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