cluster))
5588                 delayed_refs->num_heads_ready--;
5589
5590         list_del_init(&head->cluster);
5591         spin_unlock(&delayed_refs->lock);
5592
5593         BUG_ON(head->extent_op);
5594         if (head->must_insert_reserved)
5595                 ret = 1;
5596
5597         mutex_unlock(&head->mutex);
5598         btrfs_put_delayed_ref(&head->node);
5599         return ret;
5600 out:
5601         spin_unlock(&delayed_refs->lock);
5602         return 0;
5603 }
5604
5605 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5606                            struct btrfs_root *root,
5607                            struct extent_buffer *buf,
5608                            u64 parent, int last_ref)
5609 {
5610         struct btrfs_block_group_cache *cache = NULL;
5611         int ret;
5612
5613         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5614                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5615                                         buf->start, buf->len,
5616                                         parent, root->root_key.objectid,
5617                                         btrfs_header_level(buf),
5618                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5619                 BUG_ON(ret); /* -ENOMEM */
5620         }
5621
5622         if (!last_ref)
5623                 return;
5624
5625         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5626
5627         if (btrfs_header_generation(buf) == trans->transid) {
5628                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5629                         ret = check_ref_cleanup(trans, root, buf->start);
5630                         if (!ret)
5631                                 goto out;
5632                 }
5633
5634                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5635                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5636                         goto out;
5637                 }
5638
5639                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5640
5641                 btrfs_add_free_space(cache, buf->start, buf->len);
5642                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5643         }
5644 out:
5645         /*
5646          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5647          * anymore.
5648          */
5649         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5650         btrfs_put_block_group(cache);
5651 }
5652
5653 /* Can return -ENOMEM */
5654 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5655                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5656                       u64 owner, u64 offset, int for_cow)
5657 {
5658         int ret;
5659         struct btrfs_fs_info *fs_info = root->fs_info;
5660
5661         /*
5662          * tree log blocks never actually go into the extent allocation
5663          * tree, just update pinning info and exit early.
5664          */
5665         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5666                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5667                 /* unlocks the pinned mutex */
5668                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5669                 ret = 0;
5670         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5671                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5672                                         num_bytes,
5673                                         parent, root_objectid, (int)owner,
5674                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5675         } else {
5676                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5677                                                 num_bytes,
5678                                                 parent, root_objectid, owner,
5679                                                 offset, BTRFS_DROP_DELAYED_REF,
5680                                                 NULL, for_cow);
5681         }
5682         return ret;
5683 }
5684
5685 static u64 stripe_align(struct btrfs_root *root,
5686                         struct btrfs_block_group_cache *cache,
5687                         u64 val, u64 num_bytes)
5688 {
5689         u64 ret = ALIGN(val, root->stripesize);
5690         return ret;
5691 }
5692
5693 /*
5694  * when we wait for progress in the block group caching, its because
5695  * our allocation attempt failed at least once.  So, we must sleep
5696  * and let some progress happen before we try again.
5697  *
5698  * This function will sleep at least once waiting for new free space to
5699  * show up, and then it will check the block group free space numbers
5700  * for our min num_bytes.  Another option is to have it go ahead
5701  * and look in the rbtree for a free extent of a given size, but this
5702  * is a good start.
5703  */
5704 static noinline int
5705 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5706                                 u64 num_bytes)
5707 {
5708         struct btrfs_caching_control *caching_ctl;
5709
5710         caching_ctl = get_caching_control(cache);
5711         if (!caching_ctl)
5712                 return 0;
5713
5714         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5715                    (cache->free_space_ctl->free_space >= num_bytes));
5716
5717         put_caching_control(caching_ctl);
5718         return 0;
5719 }
5720
5721 static noinline int
5722 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5723 {
5724         struct btrfs_caching_control *caching_ctl;
5725
5726         caching_ctl = get_caching_control(cache);
5727         if (!caching_ctl)
5728                 return 0;
5729
5730         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5731
5732         put_caching_control(caching_ctl);
5733         return 0;
5734 }
5735
5736 int __get_raid_index(u64 flags)
5737 {
5738         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5739                 return BTRFS_RAID_RAID10;
5740         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5741                 return BTRFS_RAID_RAID1;
5742         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5743                 return BTRFS_RAID_DUP;
5744         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5745                 return BTRFS_RAID_RAID0;
5746         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
5747                 return BTRFS_RAID_RAID5;
5748         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
5749                 return BTRFS_RAID_RAID6;
5750
5751         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
5752 }
5753
5754 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5755 {
5756         return __get_raid_index(cache->flags);
5757 }
5758
5759 enum btrfs_loop_type {
5760         LOOP_CACHING_NOWAIT = 0,
5761         LOOP_CACHING_WAIT = 1,
5762         LOOP_ALLOC_CHUNK = 2,
5763         LOOP_NO_EMPTY_SIZE = 3,
5764 };
5765
5766 /*
5767  * walks the btree of allocated extents and find a hole of a given size.
5768  * The key ins is changed to record the hole:
5769  * ins->objectid == block start
5770  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5771  * ins->offset == number of blocks
5772  * Any available blocks before search_start are skipped.
5773  */
5774 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5775                                      struct btrfs_root *orig_root,
5776                                      u64 num_bytes, u64 empty_size,
5777                                      u64 hint_byte, struct btrfs_key *ins,
5778                                      u64 data)
5779 {
5780         int ret = 0;
5781         struct btrfs_root *root = orig_root->fs_info->extent_root;
5782         struct btrfs_free_cluster *last_ptr = NULL;
5783         struct btrfs_block_group_cache *block_group = NULL;
5784         struct btrfs_block_group_cache *used_block_group;
5785         u64 search_start = 0;
5786         int empty_cluster = 2 * 1024 * 1024;
5787         struct btrfs_space_info *space_info;
5788         int loop = 0;
5789         int index = __get_raid_index(data);
5790         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5791                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5792         bool found_uncached_bg = false;
5793         bool failed_cluster_refill = false;
5794         bool failed_alloc = false;
5795         bool use_cluster = true;
5796         bool have_caching_bg = false;
5797
5798         WARN_ON(num_bytes < root->sectorsize);
5799         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5800         ins->objectid = 0;
5801         ins->offset = 0;
5802
5803         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5804
5805         space_info = __find_space_info(root->fs_info, data);
5806         if (!space_info) {
5807                 printk(KERN_ERR "No space info for %llu\n", data);
5808                 return -ENOSPC;
5809         }
5810
5811         /*
5812          * If the space info is for both data and metadata it means we have a
5813          * small filesystem and we can't use the clustering stuff.
5814          */
5815         if (btrfs_mixed_space_info(space_info))
5816                 use_cluster = false;
5817
5818         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5819                 last_ptr = &root->fs_info->meta_alloc_cluster;
5820                 if (!btrfs_test_opt(root, SSD))
5821                         empty_cluster = 64 * 1024;
5822         }
5823
5824         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5825             btrfs_test_opt(root, SSD)) {
5826                 last_ptr = &root->fs_info->data_alloc_cluster;
5827         }
5828
5829         if (last_ptr) {
5830                 spin_lock(&last_ptr->lock);
5831                 if (last_ptr->block_group)
5832                         hint_byte = last_ptr->window_start;
5833                 spin_unlock(&last_ptr->lock);
5834         }
5835
5836         search_start = max(search_start, first_logical_byte(root, 0));
5837         search_start = max(search_start, hint_byte);
5838
5839         if (!last_ptr)
5840                 empty_cluster = 0;
5841
5842         if (search_start == hint_byte) {
5843                 block_group = btrfs_lookup_block_group(root->fs_info,
5844                                                        search_start);
5845                 used_block_group = block_group;
5846                 /*
5847                  * we don't want to use the block group if it doesn't match our
5848                  * allocation bits, or if its not cached.
5849                  *
5850                  * However if we are re-searching with an ideal block group
5851                  * picked out then we don't care that the block group is cached.
5852                  */
5853                 if (block_group && block_group_bits(block_group, data) &&
5854                     block_group->cached != BTRFS_CACHE_NO) {
5855                         down_read(&space_info->groups_sem);
5856                         if (list_empty(&block_group->list) ||
5857                             block_group->ro) {
5858                                 /*
5859                                  * someone is removing this block group,
5860                                  * we can't jump into the have_block_group
5861                                  * target because our list pointers are not
5862                                  * valid
5863                                  */
5864                                 btrfs_put_block_group(block_group);
5865                                 up_read(&space_info->groups_sem);
5866                         } else {
5867                                 index = get_block_group_index(block_group);
5868                                 goto have_block_group;
5869                         }
5870                 } else if (block_group) {
5871                         btrfs_put_block_group(block_group);
5872                 }
5873         }
5874 search:
5875         have_caching_bg = false;
5876         down_read(&space_info->groups_sem);
5877         list_for_each_entry(block_group, &space_info->block_groups[index],
5878                             list) {
5879                 u64 offset;
5880                 int cached;
5881
5882                 used_block_group = block_group;
5883                 btrfs_get_block_group(block_group);
5884                 search_start = block_group->key.objectid;
5885
5886                 /*
5887                  * this can happen if we end up cycling through all the
5888                  * raid types, but we want to make sure we only allocate
5889                  * for the proper type.
5890                  */
5891                 if (!block_group_bits(block_group, data)) {
5892                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5893                                 BTRFS_BLOCK_GROUP_RAID1 |
5894                                 BTRFS_BLOCK_GROUP_RAID5 |
5895                                 BTRFS_BLOCK_GROUP_RAID6 |
5896                                 BTRFS_BLOCK_GROUP_RAID10;
5897
5898                         /*
5899                          * if they asked for extra copies and this block group
5900                          * doesn't provide them, bail.  This does allow us to
5901                          * fill raid0 from raid1.
5902                          */
5903                         if ((data & extra) && !(block_group->flags & extra))
5904                                 goto loop;
5905                 }
5906
5907 have_block_group:
5908                 cached = block_group_cache_done(block_group);
5909                 if (unlikely(!cached)) {
5910                         found_uncached_bg = true;
5911                         ret = cache_block_group(block_group, 0);
5912                         BUG_ON(ret < 0);
5913                         ret = 0;
5914                 }
5915
5916                 if (unlikely(block_group->ro))
5917                         goto loop;
5918
5919                 /*
5920                  * Ok we want to try and use the cluster allocator, so
5921                  * lets look there
5922                  */
5923                 if (last_ptr) {
5924                         unsigned long aligned_cluster;
5925                         /*
5926                          * the refill lock keeps out other
5927                          * people trying to start a new cluster
5928                          */
5929                         spin_lock(&last_ptr->refill_lock);
5930                         used_block_group = last_ptr->block_group;
5931                         if (used_block_group != block_group &&
5932                             (!used_block_group ||
5933                              used_block_group->ro ||
5934                              !block_group_bits(used_block_group, data))) {
5935                                 used_block_group = block_group;
5936                                 goto refill_cluster;
5937                         }
5938
5939                         if (used_block_group != block_group)
5940                                 btrfs_get_block_group(used_block_group);
5941
5942                         offset = btrfs_alloc_from_cluster(used_block_group,
5943                           last_ptr, num_bytes, used_block_group->key.objectid);
5944                         if (offset) {
5945                                 /* we have a block, we're done */
5946                                 spin_unlock(&last_ptr->refill_lock);
5947                                 trace_btrfs_reserve_extent_cluster(root,
5948                                         block_group, search_start, num_bytes);
5949                                 goto checks;
5950                         }
5951
5952                         WARN_ON(last_ptr->block_group != used_block_group);
5953                         if (used_block_group != block_group) {
5954                                 btrfs_put_block_group(used_block_group);
5955                                 used_block_group = block_group;
5956                         }
5957 refill_cluster:
5958                         BUG_ON(used_block_group != block_group);
5959                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5960                          * set up a new clusters, so lets just skip it
5961                          * and let the allocator find whatever block
5962                          * it can find.  If we reach this point, we
5963                          * will have tried the cluster allocator
5964                          * plenty of times and not have found
5965                          * anything, so we are likely way too
5966                          * fragmented for the clustering stuff to find
5967                          * anything.
5968                          *
5969                          * However, if the cluster is taken from the
5970                          * current block group, release the cluster
5971                          * first, so that we stand a better chance of
5972                          * succeeding in the unclustered
5973                          * allocation.  */
5974                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5975                             last_ptr->block_group != block_group) {
5976                                 spin_unlock(&last_ptr->refill_lock);
5977                                 goto unclustered_alloc;
5978                         }
5979
5980                         /*
5981                          * this cluster didn't work out, free it and
5982                          * start over
5983                          */
5984                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5985
5986                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5987                                 spin_unlock(&last_ptr->refill_lock);
5988                                 goto unclustered_alloc;
5989                         }
5990
5991                         aligned_cluster = max_t(unsigned long,
5992                                                 empty_cluster + empty_size,
5993                                               block_group->full_stripe_len);
5994
5995                         /* allocate a cluster in this block group */
5996                         ret = btrfs_find_space_cluster(trans, root,
5997                                                block_group, last_ptr,
5998                                                search_start, num_bytes,
5999                                                aligned_cluster);
6000                         if (ret == 0) {
6001                                 /*
6002                                  * now pull our allocation out of this
6003                                  * cluster
6004                                  */
6005                                 offset = btrfs_alloc_from_cluster(block_group,
6006                                                   last_ptr, num_bytes,
6007                                                   search_start);
6008                                 if (offset) {
6009                                         /* we found one, proceed */
6010                                         spin_unlock(&last_ptr->refill_lock);
6011                                         trace_btrfs_reserve_extent_cluster(root,
6012                                                 block_group, search_start,
6013                                                 num_bytes);
6014                                         goto checks;
6015                                 }
6016                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6017                                    && !failed_cluster_refill) {
6018                                 spin_unlock(&last_ptr->refill_lock);
6019
6020                                 failed_cluster_refill = true;
6021                                 wait_block_group_cache_progress(block_group,
6022                                        num_bytes + empty_cluster + empty_size);
6023                                 goto have_block_group;
6024                         }
6025
6026                         /*
6027                          * at this point we either didn't find a cluster
6028                          * or we weren't able to allocate a block from our
6029                          * cluster.  Free the cluster we've been trying
6030                          * to use, and go to the next block group
6031                          */
6032                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6033                         spin_unlock(&last_ptr->refill_lock);
6034                         goto loop;
6035                 }
6036
6037 unclustered_alloc:
6038                 spin_lock(&block_group->free_space_ctl->tree_lock);
6039                 if (cached &&
6040                     block_group->free_space_ctl->free_space <
6041                     num_bytes + empty_cluster + empty_size) {
6042                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6043                         goto loop;
6044                 }
6045                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6046
6047                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6048                                                     num_bytes, empty_size);
6049                 /*
6050                  * If we didn't find a chunk, and we haven't failed on this
6051                  * block group before, and this block group is in the middle of
6052                  * caching and we are ok with waiting, then go ahead and wait
6053                  * for progress to be made, and set failed_alloc to true.
6054                  *
6055                  * If failed_alloc is true then we've already waited on this
6056                  * block group once and should move on to the next block group.
6057                  */
6058                 if (!offset && !failed_alloc && !cached &&
6059                     loop > LOOP_CACHING_NOWAIT) {
6060                         wait_block_group_cache_progress(block_group,
6061                                                 num_bytes + empty_size);
6062                         failed_alloc = true;
6063                         goto have_block_group;
6064                 } else if (!offset) {
6065                         if (!cached)
6066                                 have_caching_bg = true;
6067                         goto loop;
6068                 }
6069 checks:
6070                 search_start = stripe_align(root, used_block_group,
6071                                             offset, num_bytes);
6072
6073                 /* move on to the next group */
6074                 if (search_start + num_bytes >
6075                     used_block_group->key.objectid + used_block_group->key.offset) {
6076                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6077                         goto loop;
6078                 }
6079
6080                 if (offset < search_start)
6081                         btrfs_add_free_space(used_block_group, offset,
6082                                              search_start - offset);
6083                 BUG_ON(offset > search_start);
6084
6085                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
6086                                                   alloc_type);
6087                 if (ret == -EAGAIN) {
6088                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6089                         goto loop;
6090                 }
6091
6092                 /* we are all good, lets return */
6093                 ins->objectid = search_start;
6094                 ins->offset = num_bytes;
6095
6096                 trace_btrfs_reserve_extent(orig_root, block_group,
6097                                            search_start, num_bytes);
6098                 if (used_block_group != block_group)
6099                         btrfs_put_block_group(used_block_group);
6100                 btrfs_put_block_group(block_group);
6101                 break;
6102 loop:
6103                 failed_cluster_refill = false;
6104                 failed_alloc = false;
6105                 BUG_ON(index != get_block_group_index(block_group));
6106                 if (used_block_group != block_group)
6107                         btrfs_put_block_group(used_block_group);
6108                 btrfs_put_block_group(block_group);
6109         }
6110         up_read(&space_info->groups_sem);
6111
6112         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6113                 goto search;
6114
6115         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6116                 goto search;
6117
6118         /*
6119          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6120          *                      caching kthreads as we move along
6121          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6122          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6123          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6124          *                      again
6125          */
6126         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6127                 index = 0;
6128                 loop++;
6129                 if (loop == LOOP_ALLOC_CHUNK) {
6130                         ret = do_chunk_alloc(trans, root, data,
6131                                              CHUNK_ALLOC_FORCE);
6132                         /*
6133                          * Do not bail out on ENOSPC since we
6134                          * can do more things.
6135                          */
6136                         if (ret < 0 && ret != -ENOSPC) {
6137                                 btrfs_abort_transaction(trans,
6138                                                         root, ret);
6139                                 goto out;
6140                         }
6141                 }
6142
6143                 if (loop == LOOP_NO_EMPTY_SIZE) {
6144                         empty_size = 0;
6145                         empty_cluster = 0;
6146                 }
6147
6148                 goto search;
6149         } else if (!ins->objectid) {
6150                 ret = -ENOSPC;
6151         } else if (ins->objectid) {
6152                 ret = 0;
6153         }
6154 out:
6155
6156         return ret;
6157 }
6158
6159 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6160                             int dump_block_groups)
6161 {
6162         struct btrfs_block_group_cache *cache;
6163         int index = 0;
6164
6165         spin_lock(&info->lock);
6166         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
6167                (unsigned long long)info->flags,
6168                (unsigned long long)(info->total_bytes - info->bytes_used -
6169                                     info->bytes_pinned - info->bytes_reserved -
6170                                     info->bytes_readonly),
6171                (info->full) ? "" : "not ");
6172         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
6173                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6174                (unsigned long long)info->total_bytes,
6175                (unsigned long long)info->bytes_used,
6176                (unsigned long long)info->bytes_pinned,
6177                (unsigned long long)info->bytes_reserved,
6178                (unsigned long long)info->bytes_may_use,
6179                (unsigned long long)info->bytes_readonly);
6180         spin_unlock(&info->lock);
6181
6182         if (!dump_block_groups)
6183                 return;
6184
6185         down_read(&info->groups_sem);
6186 again:
6187         list_for_each_entry(cache, &info->block_groups[index], list) {
6188                 spin_lock(&cache->lock);
6189                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
6190                        (unsigned long long)cache->key.objectid,
6191                        (unsigned long long)cache->key.offset,
6192                        (unsigned long long)btrfs_block_group_used(&cache->item),
6193                        (unsigned long long)cache->pinned,
6194                        (unsigned long long)cache->reserved,
6195                        cache->ro ? "[readonly]" : "");
6196                 btrfs_dump_free_space(cache, bytes);
6197                 spin_unlock(&cache->lock);
6198         }
6199         if (++index < BTRFS_NR_RAID_TYPES)
6200                 goto again;
6201         up_read(&info->groups_sem);
6202 }
6203
6204 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
6205                          struct btrfs_root *root,
6206                          u64 num_bytes, u64 min_alloc_size,
6207                          u64 empty_size, u64 hint_byte,
6208                          struct btrfs_key *ins, u64 data)
6209 {
6210         bool final_tried = false;
6211         int ret;
6212
6213         data = btrfs_get_alloc_profile(root, data);
6214 again:
6215         WARN_ON(num_bytes < root->sectorsize);
6216         ret = find_free_extent(trans, root, num_bytes, empty_size,
6217                                hint_byte, ins, data);
6218
6219         if (ret == -ENOSPC) {
6220                 if (!final_tried) {
6221                         num_bytes = num_bytes >> 1;
6222                         num_bytes = round_down(num_bytes, root->sectorsize);
6223                         num_bytes = max(num_bytes, min_alloc_size);
6224                         if (num_bytes == min_alloc_size)
6225                                 final_tried = true;
6226                         goto again;
6227                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6228                         struct btrfs_space_info *sinfo;
6229
6230                         sinfo = __find_space_info(root->fs_info, data);
6231                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
6232                                "wanted %llu\n", (unsigned long long)data,
6233                                (unsigned long long)num_bytes);
6234                         if (sinfo)
6235                                 dump_space_info(sinfo, num_bytes, 1);
6236                 }
6237         }
6238
6239         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6240
6241         return ret;
6242 }
6243
6244 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6245                                         u64 start, u64 len, int pin)
6246 {
6247         struct btrfs_block_group_cache *cache;
6248         int ret = 0;
6249
6250         cache = btrfs_lookup_block_group(root->fs_info, start);
6251         if (!cache) {
6252                 printk(KERN_ERR "Unable to find block group for %llu\n",
6253                        (unsigned long long)start);
6254                 return -ENOSPC;
6255         }
6256
6257         if (btrfs_test_opt(root, DISCARD))
6258                 ret = btrfs_discard_extent(root, start, len, NULL);
6259
6260         if (pin)
6261                 pin_down_extent(root, cache, start, len, 1);
6262         else {
6263                 btrfs_add_free_space(cache, start, len);
6264                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6265         }
6266         btrfs_put_block_group(cache);
6267
6268         trace_btrfs_reserved_extent_free(root, start, len);
6269
6270         return ret;
6271 }
6272
6273 int btrfs_free_reserved_extent(struct btrfs_root *root,
6274                                         u64 start, u64 len)
6275 {
6276         return __btrfs_free_reserved_extent(root, start, len, 0);
6277 }
6278
6279 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6280                                        u64 start, u64 len)
6281 {
6282         return __btrfs_free_reserved_extent(root, start, len, 1);
6283 }
6284
6285 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6286                                       struct btrfs_root *root,
6287                                       u64 parent, u64 root_objectid,
6288                                       u64 flags, u64 owner, u64 offset,
6289                                       struct btrfs_key *ins, int ref_mod)
6290 {
6291         int ret;
6292         struct btrfs_fs_info *fs_info = root->fs_info;
6293         struct btrfs_extent_item *extent_item;
6294         struct btrfs_extent_inline_ref *iref;
6295         struct btrfs_path *path;
6296         struct extent_buffer *leaf;
6297         int type;
6298         u32 size;
6299
6300         if (parent > 0)
6301                 type = BTRFS_SHARED_DATA_REF_KEY;
6302         else
6303                 type = BTRFS_EXTENT_DATA_REF_KEY;
6304
6305         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6306
6307         path = btrfs_alloc_path();
6308         if (!path)
6309                 return -ENOMEM;
6310
6311         path->leave_spinning = 1;
6312         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6313                                       ins, size);
6314         if (ret) {
6315                 btrfs_free_path(path);
6316                 return ret;
6317         }
6318
6319         leaf = path->nodes[0];
6320         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6321                                      struct btrfs_extent_item);
6322         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6323         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6324         btrfs_set_extent_flags(leaf, extent_item,
6325                                flags | BTRFS_EXTENT_FLAG_DATA);
6326
6327         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6328         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6329         if (parent > 0) {
6330                 struct btrfs_shared_data_ref *ref;
6331                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6332                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6333                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6334         } else {
6335                 struct btrfs_extent_data_ref *ref;
6336                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6337                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6338                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6339                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6340                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6341         }
6342
6343         btrfs_mark_buffer_dirty(path->nodes[0]);
6344         btrfs_free_path(path);
6345
6346         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6347         if (ret) { /* -ENOENT, logic error */
6348                 printk(KERN_ERR "btrfs update block group failed for %llu "
6349                        "%llu\n", (unsigned long long)ins->objectid,
6350                        (unsigned long long)ins->offset);
6351                 BUG();
6352         }
6353         return ret;
6354 }
6355
6356 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6357                                      struct btrfs_root *root,
6358                                      u64 parent, u64 root_objectid,
6359                                      u64 flags, struct btrfs_disk_key *key,
6360                                      int level, struct btrfs_key *ins)
6361 {
6362         int ret;
6363         struct btrfs_fs_info *fs_info = root->fs_info;
6364         struct btrfs_extent_item *extent_item;
6365         struct btrfs_tree_block_info *block_info;
6366         struct btrfs_extent_inline_ref *iref;
6367         struct btrfs_path *path;
6368         struct extent_buffer *leaf;
6369         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6370
6371         path = btrfs_alloc_path();
6372         if (!path)
6373                 return -ENOMEM;
6374
6375         path->leave_spinning = 1;
6376         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6377                                       ins, size);
6378         if (ret) {
6379                 btrfs_free_path(path);
6380                 return ret;
6381         }
6382
6383         leaf = path->nodes[0];
6384         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6385                                      struct btrfs_extent_item);
6386         btrfs_set_extent_refs(leaf, extent_item, 1);
6387         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6388         btrfs_set_extent_flags(leaf, extent_item,
6389                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6390         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6391
6392         btrfs_set_tree_block_key(leaf, block_info, key);
6393         btrfs_set_tree_block_level(leaf, block_info, level);
6394
6395         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6396         if (parent > 0) {
6397                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6398                 btrfs_set_extent_inline_ref_type(leaf, iref,
6399                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6400                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6401         } else {
6402                 btrfs_set_extent_inline_ref_type(leaf, iref,
6403                                                  BTRFS_TREE_BLOCK_REF_KEY);
6404                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6405         }
6406
6407         btrfs_mark_buffer_dirty(leaf);
6408         btrfs_free_path(path);
6409
6410         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6411         if (ret) { /* -ENOENT, logic error */
6412                 printk(KERN_ERR "btrfs update block group failed for %llu "
6413                        "%llu\n", (unsigned long long)ins->objectid,
6414                        (unsigned long long)ins->offset);
6415                 BUG();
6416         }
6417         return ret;
6418 }
6419
6420 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6421                                      struct btrfs_root *root,
6422                                      u64 root_objectid, u64 owner,
6423                                      u64 offset, struct btrfs_key *ins)
6424 {
6425         int ret;
6426
6427         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6428
6429         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6430                                          ins->offset, 0,
6431                                          root_objectid, owner, offset,
6432                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6433         return ret;
6434 }
6435
6436 /*
6437  * this is used by the tree logging recovery code.  It records that
6438  * an extent has been allocated and makes sure to clear the free
6439  * space cache bits as well
6440  */
6441 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6442                                    struct btrfs_root *root,
6443                                    u64 root_objectid, u64 owner, u64 offset,
6444                                    struct btrfs_key *ins)
6445 {
6446         int ret;
6447         struct btrfs_block_group_cache *block_group;
6448         struct btrfs_caching_control *caching_ctl;
6449         u64 start = ins->objectid;
6450         u64 num_bytes = ins->offset;
6451
6452         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6453         cache_block_group(block_group, 0);
6454         caching_ctl = get_caching_control(block_group);
6455
6456         if (!caching_ctl) {
6457                 BUG_ON(!block_group_cache_done(block_group));
6458                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6459                 BUG_ON(ret); /* -ENOMEM */
6460         } else {
6461                 mutex_lock(&caching_ctl->mutex);
6462
6463                 if (start >= caching_ctl->progress) {
6464                         ret = add_excluded_extent(root, start, num_bytes);
6465                         BUG_ON(ret); /* -ENOMEM */
6466                 } else if (start + num_bytes <= caching_ctl->progress) {
6467                         ret = btrfs_remove_free_space(block_group,
6468                                                       start, num_bytes);
6469                         BUG_ON(ret); /* -ENOMEM */
6470                 } else {
6471                         num_bytes = caching_ctl->progress - start;
6472                         ret = btrfs_remove_free_space(block_group,
6473                                                       start, num_bytes);
6474                         BUG_ON(ret); /* -ENOMEM */
6475
6476                         start = caching_ctl->progress;
6477                         num_bytes = ins->objectid + ins->offset -
6478                                     caching_ctl->progress;
6479                         ret = add_excluded_extent(root, start, num_bytes);
6480                         BUG_ON(ret); /* -ENOMEM */
6481                 }
6482
6483                 mutex_unlock(&caching_ctl->mutex);
6484                 put_caching_control(caching_ctl);
6485         }
6486
6487         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6488                                           RESERVE_ALLOC_NO_ACCOUNT);
6489         BUG_ON(ret); /* logic error */
6490         btrfs_put_block_group(block_group);
6491         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6492                                          0, owner, offset, ins, 1);
6493         return ret;
6494 }
6495
6496 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6497                                             struct btrfs_root *root,
6498                                             u64 bytenr, u32 blocksize,
6499                                             int level)
6500 {
6501         struct extent_buffer *buf;
6502
6503         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6504         if (!buf)
6505                 return ERR_PTR(-ENOMEM);
6506         btrfs_set_header_generation(buf, trans->transid);
6507         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6508         btrfs_tree_lock(buf);
6509         clean_tree_block(trans, root, buf);
6510         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6511
6512         btrfs_set_lock_blocking(buf);
6513         btrfs_set_buffer_uptodate(buf);
6514
6515         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6516                 /*
6517                  * we allow two log transactions at a time, use different
6518                  * EXENT bit to differentiate dirty pages.
6519                  */
6520                 if (root->log_transid % 2 == 0)
6521                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6522                                         buf->start + buf->len - 1, GFP_NOFS);
6523                 else
6524                         set_extent_new(&root->dirty_log_pages, buf->start,
6525                                         buf->start + buf->len - 1, GFP_NOFS);
6526         } else {
6527                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6528                          buf->start + buf->len - 1, GFP_NOFS);
6529         }
6530         trans->blocks_used++;
6531         /* this returns a buffer locked for blocking */
6532         return buf;
6533 }
6534
6535 static struct btrfs_block_rsv *
6536 use_block_rsv(struct btrfs_trans_handle *trans,
6537               struct btrfs_root *root, u32 blocksize)
6538 {
6539         struct btrfs_block_rsv *block_rsv;
6540         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6541         int ret;
6542
6543         block_rsv = get_block_rsv(trans, root);
6544
6545         if (block_rsv->size == 0) {
6546                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6547                                              BTRFS_RESERVE_NO_FLUSH);
6548                 /*
6549                  * If we couldn't reserve metadata bytes try and use some from
6550                  * the global reserve.
6551                  */
6552                 if (ret && block_rsv != global_rsv) {
6553                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6554                         if (!ret)
6555                                 return global_rsv;
6556                         return ERR_PTR(ret);
6557                 } else if (ret) {
6558                         return ERR_PTR(ret);
6559                 }
6560                 return block_rsv;
6561         }
6562
6563         ret = block_rsv_use_bytes(block_rsv, blocksize);
6564         if (!ret)
6565                 return block_rsv;
6566         if (ret && !block_rsv->failfast) {
6567                 if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6568                         static DEFINE_RATELIMIT_STATE(_rs,
6569                                         DEFAULT_RATELIMIT_INTERVAL * 10,
6570                                         /*DEFAULT_RATELIMIT_BURST*/ 1);
6571                         if (__ratelimit(&_rs))
6572                                 WARN(1, KERN_DEBUG
6573                                         "btrfs: block rsv returned %d\n", ret);
6574                 }
6575                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6576                                              BTRFS_RESERVE_NO_FLUSH);
6577                 if (!ret) {
6578                         return block_rsv;
6579                 } else if (ret && block_rsv != global_rsv) {
6580                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6581                         if (!ret)
6582                                 return global_rsv;
6583                 }
6584         }
6585
6586         return ERR_PTR(-ENOSPC);
6587 }
6588
6589 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6590                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6591 {
6592         block_rsv_add_bytes(block_rsv, blocksize, 0);
6593         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6594 }
6595
6596 /*
6597  * finds a free extent and does all the dirty work required for allocation
6598  * returns the key for the extent through ins, and a tree buffer for
6599  * the first block of the extent through buf.
6600  *
6601  * returns the tree buffer or NULL.
6602  */
6603 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6604                                         struct btrfs_root *root, u32 blocksize,
6605                                         u64 parent, u64 root_objectid,
6606                                         struct btrfs_disk_key *key, int level,
6607                                         u64 hint, u64 empty_size)
6608 {
6609         struct btrfs_key ins;
6610         struct btrfs_block_rsv *block_rsv;
6611         struct extent_buffer *buf;
6612         u64 flags = 0;
6613         int ret;
6614
6615
6616         block_rsv = use_block_rsv(trans, root, blocksize);
6617         if (IS_ERR(block_rsv))
6618                 return ERR_CAST(block_rsv);
6619
6620         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6621                                    empty_size, hint, &ins, 0);
6622         if (ret) {
6623                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6624                 return ERR_PTR(ret);
6625         }
6626
6627         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6628                                     blocksize, level);
6629         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6630
6631         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6632                 if (parent == 0)
6633                         parent = ins.objectid;
6634                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6635         } else
6636                 BUG_ON(parent > 0);
6637
6638         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6639                 struct btrfs_delayed_extent_op *extent_op;
6640                 extent_op = btrfs_alloc_delayed_extent_op();
6641                 BUG_ON(!extent_op); /* -ENOMEM */
6642                 if (key)
6643                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6644                 else
6645                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6646                 extent_op->flags_to_set = flags;
6647                 extent_op->update_key = 1;
6648                 extent_op->update_flags = 1;
6649                 extent_op->is_data = 0;
6650
6651                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6652                                         ins.objectid,
6653                                         ins.offset, parent, root_objectid,
6654                                         level, BTRFS_ADD_DELAYED_EXTENT,
6655                                         extent_op, 0);
6656                 BUG_ON(ret); /* -ENOMEM */
6657         }
6658         return buf;
6659 }
6660
6661 struct walk_control {
6662         u64 refs[BTRFS_MAX_LEVEL];
6663         u64 flags[BTRFS_MAX_LEVEL];
6664         struct btrfs_key update_progress;
6665         int stage;
6666         int level;
6667         int shared_level;
6668         int update_ref;
6669         int keep_locks;
6670         int reada_slot;
6671         int reada_count;
6672         int for_reloc;
6673 };
6674
6675 #define DROP_REFERENCE  1
6676 #define UPDATE_BACKREF  2
6677
6678 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6679                                      struct btrfs_root *root,
6680                                      struct walk_control *wc,
6681                                      struct btrfs_path *path)
6682 {
6683         u64 bytenr;
6684         u64 generation;
6685         u64 refs;
6686         u64 flags;
6687         u32 nritems;
6688         u32 blocksize;
6689         struct btrfs_key key;
6690         struct extent_buffer *eb;
6691         int ret;
6692         int slot;
6693         int nread = 0;
6694
6695         if (path->slots[wc->level] < wc->reada_slot) {
6696                 wc->reada_count = wc->reada_count * 2 / 3;
6697                 wc->reada_count = max(wc->reada_count, 2);
6698         } else {
6699                 wc->reada_count = wc->reada_count * 3 / 2;
6700                 wc->reada_count = min_t(int, wc->reada_count,
6701                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6702         }
6703
6704         eb = path->nodes[wc->level];
6705         nritems = btrfs_header_nritems(eb);
6706         blocksize = btrfs_level_size(root, wc->level - 1);
6707
6708         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6709                 if (nread >= wc->reada_count)
6710                         break;
6711
6712                 cond_resched();
6713                 bytenr = btrfs_node_blockptr(eb, slot);
6714                 generation = btrfs_node_ptr_generation(eb, slot);
6715
6716                 if (slot == path->slots[wc->level])
6717                         goto reada;
6718
6719                 if (wc->stage == UPDATE_BACKREF &&
6720                     generation <= root->root_key.offset)
6721                         continue;
6722
6723                 /* We don't lock the tree block, it's OK to be racy here */
6724                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6725                                                &refs, &flags);
6726                 /* We don't care about errors in readahead. */
6727                 if (ret < 0)
6728                         continue;
6729                 BUG_ON(refs == 0);
6730
6731                 if (wc->stage == DROP_REFERENCE) {
6732                         if (refs == 1)
6733                                 goto reada;
6734
6735                         if (wc->level == 1 &&
6736                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6737                                 continue;
6738                         if (!wc->update_ref ||
6739                             generation <= root->root_key.offset)
6740                                 continue;
6741                         btrfs_node_key_to_cpu(eb, &key, slot);
6742                         ret = btrfs_comp_cpu_keys(&key,
6743                                                   &wc->update_progress);
6744                         if (ret < 0)
6745                                 continue;
6746                 } else {
6747                         if (wc->level == 1 &&
6748                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6749                                 continue;
6750                 }
6751 reada:
6752                 ret = readahead_tree_block(root, bytenr, blocksize,
6753                                            generation);
6754                 if (ret)
6755                         break;
6756                 nread++;
6757         }
6758         wc->reada_slot = slot;
6759 }
6760
6761 /*
6762  * hepler to process tree block while walking down the tree.
6763  *
6764  * when wc->stage == UPDATE_BACKREF, this function updates
6765  * back refs for pointers in the block.
6766  *
6767  * NOTE: return value 1 means we should stop walking down.
6768  */
6769 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6770                                    struct btrfs_root *root,
6771                                    struct btrfs_path *path,
6772                                    struct walk_control *wc, int lookup_info)
6773 {
6774         int level = wc->level;
6775         struct extent_buffer *eb = path->nodes[level];
6776         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6777         int ret;
6778
6779         if (wc->stage == UPDATE_BACKREF &&
6780             btrfs_header_owner(eb) != root->root_key.objectid)
6781                 return 1;
6782
6783         /*
6784          * when reference count of tree block is 1, it won't increase
6785          * again. once full backref flag is set, we never clear it.
6786          */
6787         if (lookup_info &&
6788             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6789              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6790                 BUG_ON(!path->locks[level]);
6791                 ret = btrfs_lookup_extent_info(trans, root,
6792                                                eb->start, eb->len,
6793                                                &wc->refs[level],
6794                                                &wc->flags[level]);
6795                 BUG_ON(ret == -ENOMEM);
6796                 if (ret)
6797                         return ret;
6798                 BUG_ON(wc->refs[level] == 0);
6799         }
6800
6801         if (wc->stage == DROP_REFERENCE) {
6802                 if (wc->refs[level] > 1)
6803                         return 1;
6804
6805                 if (path->locks[level] && !wc->keep_locks) {
6806                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6807                         path->locks[level] = 0;
6808                 }
6809                 return 0;
6810         }
6811
6812         /* wc->stage == UPDATE_BACKREF */
6813         if (!(wc->flags[level] & flag)) {
6814                 BUG_ON(!path->locks[level]);
6815                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6816                 BUG_ON(ret); /* -ENOMEM */
6817                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6818                 BUG_ON(ret); /* -ENOMEM */
6819                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6820                                                   eb->len, flag, 0);
6821                 BUG_ON(ret); /* -ENOMEM */
6822                 wc->flags[level] |= flag;
6823         }
6824
6825         /*
6826          * the block is shared by multiple trees, so it's not good to
6827          * keep the tree lock
6828          */
6829         if (path->locks[level] && level > 0) {
6830                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6831                 path->locks[level] = 0;
6832         }
6833         return 0;
6834 }
6835
6836 /*
6837  * hepler to process tree block pointer.
6838  *
6839  * when wc->stage == DROP_REFERENCE, this function checks
6840  * reference count of the block pointed to. if the block
6841  * is shared and we need update back refs for the subtree
6842  * rooted at the block, this function changes wc->stage to
6843  * UPDATE_BACKREF. if the block is shared and there is no
6844  * need to update back, this function drops the reference
6845  * to the block.
6846  *
6847  * NOTE: return value 1 means we should stop walking down.
6848  */
6849 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6850                                  struct btrfs_root *root,
6851                                  struct btrfs_path *path,
6852                                  struct walk_control *wc, int *lookup_info)
6853 {
6854         u64 bytenr;
6855         u64 generation;
6856         u64 parent;
6857         u32 blocksize;
6858         struct btrfs_key key;
6859         struct extent_buffer *next;
6860         int level = wc->level;
6861         int reada = 0;
6862         int ret = 0;
6863
6864         generation = btrfs_node_ptr_generation(path->nodes[level],
6865                                                path->slots[level]);
6866         /*
6867          * if the lower level block was created before the snapshot
6868          * was created, we know there is no need to update back refs
6869          * for the subtree
6870          */
6871         if (wc->stage == UPDATE_BACKREF &&
6872             generation <= root->root_key.offset) {
6873                 *lookup_info = 1;
6874                 return 1;
6875         }
6876
6877         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6878         blocksize = btrfs_level_size(root, level - 1);
6879
6880         next = btrfs_find_tree_block(root, bytenr, blocksize);
6881         if (!next) {
6882                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6883                 if (!next)
6884                         return -ENOMEM;
6885                 reada = 1;
6886         }
6887         btrfs_tree_lock(next);
6888         btrfs_set_lock_blocking(next);
6889
6890         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6891                                        &wc->refs[level - 1],
6892                                        &wc->flags[level - 1]);
6893         if (ret < 0) {
6894                 btrfs_tree_unlock(next);
6895                 return ret;
6896         }
6897
6898         BUG_ON(wc->refs[level - 1] == 0);
6899         *lookup_info = 0;
6900
6901         if (wc->stage == DROP_REFERENCE) {
6902                 if (wc->refs[level - 1] > 1) {
6903                         if (level == 1 &&
6904                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6905                                 goto skip;
6906
6907                         if (!wc->update_ref ||
6908                             generation <= root->root_key.offset)
6909                                 goto skip;
6910
6911                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6912                                               path->slots[level]);
6913                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6914                         if (ret < 0)
6915                                 goto skip;
6916
6917                         wc->stage = UPDATE_BACKREF;
6918                         wc->shared_level = level - 1;
6919                 }
6920         } else {
6921                 if (level == 1 &&
6922                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6923                         goto skip;
6924         }
6925
6926         if (!btrfs_buffer_uptodate(next, generation, 0)) {
6927                 btrfs_tree_unlock(next);
6928                 free_extent_buffer(next);
6929                 next = NULL;
6930                 *lookup_info = 1;
6931         }
6932
6933         if (!next) {
6934                 if (reada && level == 1)
6935                         reada_walk_down(trans, root, wc, path);
6936                 next = read_tree_block(root, bytenr, blocksize, generation);
6937                 if (!next)
6938                         return -EIO;
6939                 btrfs_tree_lock(next);
6940                 btrfs_set_lock_blocking(next);
6941         }
6942
6943         level--;
6944         BUG_ON(level != btrfs_header_level(next));
6945         path->nodes[level] = next;
6946         path->slots[level] = 0;
6947         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6948         wc->level = level;
6949         if (wc->level == 1)
6950                 wc->reada_slot = 0;
6951         return 0;
6952 skip:
6953         wc->refs[level - 1] = 0;
6954         wc->flags[level - 1] = 0;
6955         if (wc->stage == DROP_REFERENCE) {
6956                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6957                         parent = path->nodes[level]->start;
6958                 } else {
6959                         BUG_ON(root->root_key.objectid !=
6960                                btrfs_header_owner(path->nodes[level]));
6961                         parent = 0;
6962                 }
6963
6964                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6965                                 root->root_key.objectid, level - 1, 0, 0);
6966                 BUG_ON(ret); /* -ENOMEM */
6967         }
6968         btrfs_tree_unlock(next);
6969         free_extent_buffer(next);
6970         *lookup_info = 1;
6971         return 1;
6972 }
6973
6974 /*
6975  * hepler to process tree block while walking up the tree.
6976  *
6977  * when wc->stage == DROP_REFERENCE, this function drops
6978  * reference count on the block.
6979  *
6980  * when wc->stage == UPDATE_BACKREF, this function changes
6981  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6982  * to UPDATE_BACKREF previously while processing the block.
6983  *
6984  * NOTE: return value 1 means we should stop walking up.
6985  */
6986 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6987                                  struct btrfs_root *root,
6988                                  struct btrfs_path *path,
6989                                  struct walk_control *wc)
6990 {
6991         int ret;
6992         int level = wc->level;
6993         struct extent_buffer *eb = path->nodes[level];
6994         u64 parent = 0;
6995
6996         if (wc->stage == UPDATE_BACKREF) {
6997                 BUG_ON(wc->shared_level < level);
6998                 if (level < wc->shared_level)
6999                         goto out;
7000
7001                 ret = find_next_key(path, level + 1, &wc->update_progress);
7002                 if (ret > 0)
7003                         wc->update_ref = 0;
7004
7005                 wc->stage = DROP_REFERENCE;
7006                 wc->shared_level = -1;
7007                 path->slots[level] = 0;
7008
7009                 /*
7010                  * check reference count again if the block isn't locked.
7011                  * we should start walking down the tree again if reference
7012                  * count is one.
7013                  */
7014                 if (!path->locks[level]) {
7015                         BUG_ON(level == 0);
7016                         btrfs_tree_lock(eb);
7017                         btrfs_set_lock_blocking(eb);
7018                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7019
7020                         ret = btrfs_lookup_extent_info(trans, root,
7021                                                        eb->start, eb->len,
7022                                                        &wc->refs[level],
7023                                                        &wc->flags[level]);
7024                         if (ret < 0) {
7025                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7026                                 path->locks[level] = 0;
7027                                 return ret;
7028                         }
7029                         BUG_ON(wc->refs[level] == 0);
7030                         if (wc->refs[level] == 1) {
7031                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7032                                 path->locks[level] = 0;
7033                                 return 1;
7034                         }
7035                 }
7036         }
7037
7038         /* wc->stage == DROP_REFERENCE */
7039         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7040
7041         if (wc->refs[level] == 1) {
7042                 if (level == 0) {
7043                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7044                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7045                                                     wc->for_reloc);
7046                         else
7047                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7048                                                     wc->for_reloc);
7049                         BUG_ON(ret); /* -ENOMEM */
7050                 }
7051                 /* make block locked assertion in clean_tree_block happy */
7052                 if (!path->locks[level] &&
7053                     btrfs_header_generation(eb) == trans->transid) {
7054                         btrfs_tree_lock(eb);
7055                         btrfs_set_lock_blocking(eb);
7056                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7057                 }
7058                 clean_tree_block(trans, root, eb);
7059         }
7060
7061         if (eb == root->node) {
7062                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7063                         parent = eb->start;
7064                 else
7065                         BUG_ON(root->root_key.objectid !=
7066                                btrfs_header_owner(eb));
7067         } else {
7068                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7069                         parent = path->nodes[level + 1]->start;
7070                 else
7071                         BUG_ON(root->root_key.objectid !=
7072                                btrfs_header_owner(path->nodes[level + 1]));
7073         }
7074
7075         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7076 out:
7077         wc->refs[level] = 0;
7078         wc->flags[level] = 0;
7079         return 0;
7080 }
7081
7082 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7083                                    struct btrfs_root *root,
7084                                    struct btrfs_path *path,
7085                                    struct walk_control *wc)
7086 {
7087         int level = wc->level;
7088         int lookup_info = 1;
7089         int ret;
7090
7091         while (level >= 0) {
7092                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7093                 if (ret > 0)
7094                         break;
7095
7096                 if (level == 0)
7097                         break;
7098
7099                 if (path->slots[level] >=
7100                     btrfs_header_nritems(path->nodes[level]))
7101                         break;
7102
7103                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7104                 if (ret > 0) {
7105                         path->slots[level]++;
7106                         continue;
7107                 } else if (ret < 0)
7108                         return ret;
7109                 level = wc->level;
7110         }
7111         return 0;
7112 }
7113
7114 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7115                                  struct btrfs_root *root,
7116                                  struct btrfs_path *path,
7117                                  struct walk_control *wc, int max_level)
7118 {
7119         int level = wc->level;
7120         int ret;
7121
7122         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7123         while (level < max_level && path->nodes[level]) {
7124                 wc->level = level;
7125                 if (path->slots[level] + 1 <
7126                     btrfs_header_nritems(path->nodes[level])) {
7127                         path->slots[level]++;
7128                         return 0;
7129                 } else {
7130                         ret = walk_up_proc(trans, root, path, wc);
7131                         if (ret > 0)
7132                                 return 0;
7133
7134                         if (path->locks[level]) {
7135                                 btrfs_tree_unlock_rw(path->nodes[level],
7136                                                      path->locks[level]);
7137                                 path->locks[level] = 0;
7138                         }
7139                         free_extent_buffer(path->nodes[level]);
7140                         path->nodes[level] = NULL;
7141                         level++;
7142                 }
7143         }
7144         return 1;
7145 }
7146
7147 /*
7148  * drop a subvolume tree.
7149  *
7150  * this function traverses the tree freeing any blocks that only
7151  * referenced by the tree.
7152  *
7153  * when a shared tree block is found. this function decreases its
7154  * reference count by one. if update_ref is true, this function
7155  * also make sure backrefs for the shared block and all lower level
7156  * blocks are properly updated.
7157  */
7158 int btrfs_drop_snapshot(struct btrfs_root *root,
7159                          struct btrfs_block_rsv *block_rsv, int update_ref,
7160                          int for_reloc)
7161 {
7162         struct btrfs_path *path;
7163         struct btrfs_trans_handle *trans;
7164         struct btrfs_root *tree_root = root->fs_info->tree_root;
7165         struct btrfs_root_item *root_item = &root->root_item;
7166         struct walk_control *wc;
7167         struct btrfs_key key;
7168         int err = 0;
7169         int ret;
7170         int level;
7171
7172         path = btrfs_alloc_path();
7173         if (!path) {
7174                 err = -ENOMEM;
7175                 goto out;
7176         }
7177
7178         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7179         if (!wc) {
7180                 btrfs_free_path(path);
7181                 err = -ENOMEM;
7182                 goto out;
7183         }
7184
7185         trans = btrfs_start_transaction(tree_root, 0);
7186         if (IS_ERR(trans)) {
7187                 err = PTR_ERR(trans);
7188                 goto out_free;
7189         }
7190
7191         if (block_rsv)
7192                 trans->block_rsv = block_rsv;
7193
7194         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7195                 level = btrfs_header_level(root->node);
7196                 path->nodes[level] = btrfs_lock_root_node(root);
7197                 btrfs_set_lock_blocking(path->nodes[level]);
7198                 path->slots[level] = 0;
7199                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7200                 memset(&wc->update_progress, 0,
7201                        sizeof(wc->update_progress));
7202         } else {
7203                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7204                 memcpy(&wc->update_progress, &key,
7205                        sizeof(wc->update_progress));
7206
7207                 level = root_item->drop_level;
7208                 BUG_ON(level == 0);
7209                 path->lowest_level = level;
7210                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7211                 path->lowest_level = 0;
7212                 if (ret < 0) {
7213                         err = ret;
7214                         goto out_end_trans;
7215                 }
7216                 WARN_ON(ret > 0);
7217
7218                 /*
7219                  * unlock our path, this is safe because only this
7220                  * function is allowed to delete this snapshot
7221                  */
7222                 btrfs_unlock_up_safe(path, 0);
7223
7224                 level = btrfs_header_level(root->node);
7225                 while (1) {
7226                         btrfs_tree_lock(path->nodes[level]);
7227                         btrfs_set_lock_blocking(path->nodes[level]);
7228
7229                         ret = btrfs_lookup_extent_info(trans, root,
7230                                                 path->nodes[level]->start,
7231                                                 path->nodes[level]->len,
7232                                                 &wc->refs[level],
7233                                                 &wc->flags[level]);
7234                         if (ret < 0) {
7235                                 err = ret;
7236                                 goto out_end_trans;
7237                         }
7238                         BUG_ON(wc->refs[level] == 0);
7239
7240                         if (level == root_item->drop_level)
7241                                 break;
7242
7243                         btrfs_tree_unlock(path->nodes[level]);
7244                         WARN_ON(wc->refs[level] != 1);
7245                         level--;
7246                 }
7247         }
7248
7249         wc->level = level;
7250         wc->shared_level = -1;
7251         wc->stage = DROP_REFERENCE;
7252         wc->update_ref = update_ref;
7253         wc->keep_locks = 0;
7254         wc->for_reloc = for_reloc;
7255         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7256
7257         while (1) {
7258                 ret = walk_down_tree(trans, root, path, wc);
7259                 if (ret < 0) {
7260                         err = ret;
7261                         break;
7262                 }
7263
7264                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7265                 if (ret < 0) {
7266                         err = ret;
7267                         break;
7268                 }
7269
7270                 if (ret > 0) {
7271                         BUG_ON(wc->stage != DROP_REFERENCE);
7272                         break;
7273                 }
7274
7275                 if (wc->stage == DROP_REFERENCE) {
7276                         level = wc->level;
7277                         btrfs_node_key(path->nodes[level],
7278                                        &root_item->drop_progress,
7279                                        path->slots[level]);
7280                         root_item->drop_level = level;
7281                 }
7282
7283                 BUG_ON(wc->level == 0);
7284                 if (btrfs_should_end_transaction(trans, tree_root)) {
7285                         ret = btrfs_update_root(trans, tree_root,
7286                                                 &root->root_key,
7287                                                 root_item);
7288                         if (ret) {
7289                                 btrfs_abort_transaction(trans, tree_root, ret);
7290                                 err = ret;
7291                                 goto out_end_trans;
7292                         }
7293
7294                         btrfs_end_transaction_throttle(trans, tree_root);
7295                         trans = btrfs_start_transaction(tree_root, 0);
7296                         if (IS_ERR(trans)) {
7297                                 err = PTR_ERR(trans);
7298                                 goto out_free;
7299                         }
7300                         if (block_rsv)
7301                                 trans->block_rsv = block_rsv;
7302                 }
7303         }
7304         btrfs_release_path(path);
7305         if (err)
7306                 goto out_end_trans;
7307
7308         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7309         if (ret) {
7310                 btrfs_abort_transaction(trans, tree_root, ret);
7311                 goto out_end_trans;
7312         }
7313
7314         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7315                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
7316                                            NULL, NULL);
7317                 if (ret < 0) {
7318                         btrfs_abort_transaction(trans, tree_root, ret);
7319                         err = ret;
7320                         goto out_end_trans;
7321                 } else if (ret > 0) {
7322                         /* if we fail to delete the orphan item this time
7323                          * around, it'll get picked up the next time.
7324                          *
7325                          * The most common failure here is just -ENOENT.
7326                          */
7327                         btrfs_del_orphan_item(trans, tree_root,
7328                                               root->root_key.objectid);
7329                 }
7330         }
7331
7332         if (root->in_radix) {
7333                 btrfs_free_fs_root(tree_root->fs_info, root);
7334         } else {
7335                 free_extent_buffer(root->node);
7336                 free_extent_buffer(root->commit_root);
7337                 kfree(root);
7338         }
7339 out_end_trans:
7340         btrfs_end_transaction_throttle(trans, tree_root);
7341 out_free:
7342         kfree(wc);
7343         btrfs_free_path(path);
7344 out:
7345         if (err)
7346                 btrfs_std_error(root->fs_info, err);
7347         return err;
7348 }
7349
7350 /*
7351  * drop subtree rooted at tree block 'node'.
7352  *
7353  * NOTE: this function will unlock and release tree block 'node'
7354  * only used by relocation code
7355  */
7356 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7357                         struct btrfs_root *root,
7358                         struct extent_buffer *node,
7359                         struct extent_buffer *parent)
7360 {
7361         struct btrfs_path *path;
7362         struct walk_control *wc;
7363         int level;
7364         int parent_level;
7365         int ret = 0;
7366         int wret;
7367
7368         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7369
7370         path = btrfs_alloc_path();
7371         if (!path)
7372                 return -ENOMEM;
7373
7374         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7375         if (!wc) {
7376                 btrfs_free_path(path);
7377                 return -ENOMEM;
7378         }
7379
7380         btrfs_assert_tree_locked(parent);
7381         parent_level = btrfs_header_level(parent);
7382         extent_buffer_get(parent);
7383         path->nodes[parent_level] = parent;
7384         path->slots[parent_level] = btrfs_header_nritems(parent);
7385
7386         btrfs_assert_tree_locked(node);
7387         level = btrfs_header_level(node);
7388         path->nodes[level] = node;
7389         path->slots[level] = 0;
7390         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7391
7392         wc->refs[parent_level] = 1;
7393         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7394         wc->level = level;
7395         wc->shared_level = -1;
7396         wc->stage = DROP_REFERENCE;
7397         wc->update_ref = 0;
7398         wc->keep_locks = 1;
7399         wc->for_reloc = 1;
7400         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7401
7402         while (1) {
7403                 wret = walk_down_tree(trans, root, path, wc);
7404                 if (wret < 0) {
7405                         ret = wret;
7406                         break;
7407                 }
7408
7409                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7410                 if (wret < 0)
7411                         ret = wret;
7412                 if (wret != 0)
7413                         break;
7414         }
7415
7416         kfree(wc);
7417         btrfs_free_path(path);
7418         return ret;
7419 }
7420
7421 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7422 {
7423         u64 num_devices;
7424         u64 stripped;
7425
7426         /*
7427          * if restripe for this chunk_type is on pick target profile and
7428          * return, otherwise do the usual balance
7429          */
7430         stripped = get_restripe_target(root->fs_info, flags);
7431         if (stripped)
7432                 return extended_to_chunk(stripped);
7433
7434         /*
7435          * we add in the count of missing devices because we want
7436          * to make sure that any RAID levels on a degraded FS
7437          * continue to be honored.
7438          */
7439         num_devices = root->fs_info->fs_devices->rw_devices +
7440                 root->fs_info->fs_devices->missing_devices;
7441
7442         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7443                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7444                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7445
7446         if (num_devices == 1) {
7447                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7448                 stripped = flags & ~stripped;
7449
7450                 /* turn raid0 into single device chunks */
7451                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7452                         return stripped;
7453
7454                 /* turn mirroring into duplication */
7455                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7456                              BTRFS_BLOCK_GROUP_RAID10))
7457                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7458         } else {
7459                 /* they already had raid on here, just return */
7460                 if (flags & stripped)
7461                         return flags;
7462
7463                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7464                 stripped = flags & ~stripped;
7465
7466                 /* switch duplicated blocks with raid1 */
7467                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7468                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7469
7470                 /* this is drive concat, leave it alone */
7471         }
7472
7473         return flags;
7474 }
7475
7476 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7477 {
7478         struct btrfs_space_info *sinfo = cache->space_info;
7479         u64 num_bytes;
7480         u64 min_allocable_bytes;
7481         int ret = -ENOSPC;
7482
7483
7484         /*
7485          * We need some metadata space and system metadata space for
7486          * allocating chunks in some corner cases until we force to set
7487          * it to be readonly.
7488          */
7489         if ((sinfo->flags &
7490              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7491             !force)
7492                 min_allocable_bytes = 1 * 1024 * 1024;
7493         else
7494                 min_allocable_bytes = 0;
7495
7496         spin_lock(&sinfo->lock);
7497         spin_lock(&cache->lock);
7498
7499         if (cache->ro) {
7500                 ret = 0;
7501                 goto out;
7502         }
7503
7504         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7505                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7506
7507         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7508             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7509             min_allocable_bytes <= sinfo->total_bytes) {
7510                 sinfo->bytes_readonly += num_bytes;
7511                 cache->ro = 1;
7512                 ret = 0;
7513         }
7514 out:
7515         spin_unlock(&cache->lock);
7516         spin_unlock(&sinfo->lock);
7517         return ret;
7518 }
7519
7520 int btrfs_set_block_group_ro(struct btrfs_root *root,
7521                              struct btrfs_block_group_cache *cache)
7522
7523 {
7524         struct btrfs_trans_handle *trans;
7525         u64 alloc_flags;
7526         int ret;
7527
7528         BUG_ON(cache->ro);
7529
7530         trans = btrfs_join_transaction(root);
7531         if (IS_ERR(trans))
7532                 return PTR_ERR(trans);
7533
7534         alloc_flags = update_block_group_flags(root, cache->flags);
7535         if (alloc_flags != cache->flags) {
7536                 ret = do_chunk_alloc(trans, root, alloc_flags,
7537                                      CHUNK_ALLOC_FORCE);
7538                 if (ret < 0)
7539                         goto out;
7540         }
7541
7542         ret = set_block_group_ro(cache, 0);
7543         if (!ret)
7544                 goto out;
7545         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7546         ret = do_chunk_alloc(trans, root, alloc_flags,
7547                              CHUNK_ALLOC_FORCE);
7548         if (ret < 0)
7549                 goto out;
7550         ret = set_block_group_ro(cache, 0);
7551 out:
7552         btrfs_end_transaction(trans, root);
7553         return ret;
7554 }
7555
7556 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7557                             struct btrfs_root *root, u64 type)
7558 {
7559         u64 alloc_flags = get_alloc_profile(root, type);
7560         return do_chunk_alloc(trans, root, alloc_flags,
7561                               CHUNK_ALLOC_FORCE);
7562 }
7563
7564 /*
7565  * helper to account the unused space of all the readonly block group in the
7566  * list. takes mirrors into account.
7567  */
7568 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7569 {
7570         struct btrfs_block_group_cache *block_group;
7571         u64 free_bytes = 0;
7572         int factor;
7573
7574         list_for_each_entry(block_group, groups_list, list) {
7575                 spin_lock(&block_group->lock);
7576
7577                 if (!block_group->ro) {
7578                         spin_unlock(&block_group->lock);
7579                         continue;
7580                 }
7581
7582                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7583                                           BTRFS_BLOCK_GROUP_RAID10 |
7584                                           BTRFS_BLOCK_GROUP_DUP))
7585                         factor = 2;
7586                 else
7587                         factor = 1;
7588
7589                 free_bytes += (block_group->key.offset -
7590                                btrfs_block_group_used(&block_group->item)) *
7591                                factor;
7592
7593                 spin_unlock(&block_group->lock);
7594         }
7595
7596         return free_bytes;
7597 }
7598
7599 /*
7600  * helper to account the unused space of all the readonly block group in the
7601  * space_info. takes mirrors into account.
7602  */
7603 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7604 {
7605         int i;
7606         u64 free_bytes = 0;
7607
7608         spin_lock(&sinfo->lock);
7609
7610         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7611                 if (!list_empty(&sinfo->block_groups[i]))
7612                         free_bytes += __btrfs_get_ro_block_group_free_space(
7613                                                 &sinfo->block_groups[i]);
7614
7615         spin_unlock(&sinfo->lock);
7616
7617         return free_bytes;
7618 }
7619
7620 void btrfs_set_block_group_rw(struct btrfs_root *root,
7621                               struct btrfs_block_group_cache *cache)
7622 {
7623         struct btrfs_space_info *sinfo = cache->space_info;
7624         u64 num_bytes;
7625
7626         BUG_ON(!cache->ro);
7627
7628         spin_lock(&sinfo->lock);
7629         spin_lock(&cache->lock);
7630         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7631                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7632         sinfo->bytes_readonly -= num_bytes;
7633         cache->ro = 0;
7634         spin_unlock(&cache->lock);
7635         spin_unlock(&sinfo->lock);
7636 }
7637
7638 /*
7639  * checks to see if its even possible to relocate this block group.
7640  *
7641  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7642  * ok to go ahead and try.
7643  */
7644 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7645 {
7646         struct btrfs_block_group_cache *block_group;
7647         struct btrfs_space_info *space_info;
7648         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7649         struct btrfs_device *device;
7650         u64 min_free;
7651         u64 dev_min = 1;
7652         u64 dev_nr = 0;
7653         u64 target;
7654         int index;
7655         int full = 0;
7656         int ret = 0;
7657
7658         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7659
7660         /* odd, couldn't find the block group, leave it alone */
7661         if (!block_group)
7662                 return -1;
7663
7664         min_free = btrfs_block_group_used(&block_group->item);
7665
7666         /* no bytes used, we're good */
7667         if (!min_free)
7668                 goto out;
7669
7670         space_info = block_group->space_info;
7671         spin_lock(&space_info->lock);
7672
7673         full = space_info->full;
7674
7675         /*
7676          * if this is the last block group we have in this space, we can't
7677          * relocate it unless we're able to allocate a new chunk below.
7678          *
7679          * Otherwise, we need to make sure we have room in the space to handle
7680          * all of the extents from this block group.  If we can, we're good
7681          */
7682         if ((space_info->total_bytes != block_group->key.offset) &&
7683             (space_info->bytes_used + space_info->bytes_reserved +
7684              space_info->bytes_pinned + space_info->bytes_readonly +
7685              min_free < space_info->total_bytes)) {
7686                 spin_unlock(&space_info->lock);
7687                 goto out;
7688         }
7689         spin_unlock(&space_info->lock);
7690
7691         /*
7692          * ok we don't have enough space, but maybe we have free space on our
7693          * devices to allocate new chunks for relocation, so loop through our
7694          * alloc devices and guess if we have enough space.  if this block
7695          * group is going to be restriped, run checks against the target
7696          * profile instead of the current one.
7697          */
7698         ret = -1;
7699
7700         /*
7701          * index:
7702          *      0: raid10
7703          *      1: raid1
7704          *      2: dup
7705          *      3: raid0
7706          *      4: single
7707          */
7708         target = get_restripe_target(root->fs_info, block_group->flags);
7709         if (target) {
7710                 index = __get_raid_index(extended_to_chunk(target));
7711         } else {
7712                 /*
7713                  * this is just a balance, so if we were marked as full
7714                  * we know there is no space for a new chunk
7715                  */
7716                 if (full)
7717                         goto out;
7718
7719                 index = get_block_group_index(block_group);
7720         }
7721
7722         if (index == BTRFS_RAID_RAID10) {
7723                 dev_min = 4;
7724                 /* Divide by 2 */
7725                 min_free >>= 1;
7726         } else if (index == BTRFS_RAID_RAID1) {
7727                 dev_min = 2;
7728         } else if (index == BTRFS_RAID_DUP) {
7729                 /* Multiply by 2 */
7730                 min_free <<= 1;
7731         } else if (index == BTRFS_RAID_RAID0) {
7732                 dev_min = fs_devices->rw_devices;
7733                 do_div(min_free, dev_min);
7734         }
7735
7736         mutex_lock(&root->fs_info->chunk_mutex);
7737         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7738                 u64 dev_offset;
7739
7740                 /*
7741                  * check to make sure we can actually find a chunk with enough
7742                  * space to fit our block group in.
7743                  */
7744                 if (device->total_bytes > device->bytes_used + min_free &&
7745                     !device->is_tgtdev_for_dev_replace) {
7746                         ret = find_free_dev_extent(device, min_free,
7747                                                    &dev_offset, NULL);
7748                         if (!ret)
7749                                 dev_nr++;
7750
7751                         if (dev_nr >= dev_min)
7752                                 break;
7753
7754                         ret = -1;
7755                 }
7756         }
7757         mutex_unlock(&root->fs_info->chunk_mutex);
7758 out:
7759         btrfs_put_block_group(block_group);
7760         return ret;
7761 }
7762
7763 static int find_first_block_group(struct btrfs_root *root,
7764                 struct btrfs_path *path, struct btrfs_key *key)
7765 {
7766         int ret = 0;
7767         struct btrfs_key found_key;
7768         struct extent_buffer *leaf;
7769         int slot;
7770
7771         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7772         if (ret < 0)
7773                 goto out;
7774
7775         while (1) {
7776                 slot = path->slots[0];
7777                 leaf = path->nodes[0];
7778                 if (slot >= btrfs_header_nritems(leaf)) {
7779                         ret = btrfs_next_leaf(root, path);
7780                         if (ret == 0)
7781                                 continue;
7782                         if (ret < 0)
7783                                 goto out;
7784                         break;
7785                 }
7786                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7787
7788                 if (found_key.objectid >= key->objectid &&
7789                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7790                         ret = 0;
7791                         goto out;
7792                 }
7793                 path->slots[0]++;
7794         }
7795 out:
7796         return ret;
7797 }
7798
7799 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7800 {
7801         struct btrfs_block_group_cache *block_group;
7802         u64 last = 0;
7803
7804         while (1) {
7805                 struct inode *inode;
7806
7807                 block_group = btrfs_lookup_first_block_group(info, last);
7808                 while (block_group) {
7809                         spin_lock(&block_group->lock);
7810                         if (block_group->iref)
7811                                 break;
7812                         spin_unlock(&block_group->lock);
7813                         block_group = next_block_group(info->tree_root,
7814                                                        block_group);
7815                 }
7816                 if (!block_group) {
7817                         if (last == 0)
7818                                 break;
7819                         last = 0;
7820                         continue;
7821                 }
7822
7823                 inode = block_group->inode;
7824                 block_group->iref = 0;
7825                 block_group->inode = NULL;
7826                 spin_unlock(&block_group->lock);
7827                 iput(inode);
7828                 last = block_group->key.objectid + block_group->key.offset;
7829                 btrfs_put_block_group(block_group);
7830         }
7831 }
7832
7833 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7834 {
7835         struct btrfs_block_group_cache *block_group;
7836         struct btrfs_space_info *space_info;
7837         struct btrfs_caching_control *caching_ctl;
7838         struct rb_node *n;
7839
7840         down_write(&info->extent_commit_sem);
7841         while (!list_empty(&info->caching_block_groups)) {
7842                 caching_ctl = list_entry(info->caching_block_groups.next,
7843                                          struct btrfs_caching_control, list);
7844                 list_del(&caching_ctl->list);
7845                 put_caching_control(caching_ctl);
7846         }
7847         up_write(&info->extent_commit_sem);
7848
7849         spin_lock(&info->block_group_cache_lock);
7850         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7851                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7852                                        cache_node);
7853                 rb_erase(&block_group->cache_node,
7854                          &info->block_group_cache_tree);
7855                 spin_unlock(&info->block_group_cache_lock);
7856
7857                 down_write(&block_group->space_info->groups_sem);
7858                 list_del(&block_group->list);
7859                 up_write(&block_group->space_info->groups_sem);
7860
7861                 if (block_group->cached == BTRFS_CACHE_STARTED)
7862                         wait_block_group_cache_done(block_group);
7863
7864                 /*
7865                  * We haven't cached this block group, which means we could
7866                  * possibly have excluded extents on this block group.
7867                  */
7868                 if (block_group->cached == BTRFS_CACHE_NO)
7869                         free_excluded_extents(info->extent_root, block_group);
7870
7871                 btrfs_remove_free_space_cache(block_group);
7872                 btrfs_put_block_group(block_group);
7873
7874                 spin_lock(&info->block_group_cache_lock);
7875         }
7876         spin_unlock(&info->block_group_cache_lock);
7877
7878         /* now that all the block groups are freed, go through and
7879          * free all the space_info structs.  This is only called during
7880          * the final stages of unmount, and so we know nobody is
7881          * using them.  We call synchronize_rcu() once before we start,
7882          * just to be on the safe side.
7883          */
7884         synchronize_rcu();
7885
7886         release_global_block_rsv(info);
7887
7888         while(!list_empty(&info->space_info)) {
7889                 space_info = list_entry(info->space_info.next,
7890                                         struct btrfs_space_info,
7891                                         list);
7892                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
7893                         if (space_info->bytes_pinned > 0 ||
7894                             space_info->bytes_reserved > 0 ||
7895                             space_info->bytes_may_use > 0) {
7896                                 WARN_ON(1);
7897                                 dump_space_info(space_info, 0, 0);
7898                         }
7899                 }
7900                 list_del(&space_info->list);
7901                 kfree(space_info);
7902         }
7903         return 0;
7904 }
7905
7906 static void __link_block_group(struct btrfs_space_info *space_info,
7907                                struct btrfs_block_group_cache *cache)
7908 {
7909         int index = get_block_group_index(cache);
7910
7911         down_write(&space_info->groups_sem);
7912         list_add_tail(&cache->list, &space_info->block_groups[index]);
7913         up_write(&space_info->groups_sem);
7914 }
7915
7916 int btrfs_read_block_groups(struct btrfs_root *root)
7917 {
7918         struct btrfs_path *path;
7919         int ret;
7920         struct btrfs_block_group_cache *cache;
7921         struct btrfs_fs_info *info = root->fs_info;
7922         struct btrfs_space_info *space_info;
7923         struct btrfs_key key;
7924         struct btrfs_key found_key;
7925         struct extent_buffer *leaf;
7926         int need_clear = 0;
7927         u64 cache_gen;
7928
7929         root = info->extent_root;
7930         key.objectid = 0;
7931         key.offset = 0;
7932         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7933         path = btrfs_alloc_path();
7934         if (!path)
7935                 return -ENOMEM;
7936         path->reada = 1;
7937
7938         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7939         if (btrfs_test_opt(root, SPACE_CACHE) &&
7940             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7941                 need_clear = 1;
7942         if (btrfs_test_opt(root, CLEAR_CACHE))
7943                 need_clear = 1;
7944
7945         while (1) {
7946                 ret = find_first_block_group(root, path, &key);
7947                 if (ret > 0)
7948                         break;
7949                 if (ret != 0)
7950                         goto error;
7951                 leaf = path->nodes[0];
7952                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7953                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7954                 if (!cache) {
7955                         ret = -ENOMEM;
7956                         goto error;
7957                 }
7958                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7959                                                 GFP_NOFS);
7960                 if (!cache->free_space_ctl) {
7961                         kfree(cache);
7962                         ret = -ENOMEM;
7963                         goto error;
7964                 }
7965
7966                 atomic_set(&cache->count, 1);
7967                 spin_lock_init(&cache->lock);
7968                 cache->fs_info = info;
7969                 INIT_LIST_HEAD(&cache->list);
7970                 INIT_LIST_HEAD(&cache->cluster_list);
7971
7972                 if (need_clear) {
7973                         /*
7974                          * When we mount with old space cache, we need to
7975                          * set BTRFS_DC_CLEAR and set dirty flag.
7976                          *
7977                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
7978                          *    truncate the old free space cache inode and
7979                          *    setup a new one.
7980                          * b) Setting 'dirty flag' makes sure that we flush
7981                          *    the new space cache info onto disk.
7982                          */
7983                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7984                         if (btrfs_test_opt(root, SPACE_CACHE))
7985                                 cache->dirty = 1;
7986                 }
7987
7988                 read_extent_buffer(leaf, &cache->item,
7989                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7990                                    sizeof(cache->item));
7991                 memcpy(&cache->key, &found_key, sizeof(found_key));
7992
7993                 key.objectid = found_key.objectid + found_key.offset;
7994                 btrfs_release_path(path);
7995                 cache->flags = btrfs_block_group_flags(&cache->item);
7996                 cache->sectorsize = root->sectorsize;
7997                 cache->full_stripe_len = btrfs_full_stripe_len(root,
7998                                                &root->fs_info->mapping_tree,
7999                                                found_key.objectid);
8000                 btrfs_init_free_space_ctl(cache);
8001
8002                 /*
8003                  * We need to exclude the super stripes now so that the space
8004                  * info has super bytes accounted for, otherwise we'll think
8005                  * we have more space than we actually do.
8006                  */
8007                 ret = exclude_super_stripes(root, cache);
8008                 if (ret) {
8009                         /*
8010                          * We may have excluded something, so call this just in
8011                          * case.
8012                          */
8013                         free_excluded_extents(root, cache);
8014                         kfree(cache->free_space_ctl);
8015                         kfree(cache);
8016                         goto error;
8017                 }
8018
8019                 /*
8020                  * check for two cases, either we are full, and therefore
8021                  * don't need to bother with the caching work since we won't
8022                  * find any space, or we are empty, and we can just add all
8023                  * the space in and be done with it.  This saves us _alot_ of
8024                  * time, particularly in the full case.
8025                  */
8026                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8027                         cache->last_byte_to_unpin = (u64)-1;
8028                         cache->cached = BTRFS_CACHE_FINISHED;
8029                         free_excluded_extents(root, cache);
8030                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8031                         cache->last_byte_to_unpin = (u64)-1;
8032                         cache->cached = BTRFS_CACHE_FINISHED;
8033                         add_new_free_space(cache, root->fs_info,
8034                                            found_key.objectid,
8035                                            found_key.objectid +
8036                                            found_key.offset);
8037                         free_excluded_extents(root, cache);
8038                 }
8039
8040                 ret = update_space_info(info, cache->flags, found_key.offset,
8041                                         btrfs_block_group_used(&cache->item),
8042                                         &space_info);
8043                 BUG_ON(ret); /* -ENOMEM */
8044                 cache->space_info = space_info;
8045                 spin_lock(&cache->space_info->lock);
8046                 cache->space_info->bytes_readonly += cache->bytes_super;
8047                 spin_unlock(&cache->space_info->lock);
8048
8049                 __link_block_group(space_info, cache);
8050
8051                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8052                 BUG_ON(ret); /* Logic error */
8053
8054                 set_avail_alloc_bits(root->fs_info, cache->flags);
8055                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8056                         set_block_group_ro(cache, 1);
8057         }
8058
8059         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8060                 if (!(get_alloc_profile(root, space_info->flags) &
8061                       (BTRFS_BLOCK_GROUP_RAID10 |
8062                        BTRFS_BLOCK_GROUP_RAID1 |
8063                        BTRFS_BLOCK_GROUP_RAID5 |
8064                        BTRFS_BLOCK_GROUP_RAID6 |
8065                        BTRFS_BLOCK_GROUP_DUP)))
8066                         continue;
8067                 /*
8068                  * avoid allocating from un-mirrored block group if there are
8069                  * mirrored block groups.
8070                  */
8071                 list_for_each_entry(cache, &space_info->block_groups[3], list)
8072                         set_block_group_ro(cache, 1);
8073                 list_for_each_entry(cache, &space_info->block_groups[4], list)
8074                         set_block_group_ro(cache, 1);
8075         }
8076
8077         init_global_block_rsv(info);
8078         ret = 0;
8079 error:
8080         btrfs_free_path(path);
8081         return ret;
8082 }
8083
8084 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8085                                        struct btrfs_root *root)
8086 {
8087         struct btrfs_block_group_cache *block_group, *tmp;
8088         struct btrfs_root *extent_root = root->fs_info->extent_root;
8089         struct btrfs_block_group_item item;
8090         struct btrfs_key key;
8091         int ret = 0;
8092
8093         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8094                                  new_bg_list) {
8095                 list_del_init(&block_group->new_bg_list);
8096
8097                 if (ret)
8098                         continue;
8099
8100                 spin_lock(&block_group->lock);
8101                 memcpy(&item, &block_group->item, sizeof(item));
8102                 memcpy(&key, &block_group->key, sizeof(key));
8103                 spin_unlock(&block_group->lock);
8104
8105                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8106                                         sizeof(item));
8107                 if (ret)
8108                         btrfs_abort_transaction(trans, extent_root, ret);
8109         }
8110 }
8111
8112 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8113                            struct btrfs_root *root, u64 bytes_used,
8114                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8115                            u64 size)
8116 {
8117         int ret;
8118         struct btrfs_root *extent_root;
8119         struct btrfs_block_group_cache *cache;
8120
8121         extent_root = root->fs_info->extent_root;
8122
8123         root->fs_info->last_trans_log_full_commit = trans->transid;
8124
8125         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8126         if (!cache)
8127                 return -ENOMEM;
8128         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8129                                         GFP_NOFS);
8130         if (!cache->free_space_ctl) {
8131                 kfree(cache);
8132                 return -ENOMEM;
8133         }
8134
8135         cache->key.objectid = chunk_offset;
8136         cache->key.offset = size;
8137         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8138         cache->sectorsize = root->sectorsize;
8139         cache->fs_info = root->fs_info;
8140         cache->full_stripe_len = btrfs_full_stripe_len(root,
8141                                                &root->fs_info->mapping_tree,
8142                                                chunk_offset);
8143
8144         atomic_set(&cache->count, 1);
8145         spin_lock_init(&cache->lock);
8146         INIT_LIST_HEAD(&cache->list);
8147         INIT_LIST_HEAD(&cache->cluster_list);
8148         INIT_LIST_HEAD(&cache->new_bg_list);
8149
8150         btrfs_init_free_space_ctl(cache);
8151
8152         btrfs_set_block_group_used(&cache->item, bytes_used);
8153         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8154         cache->flags = type;
8155         btrfs_set_block_group_flags(&cache->item, type);
8156
8157         cache->last_byte_to_unpin = (u64)-1;
8158         cache->cached = BTRFS_CACHE_FINISHED;
8159         ret = exclude_super_stripes(root, cache);
8160         if (ret) {
8161                 /*
8162                  * We may have excluded something, so call this just in
8163                  * case.
8164                  */
8165                 free_excluded_extents(root, cache);
8166                 kfree(cache->free_space_ctl);
8167                 kfree(cache);
8168                 return ret;
8169         }
8170
8171         add_new_free_space(cache, root->fs_info, chunk_offset,
8172                            chunk_offset + size);
8173
8174         free_excluded_extents(root, cache);
8175
8176         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8177                                 &cache->space_info);
8178         BUG_ON(ret); /* -ENOMEM */
8179         update_global_block_rsv(root->fs_info);
8180
8181         spin_lock(&cache->space_info->lock);
8182         cache->space_info->bytes_readonly += cache->bytes_super;
8183         spin_unlock(&cache->space_info->lock);
8184
8185         __link_block_group(cache->space_info, cache);
8186
8187         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8188         BUG_ON(ret); /* Logic error */
8189
8190         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8191
8192         set_avail_alloc_bits(extent_root->fs_info, type);
8193
8194         return 0;
8195 }
8196
8197 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8198 {
8199         u64 extra_flags = chunk_to_extended(flags) &
8200                                 BTRFS_EXTENDED_PROFILE_MASK;
8201
8202         write_seqlock(&fs_info->profiles_lock);
8203         if (flags & BTRFS_BLOCK_GROUP_DATA)
8204                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8205         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8206                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8207         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8208                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8209         write_sequnlock(&fs_info->profiles_lock);
8210 }
8211
8212 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8213                              struct btrfs_root *root, u64 group_start)
8214 {
8215         struct btrfs_path *path;
8216         struct btrfs_block_group_cache *block_group;
8217         struct btrfs_free_cluster *cluster;
8218         struct btrfs_root *tree_root = root->fs_info->tree_root;
8219         struct btrfs_key key;
8220         struct inode *inode;
8221         int ret;
8222         int index;
8223         int factor;
8224
8225         root = root->fs_info->extent_root;
8226
8227         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8228         BUG_ON(!block_group);
8229         BUG_ON(!block_group->ro);
8230
8231         /*
8232          * Free the reserved super bytes from this block group before
8233          * remove it.
8234          */
8235         free_excluded_extents(root, block_group);
8236
8237         memcpy(&key, &block_group->key, sizeof(key));
8238         index = get_block_group_index(block_group);
8239         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8240                                   BTRFS_BLOCK_GROUP_RAID1 |
8241                                   BTRFS_BLOCK_GROUP_RAID10))
8242                 factor = 2;
8243         else
8244                 factor = 1;
8245
8246         /* make sure this block group isn't part of an allocation cluster */
8247         cluster = &root->fs_info->data_alloc_cluster;
8248         spin_lock(&cluster->refill_lock);
8249         btrfs_return_cluster_to_free_space(block_group, cluster);
8250         spin_unlock(&cluster->refill_lock);
8251
8252         /*
8253          * make sure this block group isn't part of a metadata
8254          * allocation cluster
8255          */
8256         cluster = &root->fs_info->meta_alloc_cluster;
8257         spin_lock(&cluster->refill_lock);
8258         btrfs_return_cluster_to_free_space(block_group, cluster);
8259         spin_unlock(&cluster->refill_lock);
8260
8261         path = btrfs_alloc_path();
8262         if (!path) {
8263                 ret = -ENOMEM;
8264                 goto out;
8265         }
8266
8267         inode = lookup_free_space_inode(tree_root, block_group, path);
8268         if (!IS_ERR(inode)) {
8269                 ret = btrfs_orphan_add(trans, inode);
8270                 if (ret) {
8271                         btrfs_add_delayed_iput(inode);
8272                         goto out;
8273                 }
8274                 clear_nlink(inode);
8275                 /* One for the block groups ref */
8276                 spin_lock(&block_group->lock);
8277                 if (block_group->iref) {
8278                         block_group->iref = 0;
8279                         block_group->inode = NULL;
8280                         spin_unlock(&block_group->lock);
8281                         iput(inode);
8282                 } else {
8283                         spin_unlock(&block_group->lock);
8284                 }
8285                 /* One for our lookup ref */
8286                 btrfs_add_delayed_iput(inode);
8287         }
8288
8289         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8290         key.offset = block_group->key.objectid;
8291         key.type = 0;
8292
8293         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8294         if (ret < 0)
8295                 goto out;
8296         if (ret > 0)
8297                 btrfs_release_path(path);
8298         if (ret == 0) {
8299                 ret = btrfs_del_item(trans, tree_root, path);
8300                 if (ret)
8301                         goto out;
8302                 btrfs_release_path(path);
8303         }
8304
8305         spin_lock(&root->fs_info->block_group_cache_lock);
8306         rb_erase(&block_group->cache_node,
8307                  &root->fs_info->block_group_cache_tree);
8308
8309         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8310                 root->fs_info->first_logical_byte = (u64)-1;
8311         spin_unlock(&root->fs_info->block_group_cache_lock);
8312
8313         down_write(&block_group->space_info->groups_sem);
8314         /*
8315          * we must use list_del_init so people can check to see if they
8316          * are still on the list after taking the semaphore
8317          */
8318         list_del_init(&block_group->list);
8319         if (list_empty(&block_group->space_info->block_groups[index]))
8320                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8321         up_write(&block_group->space_info->groups_sem);
8322
8323         if (block_group->cached == BTRFS_CACHE_STARTED)
8324                 wait_block_group_cache_done(block_group);
8325
8326         btrfs_remove_free_space_cache(block_group);
8327
8328         spin_lock(&block_group->space_info->lock);
8329         block_group->space_info->total_bytes -= block_group->key.offset;
8330         block_group->space_info->bytes_readonly -= block_group->key.offset;
8331         block_group->space_info->disk_total -= block_group->key.offset * factor;
8332         spin_unlock(&block_group->space_info->lock);
8333
8334         memcpy(&key, &block_group->key, sizeof(key));
8335
8336         btrfs_clear_space_info_full(root->fs_info);
8337
8338         btrfs_put_block_group(block_group);
8339         btrfs_put_block_group(block_group);
8340
8341         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8342         if (ret > 0)
8343                 ret = -EIO;
8344         if (ret < 0)
8345                 goto out;
8346
8347         ret = btrfs_del_item(trans, root, path);
8348 out:
8349         btrfs_free_path(path);
8350         return ret;
8351 }
8352
8353 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8354 {
8355         struct btrfs_space_info *space_info;
8356         struct btrfs_super_block *disk_super;
8357         u64 features;
8358         u64 flags;
8359         int mixed = 0;
8360         int ret;
8361
8362         disk_super = fs_info->super_copy;
8363         if (!btrfs_super_root(disk_super))
8364                 return 1;
8365
8366         features = btrfs_super_incompat_flags(disk_super);
8367         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8368                 mixed = 1;
8369
8370         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8371         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8372         if (ret)
8373                 goto out;
8374
8375         if (mixed) {
8376                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8377                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8378         } else {
8379                 flags = BTRFS_BLOCK_GROUP_METADATA;
8380                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8381                 if (ret)
8382                         goto out;
8383
8384                 flags = BTRFS_BLOCK_GROUP_DATA;
8385                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8386         }
8387 out:
8388         return ret;
8389 }
8390
8391 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8392 {
8393         return unpin_extent_range(root, start, end);
8394 }
8395
8396 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8397                                u64 num_bytes, u64 *actual_bytes)
8398 {
8399         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8400 }
8401
8402 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8403 {
8404         struct btrfs_fs_info *fs_info = root->fs_info;
8405         struct btrfs_block_group_cache *cache = NULL;
8406         u64 group_trimmed;
8407         u64 start;
8408         u64 end;
8409         u64 trimmed = 0;
8410         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8411         int ret = 0;
8412
8413         /*
8414          * try to trim all FS space, our block group may start from non-zero.
8415          */
8416         if (range->len == total_bytes)
8417                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8418         else
8419                 cache = btrfs_lookup_block_group(fs_info, range->start);
8420
8421         while (cache) {
8422                 if (cache->key.objectid >= (range->start + range->len)) {
8423                         btrfs_put_block_group(cache);
8424                         break;
8425                 }
8426
8427                 start = max(range->start, cache->key.objectid);
8428                 end = min(range->start + range->len,
8429                                 cache->key.objectid + cache->key.offset);
8430
8431                 if (end - start >= range->minlen) {
8432                         if (!block_group_cache_done(cache)) {
8433                                 ret = cache_block_group(cache, 0);
8434                                 if (!ret)
8435                                         wait_block_group_cache_done(cache);
8436                         }
8437                         ret = btrfs_trim_block_group(cache,
8438                                                      &group_trimmed,
8439                                                      start,
8440                                                      end,
8441                                                      range->minlen);
8442
8443                         trimmed += group_trimmed;
8444                         if (ret) {
8445                                 btrfs_put_block_group(cache);
8446                                 break;
8447                         }
8448                 }
8449
8450                 cache = next_block_group(fs_info->tree_root, cache);
8451         }
8452
8453         range->len = trimmed;
8454         return ret;
8455 }