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                 return ret;
2147         }
2148
2149         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2150             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2151                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2152                                            insert_reserved);
2153         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2154                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2155                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2156                                            insert_reserved);
2157         else
2158                 BUG();
2159         return ret;
2160 }
2161
2162 static noinline struct btrfs_delayed_ref_node *
2163 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2164 {
2165         struct rb_node *node;
2166         struct btrfs_delayed_ref_node *ref;
2167         int action = BTRFS_ADD_DELAYED_REF;
2168 again:
2169         /*
2170          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2171          * this prevents ref count from going down to zero when
2172          * there still are pending delayed ref.
2173          */
2174         node = rb_prev(&head->node.rb_node);
2175         while (1) {
2176                 if (!node)
2177                         break;
2178                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2179                                 rb_node);
2180                 if (ref->bytenr != head->node.bytenr)
2181                         break;
2182                 if (ref->action == action)
2183                         return ref;
2184                 node = rb_prev(node);
2185         }
2186         if (action == BTRFS_ADD_DELAYED_REF) {
2187                 action = BTRFS_DROP_DELAYED_REF;
2188                 goto again;
2189         }
2190         return NULL;
2191 }
2192
2193 /*
2194  * Returns 0 on success or if called with an already aborted transaction.
2195  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2196  */
2197 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2198                                        struct btrfs_root *root,
2199                                        struct list_head *cluster)
2200 {
2201         struct btrfs_delayed_ref_root *delayed_refs;
2202         struct btrfs_delayed_ref_node *ref;
2203         struct btrfs_delayed_ref_head *locked_ref = NULL;
2204         struct btrfs_delayed_extent_op *extent_op;
2205         struct btrfs_fs_info *fs_info = root->fs_info;
2206         int ret;
2207         int count = 0;
2208         int must_insert_reserved = 0;
2209
2210         delayed_refs = &trans->transaction->delayed_refs;
2211         while (1) {
2212                 if (!locked_ref) {
2213                         /* pick a new head ref from the cluster list */
2214                         if (list_empty(cluster))
2215                                 break;
2216
2217                         locked_ref = list_entry(cluster->next,
2218                                      struct btrfs_delayed_ref_head, cluster);
2219
2220                         /* grab the lock that says we are going to process
2221                          * all the refs for this head */
2222                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2223
2224                         /*
2225                          * we may have dropped the spin lock to get the head
2226                          * mutex lock, and that might have given someone else
2227                          * time to free the head.  If that's true, it has been
2228                          * removed from our list and we can move on.
2229                          */
2230                         if (ret == -EAGAIN) {
2231                                 locked_ref = NULL;
2232                                 count++;
2233                                 continue;
2234                         }
2235                 }
2236
2237                 /*
2238                  * We need to try and merge add/drops of the same ref since we
2239                  * can run into issues with relocate dropping the implicit ref
2240                  * and then it being added back again before the drop can
2241                  * finish.  If we merged anything we need to re-loop so we can
2242                  * get a good ref.
2243                  */
2244                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2245                                          locked_ref);
2246
2247                 /*
2248                  * locked_ref is the head node, so we have to go one
2249                  * node back for any delayed ref updates
2250                  */
2251                 ref = select_delayed_ref(locked_ref);
2252
2253                 if (ref && ref->seq &&
2254                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2255                         /*
2256                          * there are still refs with lower seq numbers in the
2257                          * process of being added. Don't run this ref yet.
2258                          */
2259                         list_del_init(&locked_ref->cluster);
2260                         btrfs_delayed_ref_unlock(locked_ref);
2261                         locked_ref = NULL;
2262                         delayed_refs->num_heads_ready++;
2263                         spin_unlock(&delayed_refs->lock);
2264                         cond_resched();
2265                         spin_lock(&delayed_refs->lock);
2266                         continue;
2267                 }
2268
2269                 /*
2270                  * record the must insert reserved flag before we
2271                  * drop the spin lock.
2272                  */
2273                 must_insert_reserved = locked_ref->must_insert_reserved;
2274                 locked_ref->must_insert_reserved = 0;
2275
2276                 extent_op = locked_ref->extent_op;
2277                 locked_ref->extent_op = NULL;
2278
2279                 if (!ref) {
2280                         /* All delayed refs have been processed, Go ahead
2281                          * and send the head node to run_one_delayed_ref,
2282                          * so that any accounting fixes can happen
2283                          */
2284                         ref = &locked_ref->node;
2285
2286                         if (extent_op && must_insert_reserved) {
2287                                 btrfs_free_delayed_extent_op(extent_op);
2288                                 extent_op = NULL;
2289                         }
2290
2291                         if (extent_op) {
2292                                 spin_unlock(&delayed_refs->lock);
2293
2294                                 ret = run_delayed_extent_op(trans, root,
2295                                                             ref, extent_op);
2296                                 btrfs_free_delayed_extent_op(extent_op);
2297
2298                                 if (ret) {
2299                                         printk(KERN_DEBUG
2300                                                "btrfs: run_delayed_extent_op "
2301                                                "returned %d\n", ret);
2302                                         spin_lock(&delayed_refs->lock);
2303                                         btrfs_delayed_ref_unlock(locked_ref);
2304                                         return ret;
2305                                 }
2306
2307                                 goto next;
2308                         }
2309                 }
2310
2311                 ref->in_tree = 0;
2312                 rb_erase(&ref->rb_node, &delayed_refs->root);
2313                 delayed_refs->num_entries--;
2314                 if (!btrfs_delayed_ref_is_head(ref)) {
2315                         /*
2316                          * when we play the delayed ref, also correct the
2317                          * ref_mod on head
2318                          */
2319                         switch (ref->action) {
2320                         case BTRFS_ADD_DELAYED_REF:
2321                         case BTRFS_ADD_DELAYED_EXTENT:
2322                                 locked_ref->node.ref_mod -= ref->ref_mod;
2323                                 break;
2324                         case BTRFS_DROP_DELAYED_REF:
2325                                 locked_ref->node.ref_mod += ref->ref_mod;
2326                                 break;
2327                         default:
2328                                 WARN_ON(1);
2329                         }
2330                 }
2331                 spin_unlock(&delayed_refs->lock);
2332
2333                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2334                                           must_insert_reserved);
2335
2336                 btrfs_free_delayed_extent_op(extent_op);
2337                 if (ret) {
2338                         btrfs_delayed_ref_unlock(locked_ref);
2339                         btrfs_put_delayed_ref(ref);
2340                         printk(KERN_DEBUG
2341                                "btrfs: run_one_delayed_ref returned %d\n", ret);
2342                         spin_lock(&delayed_refs->lock);
2343                         return ret;
2344                 }
2345
2346                 /*
2347                  * If this node is a head, that means all the refs in this head
2348                  * have been dealt with, and we will pick the next head to deal
2349                  * with, so we must unlock the head and drop it from the cluster
2350                  * list before we release it.
2351                  */
2352                 if (btrfs_delayed_ref_is_head(ref)) {
2353                         list_del_init(&locked_ref->cluster);
2354                         btrfs_delayed_ref_unlock(locked_ref);
2355                         locked_ref = NULL;
2356                 }
2357                 btrfs_put_delayed_ref(ref);
2358                 count++;
2359 next:
2360                 cond_resched();
2361                 spin_lock(&delayed_refs->lock);
2362         }
2363         return count;
2364 }
2365
2366 #ifdef SCRAMBLE_DELAYED_REFS
2367 /*
2368  * Normally delayed refs get processed in ascending bytenr order. This
2369  * correlates in most cases to the order added. To expose dependencies on this
2370  * order, we start to process the tree in the middle instead of the beginning
2371  */
2372 static u64 find_middle(struct rb_root *root)
2373 {
2374         struct rb_node *n = root->rb_node;
2375         struct btrfs_delayed_ref_node *entry;
2376         int alt = 1;
2377         u64 middle;
2378         u64 first = 0, last = 0;
2379
2380         n = rb_first(root);
2381         if (n) {
2382                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2383                 first = entry->bytenr;
2384         }
2385         n = rb_last(root);
2386         if (n) {
2387                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2388                 last = entry->bytenr;
2389         }
2390         n = root->rb_node;
2391
2392         while (n) {
2393                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2394                 WARN_ON(!entry->in_tree);
2395
2396                 middle = entry->bytenr;
2397
2398                 if (alt)
2399                         n = n->rb_left;
2400                 else
2401                         n = n->rb_right;
2402
2403                 alt = 1 - alt;
2404         }
2405         return middle;
2406 }
2407 #endif
2408
2409 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2410                                          struct btrfs_fs_info *fs_info)
2411 {
2412         struct qgroup_update *qgroup_update;
2413         int ret = 0;
2414
2415         if (list_empty(&trans->qgroup_ref_list) !=
2416             !trans->delayed_ref_elem.seq) {
2417                 /* list without seq or seq without list */
2418                 printk(KERN_ERR "btrfs: qgroup accounting update error, list is%s empty, seq is %llu\n",
2419                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2420                         trans->delayed_ref_elem.seq);
2421                 BUG();
2422         }
2423
2424         if (!trans->delayed_ref_elem.seq)
2425                 return 0;
2426
2427         while (!list_empty(&trans->qgroup_ref_list)) {
2428                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2429                                                  struct qgroup_update, list);
2430                 list_del(&qgroup_update->list);
2431                 if (!ret)
2432                         ret = btrfs_qgroup_account_ref(
2433                                         trans, fs_info, qgroup_update->node,
2434                                         qgroup_update->extent_op);
2435                 kfree(qgroup_update);
2436         }
2437
2438         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2439
2440         return ret;
2441 }
2442
2443 /*
2444  * this starts processing the delayed reference count updates and
2445  * extent insertions we have queued up so far.  count can be
2446  * 0, which means to process everything in the tree at the start
2447  * of the run (but not newly added entries), or it can be some target
2448  * number you'd like to process.
2449  *
2450  * Returns 0 on success or if called with an aborted transaction
2451  * Returns <0 on error and aborts the transaction
2452  */
2453 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2454                            struct btrfs_root *root, unsigned long count)
2455 {
2456         struct rb_node *node;
2457         struct btrfs_delayed_ref_root *delayed_refs;
2458         struct btrfs_delayed_ref_node *ref;
2459         struct list_head cluster;
2460         int ret;
2461         u64 delayed_start;
2462         int run_all = count == (unsigned long)-1;
2463         int run_most = 0;
2464         int loops;
2465
2466         /* We'll clean this up in btrfs_cleanup_transaction */
2467         if (trans->aborted)
2468                 return 0;
2469
2470         if (root == root->fs_info->extent_root)
2471                 root = root->fs_info->tree_root;
2472
2473         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2474
2475         delayed_refs = &trans->transaction->delayed_refs;
2476         INIT_LIST_HEAD(&cluster);
2477 again:
2478         loops = 0;
2479         spin_lock(&delayed_refs->lock);
2480
2481 #ifdef SCRAMBLE_DELAYED_REFS
2482         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2483 #endif
2484
2485         if (count == 0) {
2486                 count = delayed_refs->num_entries * 2;
2487                 run_most = 1;
2488         }
2489         while (1) {
2490                 if (!(run_all || run_most) &&
2491                     delayed_refs->num_heads_ready < 64)
2492                         break;
2493
2494                 /*
2495                  * go find something we can process in the rbtree.  We start at
2496                  * the beginning of the tree, and then build a cluster
2497                  * of refs to process starting at the first one we are able to
2498                  * lock
2499                  */
2500                 delayed_start = delayed_refs->run_delayed_start;
2501                 ret = btrfs_find_ref_cluster(trans, &cluster,
2502                                              delayed_refs->run_delayed_start);
2503                 if (ret)
2504                         break;
2505
2506                 ret = run_clustered_refs(trans, root, &cluster);
2507                 if (ret < 0) {
2508                         btrfs_release_ref_cluster(&cluster);
2509                         spin_unlock(&delayed_refs->lock);
2510                         btrfs_abort_transaction(trans, root, ret);
2511                         return ret;
2512                 }
2513
2514                 count -= min_t(unsigned long, ret, count);
2515
2516                 if (count == 0)
2517                         break;
2518
2519                 if (delayed_start >= delayed_refs->run_delayed_start) {
2520                         if (loops == 0) {
2521                                 /*
2522                                  * btrfs_find_ref_cluster looped. let's do one
2523                                  * more cycle. if we don't run any delayed ref
2524                                  * during that cycle (because we can't because
2525                                  * all of them are blocked), bail out.
2526                                  */
2527                                 loops = 1;
2528                         } else {
2529                                 /*
2530                                  * no runnable refs left, stop trying
2531                                  */
2532                                 BUG_ON(run_all);
2533                                 break;
2534                         }
2535                 }
2536                 if (ret) {
2537                         /* refs were run, let's reset staleness detection */
2538                         loops = 0;
2539                 }
2540         }
2541
2542         if (run_all) {
2543                 if (!list_empty(&trans->new_bgs)) {
2544                         spin_unlock(&delayed_refs->lock);
2545                         btrfs_create_pending_block_groups(trans, root);
2546                         spin_lock(&delayed_refs->lock);
2547                 }
2548
2549                 node = rb_first(&delayed_refs->root);
2550                 if (!node)
2551                         goto out;
2552                 count = (unsigned long)-1;
2553
2554                 while (node) {
2555                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2556                                        rb_node);
2557                         if (btrfs_delayed_ref_is_head(ref)) {
2558                                 struct btrfs_delayed_ref_head *head;
2559
2560                                 head = btrfs_delayed_node_to_head(ref);
2561                                 atomic_inc(&ref->refs);
2562
2563                                 spin_unlock(&delayed_refs->lock);
2564                                 /*
2565                                  * Mutex was contended, block until it's
2566                                  * released and try again
2567                                  */
2568                                 mutex_lock(&head->mutex);
2569                                 mutex_unlock(&head->mutex);
2570
2571                                 btrfs_put_delayed_ref(ref);
2572                                 cond_resched();
2573                                 goto again;
2574                         }
2575                         node = rb_next(node);
2576                 }
2577                 spin_unlock(&delayed_refs->lock);
2578                 schedule_timeout(1);
2579                 goto again;
2580         }
2581 out:
2582         spin_unlock(&delayed_refs->lock);
2583         assert_qgroups_uptodate(trans);
2584         return 0;
2585 }
2586
2587 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2588                                 struct btrfs_root *root,
2589                                 u64 bytenr, u64 num_bytes, u64 flags,
2590                                 int is_data)
2591 {
2592         struct btrfs_delayed_extent_op *extent_op;
2593         int ret;
2594
2595         extent_op = btrfs_alloc_delayed_extent_op();
2596         if (!extent_op)
2597                 return -ENOMEM;
2598
2599         extent_op->flags_to_set = flags;
2600         extent_op->update_flags = 1;
2601         extent_op->update_key = 0;
2602         extent_op->is_data = is_data ? 1 : 0;
2603
2604         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2605                                           num_bytes, extent_op);
2606         if (ret)
2607                 btrfs_free_delayed_extent_op(extent_op);
2608         return ret;
2609 }
2610
2611 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2612                                       struct btrfs_root *root,
2613                                       struct btrfs_path *path,
2614                                       u64 objectid, u64 offset, u64 bytenr)
2615 {
2616         struct btrfs_delayed_ref_head *head;
2617         struct btrfs_delayed_ref_node *ref;
2618         struct btrfs_delayed_data_ref *data_ref;
2619         struct btrfs_delayed_ref_root *delayed_refs;
2620         struct rb_node *node;
2621         int ret = 0;
2622
2623         ret = -ENOENT;
2624         delayed_refs = &trans->transaction->delayed_refs;
2625         spin_lock(&delayed_refs->lock);
2626         head = btrfs_find_delayed_ref_head(trans, bytenr);
2627         if (!head)
2628                 goto out;
2629
2630         if (!mutex_trylock(&head->mutex)) {
2631                 atomic_inc(&head->node.refs);
2632                 spin_unlock(&delayed_refs->lock);
2633
2634                 btrfs_release_path(path);
2635
2636                 /*
2637                  * Mutex was contended, block until it's released and let
2638                  * caller try again
2639                  */
2640                 mutex_lock(&head->mutex);
2641                 mutex_unlock(&head->mutex);
2642                 btrfs_put_delayed_ref(&head->node);
2643                 return -EAGAIN;
2644         }
2645
2646         node = rb_prev(&head->node.rb_node);
2647         if (!node)
2648                 goto out_unlock;
2649
2650         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2651
2652         if (ref->bytenr != bytenr)
2653                 goto out_unlock;
2654
2655         ret = 1;
2656         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2657                 goto out_unlock;
2658
2659         data_ref = btrfs_delayed_node_to_data_ref(ref);
2660
2661         node = rb_prev(node);
2662         if (node) {
2663                 int seq = ref->seq;
2664
2665                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2666                 if (ref->bytenr == bytenr && ref->seq == seq)
2667                         goto out_unlock;
2668         }
2669
2670         if (data_ref->root != root->root_key.objectid ||
2671             data_ref->objectid != objectid || data_ref->offset != offset)
2672                 goto out_unlock;
2673
2674         ret = 0;
2675 out_unlock:
2676         mutex_unlock(&head->mutex);
2677 out:
2678         spin_unlock(&delayed_refs->lock);
2679         return ret;
2680 }
2681
2682 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2683                                         struct btrfs_root *root,
2684                                         struct btrfs_path *path,
2685                                         u64 objectid, u64 offset, u64 bytenr)
2686 {
2687         struct btrfs_root *extent_root = root->fs_info->extent_root;
2688         struct extent_buffer *leaf;
2689         struct btrfs_extent_data_ref *ref;
2690         struct btrfs_extent_inline_ref *iref;
2691         struct btrfs_extent_item *ei;
2692         struct btrfs_key key;
2693         u32 item_size;
2694         int ret;
2695
2696         key.objectid = bytenr;
2697         key.offset = (u64)-1;
2698         key.type = BTRFS_EXTENT_ITEM_KEY;
2699
2700         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2701         if (ret < 0)
2702                 goto out;
2703         BUG_ON(ret == 0); /* Corruption */
2704
2705         ret = -ENOENT;
2706         if (path->slots[0] == 0)
2707                 goto out;
2708
2709         path->slots[0]--;
2710         leaf = path->nodes[0];
2711         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2712
2713         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2714                 goto out;
2715
2716         ret = 1;
2717         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2718 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2719         if (item_size < sizeof(*ei)) {
2720                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2721                 goto out;
2722         }
2723 #endif
2724         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2725
2726         if (item_size != sizeof(*ei) +
2727             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2728                 goto out;
2729
2730         if (btrfs_extent_generation(leaf, ei) <=
2731             btrfs_root_last_snapshot(&root->root_item))
2732                 goto out;
2733
2734         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2735         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2736             BTRFS_EXTENT_DATA_REF_KEY)
2737                 goto out;
2738
2739         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2740         if (btrfs_extent_refs(leaf, ei) !=
2741             btrfs_extent_data_ref_count(leaf, ref) ||
2742             btrfs_extent_data_ref_root(leaf, ref) !=
2743             root->root_key.objectid ||
2744             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2745             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2746                 goto out;
2747
2748         ret = 0;
2749 out:
2750         return ret;
2751 }
2752
2753 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2754                           struct btrfs_root *root,
2755                           u64 objectid, u64 offset, u64 bytenr)
2756 {
2757         struct btrfs_path *path;
2758         int ret;
2759         int ret2;
2760
2761         path = btrfs_alloc_path();
2762         if (!path)
2763                 return -ENOENT;
2764
2765         do {
2766                 ret = check_committed_ref(trans, root, path, objectid,
2767                                           offset, bytenr);
2768                 if (ret && ret != -ENOENT)
2769                         goto out;
2770
2771                 ret2 = check_delayed_ref(trans, root, path, objectid,
2772                                          offset, bytenr);
2773         } while (ret2 == -EAGAIN);
2774
2775         if (ret2 && ret2 != -ENOENT) {
2776                 ret = ret2;
2777                 goto out;
2778         }
2779
2780         if (ret != -ENOENT || ret2 != -ENOENT)
2781                 ret = 0;
2782 out:
2783         btrfs_free_path(path);
2784         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2785                 WARN_ON(ret > 0);
2786         return ret;
2787 }
2788
2789 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2790                            struct btrfs_root *root,
2791                            struct extent_buffer *buf,
2792                            int full_backref, int inc, int for_cow)
2793 {
2794         u64 bytenr;
2795         u64 num_bytes;
2796         u64 parent;
2797         u64 ref_root;
2798         u32 nritems;
2799         struct btrfs_key key;
2800         struct btrfs_file_extent_item *fi;
2801         int i;
2802         int level;
2803         int ret = 0;
2804         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2805                             u64, u64, u64, u64, u64, u64, int);
2806
2807         ref_root = btrfs_header_owner(buf);
2808         nritems = btrfs_header_nritems(buf);
2809         level = btrfs_header_level(buf);
2810
2811         if (!root->ref_cows && level == 0)
2812                 return 0;
2813
2814         if (inc)
2815                 process_func = btrfs_inc_extent_ref;
2816         else
2817                 process_func = btrfs_free_extent;
2818
2819         if (full_backref)
2820                 parent = buf->start;
2821         else
2822                 parent = 0;
2823
2824         for (i = 0; i < nritems; i++) {
2825                 if (level == 0) {
2826                         btrfs_item_key_to_cpu(buf, &key, i);
2827                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2828                                 continue;
2829                         fi = btrfs_item_ptr(buf, i,
2830                                             struct btrfs_file_extent_item);
2831                         if (btrfs_file_extent_type(buf, fi) ==
2832                             BTRFS_FILE_EXTENT_INLINE)
2833                                 continue;
2834                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2835                         if (bytenr == 0)
2836                                 continue;
2837
2838                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2839                         key.offset -= btrfs_file_extent_offset(buf, fi);
2840                         ret = process_func(trans, root, bytenr, num_bytes,
2841                                            parent, ref_root, key.objectid,
2842                                            key.offset, for_cow);
2843                         if (ret)
2844                                 goto fail;
2845                 } else {
2846                         bytenr = btrfs_node_blockptr(buf, i);
2847                         num_bytes = btrfs_level_size(root, level - 1);
2848                         ret = process_func(trans, root, bytenr, num_bytes,
2849                                            parent, ref_root, level - 1, 0,
2850                                            for_cow);
2851                         if (ret)
2852                                 goto fail;
2853                 }
2854         }
2855         return 0;
2856 fail:
2857         return ret;
2858 }
2859
2860 int btrfs_inc_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, 1, for_cow);
2864 }
2865
2866 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2867                   struct extent_buffer *buf, int full_backref, int for_cow)
2868 {
2869         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2870 }
2871
2872 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2873                                  struct btrfs_root *root,
2874                                  struct btrfs_path *path,
2875                                  struct btrfs_block_group_cache *cache)
2876 {
2877         int ret;
2878         struct btrfs_root *extent_root = root->fs_info->extent_root;
2879         unsigned long bi;
2880         struct extent_buffer *leaf;
2881
2882         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2883         if (ret < 0)
2884                 goto fail;
2885         BUG_ON(ret); /* Corruption */
2886
2887         leaf = path->nodes[0];
2888         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2889         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2890         btrfs_mark_buffer_dirty(leaf);
2891         btrfs_release_path(path);
2892 fail:
2893         if (ret) {
2894                 btrfs_abort_transaction(trans, root, ret);
2895                 return ret;
2896         }
2897         return 0;
2898
2899 }
2900
2901 static struct btrfs_block_group_cache *
2902 next_block_group(struct btrfs_root *root,
2903                  struct btrfs_block_group_cache *cache)
2904 {
2905         struct rb_node *node;
2906         spin_lock(&root->fs_info->block_group_cache_lock);
2907         node = rb_next(&cache->cache_node);
2908         btrfs_put_block_group(cache);
2909         if (node) {
2910                 cache = rb_entry(node, struct btrfs_block_group_cache,
2911                                  cache_node);
2912                 btrfs_get_block_group(cache);
2913         } else
2914                 cache = NULL;
2915         spin_unlock(&root->fs_info->block_group_cache_lock);
2916         return cache;
2917 }
2918
2919 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2920                             struct btrfs_trans_handle *trans,
2921                             struct btrfs_path *path)
2922 {
2923         struct btrfs_root *root = block_group->fs_info->tree_root;
2924         struct inode *inode = NULL;
2925         u64 alloc_hint = 0;
2926         int dcs = BTRFS_DC_ERROR;
2927         int num_pages = 0;
2928         int retries = 0;
2929         int ret = 0;
2930
2931         /*
2932          * If this block group is smaller than 100 megs don't bother caching the
2933          * block group.
2934          */
2935         if (block_group->key.offset < (100 * 1024 * 1024)) {
2936                 spin_lock(&block_group->lock);
2937                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2938                 spin_unlock(&block_group->lock);
2939                 return 0;
2940         }
2941
2942 again:
2943         inode = lookup_free_space_inode(root, block_group, path);
2944         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2945                 ret = PTR_ERR(inode);
2946                 btrfs_release_path(path);
2947                 goto out;
2948         }
2949
2950         if (IS_ERR(inode)) {
2951                 BUG_ON(retries);
2952                 retries++;
2953
2954                 if (block_group->ro)
2955                         goto out_free;
2956
2957                 ret = create_free_space_inode(root, trans, block_group, path);
2958                 if (ret)
2959                         goto out_free;
2960                 goto again;
2961         }
2962
2963         /* We've already setup this transaction, go ahead and exit */
2964         if (block_group->cache_generation == trans->transid &&
2965             i_size_read(inode)) {
2966                 dcs = BTRFS_DC_SETUP;
2967                 goto out_put;
2968         }
2969
2970         /*
2971          * We want to set the generation to 0, that way if anything goes wrong
2972          * from here on out we know not to trust this cache when we load up next
2973          * time.
2974          */
2975         BTRFS_I(inode)->generation = 0;
2976         ret = btrfs_update_inode(trans, root, inode);
2977         WARN_ON(ret);
2978
2979         if (i_size_read(inode) > 0) {
2980                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2981                                                       inode);
2982                 if (ret)
2983                         goto out_put;
2984         }
2985
2986         spin_lock(&block_group->lock);
2987         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2988             !btrfs_test_opt(root, SPACE_CACHE)) {
2989                 /*
2990                  * don't bother trying to write stuff out _if_
2991                  * a) we're not cached,
2992                  * b) we're with nospace_cache mount option.
2993                  */
2994                 dcs = BTRFS_DC_WRITTEN;
2995                 spin_unlock(&block_group->lock);
2996                 goto out_put;
2997         }
2998         spin_unlock(&block_group->lock);
2999
3000         /*
3001          * Try to preallocate enough space based on how big the block group is.
3002          * Keep in mind this has to include any pinned space which could end up
3003          * taking up quite a bit since it's not folded into the other space
3004          * cache.
3005          */
3006         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3007         if (!num_pages)
3008                 num_pages = 1;
3009
3010         num_pages *= 16;
3011         num_pages *= PAGE_CACHE_SIZE;
3012
3013         ret = btrfs_check_data_free_space(inode, num_pages);
3014         if (ret)
3015                 goto out_put;
3016
3017         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3018                                               num_pages, num_pages,
3019                                               &alloc_hint);
3020         if (!ret)
3021                 dcs = BTRFS_DC_SETUP;
3022         btrfs_free_reserved_data_space(inode, num_pages);
3023
3024 out_put:
3025         iput(inode);
3026 out_free:
3027         btrfs_release_path(path);
3028 out:
3029         spin_lock(&block_group->lock);
3030         if (!ret && dcs == BTRFS_DC_SETUP)
3031                 block_group->cache_generation = trans->transid;
3032         block_group->disk_cache_state = dcs;
3033         spin_unlock(&block_group->lock);
3034
3035         return ret;
3036 }
3037
3038 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3039                                    struct btrfs_root *root)
3040 {
3041         struct btrfs_block_group_cache *cache;
3042         int err = 0;
3043         struct btrfs_path *path;
3044         u64 last = 0;
3045
3046         path = btrfs_alloc_path();
3047         if (!path)
3048                 return -ENOMEM;
3049
3050 again:
3051         while (1) {
3052                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3053                 while (cache) {
3054                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3055                                 break;
3056                         cache = next_block_group(root, cache);
3057                 }
3058                 if (!cache) {
3059                         if (last == 0)
3060                                 break;
3061                         last = 0;
3062                         continue;
3063                 }
3064                 err = cache_save_setup(cache, trans, path);
3065                 last = cache->key.objectid + cache->key.offset;
3066                 btrfs_put_block_group(cache);
3067         }
3068
3069         while (1) {
3070                 if (last == 0) {
3071                         err = btrfs_run_delayed_refs(trans, root,
3072                                                      (unsigned long)-1);
3073                         if (err) /* File system offline */
3074                                 goto out;
3075                 }
3076
3077                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3078                 while (cache) {
3079                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3080                                 btrfs_put_block_group(cache);
3081                                 goto again;
3082                         }
3083
3084                         if (cache->dirty)
3085                                 break;
3086                         cache = next_block_group(root, cache);
3087                 }
3088                 if (!cache) {
3089                         if (last == 0)
3090                                 break;
3091                         last = 0;
3092                         continue;
3093                 }
3094
3095                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3096                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3097                 cache->dirty = 0;
3098                 last = cache->key.objectid + cache->key.offset;
3099
3100                 err = write_one_cache_group(trans, root, path, cache);
3101                 if (err) /* File system offline */
3102                         goto out;
3103
3104                 btrfs_put_block_group(cache);
3105         }
3106
3107         while (1) {
3108                 /*
3109                  * I don't think this is needed since we're just marking our
3110                  * preallocated extent as written, but just in case it can't
3111                  * hurt.
3112                  */
3113                 if (last == 0) {
3114                         err = btrfs_run_delayed_refs(trans, root,
3115                                                      (unsigned long)-1);
3116                         if (err) /* File system offline */
3117                                 goto out;
3118                 }
3119
3120                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3121                 while (cache) {
3122                         /*
3123                          * Really this shouldn't happen, but it could if we
3124                          * couldn't write the entire preallocated extent and
3125                          * splitting the extent resulted in a new block.
3126                          */
3127                         if (cache->dirty) {
3128                                 btrfs_put_block_group(cache);
3129                                 goto again;
3130                         }
3131                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3132                                 break;
3133                         cache = next_block_group(root, cache);
3134                 }
3135                 if (!cache) {
3136                         if (last == 0)
3137                                 break;
3138                         last = 0;
3139                         continue;
3140                 }
3141
3142                 err = btrfs_write_out_cache(root, trans, cache, path);
3143
3144                 /*
3145                  * If we didn't have an error then the cache state is still
3146                  * NEED_WRITE, so we can set it to WRITTEN.
3147                  */
3148                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3149                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3150                 last = cache->key.objectid + cache->key.offset;
3151                 btrfs_put_block_group(cache);
3152         }
3153 out:
3154
3155         btrfs_free_path(path);
3156         return err;
3157 }
3158
3159 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3160 {
3161         struct btrfs_block_group_cache *block_group;
3162         int readonly = 0;
3163
3164         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3165         if (!block_group || block_group->ro)
3166                 readonly = 1;
3167         if (block_group)
3168                 btrfs_put_block_group(block_group);
3169         return readonly;
3170 }
3171
3172 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3173                              u64 total_bytes, u64 bytes_used,
3174                              struct btrfs_space_info **space_info)
3175 {
3176         struct btrfs_space_info *found;
3177         int i;
3178         int factor;
3179
3180         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3181                      BTRFS_BLOCK_GROUP_RAID10))
3182                 factor = 2;
3183         else
3184                 factor = 1;
3185
3186         found = __find_space_info(info, flags);
3187         if (found) {
3188                 spin_lock(&found->lock);
3189                 found->total_bytes += total_bytes;
3190                 found->disk_total += total_bytes * factor;
3191                 found->bytes_used += bytes_used;
3192                 found->disk_used += bytes_used * factor;
3193                 found->full = 0;
3194                 spin_unlock(&found->lock);
3195                 *space_info = found;
3196                 return 0;
3197         }
3198         found = kzalloc(sizeof(*found), GFP_NOFS);
3199         if (!found)
3200                 return -ENOMEM;
3201
3202         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3203                 INIT_LIST_HEAD(&found->block_groups[i]);
3204         init_rwsem(&found->groups_sem);
3205         spin_lock_init(&found->lock);
3206         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3207         found->total_bytes = total_bytes;
3208         found->disk_total = total_bytes * factor;
3209         found->bytes_used = bytes_used;
3210         found->disk_used = bytes_used * factor;
3211         found->bytes_pinned = 0;
3212         found->bytes_reserved = 0;
3213         found->bytes_readonly = 0;
3214         found->bytes_may_use = 0;
3215         found->full = 0;
3216         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3217         found->chunk_alloc = 0;
3218         found->flush = 0;
3219         init_waitqueue_head(&found->wait);
3220         *space_info = found;
3221         list_add_rcu(&found->list, &info->space_info);
3222         if (flags & BTRFS_BLOCK_GROUP_DATA)
3223                 info->data_sinfo = found;
3224         return 0;
3225 }
3226
3227 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3228 {
3229         u64 extra_flags = chunk_to_extended(flags) &
3230                                 BTRFS_EXTENDED_PROFILE_MASK;
3231
3232         if (flags & BTRFS_BLOCK_GROUP_DATA)
3233                 fs_info->avail_data_alloc_bits |= extra_flags;
3234         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3235                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3236         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3237                 fs_info->avail_system_alloc_bits |= extra_flags;
3238 }
3239
3240 /*
3241  * returns target flags in extended format or 0 if restripe for this
3242  * chunk_type is not in progress
3243  *
3244  * should be called with either volume_mutex or balance_lock held
3245  */
3246 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3247 {
3248         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3249         u64 target = 0;
3250
3251         if (!bctl)
3252                 return 0;
3253
3254         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3255             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3256                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3257         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3258                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3259                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3260         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3261                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3262                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3263         }
3264
3265         return target;
3266 }
3267
3268 /*
3269  * @flags: available profiles in extended format (see ctree.h)
3270  *
3271  * Returns reduced profile in chunk format.  If profile changing is in
3272  * progress (either running or paused) picks the target profile (if it's
3273  * already available), otherwise falls back to plain reducing.
3274  */
3275 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3276 {
3277         /*
3278          * we add in the count of missing devices because we want
3279          * to make sure that any RAID levels on a degraded FS
3280          * continue to be honored.
3281          */
3282         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3283                 root->fs_info->fs_devices->missing_devices;
3284         u64 target;
3285
3286         /*
3287          * see if restripe for this chunk_type is in progress, if so
3288          * try to reduce to the target profile
3289          */
3290         spin_lock(&root->fs_info->balance_lock);
3291         target = get_restripe_target(root->fs_info, flags);
3292         if (target) {
3293                 /* pick target profile only if it's already available */
3294                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3295                         spin_unlock(&root->fs_info->balance_lock);
3296                         return extended_to_chunk(target);
3297                 }
3298         }
3299         spin_unlock(&root->fs_info->balance_lock);
3300
3301         if (num_devices == 1)
3302                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3303         if (num_devices < 4)
3304                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3305
3306         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3307             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3308                       BTRFS_BLOCK_GROUP_RAID10))) {
3309                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3310         }
3311
3312         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3313             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3314                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3315         }
3316
3317         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3318             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3319              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3320              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3321                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3322         }
3323
3324         return extended_to_chunk(flags);
3325 }
3326
3327 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3328 {
3329         if (flags & BTRFS_BLOCK_GROUP_DATA)
3330                 flags |= root->fs_info->avail_data_alloc_bits;
3331         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3332                 flags |= root->fs_info->avail_system_alloc_bits;
3333         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3334                 flags |= root->fs_info->avail_metadata_alloc_bits;
3335
3336         return btrfs_reduce_alloc_profile(root, flags);
3337 }
3338
3339 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3340 {
3341         u64 flags;
3342
3343         if (data)
3344                 flags = BTRFS_BLOCK_GROUP_DATA;
3345         else if (root == root->fs_info->chunk_root)
3346                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3347         else
3348                 flags = BTRFS_BLOCK_GROUP_METADATA;
3349
3350         return get_alloc_profile(root, flags);
3351 }
3352
3353 /*
3354  * This will check the space that the inode allocates from to make sure we have
3355  * enough space for bytes.
3356  */
3357 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3358 {
3359         struct btrfs_space_info *data_sinfo;
3360         struct btrfs_root *root = BTRFS_I(inode)->root;
3361         struct btrfs_fs_info *fs_info = root->fs_info;
3362         u64 used;
3363         int ret = 0, committed = 0, alloc_chunk = 1;
3364
3365         /* make sure bytes are sectorsize aligned */
3366         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3367
3368         if (root == root->fs_info->tree_root ||
3369             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3370                 alloc_chunk = 0;
3371                 committed = 1;
3372         }
3373
3374         data_sinfo = fs_info->data_sinfo;
3375         if (!data_sinfo)
3376                 goto alloc;
3377
3378 again:
3379         /* make sure we have enough space to handle the data first */
3380         spin_lock(&data_sinfo->lock);
3381         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3382                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3383                 data_sinfo->bytes_may_use;
3384
3385         if (used + bytes > data_sinfo->total_bytes) {
3386                 struct btrfs_trans_handle *trans;
3387
3388                 /*
3389                  * if we don't have enough free bytes in this space then we need
3390                  * to alloc a new chunk.
3391                  */
3392                 if (!data_sinfo->full && alloc_chunk) {
3393                         u64 alloc_target;
3394
3395                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3396                         spin_unlock(&data_sinfo->lock);
3397 alloc:
3398                         alloc_target = btrfs_get_alloc_profile(root, 1);
3399                         trans = btrfs_join_transaction(root);
3400                         if (IS_ERR(trans))
3401                                 return PTR_ERR(trans);
3402
3403                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3404                                              alloc_target,
3405                                              CHUNK_ALLOC_NO_FORCE);
3406                         btrfs_end_transaction(trans, root);
3407                         if (ret < 0) {
3408                                 if (ret != -ENOSPC)
3409                                         return ret;
3410                                 else
3411                                         goto commit_trans;
3412                         }
3413
3414                         if (!data_sinfo)
3415                                 data_sinfo = fs_info->data_sinfo;
3416
3417                         goto again;
3418                 }
3419
3420                 /*
3421                  * If we have less pinned bytes than we want to allocate then
3422                  * don't bother committing the transaction, it won't help us.
3423                  */
3424                 if (data_sinfo->bytes_pinned < bytes)
3425                         committed = 1;
3426                 spin_unlock(&data_sinfo->lock);
3427
3428                 /* commit the current transaction and try again */
3429 commit_trans:
3430                 if (!committed &&
3431                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3432                         committed = 1;
3433                         trans = btrfs_join_transaction(root);
3434                         if (IS_ERR(trans))
3435                                 return PTR_ERR(trans);
3436                         ret = btrfs_commit_transaction(trans, root);
3437                         if (ret)
3438                                 return ret;
3439                         goto again;
3440                 }
3441
3442                 return -ENOSPC;
3443         }
3444         data_sinfo->bytes_may_use += bytes;
3445         trace_btrfs_space_reservation(root->fs_info, "space_info",
3446                                       data_sinfo->flags, bytes, 1);
3447         spin_unlock(&data_sinfo->lock);
3448
3449         return 0;
3450 }
3451
3452 /*
3453  * Called if we need to clear a data reservation for this inode.
3454  */
3455 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3456 {
3457         struct btrfs_root *root = BTRFS_I(inode)->root;
3458         struct btrfs_space_info *data_sinfo;
3459
3460         /* make sure bytes are sectorsize aligned */
3461         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3462
3463         data_sinfo = root->fs_info->data_sinfo;
3464         spin_lock(&data_sinfo->lock);
3465         data_sinfo->bytes_may_use -= bytes;
3466         trace_btrfs_space_reservation(root->fs_info, "space_info",
3467                                       data_sinfo->flags, bytes, 0);
3468         spin_unlock(&data_sinfo->lock);
3469 }
3470
3471 static void force_metadata_allocation(struct btrfs_fs_info *info)
3472 {
3473         struct list_head *head = &info->space_info;
3474         struct btrfs_space_info *found;
3475
3476         rcu_read_lock();
3477         list_for_each_entry_rcu(found, head, list) {
3478                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3479                         found->force_alloc = CHUNK_ALLOC_FORCE;
3480         }
3481         rcu_read_unlock();
3482 }
3483
3484 static int should_alloc_chunk(struct btrfs_root *root,
3485                               struct btrfs_space_info *sinfo, int force)
3486 {
3487         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3488         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3489         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3490         u64 thresh;
3491
3492         if (force == CHUNK_ALLOC_FORCE)
3493                 return 1;
3494
3495         /*
3496          * We need to take into account the global rsv because for all intents
3497          * and purposes it's used space.  Don't worry about locking the
3498          * global_rsv, it doesn't change except when the transaction commits.
3499          */
3500         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3501                 num_allocated += global_rsv->size;
3502
3503         /*
3504          * in limited mode, we want to have some free space up to
3505          * about 1% of the FS size.
3506          */
3507         if (force == CHUNK_ALLOC_LIMITED) {
3508                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3509                 thresh = max_t(u64, 64 * 1024 * 1024,
3510                                div_factor_fine(thresh, 1));
3511
3512                 if (num_bytes - num_allocated < thresh)
3513                         return 1;
3514         }
3515
3516         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3517                 return 0;
3518         return 1;
3519 }
3520
3521 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3522 {
3523         u64 num_dev;
3524
3525         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3526             type & BTRFS_BLOCK_GROUP_RAID0)
3527                 num_dev = root->fs_info->fs_devices->rw_devices;
3528         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3529                 num_dev = 2;
3530         else
3531                 num_dev = 1;    /* DUP or single */
3532
3533         /* metadata for updaing devices and chunk tree */
3534         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3535 }
3536
3537 static void check_system_chunk(struct btrfs_trans_handle *trans,
3538                                struct btrfs_root *root, u64 type)
3539 {
3540         struct btrfs_space_info *info;
3541         u64 left;
3542         u64 thresh;
3543
3544         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3545         spin_lock(&info->lock);
3546         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3547                 info->bytes_reserved - info->bytes_readonly;
3548         spin_unlock(&info->lock);
3549
3550         thresh = get_system_chunk_thresh(root, type);
3551         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3552                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3553                        left, thresh, type);
3554                 dump_space_info(info, 0, 0);
3555         }
3556
3557         if (left < thresh) {
3558                 u64 flags;
3559
3560                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3561                 btrfs_alloc_chunk(trans, root, flags);
3562         }
3563 }
3564
3565 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3566                           struct btrfs_root *extent_root, u64 flags, int force)
3567 {
3568         struct btrfs_space_info *space_info;
3569         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3570         int wait_for_alloc = 0;
3571         int ret = 0;
3572
3573         space_info = __find_space_info(extent_root->fs_info, flags);
3574         if (!space_info) {
3575                 ret = update_space_info(extent_root->fs_info, flags,
3576                                         0, 0, &space_info);
3577                 BUG_ON(ret); /* -ENOMEM */
3578         }
3579         BUG_ON(!space_info); /* Logic error */
3580
3581 again:
3582         spin_lock(&space_info->lock);
3583         if (force < space_info->force_alloc)
3584                 force = space_info->force_alloc;
3585         if (space_info->full) {
3586                 spin_unlock(&space_info->lock);
3587                 return 0;
3588         }
3589
3590         if (!should_alloc_chunk(extent_root, space_info, force)) {
3591                 spin_unlock(&space_info->lock);
3592                 return 0;
3593         } else if (space_info->chunk_alloc) {
3594                 wait_for_alloc = 1;
3595         } else {
3596                 space_info->chunk_alloc = 1;
3597         }
3598
3599         spin_unlock(&space_info->lock);
3600
3601         mutex_lock(&fs_info->chunk_mutex);
3602
3603         /*
3604          * The chunk_mutex is held throughout the entirety of a chunk
3605          * allocation, so once we've acquired the chunk_mutex we know that the
3606          * other guy is done and we need to recheck and see if we should
3607          * allocate.
3608          */
3609         if (wait_for_alloc) {
3610                 mutex_unlock(&fs_info->chunk_mutex);
3611                 wait_for_alloc = 0;
3612                 goto again;
3613         }
3614
3615         /*
3616          * If we have mixed data/metadata chunks we want to make sure we keep
3617          * allocating mixed chunks instead of individual chunks.
3618          */
3619         if (btrfs_mixed_space_info(space_info))
3620                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3621
3622         /*
3623          * if we're doing a data chunk, go ahead and make sure that
3624          * we keep a reasonable number of metadata chunks allocated in the
3625          * FS as well.
3626          */
3627         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3628                 fs_info->data_chunk_allocations++;
3629                 if (!(fs_info->data_chunk_allocations %
3630                       fs_info->metadata_ratio))
3631                         force_metadata_allocation(fs_info);
3632         }
3633
3634         /*
3635          * Check if we have enough space in SYSTEM chunk because we may need
3636          * to update devices.
3637          */
3638         check_system_chunk(trans, extent_root, flags);
3639
3640         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3641         if (ret < 0 && ret != -ENOSPC)
3642                 goto out;
3643
3644         spin_lock(&space_info->lock);
3645         if (ret)
3646                 space_info->full = 1;
3647         else
3648                 ret = 1;
3649
3650         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3651         space_info->chunk_alloc = 0;
3652         spin_unlock(&space_info->lock);
3653 out:
3654         mutex_unlock(&fs_info->chunk_mutex);
3655         return ret;
3656 }
3657
3658 static int can_overcommit(struct btrfs_root *root,
3659                           struct btrfs_space_info *space_info, u64 bytes,
3660                           enum btrfs_reserve_flush_enum flush)
3661 {
3662         u64 profile = btrfs_get_alloc_profile(root, 0);
3663         u64 avail;
3664         u64 used;
3665
3666         used = space_info->bytes_used + space_info->bytes_reserved +
3667                 space_info->bytes_pinned + space_info->bytes_readonly +
3668                 space_info->bytes_may_use;
3669
3670         spin_lock(&root->fs_info->free_chunk_lock);
3671         avail = root->fs_info->free_chunk_space;
3672         spin_unlock(&root->fs_info->free_chunk_lock);
3673
3674         /*
3675          * If we have dup, raid1 or raid10 then only half of the free
3676          * space is actually useable.
3677          */
3678         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3679                        BTRFS_BLOCK_GROUP_RAID1 |
3680                        BTRFS_BLOCK_GROUP_RAID10))
3681                 avail >>= 1;
3682
3683         /*
3684          * If we aren't flushing all things, let us overcommit up to
3685          * 1/2th of the space. If we can flush, don't let us overcommit
3686          * too much, let it overcommit up to 1/8 of the space.
3687          */
3688         if (flush == BTRFS_RESERVE_FLUSH_ALL)
3689                 avail >>= 3;
3690         else
3691                 avail >>= 1;
3692
3693         if (used + bytes < space_info->total_bytes + avail)
3694                 return 1;
3695         return 0;
3696 }
3697
3698 static inline int writeback_inodes_sb_nr_if_idle_safe(struct super_block *sb,
3699                                                       unsigned long nr_pages,
3700                                                       enum wb_reason reason)
3701 {
3702         /* the flusher is dealing with the dirty inodes now. */
3703         if (writeback_in_progress(sb->s_bdi))
3704                 return 1;
3705
3706         if (down_read_trylock(&sb->s_umount)) {
3707                 writeback_inodes_sb_nr(sb, nr_pages, reason);
3708                 up_read(&sb->s_umount);
3709                 return 1;
3710         }
3711
3712         return 0;
3713 }
3714
3715 void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3716                                   unsigned long nr_pages)
3717 {
3718         struct super_block *sb = root->fs_info->sb;
3719         int started;
3720
3721         /* If we can not start writeback, just sync all the delalloc file. */
3722         started = writeback_inodes_sb_nr_if_idle_safe(sb, nr_pages,
3723                                                       WB_REASON_FS_FREE_SPACE);
3724         if (!started) {
3725                 /*
3726                  * We needn't worry the filesystem going from r/w to r/o though
3727                  * we don't acquire ->s_umount mutex, because the filesystem
3728                  * should guarantee the delalloc inodes list be empty after
3729                  * the filesystem is readonly(all dirty pages are written to
3730                  * the disk).
3731                  */
3732                 btrfs_start_delalloc_inodes(root, 0);
3733                 btrfs_wait_ordered_extents(root, 0);
3734         }
3735 }
3736
3737 /*
3738  * shrink metadata reservation for delalloc
3739  */
3740 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3741                             bool wait_ordered)
3742 {
3743         struct btrfs_block_rsv *block_rsv;
3744         struct btrfs_space_info *space_info;
3745         struct btrfs_trans_handle *trans;
3746         u64 delalloc_bytes;
3747         u64 max_reclaim;
3748         long time_left;
3749         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3750         int loops = 0;
3751         enum btrfs_reserve_flush_enum flush;
3752
3753         trans = (struct btrfs_trans_handle *)current->journal_info;
3754         block_rsv = &root->fs_info->delalloc_block_rsv;
3755         space_info = block_rsv->space_info;
3756
3757         smp_mb();
3758         delalloc_bytes = root->fs_info->delalloc_bytes;
3759         if (delalloc_bytes == 0) {
3760                 if (trans)
3761                         return;
3762                 btrfs_wait_ordered_extents(root, 0);
3763                 return;
3764         }
3765
3766         while (delalloc_bytes && loops < 3) {
3767                 max_reclaim = min(delalloc_bytes, to_reclaim);
3768                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3769                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
3770                 /*
3771                  * We need to wait for the async pages to actually start before
3772                  * we do anything.
3773                  */
3774                 wait_event(root->fs_info->async_submit_wait,
3775                            !atomic_read(&root->fs_info->async_delalloc_pages));
3776
3777                 if (!trans)
3778                         flush = BTRFS_RESERVE_FLUSH_ALL;
3779                 else
3780                         flush = BTRFS_RESERVE_NO_FLUSH;
3781                 spin_lock(&space_info->lock);
3782                 if (can_overcommit(root, space_info, orig, flush)) {
3783                         spin_unlock(&space_info->lock);
3784                         break;
3785                 }
3786                 spin_unlock(&space_info->lock);
3787
3788                 loops++;
3789                 if (wait_ordered && !trans) {
3790                         btrfs_wait_ordered_extents(root, 0);
3791                 } else {
3792                         time_left = schedule_timeout_killable(1);
3793                         if (time_left)
3794                                 break;
3795                 }
3796                 smp_mb();
3797                 delalloc_bytes = root->fs_info->delalloc_bytes;
3798         }
3799 }
3800
3801 /**
3802  * maybe_commit_transaction - possibly commit the transaction if its ok to
3803  * @root - the root we're allocating for
3804  * @bytes - the number of bytes we want to reserve
3805  * @force - force the commit
3806  *
3807  * This will check to make sure that committing the transaction will actually
3808  * get us somewhere and then commit the transaction if it does.  Otherwise it
3809  * will return -ENOSPC.
3810  */
3811 static int may_commit_transaction(struct btrfs_root *root,
3812                                   struct btrfs_space_info *space_info,
3813                                   u64 bytes, int force)
3814 {
3815         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3816         struct btrfs_trans_handle *trans;
3817
3818         trans = (struct btrfs_trans_handle *)current->journal_info;
3819         if (trans)
3820                 return -EAGAIN;
3821
3822         if (force)
3823                 goto commit;
3824
3825         /* See if there is enough pinned space to make this reservation */
3826         spin_lock(&space_info->lock);
3827         if (space_info->bytes_pinned >= bytes) {
3828                 spin_unlock(&space_info->lock);
3829                 goto commit;
3830         }
3831         spin_unlock(&space_info->lock);
3832
3833         /*
3834          * See if there is some space in the delayed insertion reservation for
3835          * this reservation.
3836          */
3837         if (space_info != delayed_rsv->space_info)
3838                 return -ENOSPC;
3839
3840         spin_lock(&space_info->lock);
3841         spin_lock(&delayed_rsv->lock);
3842         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3843                 spin_unlock(&delayed_rsv->lock);
3844                 spin_unlock(&space_info->lock);
3845                 return -ENOSPC;
3846         }
3847         spin_unlock(&delayed_rsv->lock);
3848         spin_unlock(&space_info->lock);
3849
3850 commit:
3851         trans = btrfs_join_transaction(root);
3852         if (IS_ERR(trans))
3853                 return -ENOSPC;
3854
3855         return btrfs_commit_transaction(trans, root);
3856 }
3857
3858 enum flush_state {
3859         FLUSH_DELAYED_ITEMS_NR  =       1,
3860         FLUSH_DELAYED_ITEMS     =       2,
3861         FLUSH_DELALLOC          =       3,
3862         FLUSH_DELALLOC_WAIT     =       4,
3863         ALLOC_CHUNK             =       5,
3864         COMMIT_TRANS            =       6,
3865 };
3866
3867 static int flush_space(struct btrfs_root *root,
3868                        struct btrfs_space_info *space_info, u64 num_bytes,
3869                        u64 orig_bytes, int state)
3870 {
3871         struct btrfs_trans_handle *trans;
3872         int nr;
3873         int ret = 0;
3874
3875         switch (state) {
3876         case FLUSH_DELAYED_ITEMS_NR:
3877         case FLUSH_DELAYED_ITEMS:
3878                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3879                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3880
3881                         nr = (int)div64_u64(num_bytes, bytes);
3882                         if (!nr)
3883                                 nr = 1;
3884                         nr *= 2;
3885                 } else {
3886                         nr = -1;
3887                 }
3888                 trans = btrfs_join_transaction(root);
3889                 if (IS_ERR(trans)) {
3890                         ret = PTR_ERR(trans);
3891                         break;
3892                 }
3893                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3894                 btrfs_end_transaction(trans, root);
3895                 break;
3896         case FLUSH_DELALLOC:
3897         case FLUSH_DELALLOC_WAIT:
3898                 shrink_delalloc(root, num_bytes, orig_bytes,
3899                                 state == FLUSH_DELALLOC_WAIT);
3900                 break;
3901         case ALLOC_CHUNK:
3902                 trans = btrfs_join_transaction(root);
3903                 if (IS_ERR(trans)) {
3904                         ret = PTR_ERR(trans);
3905                         break;
3906                 }
3907                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3908                                      btrfs_get_alloc_profile(root, 0),
3909                                      CHUNK_ALLOC_NO_FORCE);
3910                 btrfs_end_transaction(trans, root);
3911                 if (ret == -ENOSPC)
3912                         ret = 0;
3913                 break;
3914         case COMMIT_TRANS:
3915                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3916                 break;
3917         default:
3918                 ret = -ENOSPC;
3919                 break;
3920         }
3921
3922         return ret;
3923 }
3924 /**
3925  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3926  * @root - the root we're allocating for
3927  * @block_rsv - the block_rsv we're allocating for
3928  * @orig_bytes - the number of bytes we want
3929  * @flush - wether or not we can flush to make our reservation
3930  *
3931  * This will reserve orgi_bytes number of bytes from the space info associated
3932  * with the block_rsv.  If there is not enough space it will make an attempt to
3933  * flush out space to make room.  It will do this by flushing delalloc if
3934  * possible or committing the transaction.  If flush is 0 then no attempts to
3935  * regain reservations will be made and this will fail if there is not enough
3936  * space already.
3937  */
3938 static int reserve_metadata_bytes(struct btrfs_root *root,
3939                                   struct btrfs_block_rsv *block_rsv,
3940                                   u64 orig_bytes,
3941                                   enum btrfs_reserve_flush_enum flush)
3942 {
3943         struct btrfs_space_info *space_info = block_rsv->space_info;
3944         u64 used;
3945         u64 num_bytes = orig_bytes;
3946         int flush_state = FLUSH_DELAYED_ITEMS_NR;
3947         int ret = 0;
3948         bool flushing = false;
3949
3950 again:
3951         ret = 0;
3952         spin_lock(&space_info->lock);
3953         /*
3954          * We only want to wait if somebody other than us is flushing and we
3955          * are actually allowed to flush all things.
3956          */
3957         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
3958                space_info->flush) {
3959                 spin_unlock(&space_info->lock);
3960                 /*
3961                  * If we have a trans handle we can't wait because the flusher
3962                  * may have to commit the transaction, which would mean we would
3963                  * deadlock since we are waiting for the flusher to finish, but
3964                  * hold the current transaction open.
3965                  */
3966                 if (current->journal_info)
3967                         return -EAGAIN;
3968                 ret = wait_event_killable(space_info->wait, !space_info->flush);
3969                 /* Must have been killed, return */
3970                 if (ret)
3971                         return -EINTR;
3972
3973                 spin_lock(&space_info->lock);
3974         }
3975
3976         ret = -ENOSPC;
3977         used = space_info->bytes_used + space_info->bytes_reserved +
3978                 space_info->bytes_pinned + space_info->bytes_readonly +
3979                 space_info->bytes_may_use;
3980
3981         /*
3982          * The idea here is that we've not already over-reserved the block group
3983          * then we can go ahead and save our reservation first and then start
3984          * flushing if we need to.  Otherwise if we've already overcommitted
3985          * lets start flushing stuff first and then come back and try to make
3986          * our reservation.
3987          */
3988         if (used <= space_info->total_bytes) {
3989                 if (used + orig_bytes <= space_info->total_bytes) {
3990                         space_info->bytes_may_use += orig_bytes;
3991                         trace_btrfs_space_reservation(root->fs_info,
3992                                 "space_info", space_info->flags, orig_bytes, 1);
3993                         ret = 0;
3994                 } else {
3995                         /*
3996                          * Ok set num_bytes to orig_bytes since we aren't
3997                          * overocmmitted, this way we only try and reclaim what
3998                          * we need.
3999                          */
4000                         num_bytes = orig_bytes;
4001                 }
4002         } else {
4003                 /*
4004                  * Ok we're over committed, set num_bytes to the overcommitted
4005                  * amount plus the amount of bytes that we need for this
4006                  * reservation.
4007                  */
4008                 num_bytes = used - space_info->total_bytes +
4009                         (orig_bytes * 2);
4010         }
4011
4012         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4013                 space_info->bytes_may_use += orig_bytes;
4014                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4015                                               space_info->flags, orig_bytes,
4016                                               1);
4017                 ret = 0;
4018         }
4019
4020         /*
4021          * Couldn't make our reservation, save our place so while we're trying
4022          * to reclaim space we can actually use it instead of somebody else
4023          * stealing it from us.
4024          *
4025          * We make the other tasks wait for the flush only when we can flush
4026          * all things.
4027          */
4028         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4029                 flushing = true;
4030                 space_info->flush = 1;
4031         }
4032
4033         spin_unlock(&space_info->lock);
4034
4035         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4036                 goto out;
4037
4038         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4039                           flush_state);
4040         flush_state++;
4041
4042         /*
4043          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4044          * would happen. So skip delalloc flush.
4045          */
4046         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4047             (flush_state == FLUSH_DELALLOC ||
4048              flush_state == FLUSH_DELALLOC_WAIT))
4049                 flush_state = ALLOC_CHUNK;
4050
4051         if (!ret)
4052                 goto again;
4053         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4054                  flush_state < COMMIT_TRANS)
4055                 goto again;
4056         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4057                  flush_state <= COMMIT_TRANS)
4058                 goto again;
4059
4060 out:
4061         if (flushing) {
4062                 spin_lock(&space_info->lock);
4063                 space_info->flush = 0;
4064                 wake_up_all(&space_info->wait);
4065                 spin_unlock(&space_info->lock);
4066         }
4067         return ret;
4068 }
4069
4070 static struct btrfs_block_rsv *get_block_rsv(
4071                                         const struct btrfs_trans_handle *trans,
4072                                         const struct btrfs_root *root)
4073 {
4074         struct btrfs_block_rsv *block_rsv = NULL;
4075
4076         if (root->ref_cows)
4077                 block_rsv = trans->block_rsv;
4078
4079         if (root == root->fs_info->csum_root && trans->adding_csums)
4080                 block_rsv = trans->block_rsv;
4081
4082         if (!block_rsv)
4083                 block_rsv = root->block_rsv;
4084
4085         if (!block_rsv)
4086                 block_rsv = &root->fs_info->empty_block_rsv;
4087
4088         return block_rsv;
4089 }
4090
4091 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4092                                u64 num_bytes)
4093 {
4094         int ret = -ENOSPC;
4095         spin_lock(&block_rsv->lock);
4096         if (block_rsv->reserved >= num_bytes) {
4097                 block_rsv->reserved -= num_bytes;
4098                 if (block_rsv->reserved < block_rsv->size)
4099                         block_rsv->full = 0;
4100                 ret = 0;
4101         }
4102         spin_unlock(&block_rsv->lock);
4103         return ret;
4104 }
4105
4106 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4107                                 u64 num_bytes, int update_size)
4108 {
4109         spin_lock(&block_rsv->lock);
4110         block_rsv->reserved += num_bytes;
4111         if (update_size)
4112                 block_rsv->size += num_bytes;
4113         else if (block_rsv->reserved >= block_rsv->size)
4114                 block_rsv->full = 1;
4115         spin_unlock(&block_rsv->lock);
4116 }
4117
4118 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4119                                     struct btrfs_block_rsv *block_rsv,
4120                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4121 {
4122         struct btrfs_space_info *space_info = block_rsv->space_info;
4123
4124         spin_lock(&block_rsv->lock);
4125         if (num_bytes == (u64)-1)
4126                 num_bytes = block_rsv->size;
4127         block_rsv->size -= num_bytes;
4128         if (block_rsv->reserved >= block_rsv->size) {
4129                 num_bytes = block_rsv->reserved - block_rsv->size;
4130                 block_rsv->reserved = block_rsv->size;
4131                 block_rsv->full = 1;
4132         } else {
4133                 num_bytes = 0;
4134         }
4135         spin_unlock(&block_rsv->lock);
4136
4137         if (num_bytes > 0) {
4138                 if (dest) {
4139                         spin_lock(&dest->lock);
4140                         if (!dest->full) {
4141                                 u64 bytes_to_add;
4142
4143                                 bytes_to_add = dest->size - dest->reserved;
4144                                 bytes_to_add = min(num_bytes, bytes_to_add);
4145                                 dest->reserved += bytes_to_add;
4146                                 if (dest->reserved >= dest->size)
4147                                         dest->full = 1;
4148                                 num_bytes -= bytes_to_add;
4149                         }
4150                         spin_unlock(&dest->lock);
4151                 }
4152                 if (num_bytes) {
4153                         spin_lock(&space_info->lock);
4154                         space_info->bytes_may_use -= num_bytes;
4155                         trace_btrfs_space_reservation(fs_info, "space_info",
4156                                         space_info->flags, num_bytes, 0);
4157                         space_info->reservation_progress++;
4158                         spin_unlock(&space_info->lock);
4159                 }
4160         }
4161 }
4162
4163 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4164                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4165 {
4166         int ret;
4167
4168         ret = block_rsv_use_bytes(src, num_bytes);
4169         if (ret)
4170                 return ret;
4171
4172         block_rsv_add_bytes(dst, num_bytes, 1);
4173         return 0;
4174 }
4175
4176 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4177 {
4178         memset(rsv, 0, sizeof(*rsv));
4179         spin_lock_init(&rsv->lock);
4180         rsv->type = type;
4181 }
4182
4183 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4184                                               unsigned short type)
4185 {
4186         struct btrfs_block_rsv *block_rsv;
4187         struct btrfs_fs_info *fs_info = root->fs_info;
4188
4189         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4190         if (!block_rsv)
4191                 return NULL;
4192
4193         btrfs_init_block_rsv(block_rsv, type);
4194         block_rsv->space_info = __find_space_info(fs_info,
4195                                                   BTRFS_BLOCK_GROUP_METADATA);
4196         return block_rsv;
4197 }
4198
4199 void btrfs_free_block_rsv(struct btrfs_root *root,
4200                           struct btrfs_block_rsv *rsv)
4201 {
4202         if (!rsv)
4203                 return;
4204         btrfs_block_rsv_release(root, rsv, (u64)-1);
4205         kfree(rsv);
4206 }
4207
4208 int btrfs_block_rsv_add(struct btrfs_root *root,
4209                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4210                         enum btrfs_reserve_flush_enum flush)
4211 {
4212         int ret;
4213
4214         if (num_bytes == 0)
4215                 return 0;
4216
4217         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4218         if (!ret) {
4219                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4220                 return 0;
4221         }
4222
4223         return ret;
4224 }
4225
4226 int btrfs_block_rsv_check(struct btrfs_root *root,
4227                           struct btrfs_block_rsv *block_rsv, int min_factor)
4228 {
4229         u64 num_bytes = 0;
4230         int ret = -ENOSPC;
4231
4232         if (!block_rsv)
4233                 return 0;
4234
4235         spin_lock(&block_rsv->lock);
4236         num_bytes = div_factor(block_rsv->size, min_factor);
4237         if (block_rsv->reserved >= num_bytes)
4238                 ret = 0;
4239         spin_unlock(&block_rsv->lock);
4240
4241         return ret;
4242 }
4243
4244 int btrfs_block_rsv_refill(struct btrfs_root *root,
4245                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4246                            enum btrfs_reserve_flush_enum flush)
4247 {
4248         u64 num_bytes = 0;
4249         int ret = -ENOSPC;
4250
4251         if (!block_rsv)
4252                 return 0;
4253
4254         spin_lock(&block_rsv->lock);
4255         num_bytes = min_reserved;
4256         if (block_rsv->reserved >= num_bytes)
4257                 ret = 0;
4258         else
4259                 num_bytes -= block_rsv->reserved;
4260         spin_unlock(&block_rsv->lock);
4261
4262         if (!ret)
4263                 return 0;
4264
4265         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4266         if (!ret) {
4267                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4268                 return 0;
4269         }
4270
4271         return ret;
4272 }
4273
4274 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4275                             struct btrfs_block_rsv *dst_rsv,
4276                             u64 num_bytes)
4277 {
4278         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4279 }
4280
4281 void btrfs_block_rsv_release(struct btrfs_root *root,
4282                              struct btrfs_block_rsv *block_rsv,
4283                              u64 num_bytes)
4284 {
4285         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4286         if (global_rsv->full || global_rsv == block_rsv ||
4287             block_rsv->space_info != global_rsv->space_info)
4288                 global_rsv = NULL;
4289         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4290                                 num_bytes);
4291 }
4292
4293 /*
4294  * helper to calculate size of global block reservation.
4295  * the desired value is sum of space used by extent tree,
4296  * checksum tree and root tree
4297  */
4298 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4299 {
4300         struct btrfs_space_info *sinfo;
4301         u64 num_bytes;
4302         u64 meta_used;
4303         u64 data_used;
4304         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4305
4306         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4307         spin_lock(&sinfo->lock);
4308         data_used = sinfo->bytes_used;
4309         spin_unlock(&sinfo->lock);
4310
4311         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4312         spin_lock(&sinfo->lock);
4313         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4314                 data_used = 0;
4315         meta_used = sinfo->bytes_used;
4316         spin_unlock(&sinfo->lock);
4317
4318         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4319                     csum_size * 2;
4320         num_bytes += div64_u64(data_used + meta_used, 50);
4321
4322         if (num_bytes * 3 > meta_used)
4323                 num_bytes = div64_u64(meta_used, 3);
4324
4325         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4326 }
4327
4328 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4329 {
4330         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4331         struct btrfs_space_info *sinfo = block_rsv->space_info;
4332         u64 num_bytes;
4333
4334         num_bytes = calc_global_metadata_size(fs_info);
4335
4336         spin_lock(&sinfo->lock);
4337         spin_lock(&block_rsv->lock);
4338
4339         block_rsv->size = num_bytes;
4340
4341         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4342                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4343                     sinfo->bytes_may_use;
4344
4345         if (sinfo->total_bytes > num_bytes) {
4346                 num_bytes = sinfo->total_bytes - num_bytes;
4347                 block_rsv->reserved += num_bytes;
4348                 sinfo->bytes_may_use += num_bytes;
4349                 trace_btrfs_space_reservation(fs_info, "space_info",
4350                                       sinfo->flags, num_bytes, 1);
4351         }
4352
4353         if (block_rsv->reserved >= block_rsv->size) {
4354                 num_bytes = block_rsv->reserved - block_rsv->size;
4355                 sinfo->bytes_may_use -= num_bytes;
4356                 trace_btrfs_space_reservation(fs_info, "space_info",
4357                                       sinfo->flags, num_bytes, 0);
4358                 sinfo->reservation_progress++;
4359                 block_rsv->reserved = block_rsv->size;
4360                 block_rsv->full = 1;
4361         }
4362
4363         spin_unlock(&block_rsv->lock);
4364         spin_unlock(&sinfo->lock);
4365 }
4366
4367 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4368 {
4369         struct btrfs_space_info *space_info;
4370
4371         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4372         fs_info->chunk_block_rsv.space_info = space_info;
4373
4374         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4375         fs_info->global_block_rsv.space_info = space_info;
4376         fs_info->delalloc_block_rsv.space_info = space_info;
4377         fs_info->trans_block_rsv.space_info = space_info;
4378         fs_info->empty_block_rsv.space_info = space_info;
4379         fs_info->delayed_block_rsv.space_info = space_info;
4380
4381         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4382         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4383         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4384         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4385         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4386
4387         update_global_block_rsv(fs_info);
4388 }
4389
4390 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4391 {
4392         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4393                                 (u64)-1);
4394         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4395         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4396         WARN_ON(fs_info->trans_block_rsv.size > 0);
4397         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4398         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4399         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4400         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4401         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4402 }
4403
4404 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4405                                   struct btrfs_root *root)
4406 {
4407         if (!trans->block_rsv)
4408                 return;
4409
4410         if (!trans->bytes_reserved)
4411                 return;
4412
4413         trace_btrfs_space_reservation(root->fs_info, "transaction",
4414                                       trans->transid, trans->bytes_reserved, 0);
4415         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4416         trans->bytes_reserved = 0;
4417 }
4418
4419 /* Can only return 0 or -ENOSPC */
4420 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4421                                   struct inode *inode)
4422 {
4423         struct btrfs_root *root = BTRFS_I(inode)->root;
4424         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4425         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4426
4427         /*
4428          * We need to hold space in order to delete our orphan item once we've
4429          * added it, so this takes the reservation so we can release it later
4430          * when we are truly done with the orphan item.
4431          */
4432         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4433         trace_btrfs_space_reservation(root->fs_info, "orphan",
4434                                       btrfs_ino(inode), num_bytes, 1);
4435         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4436 }
4437
4438 void btrfs_orphan_release_metadata(struct inode *inode)
4439 {
4440         struct btrfs_root *root = BTRFS_I(inode)->root;
4441         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4442         trace_btrfs_space_reservation(root->fs_info, "orphan",
4443                                       btrfs_ino(inode), num_bytes, 0);
4444         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4445 }
4446
4447 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4448                                 struct btrfs_pending_snapshot *pending)
4449 {
4450         struct btrfs_root *root = pending->root;
4451         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4452         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4453         /*
4454          * two for root back/forward refs, two for directory entries,
4455          * one for root of the snapshot and one for parent inode.
4456          */
4457         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 6);
4458         dst_rsv->space_info = src_rsv->space_info;
4459         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4460 }
4461
4462 /**
4463  * drop_outstanding_extent - drop an outstanding extent
4464  * @inode: the inode we're dropping the extent for
4465  *
4466  * This is called when we are freeing up an outstanding extent, either called
4467  * after an error or after an extent is written.  This will return the number of
4468  * reserved extents that need to be freed.  This must be called with
4469  * BTRFS_I(inode)->lock held.
4470  */
4471 static unsigned drop_outstanding_extent(struct inode *inode)
4472 {
4473         unsigned drop_inode_space = 0;
4474         unsigned dropped_extents = 0;
4475
4476         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4477         BTRFS_I(inode)->outstanding_extents--;
4478
4479         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4480             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4481                                &BTRFS_I(inode)->runtime_flags))
4482                 drop_inode_space = 1;
4483
4484         /*
4485          * If we have more or the same amount of outsanding extents than we have
4486          * reserved then we need to leave the reserved extents count alone.
4487          */
4488         if (BTRFS_I(inode)->outstanding_extents >=
4489             BTRFS_I(inode)->reserved_extents)
4490                 return drop_inode_space;
4491
4492         dropped_extents = BTRFS_I(inode)->reserved_extents -
4493                 BTRFS_I(inode)->outstanding_extents;
4494         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4495         return dropped_extents + drop_inode_space;
4496 }
4497
4498 /**
4499  * calc_csum_metadata_size - return the amount of metada space that must be
4500  *      reserved/free'd for the given bytes.
4501  * @inode: the inode we're manipulating
4502  * @num_bytes: the number of bytes in question
4503  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4504  *
4505  * This adjusts the number of csum_bytes in the inode and then returns the
4506  * correct amount of metadata that must either be reserved or freed.  We
4507  * calculate how many checksums we can fit into one leaf and then divide the
4508  * number of bytes that will need to be checksumed by this value to figure out
4509  * how many checksums will be required.  If we are adding bytes then the number
4510  * may go up and we will return the number of additional bytes that must be
4511  * reserved.  If it is going down we will return the number of bytes that must
4512  * be freed.
4513  *
4514  * This must be called with BTRFS_I(inode)->lock held.
4515  */
4516 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4517                                    int reserve)
4518 {
4519         struct btrfs_root *root = BTRFS_I(inode)->root;
4520         u64 csum_size;
4521         int num_csums_per_leaf;
4522         int num_csums;
4523         int old_csums;
4524
4525         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4526             BTRFS_I(inode)->csum_bytes == 0)
4527                 return 0;
4528
4529         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4530         if (reserve)
4531                 BTRFS_I(inode)->csum_bytes += num_bytes;
4532         else
4533                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4534         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4535         num_csums_per_leaf = (int)div64_u64(csum_size,
4536                                             sizeof(struct btrfs_csum_item) +
4537                                             sizeof(struct btrfs_disk_key));
4538         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4539         num_csums = num_csums + num_csums_per_leaf - 1;
4540         num_csums = num_csums / num_csums_per_leaf;
4541
4542         old_csums = old_csums + num_csums_per_leaf - 1;
4543         old_csums = old_csums / num_csums_per_leaf;
4544
4545         /* No change, no need to reserve more */
4546         if (old_csums == num_csums)
4547                 return 0;
4548
4549         if (reserve)
4550                 return btrfs_calc_trans_metadata_size(root,
4551                                                       num_csums - old_csums);
4552
4553         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4554 }
4555
4556 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4557 {
4558         struct btrfs_root *root = BTRFS_I(inode)->root;
4559         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4560         u64 to_reserve = 0;
4561         u64 csum_bytes;
4562         unsigned nr_extents = 0;
4563         int extra_reserve = 0;
4564         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
4565         int ret = 0;
4566         bool delalloc_lock = true;
4567
4568         /* If we are a free space inode we need to not flush since we will be in
4569          * the middle of a transaction commit.  We also don't need the delalloc
4570          * mutex since we won't race with anybody.  We need this mostly to make
4571          * lockdep shut its filthy mouth.
4572          */
4573         if (btrfs_is_free_space_inode(inode)) {
4574                 flush = BTRFS_RESERVE_NO_FLUSH;
4575                 delalloc_lock = false;
4576         }
4577
4578         if (flush != BTRFS_RESERVE_NO_FLUSH &&
4579             btrfs_transaction_in_commit(root->fs_info))
4580                 schedule_timeout(1);
4581
4582         if (delalloc_lock)
4583                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4584
4585         num_bytes = ALIGN(num_bytes, root->sectorsize);
4586
4587         spin_lock(&BTRFS_I(inode)->lock);
4588         BTRFS_I(inode)->outstanding_extents++;
4589
4590         if (BTRFS_I(inode)->outstanding_extents >
4591             BTRFS_I(inode)->reserved_extents)
4592                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4593                         BTRFS_I(inode)->reserved_extents;
4594
4595         /*
4596          * Add an item to reserve for updating the inode when we complete the
4597          * delalloc io.
4598          */
4599         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4600                       &BTRFS_I(inode)->runtime_flags)) {
4601                 nr_extents++;
4602                 extra_reserve = 1;
4603         }
4604
4605         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4606         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4607         csum_bytes = BTRFS_I(inode)->csum_bytes;
4608         spin_unlock(&BTRFS_I(inode)->lock);
4609
4610         if (root->fs_info->quota_enabled)
4611                 ret = btrfs_qgroup_reserve(root, num_bytes +
4612                                            nr_extents * root->leafsize);
4613
4614         /*
4615          * ret != 0 here means the qgroup reservation failed, we go straight to
4616          * the shared error handling then.
4617          */
4618         if (ret == 0)
4619                 ret = reserve_metadata_bytes(root, block_rsv,
4620                                              to_reserve, flush);
4621
4622         if (ret) {
4623                 u64 to_free = 0;
4624                 unsigned dropped;
4625
4626                 spin_lock(&BTRFS_I(inode)->lock);
4627                 dropped = drop_outstanding_extent(inode);
4628                 /*
4629                  * If the inodes csum_bytes is the same as the original
4630                  * csum_bytes then we know we haven't raced with any free()ers
4631                  * so we can just reduce our inodes csum bytes and carry on.
4632                  * Otherwise we have to do the normal free thing to account for
4633                  * the case that the free side didn't free up its reserve
4634                  * because of this outstanding reservation.
4635                  */
4636                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4637                         calc_csum_metadata_size(inode, num_bytes, 0);
4638                 else
4639                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4640                 spin_unlock(&BTRFS_I(inode)->lock);
4641                 if (dropped)
4642                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4643
4644                 if (to_free) {
4645                         btrfs_block_rsv_release(root, block_rsv, to_free);
4646                         trace_btrfs_space_reservation(root->fs_info,
4647                                                       "delalloc",
4648                                                       btrfs_ino(inode),
4649                                                       to_free, 0);
4650                 }
4651                 if (root->fs_info->quota_enabled) {
4652                         btrfs_qgroup_free(root, num_bytes +
4653                                                 nr_extents * root->leafsize);
4654                 }
4655                 if (delalloc_lock)
4656                         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4657                 return ret;
4658         }
4659
4660         spin_lock(&BTRFS_I(inode)->lock);
4661         if (extra_reserve) {
4662                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4663                         &BTRFS_I(inode)->runtime_flags);
4664                 nr_extents--;
4665         }
4666         BTRFS_I(inode)->reserved_extents += nr_extents;
4667         spin_unlock(&BTRFS_I(inode)->lock);
4668
4669         if (delalloc_lock)
4670                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4671
4672         if (to_reserve)
4673                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4674                                               btrfs_ino(inode), to_reserve, 1);
4675         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4676
4677         return 0;
4678 }
4679
4680 /**
4681  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4682  * @inode: the inode to release the reservation for
4683  * @num_bytes: the number of bytes we're releasing
4684  *
4685  * This will release the metadata reservation for an inode.  This can be called
4686  * once we complete IO for a given set of bytes to release their metadata
4687  * reservations.
4688  */
4689 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4690 {
4691         struct btrfs_root *root = BTRFS_I(inode)->root;
4692         u64 to_free = 0;
4693         unsigned dropped;
4694
4695         num_bytes = ALIGN(num_bytes, root->sectorsize);
4696         spin_lock(&BTRFS_I(inode)->lock);
4697         dropped = drop_outstanding_extent(inode);
4698
4699         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4700         spin_unlock(&BTRFS_I(inode)->lock);
4701         if (dropped > 0)
4702                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4703
4704         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4705                                       btrfs_ino(inode), to_free, 0);
4706         if (root->fs_info->quota_enabled) {
4707                 btrfs_qgroup_free(root, num_bytes +
4708                                         dropped * root->leafsize);
4709         }
4710
4711         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4712                                 to_free);
4713 }
4714
4715 /**
4716  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4717  * @inode: inode we're writing to
4718  * @num_bytes: the number of bytes we want to allocate
4719  *
4720  * This will do the following things
4721  *
4722  * o reserve space in the data space info for num_bytes
4723  * o reserve space in the metadata space info based on number of outstanding
4724  *   extents and how much csums will be needed
4725  * o add to the inodes ->delalloc_bytes
4726  * o add it to the fs_info's delalloc inodes list.
4727  *
4728  * This will return 0 for success and -ENOSPC if there is no space left.
4729  */
4730 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4731 {
4732         int ret;
4733
4734         ret = btrfs_check_data_free_space(inode, num_bytes);
4735         if (ret)
4736                 return ret;
4737
4738         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4739         if (ret) {
4740                 btrfs_free_reserved_data_space(inode, num_bytes);
4741                 return ret;
4742         }
4743
4744         return 0;
4745 }
4746
4747 /**
4748  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4749  * @inode: inode we're releasing space for
4750  * @num_bytes: the number of bytes we want to free up
4751  *
4752  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4753  * called in the case that we don't need the metadata AND data reservations
4754  * anymore.  So if there is an error or we insert an inline extent.
4755  *
4756  * This function will release the metadata space that was not used and will
4757  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4758  * list if there are no delalloc bytes left.
4759  */
4760 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4761 {
4762         btrfs_delalloc_release_metadata(inode, num_bytes);
4763         btrfs_free_reserved_data_space(inode, num_bytes);
4764 }
4765
4766 static int update_block_group(struct btrfs_trans_handle *trans,
4767                               struct btrfs_root *root,
4768                               u64 bytenr, u64 num_bytes, int alloc)
4769 {
4770         struct btrfs_block_group_cache *cache = NULL;
4771         struct btrfs_fs_info *info = root->fs_info;
4772         u64 total = num_bytes;
4773         u64 old_val;
4774         u64 byte_in_group;
4775         int factor;
4776
4777         /* block accounting for super block */
4778         spin_lock(&info->delalloc_lock);
4779         old_val = btrfs_super_bytes_used(info->super_copy);
4780         if (alloc)
4781                 old_val += num_bytes;
4782         else
4783                 old_val -= num_bytes;
4784         btrfs_set_super_bytes_used(info->super_copy, old_val);
4785         spin_unlock(&info->delalloc_lock);
4786
4787         while (total) {
4788                 cache = btrfs_lookup_block_group(info, bytenr);
4789                 if (!cache)
4790                         return -ENOENT;
4791                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4792                                     BTRFS_BLOCK_GROUP_RAID1 |
4793                                     BTRFS_BLOCK_GROUP_RAID10))
4794                         factor = 2;
4795                 else
4796                         factor = 1;
4797                 /*
4798                  * If this block group has free space cache written out, we
4799                  * need to make sure to load it if we are removing space.  This
4800                  * is because we need the unpinning stage to actually add the
4801                  * space back to the block group, otherwise we will leak space.
4802                  */
4803                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4804                         cache_block_group(cache, trans, NULL, 1);
4805
4806                 byte_in_group = bytenr - cache->key.objectid;
4807                 WARN_ON(byte_in_group > cache->key.offset);
4808
4809                 spin_lock(&cache->space_info->lock);
4810                 spin_lock(&cache->lock);
4811
4812                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4813                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4814                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4815
4816                 cache->dirty = 1;
4817                 old_val = btrfs_block_group_used(&cache->item);
4818                 num_bytes = min(total, cache->key.offset - byte_in_group);
4819                 if (alloc) {
4820                         old_val += num_bytes;
4821                         btrfs_set_block_group_used(&cache->item, old_val);
4822                         cache->reserved -= num_bytes;
4823                         cache->space_info->bytes_reserved -= num_bytes;
4824                         cache->space_info->bytes_used += num_bytes;
4825                         cache->space_info->disk_used += num_bytes * factor;
4826                         spin_unlock(&cache->lock);
4827                         spin_unlock(&cache->space_info->lock);
4828                 } else {
4829                         old_val -= num_bytes;
4830                         btrfs_set_block_group_used(&cache->item, old_val);
4831                         cache->pinned += num_bytes;
4832                         cache->space_info->bytes_pinned += num_bytes;
4833                         cache->space_info->bytes_used -= num_bytes;
4834                         cache->space_info->disk_used -= num_bytes * factor;
4835                         spin_unlock(&cache->lock);
4836                         spin_unlock(&cache->space_info->lock);
4837
4838                         set_extent_dirty(info->pinned_extents,
4839                                          bytenr, bytenr + num_bytes - 1,
4840                                          GFP_NOFS | __GFP_NOFAIL);
4841                 }
4842                 btrfs_put_block_group(cache);
4843                 total -= num_bytes;
4844                 bytenr += num_bytes;
4845         }
4846         return 0;
4847 }
4848
4849 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4850 {
4851         struct btrfs_block_group_cache *cache;
4852         u64 bytenr;
4853
4854         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4855         if (!cache)
4856                 return 0;
4857
4858         bytenr = cache->key.objectid;
4859         btrfs_put_block_group(cache);
4860
4861         return bytenr;
4862 }
4863
4864 static int pin_down_extent(struct btrfs_root *root,
4865                            struct btrfs_block_group_cache *cache,
4866                            u64 bytenr, u64 num_bytes, int reserved)
4867 {
4868         spin_lock(&cache->space_info->lock);
4869         spin_lock(&cache->lock);
4870         cache->pinned += num_bytes;
4871         cache->space_info->bytes_pinned += num_bytes;
4872         if (reserved) {
4873                 cache->reserved -= num_bytes;
4874                 cache->space_info->bytes_reserved -= num_bytes;
4875         }
4876         spin_unlock(&cache->lock);
4877         spin_unlock(&cache->space_info->lock);
4878
4879         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4880                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4881         return 0;
4882 }
4883
4884 /*
4885  * this function must be called within transaction
4886  */
4887 int btrfs_pin_extent(struct btrfs_root *root,
4888                      u64 bytenr, u64 num_bytes, int reserved)
4889 {
4890         struct btrfs_block_group_cache *cache;
4891
4892         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4893         BUG_ON(!cache); /* Logic error */
4894
4895         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4896
4897         btrfs_put_block_group(cache);
4898         return 0;
4899 }
4900
4901 /*
4902  * this function must be called within transaction
4903  */
4904 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4905                                     struct btrfs_root *root,
4906                                     u64 bytenr, u64 num_bytes)
4907 {
4908         struct btrfs_block_group_cache *cache;
4909
4910         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4911         BUG_ON(!cache); /* Logic error */
4912
4913         /*
4914          * pull in the free space cache (if any) so that our pin
4915          * removes the free space from the cache.  We have load_only set
4916          * to one because the slow code to read in the free extents does check
4917          * the pinned extents.
4918          */
4919         cache_block_group(cache, trans, root, 1);
4920
4921         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4922
4923         /* remove us from the free space cache (if we're there at all) */
4924         btrfs_remove_free_space(cache, bytenr, num_bytes);
4925         btrfs_put_block_group(cache);
4926         return 0;
4927 }
4928
4929 /**
4930  * btrfs_update_reserved_bytes - update the block_group and space info counters
4931  * @cache:      The cache we are manipulating
4932  * @num_bytes:  The number of bytes in question
4933  * @reserve:    One of the reservation enums
4934  *
4935  * This is called by the allocator when it reserves space, or by somebody who is
4936  * freeing space that was never actually used on disk.  For example if you
4937  * reserve some space for a new leaf in transaction A and before transaction A
4938  * commits you free that leaf, you call this with reserve set to 0 in order to
4939  * clear the reservation.
4940  *
4941  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4942  * ENOSPC accounting.  For data we handle the reservation through clearing the
4943  * delalloc bits in the io_tree.  We have to do this since we could end up
4944  * allocating less disk space for the amount of data we have reserved in the
4945  * case of compression.
4946  *
4947  * If this is a reservation and the block group has become read only we cannot
4948  * make the reservation and return -EAGAIN, otherwise this function always
4949  * succeeds.
4950  */
4951 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4952                                        u64 num_bytes, int reserve)
4953 {
4954         struct btrfs_space_info *space_info = cache->space_info;
4955         int ret = 0;
4956
4957         spin_lock(&space_info->lock);
4958         spin_lock(&cache->lock);
4959         if (reserve != RESERVE_FREE) {
4960                 if (cache->ro) {
4961                         ret = -EAGAIN;
4962                 } else {
4963                         cache->reserved += num_bytes;
4964                         space_info->bytes_reserved += num_bytes;
4965                         if (reserve == RESERVE_ALLOC) {
4966                                 trace_btrfs_space_reservation(cache->fs_info,
4967                                                 "space_info", space_info->flags,
4968                                                 num_bytes, 0);
4969                                 space_info->bytes_may_use -= num_bytes;
4970                         }
4971                 }
4972         } else {
4973                 if (cache->ro)
4974                         space_info->bytes_readonly += num_bytes;
4975                 cache->reserved -= num_bytes;
4976                 space_info->bytes_reserved -= num_bytes;
4977                 space_info->reservation_progress++;
4978         }
4979         spin_unlock(&cache->lock);
4980         spin_unlock(&space_info->lock);
4981         return ret;
4982 }
4983
4984 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4985                                 struct btrfs_root *root)
4986 {
4987         struct btrfs_fs_info *fs_info = root->fs_info;
4988         struct btrfs_caching_control *next;
4989         struct btrfs_caching_control *caching_ctl;
4990         struct btrfs_block_group_cache *cache;
4991
4992         down_write(&fs_info->extent_commit_sem);
4993
4994         list_for_each_entry_safe(caching_ctl, next,
4995                                  &fs_info->caching_block_groups, list) {
4996                 cache = caching_ctl->block_group;
4997                 if (block_group_cache_done(cache)) {
4998                         cache->last_byte_to_unpin = (u64)-1;
4999                         list_del_init(&caching_ctl->list);
5000                         put_caching_control(caching_ctl);
5001                 } else {
5002                         cache->last_byte_to_unpin = caching_ctl->progress;
5003                 }
5004         }
5005
5006         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5007                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5008         else
5009                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5010
5011         up_write(&fs_info->extent_commit_sem);
5012
5013         update_global_block_rsv(fs_info);
5014 }
5015
5016 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5017 {
5018         struct btrfs_fs_info *fs_info = root->fs_info;
5019         struct btrfs_block_group_cache *cache = NULL;
5020         struct btrfs_space_info *space_info;
5021         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5022         u64 len;
5023         bool readonly;
5024
5025         while (start <= end) {
5026                 readonly = false;
5027                 if (!cache ||
5028                     start >= cache->key.objectid + cache->key.offset) {
5029                         if (cache)
5030                                 btrfs_put_block_group(cache);
5031                         cache = btrfs_lookup_block_group(fs_info, start);
5032                         BUG_ON(!cache); /* Logic error */
5033                 }
5034
5035                 len = cache->key.objectid + cache->key.offset - start;
5036                 len = min(len, end + 1 - start);
5037
5038                 if (start < cache->last_byte_to_unpin) {
5039                         len = min(len, cache->last_byte_to_unpin - start);
5040                         btrfs_add_free_space(cache, start, len);
5041                 }
5042
5043                 start += len;
5044                 space_info = cache->space_info;
5045
5046                 spin_lock(&space_info->lock);
5047                 spin_lock(&cache->lock);
5048                 cache->pinned -= len;
5049                 space_info->bytes_pinned -= len;
5050                 if (cache->ro) {
5051                         space_info->bytes_readonly += len;
5052                         readonly = true;
5053                 }
5054                 spin_unlock(&cache->lock);
5055                 if (!readonly && global_rsv->space_info == space_info) {
5056                         spin_lock(&global_rsv->lock);
5057                         if (!global_rsv->full) {
5058                                 len = min(len, global_rsv->size -
5059                                           global_rsv->reserved);
5060                                 global_rsv->reserved += len;
5061                                 space_info->bytes_may_use += len;
5062                                 if (global_rsv->reserved >= global_rsv->size)
5063                                         global_rsv->full = 1;
5064                         }
5065                         spin_unlock(&global_rsv->lock);
5066                 }
5067                 spin_unlock(&space_info->lock);
5068         }
5069
5070         if (cache)
5071                 btrfs_put_block_group(cache);
5072         return 0;
5073 }
5074
5075 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5076                                struct btrfs_root *root)
5077 {
5078         struct btrfs_fs_info *fs_info = root->fs_info;
5079         struct extent_io_tree *unpin;
5080         u64 start;
5081         u64 end;
5082         int ret;
5083
5084         if (trans->aborted)
5085                 return 0;
5086
5087         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5088                 unpin = &fs_info->freed_extents[1];
5089         else
5090                 unpin = &fs_info->freed_extents[0];
5091
5092         while (1) {
5093                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5094                                             EXTENT_DIRTY, NULL);
5095                 if (ret)
5096                         break;
5097
5098                 if (btrfs_test_opt(root, DISCARD))
5099                         ret = btrfs_discard_extent(root, start,
5100                                                    end + 1 - start, NULL);
5101
5102                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5103                 unpin_extent_range(root, start, end);
5104                 cond_resched();
5105         }
5106
5107         return 0;
5108 }
5109
5110 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5111                                 struct btrfs_root *root,
5112                                 u64 bytenr, u64 num_bytes, u64 parent,
5113                                 u64 root_objectid, u64 owner_objectid,
5114                                 u64 owner_offset, int refs_to_drop,
5115                                 struct btrfs_delayed_extent_op *extent_op)
5116 {
5117         struct btrfs_key key;
5118         struct btrfs_path *path;
5119         struct btrfs_fs_info *info = root->fs_info;
5120         struct btrfs_root *extent_root = info->extent_root;
5121         struct extent_buffer *leaf;
5122         struct btrfs_extent_item *ei;
5123         struct btrfs_extent_inline_ref *iref;
5124         int ret;
5125         int is_data;
5126         int extent_slot = 0;
5127         int found_extent = 0;
5128         int num_to_del = 1;
5129         u32 item_size;
5130         u64 refs;
5131
5132         path = btrfs_alloc_path();
5133         if (!path)
5134                 return -ENOMEM;
5135
5136         path->reada = 1;
5137         path->leave_spinning = 1;
5138
5139         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5140         BUG_ON(!is_data && refs_to_drop != 1);
5141
5142         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5143                                     bytenr, num_bytes, parent,
5144                                     root_objectid, owner_objectid,
5145                                     owner_offset);
5146         if (ret == 0) {
5147                 extent_slot = path->slots[0];
5148                 while (extent_slot >= 0) {
5149                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5150                                               extent_slot);
5151                         if (key.objectid != bytenr)
5152                                 break;
5153                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5154                             key.offset == num_bytes) {
5155                                 found_extent = 1;
5156                                 break;
5157                         }
5158                         if (path->slots[0] - extent_slot > 5)
5159                                 break;
5160                         extent_slot--;
5161                 }
5162 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5163                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5164                 if (found_extent && item_size < sizeof(*ei))
5165                         found_extent = 0;
5166 #endif
5167                 if (!found_extent) {
5168                         BUG_ON(iref);
5169                         ret = remove_extent_backref(trans, extent_root, path,
5170                                                     NULL, refs_to_drop,
5171                                                     is_data);
5172                         if (ret) {
5173                                 btrfs_abort_transaction(trans, extent_root, ret);
5174                                 goto out;
5175                         }
5176                         btrfs_release_path(path);
5177                         path->leave_spinning = 1;
5178
5179                         key.objectid = bytenr;
5180                         key.type = BTRFS_EXTENT_ITEM_KEY;
5181                         key.offset = num_bytes;
5182
5183                         ret = btrfs_search_slot(trans, extent_root,
5184                                                 &key, path, -1, 1);
5185                         if (ret) {
5186                                 printk(KERN_ERR "umm, got %d back from search"
5187                                        ", was looking for %llu\n", ret,
5188                                        (unsigned long long)bytenr);
5189                                 if (ret > 0)
5190                                         btrfs_print_leaf(extent_root,
5191                                                          path->nodes[0]);
5192                         }
5193                         if (ret < 0) {
5194                                 btrfs_abort_transaction(trans, extent_root, ret);
5195                                 goto out;
5196                         }
5197                         extent_slot = path->slots[0];
5198                 }
5199         } else if (ret == -ENOENT) {
5200                 btrfs_print_leaf(extent_root, path->nodes[0]);
5201                 WARN_ON(1);
5202                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5203                        "parent %llu root %llu  owner %llu offset %llu\n",
5204                        (unsigned long long)bytenr,
5205                        (unsigned long long)parent,
5206                        (unsigned long long)root_objectid,
5207                        (unsigned long long)owner_objectid,
5208                        (unsigned long long)owner_offset);
5209         } else {
5210                 btrfs_abort_transaction(trans, extent_root, ret);
5211                 goto out;
5212         }
5213
5214         leaf = path->nodes[0];
5215         item_size = btrfs_item_size_nr(leaf, extent_slot);
5216 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5217         if (item_size < sizeof(*ei)) {
5218                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5219                 ret = convert_extent_item_v0(trans, extent_root, path,
5220                                              owner_objectid, 0);
5221                 if (ret < 0) {
5222                         btrfs_abort_transaction(trans, extent_root, ret);
5223                         goto out;
5224                 }
5225
5226                 btrfs_release_path(path);
5227                 path->leave_spinning = 1;
5228
5229                 key.objectid = bytenr;
5230                 key.type = BTRFS_EXTENT_ITEM_KEY;
5231                 key.offset = num_bytes;
5232
5233                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5234                                         -1, 1);
5235                 if (ret) {
5236                         printk(KERN_ERR "umm, got %d back from search"
5237                                ", was looking for %llu\n", ret,
5238                                (unsigned long long)bytenr);
5239                         btrfs_print_leaf(extent_root, path->nodes[0]);
5240                 }
5241                 if (ret < 0) {
5242                         btrfs_abort_transaction(trans, extent_root, ret);
5243                         goto out;
5244                 }
5245
5246                 extent_slot = path->slots[0];
5247                 leaf = path->nodes[0];
5248                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5249         }
5250 #endif
5251         BUG_ON(item_size < sizeof(*ei));
5252         ei = btrfs_item_ptr(leaf, extent_slot,
5253                             struct btrfs_extent_item);
5254         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5255                 struct btrfs_tree_block_info *bi;
5256                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5257                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5258                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5259         }
5260
5261         refs = btrfs_extent_refs(leaf, ei);
5262         BUG_ON(refs < refs_to_drop);
5263         refs -= refs_to_drop;
5264
5265         if (refs > 0) {
5266                 if (extent_op)
5267                         __run_delayed_extent_op(extent_op, leaf, ei);
5268                 /*
5269                  * In the case of inline back ref, reference count will
5270                  * be updated by remove_extent_backref
5271                  */
5272                 if (iref) {
5273                         BUG_ON(!found_extent);
5274                 } else {
5275                         btrfs_set_extent_refs(leaf, ei, refs);
5276                         btrfs_mark_buffer_dirty(leaf);
5277                 }
5278                 if (found_extent) {
5279                         ret = remove_extent_backref(trans, extent_root, path,
5280                                                     iref, refs_to_drop,
5281                                                     is_data);
5282                         if (ret) {
5283                                 btrfs_abort_transaction(trans, extent_root, ret);
5284                                 goto out;
5285                         }
5286                 }
5287         } else {
5288                 if (found_extent) {
5289                         BUG_ON(is_data && refs_to_drop !=
5290                                extent_data_ref_count(root, path, iref));
5291                         if (iref) {
5292                                 BUG_ON(path->slots[0] != extent_slot);
5293                         } else {
5294                                 BUG_ON(path->slots[0] != extent_slot + 1);
5295                                 path->slots[0] = extent_slot;
5296                                 num_to_del = 2;
5297                         }
5298                 }
5299
5300                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5301                                       num_to_del);
5302                 if (ret) {
5303                         btrfs_abort_transaction(trans, extent_root, ret);
5304                         goto out;
5305                 }
5306                 btrfs_release_path(path);
5307
5308                 if (is_data) {
5309                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5310                         if (ret) {
5311                                 btrfs_abort_transaction(trans, extent_root, ret);
5312                                 goto out;
5313                         }
5314                 }
5315
5316                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5317                 if (ret) {
5318                         btrfs_abort_transaction(trans, extent_root, ret);
5319                         goto out;
5320                 }
5321         }
5322 out:
5323         btrfs_free_path(path);
5324         return ret;
5325 }
5326
5327 /*
5328  * when we free an block, it is possible (and likely) that we free the last
5329  * delayed ref for that extent as well.  This searches the delayed ref tree for
5330  * a given extent, and if there are no other delayed refs to be processed, it
5331  * removes it from the tree.
5332  */
5333 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5334                                       struct btrfs_root *root, u64 bytenr)
5335 {
5336         struct btrfs_delayed_ref_head *head;
5337         struct btrfs_delayed_ref_root *delayed_refs;
5338         struct btrfs_delayed_ref_node *ref;
5339         struct rb_node *node;
5340         int ret = 0;
5341
5342         delayed_refs = &trans->transaction->delayed_refs;
5343         spin_lock(&delayed_refs->lock);
5344         head = btrfs_find_delayed_ref_head(trans, bytenr);
5345         if (!head)
5346                 goto out;
5347
5348         node = rb_prev(&head->node.rb_node);
5349         if (!node)
5350                 goto out;
5351
5352         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5353
5354         /* there are still entries for this ref, we can't drop it */
5355         if (ref->bytenr == bytenr)
5356                 goto out;
5357
5358         if (head->extent_op) {
5359                 if (!head->must_insert_reserved)
5360                         goto out;
5361                 btrfs_free_delayed_extent_op(head->extent_op);
5362                 head->extent_op = NULL;
5363         }
5364
5365         /*
5366          * waiting for the lock here would deadlock.  If someone else has it
5367          * locked they are already in the process of dropping it anyway
5368          */
5369         if (!mutex_trylock(&head->mutex))
5370                 goto out;
5371
5372         /*
5373          * at this point we have a head with no other entries.  Go
5374          * ahead and process it.
5375          */
5376         head->node.in_tree = 0;
5377         rb_erase(&head->node.rb_node, &delayed_refs->root);
5378
5379         delayed_refs->num_entries--;
5380
5381         /*
5382          * we don't take a ref on the node because we're removing it from the
5383          * tree, so we just steal the ref the tree was holding.
5384          */
5385         delayed_refs->num_heads--;
5386         if (list_empty(&head->cluster))
5387                 delayed_refs->num_heads_ready--;
5388
5389         list_del_init(&head->cluster);
5390         spin_unlock(&delayed_refs->lock);
5391
5392         BUG_ON(head->extent_op);
5393         if (head->must_insert_reserved)
5394                 ret = 1;
5395
5396         mutex_unlock(&head->mutex);
5397         btrfs_put_delayed_ref(&head->node);
5398         return ret;
5399 out:
5400         spin_unlock(&delayed_refs->lock);
5401         return 0;
5402 }
5403
5404 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5405                            struct btrfs_root *root,
5406                            struct extent_buffer *buf,
5407                            u64 parent, int last_ref)
5408 {
5409         struct btrfs_block_group_cache *cache = NULL;
5410         int ret;
5411
5412         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5413                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5414                                         buf->start, buf->len,
5415                                         parent, root->root_key.objectid,
5416                                         btrfs_header_level(buf),
5417                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5418                 BUG_ON(ret); /* -ENOMEM */
5419         }
5420
5421         if (!last_ref)
5422                 return;
5423
5424         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5425
5426         if (btrfs_header_generation(buf) == trans->transid) {
5427                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5428                         ret = check_ref_cleanup(trans, root, buf->start);
5429                         if (!ret)
5430                                 goto out;
5431                 }
5432
5433                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5434                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5435                         goto out;
5436                 }
5437
5438                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5439
5440                 btrfs_add_free_space(cache, buf->start, buf->len);
5441                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5442         }
5443 out:
5444         /*
5445          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5446          * anymore.
5447          */
5448         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5449         btrfs_put_block_group(cache);
5450 }
5451
5452 /* Can return -ENOMEM */
5453 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5454                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5455                       u64 owner, u64 offset, int for_cow)
5456 {
5457         int ret;
5458         struct btrfs_fs_info *fs_info = root->fs_info;
5459
5460         /*
5461          * tree log blocks never actually go into the extent allocation
5462          * tree, just update pinning info and exit early.
5463          */
5464         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5465                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5466                 /* unlocks the pinned mutex */
5467                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5468                 ret = 0;
5469         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5470                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5471                                         num_bytes,
5472                                         parent, root_objectid, (int)owner,
5473                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5474         } else {
5475                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5476                                                 num_bytes,
5477                                                 parent, root_objectid, owner,
5478                                                 offset, BTRFS_DROP_DELAYED_REF,
5479                                                 NULL, for_cow);
5480         }
5481         return ret;
5482 }
5483
5484 static u64 stripe_align(struct btrfs_root *root, u64 val)
5485 {
5486         u64 mask = ((u64)root->stripesize - 1);
5487         u64 ret = (val + mask) & ~mask;
5488         return ret;
5489 }
5490
5491 /*
5492  * when we wait for progress in the block group caching, its because
5493  * our allocation attempt failed at least once.  So, we must sleep
5494  * and let some progress happen before we try again.
5495  *
5496  * This function will sleep at least once waiting for new free space to
5497  * show up, and then it will check the block group free space numbers
5498  * for our min num_bytes.  Another option is to have it go ahead
5499  * and look in the rbtree for a free extent of a given size, but this
5500  * is a good start.
5501  */
5502 static noinline int
5503 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5504                                 u64 num_bytes)
5505 {
5506         struct btrfs_caching_control *caching_ctl;
5507         DEFINE_WAIT(wait);
5508
5509         caching_ctl = get_caching_control(cache);
5510         if (!caching_ctl)
5511                 return 0;
5512
5513         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5514                    (cache->free_space_ctl->free_space >= num_bytes));
5515
5516         put_caching_control(caching_ctl);
5517         return 0;
5518 }
5519
5520 static noinline int
5521 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5522 {
5523         struct btrfs_caching_control *caching_ctl;
5524         DEFINE_WAIT(wait);
5525
5526         caching_ctl = get_caching_control(cache);
5527         if (!caching_ctl)
5528                 return 0;
5529
5530         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5531
5532         put_caching_control(caching_ctl);
5533         return 0;
5534 }
5535
5536 int __get_raid_index(u64 flags)
5537 {
5538         int index;
5539
5540         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5541                 index = 0;
5542         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5543                 index = 1;
5544         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5545                 index = 2;
5546         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5547                 index = 3;
5548         else
5549                 index = 4;
5550
5551         return index;
5552 }
5553
5554 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5555 {
5556         return __get_raid_index(cache->flags);
5557 }
5558
5559 enum btrfs_loop_type {
5560         LOOP_CACHING_NOWAIT = 0,
5561         LOOP_CACHING_WAIT = 1,
5562         LOOP_ALLOC_CHUNK = 2,
5563         LOOP_NO_EMPTY_SIZE = 3,
5564 };
5565
5566 /*
5567  * walks the btree of allocated extents and find a hole of a given size.
5568  * The key ins is changed to record the hole:
5569  * ins->objectid == block start
5570  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5571  * ins->offset == number of blocks
5572  * Any available blocks before search_start are skipped.
5573  */
5574 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5575                                      struct btrfs_root *orig_root,
5576                                      u64 num_bytes, u64 empty_size,
5577                                      u64 hint_byte, struct btrfs_key *ins,
5578                                      u64 data)
5579 {
5580         int ret = 0;
5581         struct btrfs_root *root = orig_root->fs_info->extent_root;
5582         struct btrfs_free_cluster *last_ptr = NULL;
5583         struct btrfs_block_group_cache *block_group = NULL;
5584         struct btrfs_block_group_cache *used_block_group;
5585         u64 search_start = 0;
5586         int empty_cluster = 2 * 1024 * 1024;
5587         struct btrfs_space_info *space_info;
5588         int loop = 0;
5589         int index = __get_raid_index(data);
5590         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5591                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5592         bool found_uncached_bg = false;
5593         bool failed_cluster_refill = false;
5594         bool failed_alloc = false;
5595         bool use_cluster = true;
5596         bool have_caching_bg = false;
5597
5598         WARN_ON(num_bytes < root->sectorsize);
5599         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5600         ins->objectid = 0;
5601         ins->offset = 0;
5602
5603         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5604
5605         space_info = __find_space_info(root->fs_info, data);
5606         if (!space_info) {
5607                 printk(KERN_ERR "No space info for %llu\n", data);
5608                 return -ENOSPC;
5609         }
5610
5611         /*
5612          * If the space info is for both data and metadata it means we have a
5613          * small filesystem and we can't use the clustering stuff.
5614          */
5615         if (btrfs_mixed_space_info(space_info))
5616                 use_cluster = false;
5617
5618         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5619                 last_ptr = &root->fs_info->meta_alloc_cluster;
5620                 if (!btrfs_test_opt(root, SSD))
5621                         empty_cluster = 64 * 1024;
5622         }
5623
5624         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5625             btrfs_test_opt(root, SSD)) {
5626                 last_ptr = &root->fs_info->data_alloc_cluster;
5627         }
5628
5629         if (last_ptr) {
5630                 spin_lock(&last_ptr->lock);
5631                 if (last_ptr->block_group)
5632                         hint_byte = last_ptr->window_start;
5633                 spin_unlock(&last_ptr->lock);
5634         }
5635
5636         search_start = max(search_start, first_logical_byte(root, 0));
5637         search_start = max(search_start, hint_byte);
5638
5639         if (!last_ptr)
5640                 empty_cluster = 0;
5641
5642         if (search_start == hint_byte) {
5643                 block_group = btrfs_lookup_block_group(root->fs_info,
5644                                                        search_start);
5645                 used_block_group = block_group;
5646                 /*
5647                  * we don't want to use the block group if it doesn't match our
5648                  * allocation bits, or if its not cached.
5649                  *
5650                  * However if we are re-searching with an ideal block group
5651                  * picked out then we don't care that the block group is cached.
5652                  */
5653                 if (block_group && block_group_bits(block_group, data) &&
5654                     block_group->cached != BTRFS_CACHE_NO) {
5655                         down_read(&space_info->groups_sem);
5656                         if (list_empty(&block_group->list) ||
5657                             block_group->ro) {
5658                                 /*
5659                                  * someone is removing this block group,
5660                                  * we can't jump into the have_block_group
5661                                  * target because our list pointers are not
5662                                  * valid
5663                                  */
5664                                 btrfs_put_block_group(block_group);
5665                                 up_read(&space_info->groups_sem);
5666                         } else {
5667                                 index = get_block_group_index(block_group);
5668                                 goto have_block_group;
5669                         }
5670                 } else if (block_group) {
5671                         btrfs_put_block_group(block_group);
5672                 }
5673         }
5674 search:
5675         have_caching_bg = false;
5676         down_read(&space_info->groups_sem);
5677         list_for_each_entry(block_group, &space_info->block_groups[index],
5678                             list) {
5679                 u64 offset;
5680                 int cached;
5681
5682                 used_block_group = block_group;
5683                 btrfs_get_block_group(block_group);
5684                 search_start = block_group->key.objectid;
5685
5686                 /*
5687                  * this can happen if we end up cycling through all the
5688                  * raid types, but we want to make sure we only allocate
5689                  * for the proper type.
5690                  */
5691                 if (!block_group_bits(block_group, data)) {
5692                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5693                                 BTRFS_BLOCK_GROUP_RAID1 |
5694                                 BTRFS_BLOCK_GROUP_RAID10;
5695
5696                         /*
5697                          * if they asked for extra copies and this block group
5698                          * doesn't provide them, bail.  This does allow us to
5699                          * fill raid0 from raid1.
5700                          */
5701                         if ((data & extra) && !(block_group->flags & extra))
5702                                 goto loop;
5703                 }
5704
5705 have_block_group:
5706                 cached = block_group_cache_done(block_group);
5707                 if (unlikely(!cached)) {
5708                         found_uncached_bg = true;
5709                         ret = cache_block_group(block_group, trans,
5710                                                 orig_root, 0);
5711                         BUG_ON(ret < 0);
5712                         ret = 0;
5713                 }
5714
5715                 if (unlikely(block_group->ro))
5716                         goto loop;
5717
5718                 /*
5719                  * Ok we want to try and use the cluster allocator, so
5720                  * lets look there
5721                  */
5722                 if (last_ptr) {
5723                         /*
5724                          * the refill lock keeps out other
5725                          * people trying to start a new cluster
5726                          */
5727                         spin_lock(&last_ptr->refill_lock);
5728                         used_block_group = last_ptr->block_group;
5729                         if (used_block_group != block_group &&
5730                             (!used_block_group ||
5731                              used_block_group->ro ||
5732                              !block_group_bits(used_block_group, data))) {
5733                                 used_block_group = block_group;
5734                                 goto refill_cluster;
5735                         }
5736
5737                         if (used_block_group != block_group)
5738                                 btrfs_get_block_group(used_block_group);
5739
5740                         offset = btrfs_alloc_from_cluster(used_block_group,
5741                           last_ptr, num_bytes, used_block_group->key.objectid);
5742                         if (offset) {
5743                                 /* we have a block, we're done */
5744                                 spin_unlock(&last_ptr->refill_lock);
5745                                 trace_btrfs_reserve_extent_cluster(root,
5746                                         block_group, search_start, num_bytes);
5747                                 goto checks;
5748                         }
5749
5750                         WARN_ON(last_ptr->block_group != used_block_group);
5751                         if (used_block_group != block_group) {
5752                                 btrfs_put_block_group(used_block_group);
5753                                 used_block_group = block_group;
5754                         }
5755 refill_cluster:
5756                         BUG_ON(used_block_group != block_group);
5757                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5758                          * set up a new clusters, so lets just skip it
5759                          * and let the allocator find whatever block
5760                          * it can find.  If we reach this point, we
5761                          * will have tried the cluster allocator
5762                          * plenty of times and not have found
5763                          * anything, so we are likely way too
5764                          * fragmented for the clustering stuff to find
5765                          * anything.
5766                          *
5767                          * However, if the cluster is taken from the
5768                          * current block group, release the cluster
5769                          * first, so that we stand a better chance of
5770                          * succeeding in the unclustered
5771                          * allocation.  */
5772                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5773                             last_ptr->block_group != block_group) {
5774                                 spin_unlock(&last_ptr->refill_lock);
5775                                 goto unclustered_alloc;
5776                         }
5777
5778                         /*
5779                          * this cluster didn't work out, free it and
5780                          * start over
5781                          */
5782                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5783
5784                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5785                                 spin_unlock(&last_ptr->refill_lock);
5786                                 goto unclustered_alloc;
5787                         }
5788
5789                         /* allocate a cluster in this block group */
5790                         ret = btrfs_find_space_cluster(trans, root,
5791                                                block_group, last_ptr,
5792                                                search_start, num_bytes,
5793                                                empty_cluster + empty_size);
5794                         if (ret == 0) {
5795                                 /*
5796                                  * now pull our allocation out of this
5797                                  * cluster
5798                                  */
5799                                 offset = btrfs_alloc_from_cluster(block_group,
5800                                                   last_ptr, num_bytes,
5801                                                   search_start);
5802                                 if (offset) {
5803                                         /* we found one, proceed */
5804                                         spin_unlock(&last_ptr->refill_lock);
5805                                         trace_btrfs_reserve_extent_cluster(root,
5806                                                 block_group, search_start,
5807                                                 num_bytes);
5808                                         goto checks;
5809                                 }
5810                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5811                                    && !failed_cluster_refill) {
5812                                 spin_unlock(&last_ptr->refill_lock);
5813
5814                                 failed_cluster_refill = true;
5815                                 wait_block_group_cache_progress(block_group,
5816                                        num_bytes + empty_cluster + empty_size);
5817                                 goto have_block_group;
5818                         }
5819
5820                         /*
5821                          * at this point we either didn't find a cluster
5822                          * or we weren't able to allocate a block from our
5823                          * cluster.  Free the cluster we've been trying
5824                          * to use, and go to the next block group
5825                          */
5826                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5827                         spin_unlock(&last_ptr->refill_lock);
5828                         goto loop;
5829                 }
5830
5831 unclustered_alloc:
5832                 spin_lock(&block_group->free_space_ctl->tree_lock);
5833                 if (cached &&
5834                     block_group->free_space_ctl->free_space <
5835                     num_bytes + empty_cluster + empty_size) {
5836                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5837                         goto loop;
5838                 }
5839                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5840
5841                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5842                                                     num_bytes, empty_size);
5843                 /*
5844                  * If we didn't find a chunk, and we haven't failed on this
5845                  * block group before, and this block group is in the middle of
5846                  * caching and we are ok with waiting, then go ahead and wait
5847                  * for progress to be made, and set failed_alloc to true.
5848                  *
5849                  * If failed_alloc is true then we've already waited on this
5850                  * block group once and should move on to the next block group.
5851                  */
5852                 if (!offset && !failed_alloc && !cached &&
5853                     loop > LOOP_CACHING_NOWAIT) {
5854                         wait_block_group_cache_progress(block_group,
5855                                                 num_bytes + empty_size);
5856                         failed_alloc = true;
5857                         goto have_block_group;
5858                 } else if (!offset) {
5859                         if (!cached)
5860                                 have_caching_bg = true;
5861                         goto loop;
5862                 }
5863 checks:
5864                 search_start = stripe_align(root, offset);
5865
5866                 /* move on to the next group */
5867                 if (search_start + num_bytes >
5868                     used_block_group->key.objectid + used_block_group->key.offset) {
5869                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5870                         goto loop;
5871                 }
5872
5873                 if (offset < search_start)
5874                         btrfs_add_free_space(used_block_group, offset,
5875                                              search_start - offset);
5876                 BUG_ON(offset > search_start);
5877
5878                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
5879                                                   alloc_type);
5880                 if (ret == -EAGAIN) {
5881                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5882                         goto loop;
5883                 }
5884
5885                 /* we are all good, lets return */
5886                 ins->objectid = search_start;
5887                 ins->offset = num_bytes;
5888
5889                 trace_btrfs_reserve_extent(orig_root, block_group,
5890                                            search_start, num_bytes);
5891                 if (used_block_group != block_group)
5892                         btrfs_put_block_group(used_block_group);
5893                 btrfs_put_block_group(block_group);
5894                 break;
5895 loop:
5896                 failed_cluster_refill = false;
5897                 failed_alloc = false;
5898                 BUG_ON(index != get_block_group_index(block_group));
5899                 if (used_block_group != block_group)
5900                         btrfs_put_block_group(used_block_group);
5901                 btrfs_put_block_group(block_group);
5902         }
5903         up_read(&space_info->groups_sem);
5904
5905         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5906                 goto search;
5907
5908         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5909                 goto search;
5910
5911         /*
5912          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5913          *                      caching kthreads as we move along
5914          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5915          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5916          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5917          *                      again
5918          */
5919         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5920                 index = 0;
5921                 loop++;
5922                 if (loop == LOOP_ALLOC_CHUNK) {
5923                         ret = do_chunk_alloc(trans, root, data,
5924                                              CHUNK_ALLOC_FORCE);
5925                         /*
5926                          * Do not bail out on ENOSPC since we
5927                          * can do more things.
5928                          */
5929                         if (ret < 0 && ret != -ENOSPC) {
5930                                 btrfs_abort_transaction(trans,
5931                                                         root, ret);
5932                                 goto out;
5933                         }
5934                 }
5935
5936                 if (loop == LOOP_NO_EMPTY_SIZE) {
5937                         empty_size = 0;
5938                         empty_cluster = 0;
5939                 }
5940
5941                 goto search;
5942         } else if (!ins->objectid) {
5943                 ret = -ENOSPC;
5944         } else if (ins->objectid) {
5945                 ret = 0;
5946         }
5947 out:
5948
5949         return ret;
5950 }
5951
5952 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5953                             int dump_block_groups)
5954 {
5955         struct btrfs_block_group_cache *cache;
5956         int index = 0;
5957
5958         spin_lock(&info->lock);
5959         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5960                (unsigned long long)info->flags,
5961                (unsigned long long)(info->total_bytes - info->bytes_used -
5962                                     info->bytes_pinned - info->bytes_reserved -
5963                                     info->bytes_readonly),
5964                (info->full) ? "" : "not ");
5965         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5966                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5967                (unsigned long long)info->total_bytes,
5968                (unsigned long long)info->bytes_used,
5969                (unsigned long long)info->bytes_pinned,
5970                (unsigned long long)info->bytes_reserved,
5971                (unsigned long long)info->bytes_may_use,
5972                (unsigned long long)info->bytes_readonly);
5973         spin_unlock(&info->lock);
5974
5975         if (!dump_block_groups)
5976                 return;
5977
5978         down_read(&info->groups_sem);
5979 again:
5980         list_for_each_entry(cache, &info->block_groups[index], list) {
5981                 spin_lock(&cache->lock);
5982                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
5983                        (unsigned long long)cache->key.objectid,
5984                        (unsigned long long)cache->key.offset,
5985                        (unsigned long long)btrfs_block_group_used(&cache->item),
5986                        (unsigned long long)cache->pinned,
5987                        (unsigned long long)cache->reserved,
5988                        cache->ro ? "[readonly]" : "");
5989                 btrfs_dump_free_space(cache, bytes);
5990                 spin_unlock(&cache->lock);
5991         }
5992         if (++index < BTRFS_NR_RAID_TYPES)
5993                 goto again;
5994         up_read(&info->groups_sem);
5995 }
5996
5997 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5998                          struct btrfs_root *root,
5999                          u64 num_bytes, u64 min_alloc_size,
6000                          u64 empty_size, u64 hint_byte,
6001                          struct btrfs_key *ins, u64 data)
6002 {
6003         bool final_tried = false;
6004         int ret;
6005
6006         data = btrfs_get_alloc_profile(root, data);
6007 again:
6008         WARN_ON(num_bytes < root->sectorsize);
6009         ret = find_free_extent(trans, root, num_bytes, empty_size,
6010                                hint_byte, ins, data);
6011
6012         if (ret == -ENOSPC) {
6013                 if (!final_tried) {
6014                         num_bytes = num_bytes >> 1;
6015                         num_bytes = num_bytes & ~(root->sectorsize - 1);
6016                         num_bytes = max(num_bytes, min_alloc_size);
6017                         if (num_bytes == min_alloc_size)
6018                                 final_tried = true;
6019                         goto again;
6020                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6021                         struct btrfs_space_info *sinfo;
6022
6023                         sinfo = __find_space_info(root->fs_info, data);
6024                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
6025                                "wanted %llu\n", (unsigned long long)data,
6026                                (unsigned long long)num_bytes);
6027                         if (sinfo)
6028                                 dump_space_info(sinfo, num_bytes, 1);
6029                 }
6030         }
6031
6032         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6033
6034         return ret;
6035 }
6036
6037 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6038                                         u64 start, u64 len, int pin)
6039 {
6040         struct btrfs_block_group_cache *cache;
6041         int ret = 0;
6042
6043         cache = btrfs_lookup_block_group(root->fs_info, start);
6044         if (!cache) {
6045                 printk(KERN_ERR "Unable to find block group for %llu\n",
6046                        (unsigned long long)start);
6047                 return -ENOSPC;
6048         }
6049
6050         if (btrfs_test_opt(root, DISCARD))
6051                 ret = btrfs_discard_extent(root, start, len, NULL);
6052
6053         if (pin)
6054                 pin_down_extent(root, cache, start, len, 1);
6055         else {
6056                 btrfs_add_free_space(cache, start, len);
6057                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6058         }
6059         btrfs_put_block_group(cache);
6060
6061         trace_btrfs_reserved_extent_free(root, start, len);
6062
6063         return ret;
6064 }
6065
6066 int btrfs_free_reserved_extent(struct btrfs_root *root,
6067                                         u64 start, u64 len)
6068 {
6069         return __btrfs_free_reserved_extent(root, start, len, 0);
6070 }
6071
6072 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6073                                        u64 start, u64 len)
6074 {
6075         return __btrfs_free_reserved_extent(root, start, len, 1);
6076 }
6077
6078 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6079                                       struct btrfs_root *root,
6080                                       u64 parent, u64 root_objectid,
6081                                       u64 flags, u64 owner, u64 offset,
6082                                       struct btrfs_key *ins, int ref_mod)
6083 {
6084         int ret;
6085         struct btrfs_fs_info *fs_info = root->fs_info;
6086         struct btrfs_extent_item *extent_item;
6087         struct btrfs_extent_inline_ref *iref;
6088         struct btrfs_path *path;
6089         struct extent_buffer *leaf;
6090         int type;
6091         u32 size;
6092
6093         if (parent > 0)
6094                 type = BTRFS_SHARED_DATA_REF_KEY;
6095         else
6096                 type = BTRFS_EXTENT_DATA_REF_KEY;
6097
6098         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6099
6100         path = btrfs_alloc_path();
6101         if (!path)
6102                 return -ENOMEM;
6103
6104         path->leave_spinning = 1;
6105         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6106                                       ins, size);
6107         if (ret) {
6108                 btrfs_free_path(path);
6109                 return ret;
6110         }
6111
6112         leaf = path->nodes[0];
6113         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6114                                      struct btrfs_extent_item);
6115         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6116         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6117         btrfs_set_extent_flags(leaf, extent_item,
6118                                flags | BTRFS_EXTENT_FLAG_DATA);
6119
6120         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6121         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6122         if (parent > 0) {
6123                 struct btrfs_shared_data_ref *ref;
6124                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6125                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6126                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6127         } else {
6128                 struct btrfs_extent_data_ref *ref;
6129                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6130                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6131                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6132                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6133                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6134         }
6135
6136         btrfs_mark_buffer_dirty(path->nodes[0]);
6137         btrfs_free_path(path);
6138
6139         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6140         if (ret) { /* -ENOENT, logic error */
6141                 printk(KERN_ERR "btrfs update block group failed for %llu "
6142                        "%llu\n", (unsigned long long)ins->objectid,
6143                        (unsigned long long)ins->offset);
6144                 BUG();
6145         }
6146         return ret;
6147 }
6148
6149 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6150                                      struct btrfs_root *root,
6151                                      u64 parent, u64 root_objectid,
6152                                      u64 flags, struct btrfs_disk_key *key,
6153                                      int level, struct btrfs_key *ins)
6154 {
6155         int ret;
6156         struct btrfs_fs_info *fs_info = root->fs_info;
6157         struct btrfs_extent_item *extent_item;
6158         struct btrfs_tree_block_info *block_info;
6159         struct btrfs_extent_inline_ref *iref;
6160         struct btrfs_path *path;
6161         struct extent_buffer *leaf;
6162         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6163
6164         path = btrfs_alloc_path();
6165         if (!path)
6166                 return -ENOMEM;
6167
6168         path->leave_spinning = 1;
6169         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6170                                       ins, size);
6171         if (ret) {
6172                 btrfs_free_path(path);
6173                 return ret;
6174         }
6175
6176         leaf = path->nodes[0];
6177         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6178                                      struct btrfs_extent_item);
6179         btrfs_set_extent_refs(leaf, extent_item, 1);
6180         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6181         btrfs_set_extent_flags(leaf, extent_item,
6182                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6183         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6184
6185         btrfs_set_tree_block_key(leaf, block_info, key);
6186         btrfs_set_tree_block_level(leaf, block_info, level);
6187
6188         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6189         if (parent > 0) {
6190                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6191                 btrfs_set_extent_inline_ref_type(leaf, iref,
6192                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6193                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6194         } else {
6195                 btrfs_set_extent_inline_ref_type(leaf, iref,
6196                                                  BTRFS_TREE_BLOCK_REF_KEY);
6197                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6198         }
6199
6200         btrfs_mark_buffer_dirty(leaf);
6201         btrfs_free_path(path);
6202
6203         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6204         if (ret) { /* -ENOENT, logic error */
6205                 printk(KERN_ERR "btrfs update block group failed for %llu "
6206                        "%llu\n", (unsigned long long)ins->objectid,
6207                        (unsigned long long)ins->offset);
6208                 BUG();
6209         }
6210         return ret;
6211 }
6212
6213 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6214                                      struct btrfs_root *root,
6215                                      u64 root_objectid, u64 owner,
6216                                      u64 offset, struct btrfs_key *ins)
6217 {
6218         int ret;
6219
6220         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6221
6222         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6223                                          ins->offset, 0,
6224                                          root_objectid, owner, offset,
6225                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6226         return ret;
6227 }
6228
6229 /*
6230  * this is used by the tree logging recovery code.  It records that
6231  * an extent has been allocated and makes sure to clear the free
6232  * space cache bits as well
6233  */
6234 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6235                                    struct btrfs_root *root,
6236                                    u64 root_objectid, u64 owner, u64 offset,
6237                                    struct btrfs_key *ins)
6238 {
6239         int ret;
6240         struct btrfs_block_group_cache *block_group;
6241         struct btrfs_caching_control *caching_ctl;
6242         u64 start = ins->objectid;
6243         u64 num_bytes = ins->offset;
6244
6245         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6246         cache_block_group(block_group, trans, NULL, 0);
6247         caching_ctl = get_caching_control(block_group);
6248
6249         if (!caching_ctl) {
6250                 BUG_ON(!block_group_cache_done(block_group));
6251                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6252                 BUG_ON(ret); /* -ENOMEM */
6253         } else {
6254                 mutex_lock(&caching_ctl->mutex);
6255
6256                 if (start >= caching_ctl->progress) {
6257                         ret = add_excluded_extent(root, start, num_bytes);
6258                         BUG_ON(ret); /* -ENOMEM */
6259                 } else if (start + num_bytes <= caching_ctl->progress) {
6260                         ret = btrfs_remove_free_space(block_group,
6261                                                       start, num_bytes);
6262                         BUG_ON(ret); /* -ENOMEM */
6263                 } else {
6264                         num_bytes = caching_ctl->progress - start;
6265                         ret = btrfs_remove_free_space(block_group,
6266                                                       start, num_bytes);
6267                         BUG_ON(ret); /* -ENOMEM */
6268
6269                         start = caching_ctl->progress;
6270                         num_bytes = ins->objectid + ins->offset -
6271                                     caching_ctl->progress;
6272                         ret = add_excluded_extent(root, start, num_bytes);
6273                         BUG_ON(ret); /* -ENOMEM */
6274                 }
6275
6276                 mutex_unlock(&caching_ctl->mutex);
6277                 put_caching_control(caching_ctl);
6278         }
6279
6280         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6281                                           RESERVE_ALLOC_NO_ACCOUNT);
6282         BUG_ON(ret); /* logic error */
6283         btrfs_put_block_group(block_group);
6284         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6285                                          0, owner, offset, ins, 1);
6286         return ret;
6287 }
6288
6289 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6290                                             struct btrfs_root *root,
6291                                             u64 bytenr, u32 blocksize,
6292                                             int level)
6293 {
6294         struct extent_buffer *buf;
6295
6296         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6297         if (!buf)
6298                 return ERR_PTR(-ENOMEM);
6299         btrfs_set_header_generation(buf, trans->transid);
6300         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6301         btrfs_tree_lock(buf);
6302         clean_tree_block(trans, root, buf);
6303         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6304
6305         btrfs_set_lock_blocking(buf);
6306         btrfs_set_buffer_uptodate(buf);
6307
6308         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6309                 /*
6310                  * we allow two log transactions at a time, use different
6311                  * EXENT bit to differentiate dirty pages.
6312                  */
6313                 if (root->log_transid % 2 == 0)
6314                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6315                                         buf->start + buf->len - 1, GFP_NOFS);
6316                 else
6317                         set_extent_new(&root->dirty_log_pages, buf->start,
6318                                         buf->start + buf->len - 1, GFP_NOFS);
6319         } else {
6320                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6321                          buf->start + buf->len - 1, GFP_NOFS);
6322         }
6323         trans->blocks_used++;
6324         /* this returns a buffer locked for blocking */
6325         return buf;
6326 }
6327
6328 static struct btrfs_block_rsv *
6329 use_block_rsv(struct btrfs_trans_handle *trans,
6330               struct btrfs_root *root, u32 blocksize)
6331 {
6332         struct btrfs_block_rsv *block_rsv;
6333         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6334         int ret;
6335
6336         block_rsv = get_block_rsv(trans, root);
6337
6338         if (block_rsv->size == 0) {
6339                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6340                                              BTRFS_RESERVE_NO_FLUSH);
6341                 /*
6342                  * If we couldn't reserve metadata bytes try and use some from
6343                  * the global reserve.
6344                  */
6345                 if (ret && block_rsv != global_rsv) {
6346                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6347                         if (!ret)
6348                                 return global_rsv;
6349                         return ERR_PTR(ret);
6350                 } else if (ret) {
6351                         return ERR_PTR(ret);
6352                 }
6353                 return block_rsv;
6354         }
6355
6356         ret = block_rsv_use_bytes(block_rsv, blocksize);
6357         if (!ret)
6358                 return block_rsv;
6359         if (ret && !block_rsv->failfast) {
6360                 static DEFINE_RATELIMIT_STATE(_rs,
6361                                 DEFAULT_RATELIMIT_INTERVAL,
6362                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
6363                 if (__ratelimit(&_rs))
6364                         WARN(1, KERN_DEBUG "btrfs: block rsv returned %d\n",
6365                              ret);
6366                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6367                                              BTRFS_RESERVE_NO_FLUSH);
6368                 if (!ret) {
6369                         return block_rsv;
6370                 } else if (ret && block_rsv != global_rsv) {
6371                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6372                         if (!ret)
6373                                 return global_rsv;
6374                 }
6375         }
6376
6377         return ERR_PTR(-ENOSPC);
6378 }
6379
6380 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6381                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6382 {
6383         block_rsv_add_bytes(block_rsv, blocksize, 0);
6384         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6385 }
6386
6387 /*
6388  * finds a free extent and does all the dirty work required for allocation
6389  * returns the key for the extent through ins, and a tree buffer for
6390  * the first block of the extent through buf.
6391  *
6392  * returns the tree buffer or NULL.
6393  */
6394 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6395                                         struct btrfs_root *root, u32 blocksize,
6396                                         u64 parent, u64 root_objectid,
6397                                         struct btrfs_disk_key *key, int level,
6398                                         u64 hint, u64 empty_size)
6399 {
6400         struct btrfs_key ins;
6401         struct btrfs_block_rsv *block_rsv;
6402         struct extent_buffer *buf;
6403         u64 flags = 0;
6404         int ret;
6405
6406
6407         block_rsv = use_block_rsv(trans, root, blocksize);
6408         if (IS_ERR(block_rsv))
6409                 return ERR_CAST(block_rsv);
6410
6411         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6412                                    empty_size, hint, &ins, 0);
6413         if (ret) {
6414                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6415                 return ERR_PTR(ret);
6416         }
6417
6418         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6419                                     blocksize, level);
6420         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6421
6422         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6423                 if (parent == 0)
6424                         parent = ins.objectid;
6425                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6426         } else
6427                 BUG_ON(parent > 0);
6428
6429         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6430                 struct btrfs_delayed_extent_op *extent_op;
6431                 extent_op = btrfs_alloc_delayed_extent_op();
6432                 BUG_ON(!extent_op); /* -ENOMEM */
6433                 if (key)
6434                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6435                 else
6436                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6437                 extent_op->flags_to_set = flags;
6438                 extent_op->update_key = 1;
6439                 extent_op->update_flags = 1;
6440                 extent_op->is_data = 0;
6441
6442                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6443                                         ins.objectid,
6444                                         ins.offset, parent, root_objectid,
6445                                         level, BTRFS_ADD_DELAYED_EXTENT,
6446                                         extent_op, 0);
6447                 BUG_ON(ret); /* -ENOMEM */
6448         }
6449         return buf;
6450 }
6451
6452 struct walk_control {
6453         u64 refs[BTRFS_MAX_LEVEL];
6454         u64 flags[BTRFS_MAX_LEVEL];
6455         struct btrfs_key update_progress;
6456         int stage;
6457         int level;
6458         int shared_level;
6459         int update_ref;
6460         int keep_locks;
6461         int reada_slot;
6462         int reada_count;
6463         int for_reloc;
6464 };
6465
6466 #define DROP_REFERENCE  1
6467 #define UPDATE_BACKREF  2
6468
6469 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6470                                      struct btrfs_root *root,
6471                                      struct walk_control *wc,
6472                                      struct btrfs_path *path)
6473 {
6474         u64 bytenr;
6475         u64 generation;
6476         u64 refs;
6477         u64 flags;
6478         u32 nritems;
6479         u32 blocksize;
6480         struct btrfs_key key;
6481         struct extent_buffer *eb;
6482         int ret;
6483         int slot;
6484         int nread = 0;
6485
6486         if (path->slots[wc->level] < wc->reada_slot) {
6487                 wc->reada_count = wc->reada_count * 2 / 3;
6488                 wc->reada_count = max(wc->reada_count, 2);
6489         } else {
6490                 wc->reada_count = wc->reada_count * 3 / 2;
6491                 wc->reada_count = min_t(int, wc->reada_count,
6492                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6493         }
6494
6495         eb = path->nodes[wc->level];
6496         nritems = btrfs_header_nritems(eb);
6497         blocksize = btrfs_level_size(root, wc->level - 1);
6498
6499         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6500                 if (nread >= wc->reada_count)
6501                         break;
6502
6503                 cond_resched();
6504                 bytenr = btrfs_node_blockptr(eb, slot);
6505                 generation = btrfs_node_ptr_generation(eb, slot);
6506
6507                 if (slot == path->slots[wc->level])
6508                         goto reada;
6509
6510                 if (wc->stage == UPDATE_BACKREF &&
6511                     generation <= root->root_key.offset)
6512                         continue;
6513
6514                 /* We don't lock the tree block, it's OK to be racy here */
6515                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6516                                                &refs, &flags);
6517                 /* We don't care about errors in readahead. */
6518                 if (ret < 0)
6519                         continue;
6520                 BUG_ON(refs == 0);
6521
6522                 if (wc->stage == DROP_REFERENCE) {
6523                         if (refs == 1)
6524                                 goto reada;
6525
6526                         if (wc->level == 1 &&
6527                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6528                                 continue;
6529                         if (!wc->update_ref ||
6530                             generation <= root->root_key.offset)
6531                                 continue;
6532                         btrfs_node_key_to_cpu(eb, &key, slot);
6533                         ret = btrfs_comp_cpu_keys(&key,
6534                                                   &wc->update_progress);
6535                         if (ret < 0)
6536                                 continue;
6537                 } else {
6538                         if (wc->level == 1 &&
6539                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6540                                 continue;
6541                 }
6542 reada:
6543                 ret = readahead_tree_block(root, bytenr, blocksize,
6544                                            generation);
6545                 if (ret)
6546                         break;
6547                 nread++;
6548         }
6549         wc->reada_slot = slot;
6550 }
6551
6552 /*
6553  * hepler to process tree block while walking down the tree.
6554  *
6555  * when wc->stage == UPDATE_BACKREF, this function updates
6556  * back refs for pointers in the block.
6557  *
6558  * NOTE: return value 1 means we should stop walking down.
6559  */
6560 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6561                                    struct btrfs_root *root,
6562                                    struct btrfs_path *path,
6563                                    struct walk_control *wc, int lookup_info)
6564 {
6565         int level = wc->level;
6566         struct extent_buffer *eb = path->nodes[level];
6567         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6568         int ret;
6569
6570         if (wc->stage == UPDATE_BACKREF &&
6571             btrfs_header_owner(eb) != root->root_key.objectid)
6572                 return 1;
6573
6574         /*
6575          * when reference count of tree block is 1, it won't increase
6576          * again. once full backref flag is set, we never clear it.
6577          */
6578         if (lookup_info &&
6579             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6580              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6581                 BUG_ON(!path->locks[level]);
6582                 ret = btrfs_lookup_extent_info(trans, root,
6583                                                eb->start, eb->len,
6584                                                &wc->refs[level],
6585                                                &wc->flags[level]);
6586                 BUG_ON(ret == -ENOMEM);
6587                 if (ret)
6588                         return ret;
6589                 BUG_ON(wc->refs[level] == 0);
6590         }
6591
6592         if (wc->stage == DROP_REFERENCE) {
6593                 if (wc->refs[level] > 1)
6594                         return 1;
6595
6596                 if (path->locks[level] && !wc->keep_locks) {
6597                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6598                         path->locks[level] = 0;
6599                 }
6600                 return 0;
6601         }
6602
6603         /* wc->stage == UPDATE_BACKREF */
6604         if (!(wc->flags[level] & flag)) {
6605                 BUG_ON(!path->locks[level]);
6606                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6607                 BUG_ON(ret); /* -ENOMEM */
6608                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6609                 BUG_ON(ret); /* -ENOMEM */
6610                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6611                                                   eb->len, flag, 0);
6612                 BUG_ON(ret); /* -ENOMEM */
6613                 wc->flags[level] |= flag;
6614         }
6615
6616         /*
6617          * the block is shared by multiple trees, so it's not good to
6618          * keep the tree lock
6619          */
6620         if (path->locks[level] && level > 0) {
6621                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6622                 path->locks[level] = 0;
6623         }
6624         return 0;
6625 }
6626
6627 /*
6628  * hepler to process tree block pointer.
6629  *
6630  * when wc->stage == DROP_REFERENCE, this function checks
6631  * reference count of the block pointed to. if the block
6632  * is shared and we need update back refs for the subtree
6633  * rooted at the block, this function changes wc->stage to
6634  * UPDATE_BACKREF. if the block is shared and there is no
6635  * need to update back, this function drops the reference
6636  * to the block.
6637  *
6638  * NOTE: return value 1 means we should stop walking down.
6639  */
6640 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6641                                  struct btrfs_root *root,
6642                                  struct btrfs_path *path,
6643                                  struct walk_control *wc, int *lookup_info)
6644 {
6645         u64 bytenr;
6646         u64 generation;
6647         u64 parent;
6648         u32 blocksize;
6649         struct btrfs_key key;
6650         struct extent_buffer *next;
6651         int level = wc->level;
6652         int reada = 0;
6653         int ret = 0;
6654
6655         generation = btrfs_node_ptr_generation(path->nodes[level],
6656                                                path->slots[level]);
6657         /*
6658          * if the lower level block was created before the snapshot
6659          * was created, we know there is no need to update back refs
6660          * for the subtree
6661          */
6662         if (wc->stage == UPDATE_BACKREF &&
6663             generation <= root->root_key.offset) {
6664                 *lookup_info = 1;
6665                 return 1;
6666         }
6667
6668         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6669         blocksize = btrfs_level_size(root, level - 1);
6670
6671         next = btrfs_find_tree_block(root, bytenr, blocksize);
6672         if (!next) {
6673                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6674                 if (!next)
6675                         return -ENOMEM;
6676                 reada = 1;
6677         }
6678         btrfs_tree_lock(next);
6679         btrfs_set_lock_blocking(next);
6680
6681         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6682                                        &wc->refs[level - 1],
6683                                        &wc->flags[level - 1]);
6684         if (ret < 0) {
6685                 btrfs_tree_unlock(next);
6686                 return ret;
6687         }
6688
6689         BUG_ON(wc->refs[level - 1] == 0);
6690         *lookup_info = 0;
6691
6692         if (wc->stage == DROP_REFERENCE) {
6693                 if (wc->refs[level - 1] > 1) {
6694                         if (level == 1 &&
6695                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6696                                 goto skip;
6697
6698                         if (!wc->update_ref ||
6699                             generation <= root->root_key.offset)
6700                                 goto skip;
6701
6702                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6703                                               path->slots[level]);
6704                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6705                         if (ret < 0)
6706                                 goto skip;
6707
6708                         wc->stage = UPDATE_BACKREF;
6709                         wc->shared_level = level - 1;
6710                 }
6711         } else {
6712                 if (level == 1 &&
6713                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6714                         goto skip;
6715         }
6716
6717         if (!btrfs_buffer_uptodate(next, generation, 0)) {
6718                 btrfs_tree_unlock(next);
6719                 free_extent_buffer(next);
6720                 next = NULL;
6721                 *lookup_info = 1;
6722         }
6723
6724         if (!next) {
6725                 if (reada && level == 1)
6726                         reada_walk_down(trans, root, wc, path);
6727                 next = read_tree_block(root, bytenr, blocksize, generation);
6728                 if (!next)
6729                         return -EIO;
6730                 btrfs_tree_lock(next);
6731                 btrfs_set_lock_blocking(next);
6732         }
6733
6734         level--;
6735         BUG_ON(level != btrfs_header_level(next));
6736         path->nodes[level] = next;
6737         path->slots[level] = 0;
6738         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6739         wc->level = level;
6740         if (wc->level == 1)
6741                 wc->reada_slot = 0;
6742         return 0;
6743 skip:
6744         wc->refs[level - 1] = 0;
6745         wc->flags[level - 1] = 0;
6746         if (wc->stage == DROP_REFERENCE) {
6747                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6748                         parent = path->nodes[level]->start;
6749                 } else {
6750                         BUG_ON(root->root_key.objectid !=
6751                                btrfs_header_owner(path->nodes[level]));
6752                         parent = 0;
6753                 }
6754
6755                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6756                                 root->root_key.objectid, level - 1, 0, 0);
6757                 BUG_ON(ret); /* -ENOMEM */
6758         }
6759         btrfs_tree_unlock(next);
6760         free_extent_buffer(next);
6761         *lookup_info = 1;
6762         return 1;
6763 }
6764
6765 /*
6766  * hepler to process tree block while walking up the tree.
6767  *
6768  * when wc->stage == DROP_REFERENCE, this function drops
6769  * reference count on the block.
6770  *
6771  * when wc->stage == UPDATE_BACKREF, this function changes
6772  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6773  * to UPDATE_BACKREF previously while processing the block.
6774  *
6775  * NOTE: return value 1 means we should stop walking up.
6776  */
6777 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6778                                  struct btrfs_root *root,
6779                                  struct btrfs_path *path,
6780                                  struct walk_control *wc)
6781 {
6782         int ret;
6783         int level = wc->level;
6784         struct extent_buffer *eb = path->nodes[level];
6785         u64 parent = 0;
6786
6787         if (wc->stage == UPDATE_BACKREF) {
6788                 BUG_ON(wc->shared_level < level);
6789                 if (level < wc->shared_level)
6790                         goto out;
6791
6792                 ret = find_next_key(path, level + 1, &wc->update_progress);
6793                 if (ret > 0)
6794                         wc->update_ref = 0;
6795
6796                 wc->stage = DROP_REFERENCE;
6797                 wc->shared_level = -1;
6798                 path->slots[level] = 0;
6799
6800                 /*
6801                  * check reference count again if the block isn't locked.
6802                  * we should start walking down the tree again if reference
6803                  * count is one.
6804                  */
6805                 if (!path->locks[level]) {
6806                         BUG_ON(level == 0);
6807                         btrfs_tree_lock(eb);
6808                         btrfs_set_lock_blocking(eb);
6809                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6810
6811                         ret = btrfs_lookup_extent_info(trans, root,
6812                                                        eb->start, eb->len,
6813                                                        &wc->refs[level],
6814                                                        &wc->flags[level]);
6815                         if (ret < 0) {
6816                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6817                                 path->locks[level] = 0;
6818                                 return ret;
6819                         }
6820                         BUG_ON(wc->refs[level] == 0);
6821                         if (wc->refs[level] == 1) {
6822                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6823                                 path->locks[level] = 0;
6824                                 return 1;
6825                         }
6826                 }
6827         }
6828
6829         /* wc->stage == DROP_REFERENCE */
6830         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6831
6832         if (wc->refs[level] == 1) {
6833                 if (level == 0) {
6834                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6835                                 ret = btrfs_dec_ref(trans, root, eb, 1,
6836                                                     wc->for_reloc);
6837                         else
6838                                 ret = btrfs_dec_ref(trans, root, eb, 0,
6839                                                     wc->for_reloc);
6840                         BUG_ON(ret); /* -ENOMEM */
6841                 }
6842                 /* make block locked assertion in clean_tree_block happy */
6843                 if (!path->locks[level] &&
6844                     btrfs_header_generation(eb) == trans->transid) {
6845                         btrfs_tree_lock(eb);
6846                         btrfs_set_lock_blocking(eb);
6847                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6848                 }
6849                 clean_tree_block(trans, root, eb);
6850         }
6851
6852         if (eb == root->node) {
6853                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6854                         parent = eb->start;
6855                 else
6856                         BUG_ON(root->root_key.objectid !=
6857                                btrfs_header_owner(eb));
6858         } else {
6859                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6860                         parent = path->nodes[level + 1]->start;
6861                 else
6862                         BUG_ON(root->root_key.objectid !=
6863                                btrfs_header_owner(path->nodes[level + 1]));
6864         }
6865
6866         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6867 out:
6868         wc->refs[level] = 0;
6869         wc->flags[level] = 0;
6870         return 0;
6871 }
6872
6873 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6874                                    struct btrfs_root *root,
6875                                    struct btrfs_path *path,
6876                                    struct walk_control *wc)
6877 {
6878         int level = wc->level;
6879         int lookup_info = 1;
6880         int ret;
6881
6882         while (level >= 0) {
6883                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6884                 if (ret > 0)
6885                         break;
6886
6887                 if (level == 0)
6888                         break;
6889
6890                 if (path->slots[level] >=
6891                     btrfs_header_nritems(path->nodes[level]))
6892                         break;
6893
6894                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6895                 if (ret > 0) {
6896                         path->slots[level]++;
6897                         continue;
6898                 } else if (ret < 0)
6899                         return ret;
6900                 level = wc->level;
6901         }
6902         return 0;
6903 }
6904
6905 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6906                                  struct btrfs_root *root,
6907                                  struct btrfs_path *path,
6908                                  struct walk_control *wc, int max_level)
6909 {
6910         int level = wc->level;
6911         int ret;
6912
6913         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6914         while (level < max_level && path->nodes[level]) {
6915                 wc->level = level;
6916                 if (path->slots[level] + 1 <
6917                     btrfs_header_nritems(path->nodes[level])) {
6918                         path->slots[level]++;
6919                         return 0;
6920                 } else {
6921                         ret = walk_up_proc(trans, root, path, wc);
6922                         if (ret > 0)
6923                                 return 0;
6924
6925                         if (path->locks[level]) {
6926                                 btrfs_tree_unlock_rw(path->nodes[level],
6927                                                      path->locks[level]);
6928                                 path->locks[level] = 0;
6929                         }
6930                         free_extent_buffer(path->nodes[level]);
6931                         path->nodes[level] = NULL;
6932                         level++;
6933                 }
6934         }
6935         return 1;
6936 }
6937
6938 /*
6939  * drop a subvolume tree.
6940  *
6941  * this function traverses the tree freeing any blocks that only
6942  * referenced by the tree.
6943  *
6944  * when a shared tree block is found. this function decreases its
6945  * reference count by one. if update_ref is true, this function
6946  * also make sure backrefs for the shared block and all lower level
6947  * blocks are properly updated.
6948  */
6949 int btrfs_drop_snapshot(struct btrfs_root *root,
6950                          struct btrfs_block_rsv *block_rsv, int update_ref,
6951                          int for_reloc)
6952 {
6953         struct btrfs_path *path;
6954         struct btrfs_trans_handle *trans;
6955         struct btrfs_root *tree_root = root->fs_info->tree_root;
6956         struct btrfs_root_item *root_item = &root->root_item;
6957         struct walk_control *wc;
6958         struct btrfs_key key;
6959         int err = 0;
6960         int ret;
6961         int level;
6962
6963         path = btrfs_alloc_path();
6964         if (!path) {
6965                 err = -ENOMEM;
6966                 goto out;
6967         }
6968
6969         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6970         if (!wc) {
6971                 btrfs_free_path(path);
6972                 err = -ENOMEM;
6973                 goto out;
6974         }
6975
6976         trans = btrfs_start_transaction(tree_root, 0);
6977         if (IS_ERR(trans)) {
6978                 err = PTR_ERR(trans);
6979                 goto out_free;
6980         }
6981
6982         if (block_rsv)
6983                 trans->block_rsv = block_rsv;
6984
6985         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6986                 level = btrfs_header_level(root->node);
6987                 path->nodes[level] = btrfs_lock_root_node(root);
6988                 btrfs_set_lock_blocking(path->nodes[level]);
6989                 path->slots[level] = 0;
6990                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6991                 memset(&wc->update_progress, 0,
6992                        sizeof(wc->update_progress));
6993         } else {
6994                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6995                 memcpy(&wc->update_progress, &key,
6996                        sizeof(wc->update_progress));
6997
6998                 level = root_item->drop_level;
6999                 BUG_ON(level == 0);
7000                 path->lowest_level = level;
7001                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7002                 path->lowest_level = 0;
7003                 if (ret < 0) {
7004                         err = ret;
7005                         goto out_end_trans;
7006                 }
7007                 WARN_ON(ret > 0);
7008
7009                 /*
7010                  * unlock our path, this is safe because only this
7011                  * function is allowed to delete this snapshot
7012                  */
7013                 btrfs_unlock_up_safe(path, 0);
7014
7015                 level = btrfs_header_level(root->node);
7016                 while (1) {
7017                         btrfs_tree_lock(path->nodes[level]);
7018                         btrfs_set_lock_blocking(path->nodes[level]);
7019
7020                         ret = btrfs_lookup_extent_info(trans, root,
7021                                                 path->nodes[level]->start,
7022                                                 path->nodes[level]->len,
7023                                                 &wc->refs[level],
7024                                                 &wc->flags[level]);
7025                         if (ret < 0) {
7026                                 err = ret;
7027                                 goto out_end_trans;
7028                         }
7029                         BUG_ON(wc->refs[level] == 0);
7030
7031                         if (level == root_item->drop_level)
7032                                 break;
7033
7034                         btrfs_tree_unlock(path->nodes[level]);
7035                         WARN_ON(wc->refs[level] != 1);
7036                         level--;
7037                 }
7038         }
7039
7040         wc->level = level;
7041         wc->shared_level = -1;
7042         wc->stage = DROP_REFERENCE;
7043         wc->update_ref = update_ref;
7044         wc->keep_locks = 0;
7045         wc->for_reloc = for_reloc;
7046         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7047
7048         while (1) {
7049                 ret = walk_down_tree(trans, root, path, wc);
7050                 if (ret < 0) {
7051                         err = ret;
7052                         break;
7053                 }
7054
7055                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7056                 if (ret < 0) {
7057                         err = ret;
7058                         break;
7059                 }
7060
7061                 if (ret > 0) {
7062                         BUG_ON(wc->stage != DROP_REFERENCE);
7063                         break;
7064                 }
7065
7066                 if (wc->stage == DROP_REFERENCE) {
7067                         level = wc->level;
7068                         btrfs_node_key(path->nodes[level],
7069                                        &root_item->drop_progress,
7070                                        path->slots[level]);
7071                         root_item->drop_level = level;
7072                 }
7073
7074                 BUG_ON(wc->level == 0);
7075                 if (btrfs_should_end_transaction(trans, tree_root)) {
7076                         ret = btrfs_update_root(trans, tree_root,
7077                                                 &root->root_key,
7078                                                 root_item);
7079                         if (ret) {
7080                                 btrfs_abort_transaction(trans, tree_root, ret);
7081                                 err = ret;
7082                                 goto out_end_trans;
7083                         }
7084
7085                         btrfs_end_transaction_throttle(trans, tree_root);
7086                         trans = btrfs_start_transaction(tree_root, 0);
7087                         if (IS_ERR(trans)) {
7088                                 err = PTR_ERR(trans);
7089                                 goto out_free;
7090                         }
7091                         if (block_rsv)
7092                                 trans->block_rsv = block_rsv;
7093                 }
7094         }
7095         btrfs_release_path(path);
7096         if (err)
7097                 goto out_end_trans;
7098
7099         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7100         if (ret) {
7101                 btrfs_abort_transaction(trans, tree_root, ret);
7102                 goto out_end_trans;
7103         }
7104
7105         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7106                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
7107                                            NULL, NULL);
7108                 if (ret < 0) {
7109                         btrfs_abort_transaction(trans, tree_root, ret);
7110                         err = ret;
7111                         goto out_end_trans;
7112                 } else if (ret > 0) {
7113                         /* if we fail to delete the orphan item this time
7114                          * around, it'll get picked up the next time.
7115                          *
7116                          * The most common failure here is just -ENOENT.
7117                          */
7118                         btrfs_del_orphan_item(trans, tree_root,
7119                                               root->root_key.objectid);
7120                 }
7121         }
7122
7123         if (root->in_radix) {
7124                 btrfs_free_fs_root(tree_root->fs_info, root);
7125         } else {
7126                 free_extent_buffer(root->node);
7127                 free_extent_buffer(root->commit_root);
7128                 kfree(root);
7129         }
7130 out_end_trans:
7131         btrfs_end_transaction_throttle(trans, tree_root);
7132 out_free:
7133         kfree(wc);
7134         btrfs_free_path(path);
7135 out:
7136         if (err)
7137                 btrfs_std_error(root->fs_info, err);
7138         return err;
7139 }
7140
7141 /*
7142  * drop subtree rooted at tree block 'node'.
7143  *
7144  * NOTE: this function will unlock and release tree block 'node'
7145  * only used by relocation code
7146  */
7147 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7148                         struct btrfs_root *root,
7149                         struct extent_buffer *node,
7150                         struct extent_buffer *parent)
7151 {
7152         struct btrfs_path *path;
7153         struct walk_control *wc;
7154         int level;
7155         int parent_level;
7156         int ret = 0;
7157         int wret;
7158
7159         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7160
7161         path = btrfs_alloc_path();
7162         if (!path)
7163                 return -ENOMEM;
7164
7165         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7166         if (!wc) {
7167                 btrfs_free_path(path);
7168                 return -ENOMEM;
7169         }
7170
7171         btrfs_assert_tree_locked(parent);
7172         parent_level = btrfs_header_level(parent);
7173         extent_buffer_get(parent);
7174         path->nodes[parent_level] = parent;
7175         path->slots[parent_level] = btrfs_header_nritems(parent);
7176
7177         btrfs_assert_tree_locked(node);
7178         level = btrfs_header_level(node);
7179         path->nodes[level] = node;
7180         path->slots[level] = 0;
7181         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7182
7183         wc->refs[parent_level] = 1;
7184         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7185         wc->level = level;
7186         wc->shared_level = -1;
7187         wc->stage = DROP_REFERENCE;
7188         wc->update_ref = 0;
7189         wc->keep_locks = 1;
7190         wc->for_reloc = 1;
7191         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7192
7193         while (1) {
7194                 wret = walk_down_tree(trans, root, path, wc);
7195                 if (wret < 0) {
7196                         ret = wret;
7197                         break;
7198                 }
7199
7200                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7201                 if (wret < 0)
7202                         ret = wret;
7203                 if (wret != 0)
7204                         break;
7205         }
7206
7207         kfree(wc);
7208         btrfs_free_path(path);
7209         return ret;
7210 }
7211
7212 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7213 {
7214         u64 num_devices;
7215         u64 stripped;
7216
7217         /*
7218          * if restripe for this chunk_type is on pick target profile and
7219          * return, otherwise do the usual balance
7220          */
7221         stripped = get_restripe_target(root->fs_info, flags);
7222         if (stripped)
7223                 return extended_to_chunk(stripped);
7224
7225         /*
7226          * we add in the count of missing devices because we want
7227          * to make sure that any RAID levels on a degraded FS
7228          * continue to be honored.
7229          */
7230         num_devices = root->fs_info->fs_devices->rw_devices +
7231                 root->fs_info->fs_devices->missing_devices;
7232
7233         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7234                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7235
7236         if (num_devices == 1) {
7237                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7238                 stripped = flags & ~stripped;
7239
7240                 /* turn raid0 into single device chunks */
7241                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7242                         return stripped;
7243
7244                 /* turn mirroring into duplication */
7245                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7246                              BTRFS_BLOCK_GROUP_RAID10))
7247                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7248         } else {
7249                 /* they already had raid on here, just return */
7250                 if (flags & stripped)
7251                         return flags;
7252
7253                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7254                 stripped = flags & ~stripped;
7255
7256                 /* switch duplicated blocks with raid1 */
7257                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7258                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7259
7260                 /* this is drive concat, leave it alone */
7261         }
7262
7263         return flags;
7264 }
7265
7266 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7267 {
7268         struct btrfs_space_info *sinfo = cache->space_info;
7269         u64 num_bytes;
7270         u64 min_allocable_bytes;
7271         int ret = -ENOSPC;
7272
7273
7274         /*
7275          * We need some metadata space and system metadata space for
7276          * allocating chunks in some corner cases until we force to set
7277          * it to be readonly.
7278          */
7279         if ((sinfo->flags &
7280              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7281             !force)
7282                 min_allocable_bytes = 1 * 1024 * 1024;
7283         else
7284                 min_allocable_bytes = 0;
7285
7286         spin_lock(&sinfo->lock);
7287         spin_lock(&cache->lock);
7288
7289         if (cache->ro) {
7290                 ret = 0;
7291                 goto out;
7292         }
7293
7294         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7295                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7296
7297         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7298             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7299             min_allocable_bytes <= sinfo->total_bytes) {
7300                 sinfo->bytes_readonly += num_bytes;
7301                 cache->ro = 1;
7302                 ret = 0;
7303         }
7304 out:
7305         spin_unlock(&cache->lock);
7306         spin_unlock(&sinfo->lock);
7307         return ret;
7308 }
7309
7310 int btrfs_set_block_group_ro(struct btrfs_root *root,
7311                              struct btrfs_block_group_cache *cache)
7312
7313 {
7314         struct btrfs_trans_handle *trans;
7315         u64 alloc_flags;
7316         int ret;
7317
7318         BUG_ON(cache->ro);
7319
7320         trans = btrfs_join_transaction(root);
7321         if (IS_ERR(trans))
7322                 return PTR_ERR(trans);
7323
7324         alloc_flags = update_block_group_flags(root, cache->flags);
7325         if (alloc_flags != cache->flags) {
7326                 ret = do_chunk_alloc(trans, root, alloc_flags,
7327                                      CHUNK_ALLOC_FORCE);
7328                 if (ret < 0)
7329                         goto out;
7330         }
7331
7332         ret = set_block_group_ro(cache, 0);
7333         if (!ret)
7334                 goto out;
7335         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7336         ret = do_chunk_alloc(trans, root, alloc_flags,
7337                              CHUNK_ALLOC_FORCE);
7338         if (ret < 0)
7339                 goto out;
7340         ret = set_block_group_ro(cache, 0);
7341 out:
7342         btrfs_end_transaction(trans, root);
7343         return ret;
7344 }
7345
7346 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7347                             struct btrfs_root *root, u64 type)
7348 {
7349         u64 alloc_flags = get_alloc_profile(root, type);
7350         return do_chunk_alloc(trans, root, alloc_flags,
7351                               CHUNK_ALLOC_FORCE);
7352 }
7353
7354 /*
7355  * helper to account the unused space of all the readonly block group in the
7356  * list. takes mirrors into account.
7357  */
7358 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7359 {
7360         struct btrfs_block_group_cache *block_group;
7361         u64 free_bytes = 0;
7362         int factor;
7363
7364         list_for_each_entry(block_group, groups_list, list) {
7365                 spin_lock(&block_group->lock);
7366
7367                 if (!block_group->ro) {
7368                         spin_unlock(&block_group->lock);
7369                         continue;
7370                 }
7371
7372                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7373                                           BTRFS_BLOCK_GROUP_RAID10 |
7374                                           BTRFS_BLOCK_GROUP_DUP))
7375                         factor = 2;
7376                 else
7377                         factor = 1;
7378
7379                 free_bytes += (block_group->key.offset -
7380                                btrfs_block_group_used(&block_group->item)) *
7381                                factor;
7382
7383                 spin_unlock(&block_group->lock);
7384         }
7385
7386         return free_bytes;
7387 }
7388
7389 /*
7390  * helper to account the unused space of all the readonly block group in the
7391  * space_info. takes mirrors into account.
7392  */
7393 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7394 {
7395         int i;
7396         u64 free_bytes = 0;
7397
7398         spin_lock(&sinfo->lock);
7399
7400         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7401                 if (!list_empty(&sinfo->block_groups[i]))
7402                         free_bytes += __btrfs_get_ro_block_group_free_space(
7403                                                 &sinfo->block_groups[i]);
7404
7405         spin_unlock(&sinfo->lock);
7406
7407         return free_bytes;
7408 }
7409
7410 void btrfs_set_block_group_rw(struct btrfs_root *root,
7411                               struct btrfs_block_group_cache *cache)
7412 {
7413         struct btrfs_space_info *sinfo = cache->space_info;
7414         u64 num_bytes;
7415
7416         BUG_ON(!cache->ro);
7417
7418         spin_lock(&sinfo->lock);
7419         spin_lock(&cache->lock);
7420         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7421                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7422         sinfo->bytes_readonly -= num_bytes;
7423         cache->ro = 0;
7424         spin_unlock(&cache->lock);
7425         spin_unlock(&sinfo->lock);
7426 }
7427
7428 /*
7429  * checks to see if its even possible to relocate this block group.
7430  *
7431  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7432  * ok to go ahead and try.
7433  */
7434 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7435 {
7436         struct btrfs_block_group_cache *block_group;
7437         struct btrfs_space_info *space_info;
7438         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7439         struct btrfs_device *device;
7440         u64 min_free;
7441         u64 dev_min = 1;
7442         u64 dev_nr = 0;
7443         u64 target;
7444         int index;
7445         int full = 0;
7446         int ret = 0;
7447
7448         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7449
7450         /* odd, couldn't find the block group, leave it alone */
7451         if (!block_group)
7452                 return -1;
7453
7454         min_free = btrfs_block_group_used(&block_group->item);
7455
7456         /* no bytes used, we're good */
7457         if (!min_free)
7458                 goto out;
7459
7460         space_info = block_group->space_info;
7461         spin_lock(&space_info->lock);
7462
7463         full = space_info->full;
7464
7465         /*
7466          * if this is the last block group we have in this space, we can't
7467          * relocate it unless we're able to allocate a new chunk below.
7468          *
7469          * Otherwise, we need to make sure we have room in the space to handle
7470          * all of the extents from this block group.  If we can, we're good
7471          */
7472         if ((space_info->total_bytes != block_group->key.offset) &&
7473             (space_info->bytes_used + space_info->bytes_reserved +
7474              space_info->bytes_pinned + space_info->bytes_readonly +
7475              min_free < space_info->total_bytes)) {
7476                 spin_unlock(&space_info->lock);
7477                 goto out;
7478         }
7479         spin_unlock(&space_info->lock);
7480
7481         /*
7482          * ok we don't have enough space, but maybe we have free space on our
7483          * devices to allocate new chunks for relocation, so loop through our
7484          * alloc devices and guess if we have enough space.  if this block
7485          * group is going to be restriped, run checks against the target
7486          * profile instead of the current one.
7487          */
7488         ret = -1;
7489
7490         /*
7491          * index:
7492          *      0: raid10
7493          *      1: raid1
7494          *      2: dup
7495          *      3: raid0
7496          *      4: single
7497          */
7498         target = get_restripe_target(root->fs_info, block_group->flags);
7499         if (target) {
7500                 index = __get_raid_index(extended_to_chunk(target));
7501         } else {
7502                 /*
7503                  * this is just a balance, so if we were marked as full
7504                  * we know there is no space for a new chunk
7505                  */
7506                 if (full)
7507                         goto out;
7508
7509                 index = get_block_group_index(block_group);
7510         }
7511
7512         if (index == 0) {
7513                 dev_min = 4;
7514                 /* Divide by 2 */
7515                 min_free >>= 1;
7516         } else if (index == 1) {
7517                 dev_min = 2;
7518         } else if (index == 2) {
7519                 /* Multiply by 2 */
7520                 min_free <<= 1;
7521         } else if (index == 3) {
7522                 dev_min = fs_devices->rw_devices;
7523                 do_div(min_free, dev_min);
7524         }
7525
7526         mutex_lock(&root->fs_info->chunk_mutex);
7527         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7528                 u64 dev_offset;
7529
7530                 /*
7531                  * check to make sure we can actually find a chunk with enough
7532                  * space to fit our block group in.
7533                  */
7534                 if (device->total_bytes > device->bytes_used + min_free &&
7535                     !device->is_tgtdev_for_dev_replace) {
7536                         ret = find_free_dev_extent(device, min_free,
7537                                                    &dev_offset, NULL);
7538                         if (!ret)
7539                                 dev_nr++;
7540
7541                         if (dev_nr >= dev_min)
7542                                 break;
7543
7544                         ret = -1;
7545                 }
7546         }
7547         mutex_unlock(&root->fs_info->chunk_mutex);
7548 out:
7549         btrfs_put_block_group(block_group);
7550         return ret;
7551 }
7552
7553 static int find_first_block_group(struct btrfs_root *root,
7554                 struct btrfs_path *path, struct btrfs_key *key)
7555 {
7556         int ret = 0;
7557         struct btrfs_key found_key;
7558         struct extent_buffer *leaf;
7559         int slot;
7560
7561         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7562         if (ret < 0)
7563                 goto out;
7564
7565         while (1) {
7566                 slot = path->slots[0];
7567                 leaf = path->nodes[0];
7568                 if (slot >= btrfs_header_nritems(leaf)) {
7569                         ret = btrfs_next_leaf(root, path);
7570                         if (ret == 0)
7571                                 continue;
7572                         if (ret < 0)
7573                                 goto out;
7574                         break;
7575                 }
7576                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7577
7578                 if (found_key.objectid >= key->objectid &&
7579                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7580                         ret = 0;
7581                         goto out;
7582                 }
7583                 path->slots[0]++;
7584         }
7585 out:
7586         return ret;
7587 }
7588
7589 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7590 {
7591         struct btrfs_block_group_cache *block_group;
7592         u64 last = 0;
7593
7594         while (1) {
7595                 struct inode *inode;
7596
7597                 block_group = btrfs_lookup_first_block_group(info, last);
7598                 while (block_group) {
7599                         spin_lock(&block_group->lock);
7600                         if (block_group->iref)
7601                                 break;
7602                         spin_unlock(&block_group->lock);
7603                         block_group = next_block_group(info->tree_root,
7604                                                        block_group);
7605                 }
7606                 if (!block_group) {
7607                         if (last == 0)
7608                                 break;
7609                         last = 0;
7610                         continue;
7611                 }
7612
7613                 inode = block_group->inode;
7614                 block_group->iref = 0;
7615                 block_group->inode = NULL;
7616                 spin_unlock(&block_group->lock);
7617                 iput(inode);
7618                 last = block_group->key.objectid + block_group->key.offset;
7619                 btrfs_put_block_group(block_group);
7620         }
7621 }
7622
7623 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7624 {
7625         struct btrfs_block_group_cache *block_group;
7626         struct btrfs_space_info *space_info;
7627         struct btrfs_caching_control *caching_ctl;
7628         struct rb_node *n;
7629
7630         down_write(&info->extent_commit_sem);
7631         while (!list_empty(&info->caching_block_groups)) {
7632                 caching_ctl = list_entry(info->caching_block_groups.next,
7633                                          struct btrfs_caching_control, list);
7634                 list_del(&caching_ctl->list);
7635                 put_caching_control(caching_ctl);
7636         }
7637         up_write(&info->extent_commit_sem);
7638
7639         spin_lock(&info->block_group_cache_lock);
7640         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7641                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7642                                        cache_node);
7643                 rb_erase(&block_group->cache_node,
7644                          &info->block_group_cache_tree);
7645                 spin_unlock(&info->block_group_cache_lock);
7646
7647                 down_write(&block_group->space_info->groups_sem);
7648                 list_del(&block_group->list);
7649                 up_write(&block_group->space_info->groups_sem);
7650
7651                 if (block_group->cached == BTRFS_CACHE_STARTED)
7652                         wait_block_group_cache_done(block_group);
7653
7654                 /*
7655                  * We haven't cached this block group, which means we could
7656                  * possibly have excluded extents on this block group.
7657                  */
7658                 if (block_group->cached == BTRFS_CACHE_NO)
7659                         free_excluded_extents(info->extent_root, block_group);
7660
7661                 btrfs_remove_free_space_cache(block_group);
7662                 btrfs_put_block_group(block_group);
7663
7664                 spin_lock(&info->block_group_cache_lock);
7665         }
7666         spin_unlock(&info->block_group_cache_lock);
7667
7668         /* now that all the block groups are freed, go through and
7669          * free all the space_info structs.  This is only called during
7670          * the final stages of unmount, and so we know nobody is
7671          * using them.  We call synchronize_rcu() once before we start,
7672          * just to be on the safe side.
7673          */
7674         synchronize_rcu();
7675
7676         release_global_block_rsv(info);
7677
7678         while(!list_empty(&info->space_info)) {
7679                 space_info = list_entry(info->space_info.next,
7680                                         struct btrfs_space_info,
7681                                         list);
7682                 if (space_info->bytes_pinned > 0 ||
7683                     space_info->bytes_reserved > 0 ||
7684                     space_info->bytes_may_use > 0) {
7685                         WARN_ON(1);
7686                         dump_space_info(space_info, 0, 0);
7687                 }
7688                 list_del(&space_info->list);
7689                 kfree(space_info);
7690         }
7691         return 0;
7692 }
7693
7694 static void __link_block_group(struct btrfs_space_info *space_info,
7695                                struct btrfs_block_group_cache *cache)
7696 {
7697         int index = get_block_group_index(cache);
7698
7699         down_write(&space_info->groups_sem);
7700         list_add_tail(&cache->list, &space_info->block_groups[index]);
7701         up_write(&space_info->groups_sem);
7702 }
7703
7704 int btrfs_read_block_groups(struct btrfs_root *root)
7705 {
7706         struct btrfs_path *path;
7707         int ret;
7708         struct btrfs_block_group_cache *cache;
7709         struct btrfs_fs_info *info = root->fs_info;
7710         struct btrfs_space_info *space_info;
7711         struct btrfs_key key;
7712         struct btrfs_key found_key;
7713         struct extent_buffer *leaf;
7714         int need_clear = 0;
7715         u64 cache_gen;
7716
7717         root = info->extent_root;
7718         key.objectid = 0;
7719         key.offset = 0;
7720         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7721         path = btrfs_alloc_path();
7722         if (!path)
7723                 return -ENOMEM;
7724         path->reada = 1;
7725
7726         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7727         if (btrfs_test_opt(root, SPACE_CACHE) &&
7728             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7729                 need_clear = 1;
7730         if (btrfs_test_opt(root, CLEAR_CACHE))
7731                 need_clear = 1;
7732
7733         while (1) {
7734                 ret = find_first_block_group(root, path, &key);
7735                 if (ret > 0)
7736                         break;
7737                 if (ret != 0)
7738                         goto error;
7739                 leaf = path->nodes[0];
7740                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7741                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7742                 if (!cache) {
7743                         ret = -ENOMEM;
7744                         goto error;
7745                 }
7746                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7747                                                 GFP_NOFS);
7748                 if (!cache->free_space_ctl) {
7749                         kfree(cache);
7750                         ret = -ENOMEM;
7751                         goto error;
7752                 }
7753
7754                 atomic_set(&cache->count, 1);
7755                 spin_lock_init(&cache->lock);
7756                 cache->fs_info = info;
7757                 INIT_LIST_HEAD(&cache->list);
7758                 INIT_LIST_HEAD(&cache->cluster_list);
7759
7760                 if (need_clear) {
7761                         /*
7762                          * When we mount with old space cache, we need to
7763                          * set BTRFS_DC_CLEAR and set dirty flag.
7764                          *
7765                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
7766                          *    truncate the old free space cache inode and
7767                          *    setup a new one.
7768                          * b) Setting 'dirty flag' makes sure that we flush
7769                          *    the new space cache info onto disk.
7770                          */
7771                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7772                         if (btrfs_test_opt(root, SPACE_CACHE))
7773                                 cache->dirty = 1;
7774                 }
7775
7776                 read_extent_buffer(leaf, &cache->item,
7777                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7778                                    sizeof(cache->item));
7779                 memcpy(&cache->key, &found_key, sizeof(found_key));
7780
7781                 key.objectid = found_key.objectid + found_key.offset;
7782                 btrfs_release_path(path);
7783                 cache->flags = btrfs_block_group_flags(&cache->item);
7784                 cache->sectorsize = root->sectorsize;
7785
7786                 btrfs_init_free_space_ctl(cache);
7787
7788                 /*
7789                  * We need to exclude the super stripes now so that the space
7790                  * info has super bytes accounted for, otherwise we'll think
7791                  * we have more space than we actually do.
7792                  */
7793                 exclude_super_stripes(root, cache);
7794
7795                 /*
7796                  * check for two cases, either we are full, and therefore
7797                  * don't need to bother with the caching work since we won't
7798                  * find any space, or we are empty, and we can just add all
7799                  * the space in and be done with it.  This saves us _alot_ of
7800                  * time, particularly in the full case.
7801                  */
7802                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7803                         cache->last_byte_to_unpin = (u64)-1;
7804                         cache->cached = BTRFS_CACHE_FINISHED;
7805                         free_excluded_extents(root, cache);
7806                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7807                         cache->last_byte_to_unpin = (u64)-1;
7808                         cache->cached = BTRFS_CACHE_FINISHED;
7809                         add_new_free_space(cache, root->fs_info,
7810                                            found_key.objectid,
7811                                            found_key.objectid +
7812                                            found_key.offset);
7813                         free_excluded_extents(root, cache);
7814                 }
7815
7816                 ret = update_space_info(info, cache->flags, found_key.offset,
7817                                         btrfs_block_group_used(&cache->item),
7818                                         &space_info);
7819                 BUG_ON(ret); /* -ENOMEM */
7820                 cache->space_info = space_info;
7821                 spin_lock(&cache->space_info->lock);
7822                 cache->space_info->bytes_readonly += cache->bytes_super;
7823                 spin_unlock(&cache->space_info->lock);
7824
7825                 __link_block_group(space_info, cache);
7826
7827                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7828                 BUG_ON(ret); /* Logic error */
7829
7830                 set_avail_alloc_bits(root->fs_info, cache->flags);
7831                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7832                         set_block_group_ro(cache, 1);
7833         }
7834
7835         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7836                 if (!(get_alloc_profile(root, space_info->flags) &
7837                       (BTRFS_BLOCK_GROUP_RAID10 |
7838                        BTRFS_BLOCK_GROUP_RAID1 |
7839                        BTRFS_BLOCK_GROUP_DUP)))
7840                         continue;
7841                 /*
7842                  * avoid allocating from un-mirrored block group if there are
7843                  * mirrored block groups.
7844                  */
7845                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7846                         set_block_group_ro(cache, 1);
7847                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7848                         set_block_group_ro(cache, 1);
7849         }
7850
7851         init_global_block_rsv(info);
7852         ret = 0;
7853 error:
7854         btrfs_free_path(path);
7855         return ret;
7856 }
7857
7858 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
7859                                        struct btrfs_root *root)
7860 {
7861         struct btrfs_block_group_cache *block_group, *tmp;
7862         struct btrfs_root *extent_root = root->fs_info->extent_root;
7863         struct btrfs_block_group_item item;
7864         struct btrfs_key key;
7865         int ret = 0;
7866
7867         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
7868                                  new_bg_list) {
7869                 list_del_init(&block_group->new_bg_list);
7870
7871                 if (ret)
7872                         continue;
7873
7874                 spin_lock(&block_group->lock);
7875                 memcpy(&item, &block_group->item, sizeof(item));
7876                 memcpy(&key, &block_group->key, sizeof(key));
7877                 spin_unlock(&block_group->lock);
7878
7879                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
7880                                         sizeof(item));
7881                 if (ret)
7882                         btrfs_abort_transaction(trans, extent_root, ret);
7883         }
7884 }
7885
7886 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7887                            struct btrfs_root *root, u64 bytes_used,
7888                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7889                            u64 size)
7890 {
7891         int ret;
7892         struct btrfs_root *extent_root;
7893         struct btrfs_block_group_cache *cache;
7894
7895         extent_root = root->fs_info->extent_root;
7896
7897         root->fs_info->last_trans_log_full_commit = trans->transid;
7898
7899         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7900         if (!cache)
7901                 return -ENOMEM;
7902         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7903                                         GFP_NOFS);
7904         if (!cache->free_space_ctl) {
7905                 kfree(cache);
7906                 return -ENOMEM;
7907         }
7908
7909         cache->key.objectid = chunk_offset;
7910         cache->key.offset = size;
7911         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7912         cache->sectorsize = root->sectorsize;
7913         cache->fs_info = root->fs_info;
7914
7915         atomic_set(&cache->count, 1);
7916         spin_lock_init(&cache->lock);
7917         INIT_LIST_HEAD(&cache->list);
7918         INIT_LIST_HEAD(&cache->cluster_list);
7919         INIT_LIST_HEAD(&cache->new_bg_list);
7920
7921         btrfs_init_free_space_ctl(cache);
7922
7923         btrfs_set_block_group_used(&cache->item, bytes_used);
7924         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7925         cache->flags = type;
7926         btrfs_set_block_group_flags(&cache->item, type);
7927
7928         cache->last_byte_to_unpin = (u64)-1;
7929         cache->cached = BTRFS_CACHE_FINISHED;
7930         exclude_super_stripes(root, cache);
7931
7932         add_new_free_space(cache, root->fs_info, chunk_offset,
7933                            chunk_offset + size);
7934
7935         free_excluded_extents(root, cache);
7936
7937         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7938                                 &cache->space_info);
7939         BUG_ON(ret); /* -ENOMEM */
7940         update_global_block_rsv(root->fs_info);
7941
7942         spin_lock(&cache->space_info->lock);
7943         cache->space_info->bytes_readonly += cache->bytes_super;
7944         spin_unlock(&cache->space_info->lock);
7945
7946         __link_block_group(cache->space_info, cache);
7947
7948         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7949         BUG_ON(ret); /* Logic error */
7950
7951         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
7952
7953         set_avail_alloc_bits(extent_root->fs_info, type);
7954
7955         return 0;
7956 }
7957
7958 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
7959 {
7960         u64 extra_flags = chunk_to_extended(flags) &
7961                                 BTRFS_EXTENDED_PROFILE_MASK;
7962
7963         if (flags & BTRFS_BLOCK_GROUP_DATA)
7964                 fs_info->avail_data_alloc_bits &= ~extra_flags;
7965         if (flags & BTRFS_BLOCK_GROUP_METADATA)
7966                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
7967         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
7968                 fs_info->avail_system_alloc_bits &= ~extra_flags;
7969 }
7970
7971 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7972                              struct btrfs_root *root, u64 group_start)
7973 {
7974         struct btrfs_path *path;
7975         struct btrfs_block_group_cache *block_group;
7976         struct btrfs_free_cluster *cluster;
7977         struct btrfs_root *tree_root = root->fs_info->tree_root;
7978         struct btrfs_key key;
7979         struct inode *inode;
7980         int ret;
7981         int index;
7982         int factor;
7983
7984         root = root->fs_info->extent_root;
7985
7986         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7987         BUG_ON(!block_group);
7988         BUG_ON(!block_group->ro);
7989
7990         /*
7991          * Free the reserved super bytes from this block group before
7992          * remove it.
7993          */
7994         free_excluded_extents(root, block_group);
7995
7996         memcpy(&key, &block_group->key, sizeof(key));
7997         index = get_block_group_index(block_group);
7998         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7999                                   BTRFS_BLOCK_GROUP_RAID1 |
8000                                   BTRFS_BLOCK_GROUP_RAID10))
8001                 factor = 2;
8002         else
8003                 factor = 1;
8004
8005         /* make sure this block group isn't part of an allocation cluster */
8006         cluster = &root->fs_info->data_alloc_cluster;
8007         spin_lock(&cluster->refill_lock);
8008         btrfs_return_cluster_to_free_space(block_group, cluster);
8009         spin_unlock(&cluster->refill_lock);
8010
8011         /*
8012          * make sure this block group isn't part of a metadata
8013          * allocation cluster
8014          */
8015         cluster = &root->fs_info->meta_alloc_cluster;
8016         spin_lock(&cluster->refill_lock);
8017         btrfs_return_cluster_to_free_space(block_group, cluster);
8018         spin_unlock(&cluster->refill_lock);
8019
8020         path = btrfs_alloc_path();
8021         if (!path) {
8022                 ret = -ENOMEM;
8023                 goto out;
8024         }
8025
8026         inode = lookup_free_space_inode(tree_root, block_group, path);
8027         if (!IS_ERR(inode)) {
8028                 ret = btrfs_orphan_add(trans, inode);
8029                 if (ret) {
8030                         btrfs_add_delayed_iput(inode);
8031                         goto out;
8032                 }
8033                 clear_nlink(inode);
8034                 /* One for the block groups ref */
8035                 spin_lock(&block_group->lock);
8036                 if (block_group->iref) {
8037                         block_group->iref = 0;
8038                         block_group->inode = NULL;
8039                         spin_unlock(&block_group->lock);
8040                         iput(inode);
8041                 } else {
8042                         spin_unlock(&block_group->lock);
8043                 }
8044                 /* One for our lookup ref */
8045                 btrfs_add_delayed_iput(inode);
8046         }
8047
8048         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8049         key.offset = block_group->key.objectid;
8050         key.type = 0;
8051
8052         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8053         if (ret < 0)
8054                 goto out;
8055         if (ret > 0)
8056                 btrfs_release_path(path);
8057         if (ret == 0) {
8058                 ret = btrfs_del_item(trans, tree_root, path);
8059                 if (ret)
8060                         goto out;
8061                 btrfs_release_path(path);
8062         }
8063
8064         spin_lock(&root->fs_info->block_group_cache_lock);
8065         rb_erase(&block_group->cache_node,
8066                  &root->fs_info->block_group_cache_tree);
8067         spin_unlock(&root->fs_info->block_group_cache_lock);
8068
8069         down_write(&block_group->space_info->groups_sem);
8070         /*
8071          * we must use list_del_init so people can check to see if they
8072          * are still on the list after taking the semaphore
8073          */
8074         list_del_init(&block_group->list);
8075         if (list_empty(&block_group->space_info->block_groups[index]))
8076                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8077         up_write(&block_group->space_info->groups_sem);
8078
8079         if (block_group->cached == BTRFS_CACHE_STARTED)
8080                 wait_block_group_cache_done(block_group);
8081
8082         btrfs_remove_free_space_cache(block_group);
8083
8084         spin_lock(&block_group->space_info->lock);
8085         block_group->space_info->total_bytes -= block_group->key.offset;
8086         block_group->space_info->bytes_readonly -= block_group->key.offset;
8087         block_group->space_info->disk_total -= block_group->key.offset * factor;
8088         spin_unlock(&block_group->space_info->lock);
8089
8090         memcpy(&key, &block_group->key, sizeof(key));
8091
8092         btrfs_clear_space_info_full(root->fs_info);
8093
8094         btrfs_put_block_group(block_group);
8095         btrfs_put_block_group(block_group);
8096
8097         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8098         if (ret > 0)
8099                 ret = -EIO;
8100         if (ret < 0)
8101                 goto out;
8102
8103         ret = btrfs_del_item(trans, root, path);
8104 out:
8105         btrfs_free_path(path);
8106         return ret;
8107 }
8108
8109 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8110 {
8111         struct btrfs_space_info *space_info;
8112         struct btrfs_super_block *disk_super;
8113         u64 features;
8114         u64 flags;
8115         int mixed = 0;
8116         int ret;
8117
8118         disk_super = fs_info->super_copy;
8119         if (!btrfs_super_root(disk_super))
8120                 return 1;
8121
8122         features = btrfs_super_incompat_flags(disk_super);
8123         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8124                 mixed = 1;
8125
8126         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8127         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8128         if (ret)
8129                 goto out;
8130
8131         if (mixed) {
8132                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8133                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8134         } else {
8135                 flags = BTRFS_BLOCK_GROUP_METADATA;
8136                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8137                 if (ret)
8138                         goto out;
8139
8140                 flags = BTRFS_BLOCK_GROUP_DATA;
8141                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8142         }
8143 out:
8144         return ret;
8145 }
8146
8147 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8148 {
8149         return unpin_extent_range(root, start, end);
8150 }
8151
8152 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8153                                u64 num_bytes, u64 *actual_bytes)
8154 {
8155         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8156 }
8157
8158 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8159 {
8160         struct btrfs_fs_info *fs_info = root->fs_info;
8161         struct btrfs_block_group_cache *cache = NULL;
8162         u64 group_trimmed;
8163         u64 start;
8164         u64 end;
8165         u64 trimmed = 0;
8166         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8167         int ret = 0;
8168
8169         /*
8170          * try to trim all FS space, our block group may start from non-zero.
8171          */
8172         if (range->len == total_bytes)
8173                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8174         else
8175                 cache = btrfs_lookup_block_group(fs_info, range->start);
8176
8177         while (cache) {
8178                 if (cache->key.objectid >= (range->start + range->len)) {
8179                         btrfs_put_block_group(cache);
8180                         break;
8181                 }
8182
8183                 start = max(range->start, cache->key.objectid);
8184                 end = min(range->start + range->len,
8185                                 cache->key.objectid + cache->key.offset);
8186
8187                 if (end - start >= range->minlen) {
8188                         if (!block_group_cache_done(cache)) {
8189                                 ret = cache_block_group(cache, NULL, root, 0);
8190                                 if (!ret)
8191                                         wait_block_group_cache_done(cache);
8192                         }
8193                         ret = btrfs_trim_block_group(cache,
8194                                                      &group_trimmed,
8195                                                      start,
8196                                                      end,
8197                                                      range->minlen);
8198
8199                         trimmed += group_trimmed;
8200                         if (ret) {
8201                                 btrfs_put_block_group(cache);
8202                                 break;
8203                         }
8204                 }
8205
8206                 cache = next_block_group(fs_info->tree_root, cache);
8207         }
8208
8209         range->len = trimmed;
8210         return ret;
8211 }